Essays in the Industrial Organization of the
Pharmaceutical Industry
by
Bradley T. Shapiro
M.S. Mathematics, Virginia Polytechnic Institute & State University (2009)
B.S. Mathematics, B.A. Economics, Virginia Polytechnic Institute & State University (2007)
SUBMITTED TO THE DEPARTMENT OF ECONOMICS IN PARTIAL
FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY IN ECONOMICS
AT
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
'MASSACHUSETTS INSTITUTE
OP TEC,4Njrjy
APRIL 2014
JUN 1 12014
LIBRARIESj
©2014 Bradley T. Shapiro. All rights reserved.
The author hereby grants to MIT permission to reproduce and to distribute
publicly paper and electronic copies of this thesis document in whole
or in part in any medium known or hereafter created.
Signature of Author:
Certified By:
Signature redacted
Signature redacted
/
Certified By:
0iyatuie
&
Department of Economics
April 2, 2014
Nancy Rose
Charles P. Kindleberger Professor of Applied Economics
Thesis Supervisor
FUaLAUd
Ernst Berndt
Louis E. Seley Professor in Applied Economics
A
Ar
Signature redact
Thesis Supervisor
Accepted By:
Michael Greenstone
3M Professor of Environmental Economics
Chairman, Departmental Committee on Graduate Studies
1
I
Essays in the Industrial Organization of the
Pharmaceutical Industry
by
Bradley T. Shapiro
Submitted to the Department of Economics on April 2, 2014 in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy in Economics
ABSTRACT
This dissertation comprises of three chapters, each exploring different issues in the industrial organization
of the pharmaceutical industry.
In the first chapter, I study the effects of television advertising of antidepressants using a novel method
comparing households near the borders of television markets. I find that advertising has positive benefits
on the market as a whole, including competitors. Using this information and a simple supply model, I
consider the counterfactual whereby firms work together in a co-operative. In the co-operative
equilibrium, firms advertise more than five times as much as is observed in competitive equilibrium and
profits would increase by about twenty percent.
In the second chapter, the documented phenomenon of strategic entry delay is analyzed in the sleep aid
industry in order to measure the cost to consumers of that delay. With the Hatch-Waxman Act of 1984,
the FDA included an unchallengeable exclusivity period for new approved drugs, independent of patents.
This generates an incentive for firms to strategically delay the introduction of new versions of drugs until
just before patent expiration of the originals in order to take market share away from new generics rather
than its own original product in its time of FDA exclusivity. Using detailed prescribing and pricing data, I
document that delay and estimate cost to consumers in the prescription sleep aid market at about $644
million over seven years.
In the final chapter, we examine firm behavior following the loss of exclusivity (LOE) of six molecules
between June 2009 and May 2013 that were among the 50 most prescribed molecules in May 2013. We
analyze speed of generic firm entry, prices, generic and penetration separately by four payer types (cash,
Medicare Part D, Medicaid, and other third party payer -TPP) and by age (under vs. over 65). We find
modest price decreases following LOE but very rapid and complete penetration of generics. While on
average molecule prices decrease, the branded versions continue to increase prices after LOE. Expansion
of molecule sales is increasingly common. Generic penetration rates are typically highest and most rapid
for TPPs, and lowest and slowest for Medicaid. Cash customers and seniors generally pay the highest
prices, TPPs and those under 65 pay the lowest prices. The presence of an authorized generic is also
analyzed and found to have effects that vary across molecule.
Thesis Supervisor: Nancy Rose
Title: Charles P. Kindleberger Professor of Applied Economics
2
Acknowledgements
I would like to thank my advisors Nancy Rose, Ernst Berndt and Stephen Ryan for their guidance
and mentorship through the course of my education at MIT.
Nancy, it would be stating it mildly to say that I hit the advisor jackpot. You always knew when I
needed a nudge in the right direction, an encouraging word, or a kick in the pants. You were there to
listen to my issues, personal or professional. I will always look up to you. One day I hope to be half the
advisor to my students as you are to me.
Ernie, meeting you at the end of my second year was perhaps the most significant moment in my
time at MIT, as it led me down a path to a deeper understanding of the pharmaceutical industry and what
it means to be a scholar. I will always be thankful for our friendship and for your help in getting me
through graduate school.
Stephen, your enthusiasm and support made my transition from classroom student to researcher
as smooth as butter, and it was awesome. You always help me find the exciting in what might first seem
mundane. Thank you for all of your encouragement.
I must also acknowledge the others that led me down the path to being a scholar, of which the
most significant were my undergraduate advisors Bud Brown, William Floyd, Amoz Kats, and Nicolaus
Tideman
I would like to thank the faculty at MIT who helped shape my dissertation with their thoughtful
comments in lunches and seminars. In particular, I would like to thank Sara Ellison, Glenn Ellison,
Michael Winston, Paulo Somaini, Nikhil Agarwal, Chris Knittel, Dick Schmalensee and Heidi Williams.
In addition, I am grateful for the students in the 10 group, particularly Brad Larsen, Ashley Swanson,
Gaston Illanes and Sarah Moshary.
My MIT experience would not have been as fulfilling or as challenging without my amazing
classmates. I thank my first year roommates Dan Rees and Maria Polyakova for their support in
surviving the first year. Next, I thank Michael Powell for taking me under his wing and being a mentor to
me from the very beginning. I thank Greg Leiserson and Xiao Yu Wang not only for giving my job
market paper a very thorough and critical read through, but also for being amazing friends. I thank my
Muddy Charles crew, Dan Barron, Isaiah Andrews, Miikka Rokkanen and Conrad Miller for always
being ready to have a good time. I thank Henry Swift, Ben Feigenberg and Chris Walters for helping to
distract me from the stress of work by feeding my fantasy football addiction. I thank Adam Sacarny for
always being ready to listen at the drop of hat. Finally, I thank Annalisa Scognamiglio for being the best
officemate, colleague and friend anyone could ever ask for.
I have had no greater supporters than my family. To my parents, Mary and Jeff Shapiro, thank
you for always encouraging me to follow my dreams. Thank you for instilling in me the desire to
continue challenging myself throughout my entire life. To my sister, Allison, thank you for putting up
with me for all of these years and for being a great friend. Finally, to the love of my life, Elaine, thank
you for being there with me every step of the way. I can't imagine surviving graduate school without your
love and support. I'm thankful to have you as my best friend and partner in life. I look forward to our
next chapter together.
3
4
Contents
1
Positive Spillovers and Free Riding in Advertising of Prescription Pharmaceuticals:
The Case of Antidepressants
10
1.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.2
Empirical Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.3
1.4
1.5
1.6
1.2.1
Prescription Drugs and Advertising
1.2.2
Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
Reduced Form Evidence
. . . . . . . . . . . . . . . . . . . . . 14
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.3.1
Empirical Identification Strategy - Border Strategy . . . . . . . . . . . . . 19
1.3.2
Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.4.1
Demand . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.4.2
Supply
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Empirical Specification and Estimation of the Model . . . . . . . . . . . . . . . . 3 0
1.5.1
Demand Specification
1.5.2
Transforming to a Linear Problem . . . . . . . . . . . . . . . . . . . . . . 3 1
1.5.3
Identification and Estimation Strategy . . . . . . . . . . . . . . . . . . . . 3 3
1.5.4
Demand Results
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
Supply and Counterfactual . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 5
1.6.1
Supply Implications of Positive Spillovers . . . . . . . . . . . . . . . . . . 3 5
1.6.2
Cost Computation
1.6.3
Counterfactual Assumptions
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
37
5
1.7
2
1.6.4
Advertising . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
1.6.5
Quantities and Profits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
1.6.6
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
Estimating the Cost of Strategic Entry Delay in Pharmaceuticals: The Case of Am65
bien CR
......................................
65
2.1
Introduction ........
2.2
Regulatory Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
2.2.1
Patents
2.2.2
FDA Exclusivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
2.2.3
Policy Proposal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
2.3
The Prescription Sleep Aid Market . . . . . . . . . . . . . . . . . . . . . . . . . . 69
2.4
Theoretical Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
2.5
2.6
2.7
2.4.1
Random Utility Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
2.4.2
Alternative Demand Models . . . . . . . . . . . . . . . . . . . . . . . . . 72
2.4.3
Counterfactuals and Welfare . . . . . . . . . . . . . . . . . . . . . . . . . 73
D ata . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
2.5.1
The Prescription Sleep Aid Market, The Outside Good and Market Shares . 74
2.5.2
Product Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
2.5.3
Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
R esults . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
2.6.1
Demand Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
2.6.2
Welfare Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Conclusion
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
6
3
The Regulation of Prescription Drug Competition and Market Responses: Patterns
94
in Prices and Sales Following Loss of Exclusivity
3.1
Introduction and Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
3.2
Data and M ethods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
3.3
Findings: Characteristics of Six Initial Generic launches, June 2009-May 2013
3.4
R esults . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
3.5
. . 98
3.4.1
Generic Penetration Rates, Over All and By Payer Type and Age . . . . . . 99
3.4.2
Quantities Post-LOE Relative to Pre-LOE . . . . . . . . . . . . . . . . . . 102
3.4.3
Number Extended Units Per Prescription
. . . . . . . . . . . . . . . . . . 104
3.4.4
Price Per Day of Therapy By Payer Type
. . . . . . . . . . . . . . . . . . 105
3.4.5
Prices and Prescription Shares During the 180-Day Triopoly . . . . . . . . 106
3.4.6
Cash vs. Full Sample Average Price Levels and Growth Rates
Summary and Concluding Remarks
. . . . . . . 108
. . . . . . . . . . . . . . . . . . . . . . . . . 109
List of Figures
1.1
Antidepressant Commercials Relative to FDA Memo
. . . . . . . . . . . . . . 45
1.2
Antidepressant Revenues 1996-2008 . . . . . . . . .
. . . . . . . . . . . . . . 45
1.3
Variation Across Three Markets in Advertising
. . .
. . . . . . . . . . . . . . 46
1.4
Two Different Advertising Measures: National
. . .
. . . . . . . . . . . . . . 46
1.5
Quantity Weighted Average Margins in Boston
. . .
. . . . . . . . . . . . . . 47
1.6
Full Sample: Top 101 DMAs . . . . . . . . . . . . .
. . . . . . . . . . . . . . 47
1.7
Ohio and Its DMAs . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . 48
7
1.8
Border Sample: Counties on the Borders of the Top 101 DMAs . . . . . . . . . . . 48
1.9
Variation in Log DTC Net of Fixed Effects, 14% Zeros . . . . . . . . . . . . . . . 49
1.10 Impulse Response Effect of Zoloft Advertisement on Own and Total Prescriptions . 51
1.11 Marginal Revenue Curves Under Various Scenarios . . . . . . .
. . . . . . . . . 51
1.12 Counterfactual versus Realized Advertising on Average . . . . .
. . . . . . . . . 52
1.13 Counterfactual versus Realized Advertising on Average, MC=$1
. . . . . . . . . 53
1.14 Counterfactual versus Realized Antidepressant Share Prescribed on Average . . . . 53
1.15 Counterfactual versus Realized Antidepressant Profits . . . . . .
. . . . . . . . . 54
2.1
Quantities over Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
2.2
Regions Used in the Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
2.3
Average Price over Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
2.4
Ambien Price Across Regions
2.5
Welfare Gains in Each Time Period from Ambien CR Presence . . . . . . . . . . . 88
2.6
Welfare Gain as a Pct of Total Revenue
2.7
Welfare Over Time With Molecule Nesting
2.8
Welfare By Pct of Total Revenue with Molecule Nesting
2.9
Welfare Over Time Counting Advertising . . . . . . . . . . . . . . . . . . . . . . 92
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
. . . . . . . . . . . . . . . . . . . . . . . 88
. . . . . . . . . . . . . . . . . . . . . 91
. . . . . . . . . . . . . . 91
2.10 Welfare By Pct. of Total Revenue Counting Advertising . . . . . . . . . . . . . . . 93
3.1
Price Per Day of Therapy By Payer Type . . . . . . . . . . . . . . . . . . . . . . . 122
8
List of Tables
1.1
Descriptive Statistics: Advertising . . . . . . . . . . . . . . . . . . . . . . . . . . 4 6
1.2
The Effect of Own and Rival Advertisements on Sales . . . . . . . . . . . . . . . . 4 9
1.3
Descriptive Statistics: Border Sample, 1997-2003 . . . . . . . . . . . . . . . . . . 5 0
1.4
Results of Base Model
1.5
Short Run Advertising Elasticities . . . . . . . . . . . . . . . . . . . . . . . . . . 5 0
1.6
Marginal Cost Distributions By Product . . . . . . . . . . . . . . . . . . . . . . . 5 2
1.7
R2 of MC Regressions
1.8
Predicting Prices with Advertising . . . . . . . . . . . . . . . . ... . . . . . . . . 5 5
1.9
Results of Model including Worst Case Detailing . . . . . . . . . . . . . . . . . . 5 6
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
1.10 Results of Base Model with Placebo . . . . . . . . . . . . . . . . . . . . . . . . . 5 7
1.11 Results of Base Model without Northeast Corridor and Other Urban Areas . . . . . 58
1.12 Results of Base Model with only Northeast Corridor and Other Urban Areas . . . . 58
1.13 Results of Base Model Without Using the Border Approach . . . . . . . . . . . . . 59
2.1
Different FDA Exclusivity Lengths By New Drug Type . . . . . . . . . . . . . . . 84
2.2
Baseline Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
2.3
Results with Hausman Instruments . . . . . . . . . . . . . . . . . . . . . . . . . . 89
2.4
Results with Molecule Nesting . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
3.1
Characteristics of Six Initial Generic Launches, June 2009-May 2013
3.2
Months to Generic Penetration Rate Thresholds, June 2009 - May 2013
3.3
Mean Quantities Post-LOE Six to Nine Months, and Post-LOE Ten to Twelve
Months Relative to Pre-LOE Quantities, June 2009 - May 2013 . . . . . . . . . . . 119
3.4
Prices and Prescription Shares During the 180-Day Exclusivity Period for Branded,
Generic and Authorized Generic Molecules . . . . . . . . . . . . . . . . . . . . . 120
3.5
Cash Versus Full Sample Mean Price Levels and Monthly Growth Rates . . . . . . 121
9
. . . . . . . 117
. . . . . . 118
1
Positive Spillovers and Free Riding in Advertising of Prescription Pharmaceuticals: The Case of Antidepressants
Abstract
Television advertising of prescription drugs is controversial, and it remains illegal in all
but.two countries. Much of the opposition stems from concerns that advertising directly to
consumers may inefficiently distort prescribing patterns toward the advertised product. Despite the controversy surrounding the practice, its effects are not well understood. Exploiting
a policy change that makes such advertising possible in the United States along with a spatial identification at the border approach, I estimate that television advertising of prescription
antidepressants exhibits significant positive spillovers on rivals' demand. I then construct and
estimate a multi-stage demand model that allows advertising to be pure category expansion,
pure business stealing or some of each. Estimated parameters indicate that advertising has
strong market level demand effects that tend to dominate business stealing effects. Spillovers
are both large and persistent. Using these estimates and a simple supply model, I explore the
consequences of the positive spillovers on firm advertising choice. In a cooperative advertising
campaign, simulations suggest that the co-operative would produce on average five and a half
times as much advertising as is observed in competitive equilibrium, resulting in a 19.2 percent
increase in category size and a 19.8 percent increase in category profits.
1.1
Introduction
How does television advertising affect the consumer choice problem? After a consumer watches
a commercial, internalizes its message and decides a product is desirable, she must take further
action to obtain the product. With groceries, she must go to the supermarket. With many consumer
products, a computer with internet will allow the consumer to make the purchase. With prescription
drugs, the consumer must go to the physician to obtain a prescription and then to the pharmacy to
purchase the drug. With many steps between the advertising incidence and purchase, at some stage
of the process, the consumer might well choose a different product from the one advertised. This
may be due to difficulty in remembering advertisements, agency problems in obtaining products
or simply because advertising convinces a consumer to go to a retailer, computer or physician. In
short, an advertisement could affect the choice process without leading the consumer to buy the
advertised product.
In this paper, I identify the existence of positive spillovers of television advertising in the market for
antidepressants. Given this, I construct and estimate a demand model which allows such spillovers.
10
To quantify the effect of spillovers on firm behavior, I conduct a supply side analysis supposing
that the firms are able to jointly decide advertising, and I compare this outcome to the realized
advertising outcome in the antidepressant market.
Television advertising of prescription drugs is contentious and has been condemned by many as
inefficiently distorting prescriptions to the advertised products. In fact, it is legal in only two countries: New Zealand and the United States. In light of the controversy, it is important to understand
the impact of these advertisements. In particular, understanding spillovers is crucial to regulators, firms and econometricians. From a regulatory perspective, the Food and Drug Administration
(FDA) regulates the content of advertisements. To the extent that advertising content is made more
informative and less brand specific, content regulation could exacerbate spillovers. Firms may
lose individual incentives to advertise as spillovers intensify. This could be either good or bad for
social welfare depending on whether or not category expansion is a public good or a public bad.
However, it is an important consideration for the regulator in either case. From a firm strategy
perspective, understanding possible channels for revenue improvement is vital. While cooperation is often difficult to enforce and non-contractible due to antitrust laws, advertising cooperatives
are precedented in other industries such as orange juice, milk and beef. Finally, from a technical
perspective, failure to model spillovers in advertising can distort estimated parameters, leading to
incorrect inferences about supply and demand.
Previous research incorporating advertising into demand analysis has frequently treated advertising of a product as affecting its probability of being in the choice set (Goeree 2008), or has
incorporated advertising into a production of goodwill that enters directly into the utility function
(Dube et. al. 2005). However, such specifications also typically exclude the possibility of positive
spillovers of advertising onto rivals. While this eliminates the complexity of modeling behavior in
the presence of possible free riding, such an exclusion may lead the researcher to miss important
strategic considerations. When deciding how much to advertise, firms do not internalize the benefit
they provide to other firms and have an incentive to free ride on their rivals' advertising efforts.
Understanding these considerations is important for marketing decision makers as well as policy
makers potentially seeking to regulate advertising.
Prescription drugs in general, and antidepressants in particular, have many characteristics which
facilitate positive spillovers in television advertising. First, the FDA regulates what firms can and
cannot say in advertisements. While the name of the product is typically prominently displayed
throughout the commercial, most of the time in each commercial is spent explaining the ailment,
the mechanism of action of the drug and its side effects. When there are several therapeutic products available, those treating the same ailments tend to share common characteristics. A consumer
11
might remember all of the things being said but forget the name of the product. Agency problems
further disrupt this link. A consumer must see a doctor to get a prescription. A physician might
have different preferences or opinions about which drugs, if any, work best for a given condition
or patient. The advertisement may lead a patient to the physician, but the physician is still the
ultimate arbiter of whether and what to prescribe.
My strategy for evaluating the extent of positive spillovers in advertising for antidepressants proceeds in three steps. First, I use discrete television market borders to determine whether advertising
does exhibit positive spillovers onto rivals. Next, I construct and estimate a model of the antidepressant market, allowing advertising to have positive spillovers on demand of horizontally differentiated products, a feature excluded by typical discrete choice specifications. Positive spillovers
are allowed, but not imposed by the model. Given the demand estimates and a model of supply,
I back out the marginal costs of advertising from a model of dynamic strategic firm interaction.
Finally, given estimates of the demand effects and marginal costs of advertising, I quantify the
importance of free riding by re-simulating the supply model, assuming that a co-operative sets
advertising for the entire industry.
Studies of Direct-to-Consumer (DTC) advertising in pharmaceuticals with varying credibility of
identification strategies have suggested the possibility that cross-advertising elasticities could be
positive, but results have been mixed. In particular, [lizuka & Jin (2005), Berndt et. al. (2004),
Wosinska (2002)], find very small estimates of advertising effects on market shares conditional on
being in the market and conclude that it might be exhibiting positive spillovers, though spillovers
are neither modeled nor tested. Wosinska (2005) and Donohue et. al. (2004) find that advertising has positive spillover effects onto drug compliance and duration of treatment. Other studies
find that advertising drives consumers to the doctor (lizuka & Jin 2004) or has class level effects
(Rosenthal et. al. 2003, Avery et. al. 2012), but they do not model any product level own or
cross-elasticities of advertising. In work that is more closely related to this study, Berndt et. al.
(1995, 1997) estimate the effect of marketing on both the size of the market and on brand shares,
focusing mostly on physician detail advertising and academic journal advertising since DTC was
extremely limited and unbranded at the time, and found some effects at both category and product
levels. In studying detailing effects Ching et. al. (2012) take advantage of the fact that identical
molecules are sometimes marketed by different firms under different names in Canada to separate
out brand versus category effects. Narayanan et. al. (2004) estimate a two level model using only
time series variation for antihistamines and do not find positive spillovers. In experimental work,
Kravitz et. al. (2005) find mixed results for patients going to their physicians asking for products
they saw on television. In a structural model, Jayawardhana (2013) imposes that television advertising must only affect class level demand and finds significant effects. Many of these studies either
12
only model a category level response or only model a conditional share level response. Those that
model both rely solely on time series data. This paper will model the full decision process and use
data with both spatial and time series variation.
The supply side of advertising in pharmaceuticals has been much less explored. If advertising
helps rivals' demand, there might well be an incentive to invest less in advertising. lizuka (2004)
finds that as the number of competitors increases, firms advertise less, leading him to suggest the
existence of a free riding problem. Ellison and Ellison (2011) find evidence that pharmaceutical
firms decrease advertising just prior to patent expiration in order to make the market smaller and
deter generics from entering. The possibility of such strategic deterrence implies the existence of
positive spillovers, at least from brand to generic. However, no research that I am aware of uses
a supply model to quantify the magnitude of the potential positive spillover effects on advertising
expenditure decisions.
Outside of the pharmaceutical literature, Sahni (2013) finds experimental evidence of positive
spillovers to rivals in restaurant advertising in India. Additionally, Lewis & Nguyen (2012) and
Anderson & Simester (2013) find evidence of positive spillovers in a number of categories. While
these studies are strong in identification, they do not spend much energy in studying the supply
side consequences of the positive spillovers that they identify.
The contributions of this paper are threefold. First, I improve upon the literature that seeks to
identify the causal effect of advertising on own and rivals' demand in pharmaceuticals by using an
identification at the border approach. That is, I will identify advertising elasticities by comparing
households that are very near to each other geographically but get different advertisements due to
the way the television market borders are drawn. In addition, I exploit a 1999 policy change making
television advertising possible in the United States. I show that advertising has significant positive
effects on rivals' sales, though smaller than its effects on own firm sales. Next, I construct and
estimate a consumer choice model, which allows advertising to influence the size of the category,
the conditional share of each subcategory in the category, and the conditional share of each product
in a subcategory. I will consider the category, the subcategory and the product levels as three
separate stages of a joint physician-consumer decision making process. At each stage, I will allow
for some inter-temporal influence. Results indicate that advertising of antidepressants affects all
levels of choice. The category effects are larger and more persistent over time than are business
stealing effects, leading to a net positive spillover. Finally, I conduct a supply side analysis to
evaluate to what extent positive spillovers suppress the incentive to advertise. I find that a cooperative deciding all advertising expenditure levels would advertise on average five times as much
as is observed in competitive equilibrium. No other research that I am aware of conducts such a
13
supply side analysis of the provision of advertising that exhibits positive spillovers. This paper
helps move us toward understanding the effects of advertising and the incentives facing the firms
who provide it, and understanding both are essential to firm profit maximization and to efficient
regulation.
1.2
1.2.1
Empirical Setting
Prescription Drugs and Advertising
Television advertising of prescription drugs did not appear in the United States until 1997. While
technically not forbidden by law, advertising was required to have much more risk information
included on all advertisements than is required today. This required risk information was similar
to the package inserts that come with prescriptions. Reading those aloud in the context of a thirty
second spot was prohibitively time consuming and costly. In the fall of 1997, the FDA issued a
draft memorandum clarifying their stance on advertising risk information, allowing advertisements
to air so long as they had a 'fair balance' of risk information, even if abbreviated. Firms had the
opportunity to submit their advertisements to the FDA for pre-approval to ensure that the 'fair
balance' condition was met. In 1999 the final copy of the FDA memorandum was circulated. The
first advertisements on television for antidepressants were seen in 1999 when GlaxoSmithKline's
brand, Paxil, began airing its first campaigns.
Figure 1.1 suggests that the FDA regulation was binding prior to 1999, and advertising did not
begin until that point.
Antidepressants Prescription antidepressants are indicated for treatment of major depressive
disorder and dysthymia, which is a more minor version of depression. Traditionally, depression
was treated with what are called tricylcic antidepressants (TCAs), which were discovered in the
1950s, but those came with significant side effects and risks. Treatment of depression took a
great leap forward in the late 1980s with the innovation of selective serotonin reuptake inhibitors
(SSRIs), the first of which was Prozac. Newer generation antidepressants are more tolerable than
the older generation TCAs and allow patients to be more safely treated and with fewer side effects
(Anderson 2000). This allows easier management of antidepressant treatment by primary care
physicians, and makes seeing a specialist less necessary.
Diagnosis and treatment of depression can be rather complicated, as with many mental disorders.
As the class of drugs has grown, so have the number of people being treated. In 1996, the industry
14
pulled in around $5 billion in revenue. By 2004, it was up to $13 billion. In 2004, an FDA black
box warning was instituted suggesting that antidepressants might lead to an increase in suicidality
among adolescents (Busch et. al 2012). Around the same time, many widely selling molecules
began to go off patent. Figure 1.2 shows the revenues of the antidepressant industry from 1996
through 2004. Since the discovery of Prozac, ten other brands, some with slightly different mechanisms, have been discovered and have entered the market. Some of those have developed extended
release versions which allow patients to have fewer doses per day.
There are six main subcategories of antidepressants: the old style TCAs, Tetracylcic (TeCA), Serotonin Antagonist and Reuptake Inhibitors (SARI), Serotonin-norepinephrine Reuptake Inhibitors
(SNRI), Norepinephrine Reuptake Inhibitors (NDRI) and SSRI. While the specific differences between these are not important to this study, it is worth noting that each subcategory has somewhat
different mechanisms, interactions and side effect profiles from the others. Deciding which subcategory of antidepressant is appropriate for a given patient is largely up to the physician, and often is
related to other medications the patient is taking. The decision between drugs within a subcategory
might depend on what is included on the patient's insurance formulary or physician preferences.
Antidepressants are characterized by a high degree of experimentation to find a good fit between
treatment and patient, as well as a low compliance rate due to the many side effects (Murphy et.
al. 2009).
Many physicians see depression as an under treated condition and some research has concluded
that restricting access to antidepressants has been associated with negative health outcomes (Busch
et. al. 2012). Given this information, it is plausible that market expansive advertising could play a
role in this market.
The Market for Advertising Firms can purchase advertising space on television in two ways.
First, there is an upfront market each summer where advertising agencies and firms make deals for
the upcoming year of television. Advertising purchased in the upfront market cannot be "returned"
and typically has minimal flexibility in terms of timing. Next, there is a spot market that is called
the 'scatter' market, where firms can purchase advertising closer to the date aired.
Additionally, there are both national and local advertisements. National advertisements are seen
by everyone in the country tuned into a particular station, while local advertisements are only seen
by households within a particular designated market area (DMA).
A DMA is a collection of counties, typically centered around a major city, and it is defined by AC
Nielsen, a global marketing research firm. The DMA location of a county determines which local
15
television stations that a consumer of cable or satellite dish gets with his or her subscription. In
addition, those who watch television over the air, are more likely to pick up stations within their
DMAs than from others. There are 210 DMAs in the United States, the largest 101 of which are
included in my data.
From informal conversations with individuals in industry, I learned that pharmaceutical companies
participate almost exclusively in the up front market. Like most consumer goods, the majority
of antidepressant spending is on national advertising, but there is a significant amount of local
advertising as well as significant variation across DMAs in the amount of local advertising.
Prices for advertisements typically are determined by projected volume and type of viewership. A
single airing of a national advertisement for antidepressants ranges from $1,600 to $23,000 from
1999-2003 and a single airing of a local advertisement ranges from $0 to $7,600 for the same
time period. Looking at each advertisement in terms of expenditure per capita, I observe that the
distribution of local advertising expenditure per capita on a single commercial looks similar to
the distribution of national advertising expenditure per capita on a single commercial. National
advertisements range from $0.0002 per 100 to $0.04 per 100 and 93% of local advertisements fall
within that range as well, with a few outliers going down to zero and up to $0.20 per 100 capita. By
scaling expenditures by potential viewing population, local and national advertising expenditures
are comparable.
1.2.2
Data
Prescribing Data Sales data for this market comes from the Xponent data set of IMS Health,
a health care market research company. The prescribing behavior of a 5% random sample of
physicians who prescribe antidepressants is followed monthly from 1997 until 2004. The data
include a rich set of physician characteristics including address of the primary practice, which
is then linked to county. The data used in this study is aggregated to the county level and ends
with 2003, thereby avoiding confounding market changes in 2004 including the FDA black box
warnings and wave of patent expirations. The sample is partially refreshed annually.
Advertising Data Product level monthly advertising data at the national and Designated Market
Area (DMA) level for the top 101 DMAs comes from Kantar Media. In addition to advertising
expenditures, the data includes number of commercials. The unit of advertising used in this study
will be expenditures per 100 capita in the viewing area. Scaling expenditures by population in the
viewing area allows me to have a comparable measure of advertising volume between national and
16
local advertising. Total advertising for a county is defined as the national advertising expenditure
scaled by the national population plus the local advertising expenditure scaled by the population of
the DMA. 1 Table 1.1 provides some descriptive statistics for the DMA level advertising variables
at the product, subcategory and category level for the period of the data where advertising is allowed: September 1999 through December 2003. The statistics are also only on the products that
ever advertise: Paxil, Paxil CR, Prozac, Prozac Weekly, Wellbutrin SR, Wellbutrin XL and Zoloft.
Figure 1.3 depicts local advertising expenditures per 100 capita in Boston, New York and Austin
as well as national as examples of what local advertising expenditures look like over time. Local
advertising for Paxil is higher in New York than it is in Boston, which in turn is higher than it is
in Austin, suggesting that there is non-trivial variation across markets in this measure. National
advertising makes up the bulk of the advertising that households see, but the local additions to
the national advertising vary a great deal. The pattern is very similar in number of commercials,
indicating that expenditures per capita are a reasonably comparable object across localities and
national. However, as commercials are likely to be priced by reach, using expenditures per 100
capita should be a reasonably comparable way to measure reach of the ads. Figure 1.4 shows that
national commercials and national expenditures per 100 capita are highly correlated.
Other Data Sources I observe prices from Medicaid reimbursement data, collected by the Centers for Medicare & Medicaid Services (CMS). Duggan and Scott-Morton (2006) argue that the
average price that Medicaid pays per prescription prior to Medicaid rebates is a good measure of
the average price of a drug on the market. As my measure of price, I use the total Medicaid prescriptions dispensed divided by the total Medicaid reimbursements during a quarter for a particular
product, deflated to 2010 dollars using the consumer price index.
CMS also collects data on the average pharmacy acquisition cost for all pharmaceutical products
(NADAC). As I will not be estimating marginal production costs empirically, these average pharmacy acquisition costs may be used as an effective upper bound on marginal production costs.
While there are markups from branded drugs, pharmacies are typically able to obtain generics at
much lower rates, particularly when there are several generic competitors (as is the case in this
market), often as low as ten cents per pill. As of 2013, all products in the sample have generic
versions available. For an upper bound on the marginal cost of each drug, I use average pharmacy acquisition cost for those generic version of the product, deflated to 2010 dollars using the
IA possible alternative measure would be to use the number of commercials at the national level plus the number of
commercials at the local level. I explored using that measure and the results were not qualitatively different. However,
as a commercial during the evening news is likely to capture far more eyeballs than a commercial during a 1:00 AM
rerun of MacGyver, using expenditures per 100 capita would seem to do a better job at measuring quality adjusted
advertising than number of commercials.
17
consumer price index. In Figure 1.5, the quantity weighted average margin in Boston is plotted
versus time for the purposes of illustration. The average margin in the market rises as more popular branded antidepressants replace old generic TCAs and falls as more newer generation generics
become widely prescribed.
Yearly county population and income data are drawn from the Current Population Survey (CPS).
Notable in the data is that there is both national and local advertising. While national advertising
makes up the majority of advertising, there is significant spatial variation in the local additions to
what households see.
Additionally, only four brands from three firms in this market advertise at all. Eli Lilly (Prozac,
Prozac Weekly), Pfizer (Zoloft) and Glaxosmithkline (Paxil, Paxil CR, Wellbutrin SR, Wellburin
XL) are the only firms advertising in this market. Notably, those firms, along with Merck, are
some of the largest advertisers among all of the pharmaceutical industry (Berndt et al. 2003). The
lack of advertising from all firms could be indicative of fixed costs of advertising at all or of free
riding. Those branded products which do not advertise either have low market share (Effexor XR,
Remeron, Serzone) or have a very small parent company which might be less likely to have an
advertising division (Celexa, Lexapro).
Finally, own firm and rival firm advertising are negatively correlated, which could be an indicator
of the positive spillovers of advertising in this setting.
1.3
Reduced Form Evidence
In this section, I explore the data to see if spillover effects exist and how they interact with own
effects. This exercise has been difficult to implement in previous research, largely due to data
limitations. Estimates show that rivals' and own advertising have a positive effect on sales, while
rivals' advertising has a smaller effect than own advertising. In addition, the cross partials indicate that rivals' advertising makes own advertising less effective, but own advertising has a larger
negative effect on the marginal own advertisement due to decreasing returns to scale.
In particular, I model sales of quantities Q of product j in time t for market m as a function of own
advertising, aO", and advertising of rivals, across:
log(Qjmt) = 2tog(Qjm,t-1)+ y aqmn +
+-3 (
qn2
d4oss )2+ 75a
ars+
Ejmt
(1.1)
18
This provides insight on whether rivals' advertisements help or hurt own demand, the nature of
decreasing returns to scale, and persistence in advertising effects.
1.3.1
Empirical Identification Strategy - Border Strategy
The endogeneity of advertising and the absence of obvious instruments pose challenges to causal
identification of the effect of advertising on demand. I identify the effects of television advertising
by taking advantage of the discrete nature of local advertising markets. That is, two households
which are directly across the television market border from one another will see different advertisements despite being otherwise very similar households. I take advantage of this comparison. This
approach is similar in spirit to that used by Dube et. al. (2012) to identify the effects of minimum
wage increases and that used by Holmes (1998) to identify the effect of right-to-work laws.
Advertising is purchased both nationally and locally. The level of total advertising that a household
gets to its television is determined by the Designated Market Area (DMA) that the household's
county belongs to, as defined by AC Nielsen. Nielsen places counties into markets by predicting
which local stations the households will be most interested in. As such, DMAs tend to be centered
at metropolitan areas. A map of all of the DMAs included in the advertising data is presented in
Figure 1.6.
To get an idea of how advertising is distributed across the country, consider the example of the
Cleveland and Columbus DMAs. Figure 1.7 depicts the state of Ohio with each DMA in a different
color 2 . Every county in the Cleveland, Ohio DMA gets the same amount of the same advertising
as every other county in the Cleveland DMA. Meanwhile, every county in the Columbus, Ohio,
DMA gets the same amount of the same advertising as every other county in the Columbus DMA,
though this might be different from the advertising in the Cleveland DMA. Meanwhile, these two
DMAs border each other. There are five counties in the Cleveland DMA which share a border with
at least one county in the Columbus DMA and five counties in the Columbus DMA which share
at least one border with a county in the Cleveland DMA. My strategy will be to consider these
ten counties as an experiment with two treatment groups (Cleveland and Columbus) in each time
period.
The data contain 153 such borders. The map of all of the counties included in this border sample is
presented in Figure 1.8. Each of these borders will be considered a separate experiment, with the
magnitude of the treatment determined by the advertising in each DMA at a given time. Only the
2
from http://www.dishuser.com/TVMarkets/Maps/ohio.gif
19
counties bordering each other will serve as controls for each other to partial out any local effects
that may be increasing or decreasing for both sides of the border.
To estimate the effects of advertising in this experiment, I will use a modified difference-indifferences estimator. My identification assumption is that along the border of two DMAs, any
differential trends in demand between the two sides of the DMA border stem from differences in
advertising. In particular, I use panel data with fixed effects. Border-time fixed effects will ensure that the common trend assumption is only enforced locally at the border between two DMAs,
allowing for spatial heterogeneity. Border-DMA fixed effects will allow systematically different
demand levels across the border. I will also include a lagged dependent variable to get at the
dynamic effects of advertising. Consider the log of quantity log(Qjbmt), at the product-borderDMA-month level. Advertising, ajmt, as mentioned before lives at the product-DMA-month level
and affects log(Qjbmt) through some function
f:
log(Qjbmt) = f (ajmt) + Ejbmt
Each product-border pair will constitute an experiment with border-markets being treatment groups.
My fixed effects specification is:
log(Qjbmt)
=
Xlog(Qjbm,t- 1)
+ f(ajmt) + ajbq + ajbm + Ejbmt
where the subscripts j and b indicate which experiment is being considered (product and border
specific), ajbq is a time effect which is used to control the experiment, which in this case will be
a quarter fixed effect, ajbm is a treatment group fixed effect, and f (ajmt) is the magnitude of the
treatment. The magnitude of the treatment is zero everywhere prior to 1999, as the FDA memo
had not yet gone into effect. To investigate persistence in demand, a lagged dependent variable is
also included.
For further intuition, again consider the Cleveland-Columbus example and the case of Zoloft advertisements. In the equation above, log(Qjbmt) is log number of prescriptions of Zoloft in the
Cleveland-Columbus border, indexed by month and which side of the border it is on. The magnitude of the treatment, f(ajmt) is a function of the Zoloft's advertising in each market. The time
effect, ajbq, is a common quarter fixed effect between the Cleveland and Columbus sides of this
border and is used to subtract out contemporaneous macro effects. The fixed effect, ajbm, allows
the different sides of the border to have systematically different levels in the outcome.
20
For this strategy to be valid, the Cleveland and Columbus sides of the border may differ by a fixed
level, but they must have common trends absent advertising differences. Is this plausible? These
counties are bordering, so they are very similar in geography. Both are sufficiently far from their
central cities. The counties on the Cleveland side are only slightly closer to Cleveland than they
are to Columbus and vice versa.
Also worth noting is that if Columbus always had a high, constant level of advertising and Cleveland always had a low, constant level of advertising, this estimation strategy would have no power
to identify the effects of interest, as the border-DMA fixed effect would subtract out this variation,
even though that advertising in Columbus might well have had an effect. Since prior to 1997, no
DMAs had any advertising, there will be at least some variation in each experiment over time.
Potential Threats to the Border Strategy One potential worry is that there would be little variation after partialling out the fixed effects. This would be the case if too much of the advertising
were national and not enough were local. Figure 1.9 displays a histogram of advertising net of
these fixed effects showing significant variation. Net of fixed effects, the log of advertising expenditures per 100 capita has a mean of zero and a standard deviation of 0.25, so there is substantial
variation even after fixed effects are partialed out.
Also potentially problematic is the lagged dependent variable, which can generate omitted variables bias in the presence of small T, as differencing mechanically induces correlation between
the lagged dependent variable and the error term. However, as T-+ o, the mechanical correlation
with the error term diminishes to zero and the fixed effects estimator is consistent. As my data is
monthly from 1997 through 2003, T=84 should be sufficiently large that any bias will be minimal.
Additionally, we might be concerned about measurement error. There are a two main possibilities
that could lead to measurement error and biased estimates:
1. Consumers watch advertisements in one DMA, but drive across the border to see their physicians.
2. Consumer watch advertisements in one DMA, but drive towards the center of the DMA to a
county not included in the border sample to see their physicians.
Both of these scenarios would lead me to understate the effect of the advertisements. To the extent
that we think that these biases are present, we can look at my estimates as lower bounds on the true
21
parameters. 3
Omitted variables bias could also be a source of bias. Prices and detailing are omitted from this
estimation. As prices tend not to vary geographically due to a very low transport cost and ease
of obtaining drugs through the mail, prices are absorbed in the product-border-time fixed effect.
Detailing is observed in my data, but only at the national product-month level. Any national
average effects of detailing are also controlled for by the product-border-time effect. Where there
might be trouble, however, is if firms strategically raise (or lower) detailing at the product-markettime level in exactly the same places where DTC to take advantage of any complementarities or
substitutabilities.
Two points help address the concern of omitted variables. First, using the border strategy, for
such detailing to be a problem, firms would have to instruct detailers to stop detailing increases
(decreases) exactly at the borders of the television market areas, which seems unlikely. Next, there
is a literature (Manchanda et. al. 2004) suggesting that detailing is largely determined by practice
size, which is effectively controlled for in the treatment group fixed effects. Finally, national
detailing is much more stable over time than is national television advertising. As a robustness
check, I have taken the national detailing data and assumed it was distributed across DMAs in
exactly the same proportions as DTC advertising. This attribution of detailing constitutes a 'worst
case scenario' in terms of how much detailing would bias any estimates of the effects of DTC.
Doing this does not significantly change the results. This robustness check is provided in the
appendix.
Finally, the identifying assumption of difference-in-differences could be violated. It could be that
the difference-in-differences model fails the parallel trends assumption, invalidating the differencein-differences design. To address this concern, I have conducted a placebo test. Using data on
DMA level television advertising of over-the-counter sleep aids as a placebo treatment, I find no
economically significant effects. Details for this robustness check are in the appendix.
Why the border strategy? A more conventional identification strategy in the discrete choice
literature is to use an instrumental variables approach, as in Berry, Levinsohn and Pakes (1995),
henceforth BLP. The main identifying assumption for the validity of the BLP instruments is that
3
However, the Dartmouth Institute has drawn primary care commuting zones which describe how far Medicare
patients travel to see their physicians. It is very rare for a commuting zone to cross DMA lines- only about 1% of
primary care commuting zones cross DMA borders at all, and those that do tend to be predominantly in only one
DMA. This should minimize the measurement error worry. Further explanation of the Dartmouth Institute commuting
zones is provided in the appendix.
22
the characteristics of competing products within a market are exogenous, thus the changing competitive structure of the market may be used as a supply side instrument for demand side choice
variables. In the market for prescription drugs, entry happens in all markets simultaneously by all
products, thus use of the BLP instruments would eliminate any spatial variation, which is a main
attribute of the data I am using. Furthermore, it might be unreasonable to think that competitor
characteristics are exogenous in this setting. It stands to reason that as consumers demand more
antidepressants with fewer of some kind of side effects that firms might well focus research and
development on that kind of product. I am more comfortable with assumptions required by the
border strategy in addition to the fact that it allows me to use the spatial variation in the data.
1.3.2
Results
Using the identification strategy at the border outlined, the estimating equation including fixed
effects becomes:
log(Qjmt)
=
)Jog(Qjm,t-1 + yja+wn
~ ijmt crooss\2
)'' +Y4
jmt (
wn cross
Yajmt aimt
ajbqg+ jbd
Tjb
+jbd
(1.2)
where ajbq is a product-border-quarter fixed effect and ajbd is a product-border-DMA fixed effect.
The ajbq effect will sweep out all variation that is not between two areas that are on opposite sides
of a DMA border.
Partialing these fixed effects out makes the identifying variation within product j local advertising
that is over and above the average on its side d of the border b and over and above the average local
advertising of product j in time period t in all counties on that either side of border b.
Results of the above regression are provided in Table 1.2. Most notable is that both rivals' and own
advertising has a positive and significant effect on demand. Rivals' advertising hits decreasing
returns to scale more slowly than does own advertising. Also, the cross partial indicates that rivals'
advertising works a firm down its marginal revenue curve with respect to advertising, but not as
much as own advertising does. Finally, there is evidence of persistence, though the persistence
parameter is not especially large. This is consistent with the idea that there is much experimentation to find the correct fit between patient and treatment in the depression space. If the category
expansive effect of advertising is more persistent, it will be another potential avenue for firms to
under-invest relative to a co-operative.
23
t
+Emt
1.4
1.4.1
Model
Demand
I propose a multi-stage choice model where advertising may affect the consumer's choice at each
stage. A consumer arrives at her desired end product through a sequence of choice problems. First,
the consumer chooses between entering the category (inside option) and the outside option. If she
chooses to enter the category, she chooses which subcategory of product she wants. Finally, given
her choice of subcategory, she chooses which product to purchase. This process can be extended,
in principle, to have any number of stages.
In the specific case of prescription antidepressants, this is plausible. A consumer first decides
whether she has a problem with depression, goes to the physician and together with the physician,
determines which class of drugs would be most suitable (perhaps considering interactions with
other drugs taken) and which product in particular is the best choice (perhaps having to do with
what is on her formulary).
I define 'utility' u of consuming the inside option, as a function of total advertising stock as well
as other market level factors:
3
Uiimt = 171 (Almt) +fPiXimt + alt + aim + lmt + Eilmt = i + Eilmt
(1.3)
In this specification, 1 denotes the inside option versus outside option, m denotes market and t
denotes time period. I define [, as an increasing function of Almt, total advertising stock of all
inside option products in market m at time t, alt is a time specific taste for the inside option, am is
a market specific taste for the inside option, and Xmt are market-time characteristics.
For the next stage, I define the relative utility vi of subcategory n conditional upon the choice of
the inside option as a function of the total advertising stock in subcategory n, Anmt, as well as other
subcategory-market-time level factors:
Vinmt = F 2 (Anmt)
+ P2Xnmt + ant + anm + 4nmt + Einmt =
3
n l + Einmt
(1.4)
Finally, relative utility wn of product j conditional upon the choice of subcategory n, is defined as
a function of advertising stock of product j, Ajmt, and other product-market-time level factors:
24
W jmt
1 3 (Ajmt) + I3Xjmt + aft + ajm + gjmt + -ijmt = 3jIn+ eijmt
(1.5)
Dynamics enter the model through advertising carry-over. That is, a consumer may remember an
advertisement from a previous period, and that advertisement may affect current period demand.
In general, advertising stock is a function of current period advertising (measured in expenditure
per 100 capita) in choice stage s, as, where s c {l, n, j}, last period's advertising stock, Asm,t-iand
a parameter governing depreciation over time, AS.
Asmt = f(As, Asm,t-1, asmt)
(1.6)
I set each disturbance term, e, to be iid extreme value type I. Given the logit errors, I compute a
closed form solution for shares. The unconditional share of product j in subcategory n is a product
of conditional shares, where market and time subscripts have been suppressed:
si = (sin)(sni)(S)
(1.7)
Those conditional shares take logit form:
1- exp(31 ;n)
" 1+EjEnexp(3jjn)
=in
exp(3 1 )
I + En exp(JIz)
sn1
s1
exp(5)
+ exp(83)
(1.9)
(1.10)
I note here that the equations at each level are independent of each other. I allow each level to
have a different persistence As, and different effects of advertising, y. I also note that while I call
the latent variables at each level 'utilities', it is not essential to interpret them literally as such.
In this paper, I will not be computing consumer welfare, and it is likely that the latent variables
contain a combination of patient and physician utility, information and persuasion. The purpose of
the choice model is to guide the firm decision problem. While it is possible that these parameters
25
could be related across levels by some kind of summing up identity (as they would if each of
the equations were only utility and consumers maximized utility), I do not restrict them to be, as
discovering the relative magnitudes of advertising effects at each level is a main question of this
paper.
For intuition, consider what happens if a single product, Zoloft, raises advertising in a market while
everything else remains constant. That advertisement may have three effects. First, it may raise
the probability that a consumer purchases any antidepressant. That effect is expressed through the
top level equation, increasing almt, which increases Almt which then in turn increases Fi (Aimt).
Next, the information in the advertisement may push the consumer towards the subcategory of
antidepressants that Zoloft is in over another, as the commercials often contain information about
mechanisms and side effects, which are highly correlated within subcategory. The Zoloft advertisement increases anmt, which increases Anmt, which in turn increases F2(Anmt). The marginal
revenue will depend on the shape of the curve and the amount of advertising done by other products in the same subcategory. Finally, the advertisement may have a pure business stealing effect.
By increasing ajmt, Ajmt and F3 (Ajmt) increase to take share away from other products within the
subcategory.
Derivatives and Elasticities Given product shares in Equation 1.7 and the logit structure, we can
get the derivative of sj which is in subcategory n with respect to new advertising, ak, of product k
which is in subcategory n' by using the chain rule and the typical logit derivatives:
dsj -Sn1nIdsl
S
a
dak
d k
snji I+n
dak
dsiln
's' dak
solving this out using our specification on shares, we get derivatives,
si [d1
s=
dk
_
dak
sj[n
+
0-sl)
(I - S
i)
( - SO) -
-(1
-sin)]
- snII) + d(
_)+
' ( - sni)
dr'
skIn]
dak
asdeisg
and advertising elasticities equal to,
26
j=k
j#
k, &n = n'
jea k&n
n'
(1.12)
Oak (ak
_k
s
(-
S;)+
Oak
a [
')+ (
1)
Snj1) -
aO
1 i--s)-
+ Q(1 - sjln)]
Sk~n
dakkn
j=k
kn'-n
(1.13)
n
j # k&n = n'
sn
From these equations, we can see that firm benefits from own advertising may flow through expansion of the category, as is denoted by the term sj
I (1 - si), through expansion of the subcategory
ini sJda210 - Sni;) and through business
- s).
da' stealing within the nest in sjQ(1
1 (I -l
SaJFr
Firm benefits
bnft
from rivals' advertising in the same subcategory may flow through expansion of the category in
s d] (1 - si) or through expansion of the subcategory in s 1H (1 - s
while this same advertising may hurt through business stealing within the subcategory in -sj daskin. Advertising from
rivals in other nests may benefit the firm only through the expansion of the inside option, but may
hurt through expansion of the other subcategory at the expense of the firm's subcategory. It is
worth noting that this structure fully allows for advertising that is a pure category expansion (i.e. if
dl=3
OV j), for advertising that is pure business stealing (i.e. if dl- - or= OVj ), or any2
daj,
daj
daj2
daj,
thing in between, including cross subcategory substitution. It is also possible that rival advertising
outside of the subcategory could help more than inside of the subcategory if da
small and
is sufficiently
ar
a 3 is sufficiently large or vice versa. What is restricted is that a firm's own advertise-
ments may not help another firm more than it helps itself in elasticity terms. In the most extreme
scenario, it is pure category expansion and helps all firms equally. Whether advertising provides
positive or negative spillovers depends on the relative strength of the market expansion and the
business stealing channels and is a result of estimation rather than an assumption of the model.
Notable is that through the category expansion channel, rivals' advertising moves a firm's marginal
revenue with respect to advertising downward.
However own advertising must move a firm's
residual marginal revenue curve even further downward, as there are decreasing returns at the
conditional share level as well. Assuming that the effect of advertising is positive at all levels, the
primary effect of own advertising is stronger than that of rivals' advertising and decreasing returns
to own advertising are more severe than decreasing returns to rival advertising. This is a testable
implication of the model.
1.4.2
Supply
Firm free riding may be an optimal strategy in a game with positive spillovers. Mixed strategy
equilibria may be an equilibrium in a specific game with a fixed cost to advertising each period.
27
To investigate the incentives generated by the demand problem above, I assume that firms play
a simultaneous game, choosing advertising each period while taking into account expectations of
rival behavior and the dynamic effects of advertising. This enables me to analyze the magnitude
of potential under-provision levels on average generated by positive spillovers.
The Firm's Problem A forward looking firm maximizes a discounted stream of future profits
with respect to advertising. Suppose advertising has a constant marginal cost kjm, the market size
is ptm , prices are pj, marginal production costs are mcj and the discount rate is #. Further, suppose
the exogenously given set of products that a firm f has in the market is denoted by 4Df and the full
set of products in the market is denoted by Uf IDf.
Per period profit for the firm will be a function of advertising stock at the product, subcategory and
category level Amt ={Ajmt, Anmt, Aimt}, which is a function of the vector of current advertising
for all products, amt, advertising stocks in the previous period Am,t-1, and persistence parameters
Xj, )n, and X. Per period profit is also a function other state variables a, a product-market-time
specific constant marginal cost of advertising kjmt, and product-market-time specific iid demand
shocks 4:
Afmt = (
(Pjt -mcjt)Amtsjmt( Amt, a,
)
kjmtajmt.
jE~f
The firm's problem is to maximize the stream of future profits for all products in its portfolio.
Advertising is set prior to the realization of any demand shocks, so only expected profits matter for
foirm choices. Expected per period profits are:
7rfmt =
E
f (Pjt -
Cjt)kmts jmt(Amt, a, 4)p(4)d - kjmtajmt.
(1.14)
jE(Df
The timing of the game is as follows. At the start of each period, the state of the market gmt, is revealed to all firms. Based on that state, firms will make advertising decisions ajm(gmt) = ajmt. The
state variables could include anything in the past history of the game, but I will restrict attention to
games where payoff-relevant state variables are the only ones that matter. Given the demand system
provided above, state variables will include the current advertising stock Amt ={Ajt, Anmt, Almt}
and the current market condition variables at ={ajt, ant, at}. These variables include information about the time to patent expiration as well as other information that is common to the products,
28
subcategories and the category nationally at a given time such as expected rival or own product entries. After the state variables are observed, the firm makes its advertising decision, the current
period demand shocks are relaized and the firms collect their profits for the current preiod.
The strategy profile a =(1, ... , a) contains the decision rules of the firms. Firm f's expected
present discounted value of all profits given the current state gmtand the full strategy profile a is
Vfm(gmt c) = E[
$ T-I 7rf(gms, Gf (gms))Igmt].
(1.15)
jE'Df T=t
Firm f chooses a sequence of advertising levels Gf , which are state dependent, to maximize
Vfm(gmt a). I will assume that firms form expectations regarding future states and thus future
strategies given current period state variables. As such, firm f will have a Markov strategy, afm :
g7f(g) = (ajm)vjE4)f. To assess the strategies of competitors, firm f makes an assumption
about a-f and chooses af.
The equilibrium concept used will be Markov perfect equilibrium. A Markov perfect equilibrium
is a set of strategies that form a subgame perfect equilibrium where strategies may only be conditioned on payoff relevant state variables and on the current state of the game. In particular,
af c argmax{Vfm(gmt Ia)}
(1.16)
for all states gmt, firms f, and actions ajm. That is, each firm will maximize the discounted sum
of expected profits given beliefs about competitors behavior, and each firm's beliefs are mutually
consistent in equilibrum. Using MPE makes the allowable strategy space relatively simple. I will
further restrict attention to pure strategies in advertising.
Other Choice Variables Other choice variables are excluded from my model for reasons of
expositional clarity and data limitation. Prices are observed only at the product-quarter level and
detailing only at the product-month (national aggregate) level. While it is indeed conceivable that
firms would maximize profit jointly with respect to both price and advertising, the purpose of this
study is to isolate the advertising decision. In many settings this may not be possible. However,
in the pharmaceutical market that I study, the institutional detail of insurance as an intermediary
makes pricing largely decided by a collection of bilateral negotiations between the firm, insurance
companies and health care providers. The resulting prices observed in the data are highly persistent
29
over time and show no correlation with observed advertising. As such, I will view the pricing
decision as orthogonal to the DTC decision, conditional on product-market and product-time fixed
effects. Under this assumption, there will be no bias in the estimation of the effects of DTC rising
from omitting price from the estimation. A discussion of pricing in this market as well as the
analysis showing that it is not correlated with advertising is available in Appendix A.4
Detail advertising to the physician is another choice variable and type of marketing pursued in
the pharmaceutical world. It is observed in the data, but only in national aggregates for each
product over time. Only having variation by product over time is not optimal, as firms could
strategically detail in response to DTC in a particular market. Due to the data limitations, in
my specifications I will view detailing as orthogonal to DTC conditional on product-time and
product-market effects. Manchanda et. al. (2004) find that high volume prescribers are detailed
more than low volume prescribers without regard to their responsiveness to detailing. This effect
is controlled for in the product-market fixed effects. Indeed, when looking at the time series of
national detailing for a product, it appears much more stable over time than DTC. However, we
might still be worried about strategic detailing. I will do robustness checks, making assumptions
about the spatial distribution of detailing and find that it is unable to explain the estimated effects.
Omission of pricing and detailing may still affect some conclusions in counterfactual simulations
and estimations and discussion of those concerns will be explored in those sections.
1.5
Empirical Specification and Estimation of the Model
1.5.1
Demand Specification
I define the advertising stock at each level s, where s E {l, n, j} is either the category level, subcategory level or product level, to be a lag of a nonlinear function of current advertising, similar to
Dube et. al. (2005).
Asmt =
02s-log(l
+asmz)
(1.17)
4
Without including price in the analysis, I will not be estimating a price elasticity. As such, the computation of
optimal advertising will be a bit different in spirit than the classic Dorfman-Steiner (1954) problem which suggests
that the advertising to sales ratio should equal to the advertising elasticity to price elasticity ratio. This theorem was
modified to allow for dynamic advertising in Nerlove and Arrow (1962). Neither of these formulations considers
positive spillover effects, and free riding incentives may make these policies sub-optimal.
30
Specifying advertising stock as a concave function of each period's advertising allows the firm's
problem to have a well behaved optimum. Other functional forms were explored and none changed
the results in any significant way.
The advertising stock enters into the utility specification linearly at each level.
]Fs(Asmt) =
sAsmt
(1.18)
I account for all product characteristics other than advertising with a rich set of fixed effects, as the
only pieces of data that vary at the choice level, DMA and time levels are shares and advertising.
Substituting equations (17) and (18) into equations (1)-(3), for the market level 1 obtain:
uilmt = yl[E'
0 2-Llog(l+
alm)]+ alt + alm+ imt + eilmt
(1.19)
The conditional utilities for the subcategory and product levels are defined analogously. From here,
it is notable that current period advertising enters the utility function in a concave manner, so the
firm maximization problem is well behaved.
1.5.2
Transforming to a Linear Problem
Following Berry (1994), at each level of the problem, I specify an 'outside good', take the log of
the market share and subtract from it the log of the outside option share. This results in a linear
form.
At the category level the outside good is naturally defined as the population not filling a prescription
for an antidepressant in month t in market m:
log(slmt) - log(somt) = y7 [Et
-- log(1 + almr)] + alt + alm + Imt
(1.20)
At the subcategory level, the outside good will be defined as the subcategory of older style TCA
antidepressants. The share of a subcategory conditional on being in the inside option follows:
log(snmtlj) - log(somt~l) = 2
%t7Zlog(1
+ anm)]+ ant + anm + gnmt
31
(1.21)
At the product level, the outside option in each nest will be the set of all products that never advertise on television. The product share equation conditional on already having chosen subcategory n
is:
log(sjmtn) - log(som,1n)
=
3(4=d,' 'log(I+ ajs)]+ ajt + ajm + jmt
(1.22)
Now, using these to solve for inside option shares shares in time t - 1, and substituting that back
into the expression for time t shares yields,
log(smt) -log(somt) =
2
(log(slm,t-i) -log(som,t-1))+ylog(1 ±amt)+Oit+
0im+ vimt (1.23)
where
Olt = alt - X, apt-I
(1.24)
is a inside option-time specific taste or quality parameter.
and
elm = alm - Alxalm
(1.25)
is the category-market specific taste parameter. Finally,
Vlmt =
lmt
-
Al'lmt
(1.26)
is a market-time specific demand shock. Equation 1.23 is precisely a lagged dependent variable
with fixed effects specification as described above, making possible the use of the border identification strategy.
Similarly, subcategory and product level share equations may be specified as:
32
log(snmtl) - og (somt=)
-
log (som,t-1| )) + 72log( + anmt) +
6
nt + Onm + Vnmt
(1.27)
log(sjmtln) - log(somtjn) = Xp(log(sjm,t-IIn) -log(sjmt-
1.5.3
In)) +73log(1
+ajmt)+Ojt +Ojm+vjmt.
(1.28)
Identification and Estimation Strategy
Since the share equations have been transformed to a linear form, estimation may be done by
OLS. A notable problem in estimating this equation is that advertising is a firm choice variable
determined in equilibrium and is thus endogenous. As such, I will take advantage of the discrete
nature of DMAs to make use of spatial variation as described in Section 1.3.
In particular, I specify the estimation equation as:
log(sjmt) - log(somt)
=
X (log(sim,t-1) - log(som,t-1)) +'1ylog(1 + aimt) + 0 lbq -
0
lbm + Vlbmt
(1.29)
where 0 lbq is a border-time fixed effect and 0 1bm is a border, DMA fixed effect. Partialling these
fixed effects out makes the identifying variation at the market level the total advertising in market
m that is over and above the average on its side d of the border b and over and above the average
local total advertising in quarter q in all counties on that either side of border b. The fixed effects
will also control for the product quality terms et and 6 m.
I identify the effects at the other two levels similarly. In the subcategory level, I include fixed
effects anbq and anbm and at the product level, I include fixed effects ajbq and ajbm. Identifying
variation will come at the subcategory level from total subcategory advertising that is above and
beyond the historical advertising in its market and above the border average in the current time
period. At the product level, identifying variation will be advertising for product j that is above
and beyond advertising for product j on the border in quarter q and above and beyond the average
over all time in market m. No between product variation in advertising will be used to identify the
advertising parameter.
33
Table 1.3 has variable definitions and summary statistics for those variables that will enter the
estimation.
1.5.4
Demand Results
Effects at Each Level Results are presented in Table 1.4. The effect of advertising stock on
demand at each stage of the decision is positive. The strongest effects are at the category level,
deciding between inside and outside option and at the product business stealing level. Effects
at the subcategory level are not significant, but it is notable that there is only advertising in two
subcategories, with most of the advertising happening in the SSRI subcategory. The small and
insignificant effect at the subcategory level is not surprising, as it seems unlikely that patients
would have good information about what separates the subcategories. Table 1.5 presents short
run demand elasticities of current advertising showing that the category expansive properties of
advertising dominate the business stealing effects and all cross advertising elasticities are positive.
This finding is consistent with the identified positive spillovers in the reduced form.
Persistence Persistence is highest at the category level- that is getting someone into consuming
antidepressants at all. The persistence parameter of 0.68 implies that 90% of the effect dissipates
within six months. Meanwhile, the 0.33 persistence parameter on the bottom level implies that
90% of the business stealing effect of an advertisement dissipates within only two months. This
makes advertising in the long run more of a category expansion than a business stealing tool.
This is consistent with the common wisdom that antidepressants are subject to a high degree of
experimentation. If a patient tries one and finds the side effects unbearable, she might well switch
to another one rather than quitting altogether. It is also consistent with a limited memory view of
advertising. Since advertising for pharmaceuticals on television usually contain a lot of information
about the condition, the mechanisms of action and the side effects and these characteristics are
highly correlated within category, a consumer might well remember seeing an advertisement about
depression without remembering which brand was advertised. This high persistence at the category
level relative to the product level is another source for potential underinvestment in advertising
relative to a co-operative.
34
1.6
1.6.1
Supply and Counterfactual
Supply Implications of Positive Spillovers
The demand results above imply that the incentive to invest in advertising is dampened by positive
spillovers for two reasons. First, advertising provides benefits to rival firms which are not internalized by the advertising firm. Second, rival advertising lessens the incentive to advertise through
the incentivize to free riding on the efforts of rivals.
Internalization
To further illustrate the effects of advertising over time on rivals, consider an
impulse response graph in Figure 1.10. The purpose of this graph is to follow the effect of a
marginal dollar per 100 capita spent by Zoloft in January of 2002 on both Zoloft and total market
prescriptions for the subsequent year. The top downward sloping curve is the marginal effect
of Zoloft advertising on total market prescriptions, while the bottom downward sloping curve is
the marginal effect on Zoloft prescriptions. The upward sloping dashed line is the ratio of the total
market effect to the Zoloft effect. A marginal dollar per 100 capita of Zoloft advertising would lead
to a contemporaneous increase of about 70,000 antidepressant commercials, only about 20,000 of
which would be Zoloft. Further, as we follow that effect through time, the effect on the total
market is more persistent, and the marginal effect of Zoloft advertising goes more and more to
other products. There is a large contemporaneous positive spillover that intensifies through time.
Zoloft has no incentive to internalize the benefits it bestows upon other firms, and thus will under
invest in advertising relative to a co-operative controlling advertising in the whole market.
Free Riding
To illustrate the free riding incentive, I consider three marginal revenue curves and
a horizontal marginal cost curve in Figure 1.11 given the demand parameters estimated in the
previous section, but for a single point in time and for a single product. In the figure, I consider the
perspective of Zoloft in January 2002 in the Boston DMA.
The top curve is the marginal revenue for Zoloft if all competitors set advertising equal to zero. Notably far below that curve, the middle curve is the marginal revenue curve of Zoloft if competitors
combine to advertise $3 per 100 capita, which is about the average competitor advertising Zoloft
sees in the Boston DMA during the time that it advertises. Finally, the lowest curve depicts the
marginal revenue with respect to advertising of Zoloft when its competitors advertise $10 per 100
capita, about the maximum it ever faces from competitors in the Boston market. Notable from the
curves is that the marginal revenue curve of Zoloft takes a significant hit as its competitors advertise
35
more. In fact, when competitors advertise up to $10 per 100 capita, it is almost not worthwhile for
Zoloft to advertise at all. There is a clear incentive for Zoloft to free ride as competitors advertise
more and more.
In the next subsection, I will more systematically explore these incentives for our realized antidepressant market.
1.6.2
Cost Computation
Given the demand estimation above, I can solve for the marginal costs associated with advertising
implied by the outlined supply model. In spirit, this means that I am assuming that firms are
behaving optimally, computing the discounted sum of all current and future marginal revenues
in each market in each month for each product given the expected behavior of rivals, demand
estimates and price and production cost and setting that equal to the expected marginal costs.
Since three firms advertise and there are multiple state variables, rather than using value function
iteration and a bellman equation approach, I will approximate the infinite horizon problem with a
finite horizon one. Firms will explicitly compute expectations of future periods and set marginal
revenue equal to marginal costs. The estimated demand system combined with firms' observed
advertising decisions provide all of the necssary information to compute marginal revenues, and
marginal costs will be a free parameter that is pinned down by these marginal revenues. Further
details of how these costs are computed are provided in Appendix E.
While at first blush, it may seem as though the marginal cost of one dollar of advertising should
be one dollar, it is possible that there are opportunity costs to firms to advertising. They could
have other product classes to advertise and limited advertising budgets set within the organization.
They could have concerns about the public or regulator perception of advertising drugs too much
(Ellison and Wolfram 2008). As in Ellison and Ellison (2011), they may be strategically trying
to deter entry by making the market seem small. Finally, the observed marginal production costs
and prices I use may provide an incomplete picture of markups. If thre are unobserved discounts
to large insurers, I will overstate the marginal revenue from an additional prescription and hence
overstate marginal costs. In principle, they could also have other marginal benefits to advertising.
There may be cross class spillovers. As such, I will allow the 'marginal cost' of advertising one
dollar to be either more or less than one dollar. Since not all firms advertise in all markets and in
all months, I will only be able to bound costs below where firms do not advertise. In addition to
potentially having a level higher than one, marginal costs may vary over market and time. This may
be due to difficulty in getting advertisements on the air in certain markets at certain times due to
36
local elections, local demand shocks for other products that causes them to want more advertising
space or organizational costs to communicating with particular stations.
Summaries of estimated marginal costs are presented in Table 1.6 for each product.
While the marginal cost of one dollar of advertising is consistently more than one, typically centered around four, it varies across product and market. In particular, to explore the variation, I
regress these computed costs on month fixed effects, DMA fixed effects and product fixed effects.
The results of the regressions are presented in Table 1.7. When product fixed effects only are inclued, the R2 of the regression is 0.20. Adding in DMA fixed effects raises the R2 to 0.43. Finally,
adding in month fixed effects raises the R2 to a whopping 0.86. This suggests that while there are
product and market specific variations in the difficulty of getting ads on the air, a simple time effect
has much more explanatory power. Given that most of the big advertisers, such as soft drinks, beer
and car insurance tend to advertise nationally, it seems likely that large demand by those firms for
specific months crowds out advertising for antidepressants.
These computed costs will be used to compute the counterfactual of co-operative advertising.
Pulsing Strategies As is noted by Dube et. al. (2005) and is evident from the figures of observed
advertising, firm advertising is often highly variable and unpredictable. While the model of Dube
et. al. rationalizes the spikes and zeros in advertising with an advertising stock function that
is s-shaped in current advertising, I rationalize those spikes and zeros by changes in costs over
time. Firms might plausibly have varying opportunity costs over time and markets due to portfolio
changes or organizational concerns.
It is possible that the supply side game could generate equilibria with entry into advertising in
mixed strategies, if for example there were fixed costs to advertising each period, which are not
specified in this model. However, as the explanation of pulsing is not the main focus of this paper,
details of such possibilities will be left to future research.
1.6.3
Counterfactual Assumptions
Given the positive spillovers of advertising, we should expect that the incentive to invest in advertising is lessened by both a failure to internalize the benefits of advertising on rivals and by
an incentive to free ride. As such, I will consider a counterfactual scenario whereby the entire
market is allowed to set advertising in a single optimization problem, thereby taking away strategic considerations. For such a scenario to work, cooperation would need to be not only allowed,
37
but enforced in some way. Antidepressants (and other pharmaceutical products) might be able to
cooperate similarly through non-profit organizations called patient advocacy groups. Such groups
are focused on educated patients on specific diseases and treatments. The specific mission of these
types of organizations is to educate patients on how, when and why to seek treatment for various
health care needs. Some examples of these types of organizations are the American Cancer Society, the Alzheimer's Association and the Depression and Bipolar Support Alliance. While they do
not tend to advertise on television currently, they might be an ideal facilitator for category level
advertising of antidepressants. As many physicians see depression as an undertreated condition,
commercials that encouraged patients to seek their physicians' advice if they encounter the symptoms of depression might be allowed by regulators. The ability to contract on coordination would
allow firms to overcome the free riding problem and provide advertising, even without a brand
level component.
For the purposes of the counterfactual, assume that the advertising firms in the antidepressant
market cooperate to make a common non-branded category advertisement for antidepressants, facilitated by a patient advocacy group. The effect of those advertisements is equal to the category
level effect of the branded advertisements estimated above.
The co-operative solves the firm's problem in each month and in each market but takes the revenues
of every product into account. It takes as given the computed marginal costs of advertising each
product from the previous section. I assume that the marginal co-operative advertisement has cost
equal to the average of the computed marginal costs of each product, weighted by the amount those
products advertise in a given period. I also assume that the co-operative can forcast margins and
populations in future periods and form expectations of future demand conditions. The co-operative
maximizes equation (15), but since all products in the market are included in Of, strategic response
is not necessary.
Similar to the strategic problem, I will approximate the co-operative's infinite horizon problem
with a finite horizon one and model future expectations. Given that, I write down a first order
condition for the firm and solve for advertising. Details of the computation algorithm are provided
in Appendix E.
Notably missing from this firm problem is the ability for each firm to readjust pricing and detailing
with the co-operative advertising decisions, as pricing and detailing are not included in the model.
As mentioned previously, prices are uncorrelated with levels of DTC in the observed world for a
given drug over time. There is little reason to suspect that they would change in the counterfactual.
This is because most price flexibility happens in negotiations with managed care organizations
which were less prevalent in the time of this sample and those negotiations happen infrequently
38
and in a disjointed way from the advertising decision. So while DTC may respond to markups in
the model, prices do not respond to DTC conditional on product-market and market-time effeects.
Detailing, on the other hand, could be more troublesome. If detailing were mainly a business
stealing device, as is suggested in Narayan et. al. (2004), it would be complementary to cooperative DTC. A firm would be happy for the co-operative to make the market large and it would
then try to compete over brand share in detailing space. Anticipating this, the co-operative would
shade down its advertising, anticipating that profits would be competed away with detailing.
If detailing were, alternatively, also mainly category expansive, it would be a substitute for DTC.
In response to the DTC, firms would be glad to reduce detailing and let the co-operative foot the
bill for market expansion. The co-operative would anticipate this and shade up its advertising.
Conversations with physicians seem to indicate that they are being detailed about as much as they
possibly can. Even if a firm wants to greatly ramp up detailing, it is unlikely that they will be
able to if for no other reason than the physician only has so many hours in the day. If physicians
are already satiated with detailing, the potential equilibrium effect of competing away all market
expansion with further detailing would not be present.
Also worth noting, since in the observed world firms under invest in advertising, the counterfactual
results in some values of advertising that are not observed in the sample. Since we only estimate
the demand curve for values of advertising within the sample, the out-of-sample specific numbers
are driven by the assumed functional form of the advertising effect in the model. 5
1.6.4
Advertising
As the business stealing incentive grows, observed total advertising is expected to increase relative
to the counterfactual advertising. As the business stealing incentive dwindles, the free riding incentive associated with the positive spillovers should lead to lower observed advertising relative to
the co-operative's ideal.
I assume that the co-operative can set a number of non-branded co-operative advertisements that
will have the market level effects and persistence estimated in the demand section. It will be able
to do so at the average cost of advertising across all products in that market and period, with
the restriction that the marginal cost of advertising one dollar may not go below one dollar plus
the industry standard 15% agency fee. The co-operative assumes stationarity of the optimization
5
Aternative specifications with quadratic and square root functions of advertising were tried and produced qualitatively and quantitatively very similar results.
39
problem and maximizes discounted future profits in each month. Generics are assumed to have
zero margin.
For illustrative purposes Figure 1.12 shows the observed flow versus co-operative choice on average across all DMAs in the co-operative scenario described above. Figures (13) is the same as
Figures (12) except that it assumes that the marginal cost of advertising one dollar must be exactly
one dollar plus the 15% agency fee.
Notably, the co-operative maintains significantly higher advertising over time than is the observed
outcome. On average, the co-operative advertises five times as much as the competitive industry.
The positive spillovers of advertising seem to be generating a free riding problem for the industry.
The exact numbers are coming from extrapolation on the functional form, as they are often out of
sample from what is estimated. Also, where the difference between counterfactual and observed
advertising is low, computed marginal costs are very high. Where the difference is large, the
computed marginal costs are low. If we assume that true costs of advertising a dollar in the cooperative advertising world do not exceed unity, differences are persistently much higher, with
counterfactual advertising on the order of ten times observed advertising.
1.6.5
Quantities and Profits
In the co-operative, the greater advertising leads to an increase in shares of the inside option by
19.2% and an increase in total industry profit of 19.8% on average. Those averages over time are
plotted in figures (14) and (15).
1.6.6
Discussion
While market expansion and increasing profits in the counterfactual would be viewed as welfare
increasing in many consumer goods markets, there are a few reasons for us to take caution in
the antidepressant market or prescription pharmaceutical markets more generally. First, many
prescriptions are covered by insurance. While many people are getting prescribed and incurring
minimal if any cost, the system is still paying out the marked price, which is quite high. It is possible that the new prescriptions are not justified by the societal cost. However, as many physicians
see depression as an under treated condition, the total welfare benefits could also be very high.
Second, if I have missed important price or detailing complementarities, it might be the case that
all increased profits are competed away after the co-operative sets the higher advertising. If this is
40
the case, the welfare effect is also ambiguous. However, as prices are not correlated with advertising expenditures conditional on product effects and detailing might well have reached satiation,
this concern might not be very large.
The actual net social welfare effect, while interesting, is not identified in this study and is certainly
worthy of further research. Category expansion is typically thought of as consumer welfare improving, as consumers would not purchase something for which the costs exceeded the benefit.
However, in health care, many levels of agency mask the costs and benefits of treatments. As such,
I am remaining agnostic on consumer welfare.
1.7
Conclusions
Using data from the antidepressant market and an identification strategy taking advantage of
both policy and spatial discontinuities, I find that television advertising has significant positive
spillovers. I construct and estimate a model to systematically explore this fact and its implications
on the supply decisions of firms. In particular, I find that the spillovers induce a commons problem
whereby observed advertising is significantly lower than the optimal strategy that a co-operative
would set if it controlled the entire market. A co-operative would set advertising five times as high
as is observed in equilibrium and would increase industry shares by 19.2% and profits by 19.8%.
These findings also speak to some of the controversy surrounding the practice of advertising pharmaceuticals on television. Contrary to some of the criticism, this type of advertising drives consumers into the market and helps all products in the category, including the low cost generics. It is
the proverbial rising tide that lifts all ships. While there is a brand effect, its effect is short lived
while the category expansion effect is persistent. Especially for conditions that are seen as under
treated, this type of advertising could be beneficial.
These findings are potentially relevant to firms, regulators, econometricians and marketers. Firms
might be able to realize gains from cooperation that might be allowed by regulators. In the absence
cooperation, it is important for firms to properly take account of spillovers when deciding advertising policy. Regulators should take into account that content regulation might reduce or eliminate
the firms' incentives to advertise. Finally, it is important for marketers and econometricians to consider the possibility of positive spillovers when building models of advertising impacts on supply
and demand.
41
References
[1] Anderson, I., (2000). Selective serotonin reuptake inhibitors versus tricyclic antidepressants:
a meta-analysis of efficacy and tolerability. Journalof Affective Disorders,58, 19-36.(1)
[2] Avery, R. J., Eisenberg, M., & Simon, K. I. (2012). The impact of direct-to-consumer television and magazine advertising on antidepressant use. Journalof Health Economics, 31(5),
705-718.
[3] Anderson, E., Simester, D., (2013). Advertising in a competitive market: the role of product
standards, consumer learning and switching costs. Working Paper.
[4] Berndt, E. R., Bui, L., Reiley, D. R., & Urban, G. L. (1995). Information, marketing, and
pricing in the US antiulcer drug market. The American Economic Review, 85(2), 100-105.
[5] Berndt, E. R., Bui, L., Reiley, D. R., & Urban, G. L. (1997). The roles of Price, quality, and
marketing in the growth and composition of the antiulcer drug market. The economics of new
goods (NBER Studies in Income and Wealth, vol. 58), Bresnahan, T. F., & Gordon, R. J.
(Eds.). (1997), (Vol. 58). University of Chicago Press., pp. 277-322.
[6] Berry, S. T. (1994). Estimating discrete-choice models of product differentiation. The RAND
Journalof Economics, 25(2), 242-262.
[7] Busch, S., Golberstein, E., & Meara, E. (2011). The FDA and ABCs: The unintended consequences of antidepressant warnings on human capital. NBER Working Paper # 17426.
[8] Ching, Andrew T., and Masakazu Ishihara. "Measuring the informative and persuasive roles
of detailing on prescribing decisions." ManagementScience 58.7 (2012): 1374-1387.
[9] Donohue, J. M., & Berndt, E. R. (2004). Effects of direct-to-consumer advertising on medication choice: The case of antidepressants. Journal of Public Policy & Marketing, 23(2),
115-127.
[10] Donohue, J. M., Berndt, E. R., Rosenthal, M., Epstein, A. M., & Frank, R. G. (2004). Effects of pharmaceutical promotion on adherence to the treatment guidelines for depression.
Medical Care, 42(12), 1176-1185.
[11] Dorfman, R., & Steiner, P. 0. (1954). Optimal advertising and optimal quality. The American
Economic Review, 44(5), 826-836.
42
[12] Dube, Arindrajit, T. William Lester, and Michael Reich. "Minimum wage effects across state
borders: Estimates using contiguous counties." The Review of Economics and Statistics 92.4
(2010): 945-964.
[13] Dub6, J. P., & Hitsch, G. J., & Manchanda, P. (2005). An empirical model of advertising
dynamics. Quantitative Marketing and Economics, 3(2), 107-144.
[14] Duggan, M., & Morton, F. M. S. (2006). The distortionary effects of government procurement: evidence from Medicaid prescription drug purchasing. The QuarterlyJournal of Economics, 121(1), 1-30.
[15] Ellison, G., & Ellison, S. F. (2011). Strategic entry deterrence and the behavior of pharmaceutical incumbents prior to patent expiration. American Economic Journal: Microeconomics,
3(1), 1-36.
[16] Ellison, S. F., & Wolfram, C. (2006). Coordinating on lower prices: pharmaceutical pricing
under political pressure. The RAND Journalof Economics, 37(2), 324-340.
[17] Goeree, M. S. (2008). Limited information and advertising in the US personal computer
industry.Econometrica, 76(5), 1017-1074.
[18] Holmes, Thomas J. "The effect of state policies on the location of manufacturing: Evidence
from state borders." Journalof PoliticalEconomy 106.4 (1998): 667-705.
[19] lizuka, T. (2004). What Explains the Use of Direct-to-Consumer Advertising of Prescription
Drugs?. The Journal of IndustrialEconomics, 52(3), 349-379.
[20] izuka, T., & Jin, G. Z. (2005). The effect of prescription drug advertising on doctor visits.
Journalof Economics & Management Strategy, 14(3), 701-727.
[21] lizuka, T., Jin, G. Z. (2005). Direct to consumer advertising and prescription choice. Working
Paper Available at SSRN 700921.
[22] Jayawardhana, J. (2013). Direct-to-consumer advertising and consumer welfare. International Journalof Industrial Organziation,31(2), 164-180.
[23] Kravitz, R. L., Epstein, R. M., Feldman, M. D., Franz, C. E., Azari, R., Wilkes, M. S., Hinton,
L., & Franks, P. (2005). Influence of patients' requests for direct-to-consumer advertised
antidepressants. JAMA: the journal of the American Medical Association, 293(16), 19952002.
43
[24] Lewis, R., Nguyen, D., (2012). Wasn't that ad for an iPad? display advertising's impact on
advertiser- and competitor-branded search. Yahoo Research Working Paper.
[25] Lewis, R. A., & Rao, J. M. (2012). On the near impossibility of measuring advertising effectiveness. Yahoo Research Working paper.
[26] Manchanda, P., Rossi, P., & Chintagunta, P. (2004). Response modeling with non-random
marketing mix variables. Journal of Marketing Research, 41(4), 467-478.
[27] Murphy, M. J., R. L. Cowan, and L. I. Sederer (2009): Blueprints: Psychiatry, Philadelphia,
PA: Lippincott Williams Wilkins, 5 ed.
[28] Narayanan, S., Desiraju, R., & Chintagunta, P. K. (2004). Return on investment implications for pharmaceutical promotional expenditures: The role of marketing-mix interactions.
Journalof Marketing, 68(4), 90-105.
[29] Nerlove, M., & Arrow, K. J. (1962). Optimal advertising policy under dynamic conditions.
Economica, 29(114), 129-142.
[30] Rosenthal, M. B., Berndt, E. R., Donohue, J. M., Epstein, A. M., & Frank, R. G. (2003).
Demand effects of recent changes in prescription drug promotion. Henry J. Kaiser Family
Foundation.
[31] Sahni, N. (2013). Advertising Spillovers: Field Experimental Evidence and Implications for
the Advertising Sales-response Curve. Working Paper.
[32] Wosinska, M. (2002). Just what the patient ordered? Direct-to-consumer advertising and the
demand for pharmaceutical products. Direct-to-Consumer advertising and the demand for
pharmaceutical products (October 2002). HBS Marketing Research Paper, (02-04).
[33] Wosinska, M. (2005). Direct-to-consumer advertising and drug therapy compliance. Journal
of Marketing Research, 42(3), 323-332.
44
Tables and Figures
Figure 1.1: Antidepressant Commercials Relative to FDA Memo
Total National Advertising Over Time
0 -
ca~
L
V
1996m1
1998m1
2000m1
date
2002m1
2004m1
Figure 1.2: Antidepressant Revenues 1996-2008
Prescription Antidepressant Industry Revenues
0
0
co
.2
M OD
1995
2000
2005
year
45
2010
Table 1.1: Descriptive Statistics: Advertising
Mean
Q25
Median
Q75
0.782
2.012
4.035
0
0
2.284
0
1.496
3.515
1.358
3.505
5.534
Mean
101
2340774
Q25
Median
Q75
903090
1469823
2622567
DTC per 100 capita
Subcategory DTC per 100 Capita
Category DTC per 100 Capita
DMAs
DMA Population
Figure 1.3: Variation Across Three Markets in Advertising
8
n
-
..........................................
...........
1997m1
..
........................
.........
1998m7
200bml
200m7
month
National Boston ..------
2003m1
2004m7
New York
Austin
Figure 1.4: Two Different Advertising Measures: National
National Advertising
-. a
r
C
CLI
.C4
199 6m1
1998m1
--
NI
A
200 1
2002ml
mnonthty date
Commerals
-
46
2004m1
DTC Per Capita
Figure 1.5: Quantity Weighted Average Margins in Boston
'
0
Mw
1996m1
1996m1
2000m1
Month
2002mi
Figure 1.6: Full Sample: Top 101 DMAs
010
maI
Mo2
47
2004m1
Figure 1.7: Ohio and Its DMAs
Figure 1.8: Border Sample: Counties on the Borders of the Top 101 DMAs
III
MbAt
-M
48
Figure 1.9: Variation in Log DTC Net of Fixed Effects, 14% Zeros
0o
01
0
=000"
U-
0D
0D
--1
-. 5
.5
0
Log DTC net Fixed Effects
1.5
Table 1.2: The Effect of Own and Rival Advertisements on Sales
(1)
log(Q)
VARIABLES
0.334***
(0.00746)
0.0240***
DTC
(0.00621)
-0.00216*
DTC 2
(0.00113)
0.0164***
DTCrival
(0.00266)
2
DTC rival
-0.000938***
(0.000252)
-0.00134**
DTCXDTCrival
(0.000631)
yes
Product-Border-Time
yes
Product-Border-DMA
316,428
Observations
0.955
R-squared
DMA clustered standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
lagged log(Q)
49
Table 1.3: Descriptive Statistics: Border Sample, 1997-2003
Mean
Q10
Median
Q90
0.190
0
0
1.033
0.447
0
0
1.682
0.817
0
0.921
1.987
Number of Border Experiments
Number of DMAs
153
97
LOGDTCproduct
LOGDTCnest
LOGDTCmarket
LOGDTC : log of one plus dtc expenditures per 100 capita
All are defined at the experiment-DMA-month level
Table 1.4: Results of Base Model
VARIABLES
Category Level
Subcategory Level
Product Level
0.0472*
(0.00665)
0.684*
(0.0299)
23,091
0.948
0.0093
(0.00719)
0.282*
(0.0116)
93,284
0.935
0.0223*
(0.00756)
0.330*
(0.0140)
60,980
0.960
adstock
persistence, X
Observations
R-squared
DMA clustered standard errors in parentheses
* p< 0 .0 1
Table 1.5: Short Run Advertising Elasticities
Products
Paxil
Paxil CR
Prozac
Prozac Weekly
Wellbutrin SR
Wellbutrin XL
Zoloft
Paxil
Paxil CR
Prozac
Prozac Weekly
Wellbutrin SR
0.037
0.016
0.0092
0.0088
0.014
0.017
0.013
0.019
0.029
0.021
0.016
0.020
0.0088
0.014
0.017
0.013
0.019
0.016
0.0080
0.018
0.012
0.017
0.013
0.020
0.012
0.0097
0.0068
0.021
0.019
0.013
-
-
0.017
0.013
50
Wellbutrin XL
Zoloft
Outside Option
-
0.021
0.016
0.0092
0.0088
0.014
0.017
0.027
-0.023
-0.015
-0.011
-0.0080
-0.014
-0.018
-0.015
0.010
-
-
0.035
0.010
Figure 1.10: Impulse Response Effect of Zoloft Advertisement on Own and Total Prescriptions
Marginal Effects of Zoloft Advertising
00
---------------------0(
400
C'
Co
20
20
4
2
4
2
6,
4-/
0
10 0
0
0m
Marinl
Mar in
l
00m
02m
602
eveuefAdvertising
$o
per10 Cait
eve
ue
fAdvertising
51
1
8
fo
ZoloftCait
10
Table 1.6: Marginal Cost Distributions By Product
Paxil
Paxil CR
Prozac
Prozac Weekly
Wellbutrin SR
Wellbutrin XL
Zoloft
Mean
Q10
Median
Q90
3.661
3.129
4.063
1.001
2.287
1.246
2.864
1.939
1.331
1.588
0.325
0.933
0.764
1.398
3.308
2.911
3.834
0.866
1.537
1.176
2.551
5.835
5.205
6.888
1.921
5.617
1.845
4.741
Table 1.7: R 2 of MC Regressions
Product FE
DMA FE
Month FE
R2
yes
yes
yes
-
-
yes
-
yes
yes
0.20
0.43
0.86
Figure 1.12: Counterfactual versus Realized Advertising on Average
0
0
C)
096m
1998m1
2000m1
Month
Co-operative Advertising
52
2002m1
2004m1
Observed Market Advertising
Figure 1. 13: Counterfactual versus Realized Advertising on Average, MC=$ 1
0
Mont
0
0
R-
0
1996ml
199 ml
2006m1
Month
Co-operative dvesg
53
2002ml
2004ml
Observed Market Savetsin
Figure 1. 15: Counterfactual versus Realized Antidepressant Profits
0D
0)
C0
60
C0
0>
0
1996ni I
1998m1
2000m1
Month
Co-operative Profits
2002m1
2004m1
Observed Market Profits
Appendix A - Orthogonality of Pricing and Detailing Decisions
A.1 Pricing
Below is the regression of television commercials on prices. As can be seen the point estimate is
very small and insignificant with respect to both own and cross advertising. Just for perspective,
the average unit price of a branded drug is about $3.60 over the course of the sample and the
average DTC per capita of those drugs that advertise over the course of the sample is $0.0055.
Raising DTC per capita by $0.01 is associated with a price decrease of $0.03. This is very small
economically. Also interesting is that just a time trend, a product fixed effect and a dummy for
patent expiration can explain prices with R squared bigger than 0.99. Prices seem quite sticky,
especially relative to advertising in this market.
54
Table 1.8: Predicting Prices with Advertising
(1)
VARIABLES
realunitprice
DTC
-0.0518
(0.0437)
-0.0332
(0.0248)
0.00963***
(0.00313)
-0.210
(0.149)
yes
DTCother
time
expired
Product FEs
Observations
1188
R-squared
0.994
Product Clustered Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
A.2 Detailing
To address the potential omitted variables bias problem in detailing, I assumed that the observed
national detailing was geographically distributed in exactly the same way as television advertising.
That is, I computed a DMA fraction of television advertising as the expenditures per capita divided
by the sum over all markets of the expenditures per capita. As detailing and television advertising
are, in general, positively correlated, I multiplied that fraction by national detailing totals to get
a 'worst case scenario' to see how much detailing could possibly affect the estimated effect of
television advertising. If detailing and television advertising were negatively correlated, I would
want to assume detailing was inversely distributed to this.
The results of this 'worst case scenario' are presented in Table 1.8. None of the estimates on
television advertising are statistically distinguishable from the baseline estimation. I will exercise
caution in interpreting the coefficients on detailing, as they are not the actual detailing numbers at
different localities.
55
Table 1.9: Results of Model including Worst Case Detailing
VARIABLES
Category Level
Subcategory Level
Product Level
0.0197***
0.0102
0.0450***
(0.00791)
(0.00731)
(0.00666)
0.0124***
0.00313
0.0316***
logdetail
(0.00946)
(0.00812)
(0.0117)
0.327***
0.280***
0.684***
persistence
(0.0143)
(0.0116)
(0.0302)
147,443
93,280
23,070
Observations
0.960
0.935
0.948
R-squared
DMA clustered standard errors in parentheses
adstock
*** p<0.01, ** p<0.05, * p<0.1
Appendix B - Placebo Test
With a difference-in-differences model, the assumption of parallel trends in the outcome variable
absent the treatment is required for a valid estimation. One way to assess the validity of this
assumption is through the use of a placebo test. In this case, I will use advertising for over-thecounter (OTC) sleep aids as a placebo treatment. This is an ideal placebo for two reasons: first,
it varies at the same level as antidepressant advertising- at the DMA month. Next, OTC sleep
aids need not be prescribed by a physician, so we should not expect a "going to the doctor" effect
of advertising to be present in OTC advertising. I will use the same identification strategy, but I
will also include OTC sleep aid advertising as a treatment. The results are below. None of the
coefficients on OTC sleep aid advertising is statistically significant at the 5% level or economically
important at any level.
56
Table 1.10: Results of Base Model with Placebo
VARIABLES
Category Level
Subcategory Level
Product Level
0.000471
0.0000565
(0.000482)
(0.000397)
0.280***
0.684***
persistence
(0.0116)
(0.0304)
93,280
22,826
Observations
0.935
0.948
R-squared
DMA clustered standard errors in parentheses
OTCSLEEPAD
0.00218*
(0.00113)
0.323***
(0.0139)
147,443
0.960
*** p<0.01, ** p<0.05, * p<0.1
Appendix C - Alternative Sample Selection
One might worry that the effect of advertising in the border sample counties differs systematically
from the non-border counties or that the effect of advertising in rural areas would be much different
than the effect of advertising in urban areas. To think about these concerns, I have repeated the
analysis with several alternative sample selections.
C.1 The Urban Rural Divide
C.1.1 Without the Northeast Corridor and Other Urban Areas
It seems the less urban areas show very similar results to the full border sample. The category and
product level effects are larger, but not significantly different from the full sample of borders.
57
Table 1.11: Results of Base Model without Northeast Corridor and Other Urban Areas
VARIABLES
Category Level
Subcategory Level
Product Level
adstock
0.0513***
0.0063
0.0279***
(0.00828)
(0.00894)
(0.00945)
persistence
0.688***
0.255***
0.313***
(0.0322)
(0.0129)
(0.0160)
Observations
17,340
67,796
42,568
R-squared
0.946
0.927
0.955
DMA clustered standard errors in parentheses
*** p<0.01, ** p<0.05, * p<O.l
C.1.2 Only the Urban Border Counties
It seems the effects of advertising are a bit smaller in the more urban areas, except at the nest level.
However, it seems more likely that the identifying assumption might fail in the more urban areas,
as the borders are much closer to the central cities than in the more rural borders.
Table 1.12: Results of Base Model with only Northeast Corridor and Other Urban Areas
VARIABLES
Category Level
Subcategory Level
Product Level
adstock
0.0345***
0.0200**
0.0071***
(0.00949)
(0.00914)
(0.0111)
persistence
0.648***
0.387***
0.381***
(0.0676)
(0.0215)
(0.0285)
Observations
5,751
25,488
18,412
R-squared
0.963
0.956
0.972
DMA clustered standard errors in parentheses
*** p<0.01, ** p<0.05, * p<O.I
C.2 Are the Borders Special? Full Sample Estimation
Next, I assume that the border sample approach is not necessary and that advertising may be viewed
as predetermined with respect to demand shocks. This might be plausible since a large fraction
of television advertising in the pharmaceutical market is purchased in the upfront market. Using
a difference-in-differences approach with a lagged dependent variable and a common trend for all
58
DMAs, I estimate the model. The point estimates for the effects of advertising are not statistically different from those in the border approach. However, the persistence parameters are a bit
larger. If we believe these persistence parameters over the ones estimated in the border approach,
the spillover problem will be even larger than was estimated. I continue to prefer the border approach, as the assumptions of common trends among similar geographies is more plausible than a
nationally common trend. In addition, it gives more modest estimates of the size of the spillover,
working against, my main finding of spillover effects.
Table 1.13: Results of Base Model Without Using the Border Approach
VARIABLES
Category Level
Subcategory Level
Product Level
adstock
0.0436***
0.0030
0.0187***
(0.00140)
(0.00248)
(0.00438)
persistence
0.747***
0.721***
0.595***
(0.0209)
(0.0124)
(0.0116)
Observations
8,216
41,465
32,710
R-squared
0.976
0.984
0.984
DMA clustered standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
Appendix D - Primary Care Service Areas
As mentioned in section (3.1.1), measurement error could bias my estimates towards zero if patients are going to the doctor in different counties than where they watch television advertisements. The Dartmouth Center for Health Policy Research has developed a Primary Care Service
Area (PCSA) project which is the first national database of primary care resources for small areas.
These areas were defined using Medicare claims data from 1999 and Census data from 2000. The
service areas include a ZIP area with one or more primary care providers and any bordering ZIPs
where the population largely gets their primary care from those physicians.
This database allows me to ask how many patients travel across DMA borders to seek their primary
care. In particular, I can match this data to my prescribing data at the ZIP level. I can then see what
percentage of each PCSA falls into a single DMA. Doing this I find that only about 1% of PCSAs
cross DMA borders at all. Of those that do cross DMA borders, they do so only minimally. That
is, the DMA holding the majority of a PCSA which crosses a border on average contains 97% of
that PCSA. As such, measurement error bias should be minimal.
59
Appendix E - Computational Details
E.1 Cost Computation
Recal that the objective function that firm
f
wants to maximize with respect to advertising is:
Vfm(gmtla) = E[ E El-'
rf(gms, uf(gms))Igmt].
jEcIf r=t
Also recall that the per period profit function is:
Irfmt =
[J(Pjt -mcjt)mtsjmt(Amt,
a, )p()d4] -kjmtajmt.
For practical implementation, the integral will turn into an empirical expectation. So, the estimated
per period profit is:
[(Pjt - mcju)pmt s jmt (Amt, a)]
Trfmt =
-
kimtakmt-
jELDf
Adding this up infinely to the future and properly discounting, we get back to the objective function. Firm f will have one first order condition for each product k E <)f:
jEf
r=t
00d
-'(pjr - mcjr)ptmr
j
L~ PZ I'
mkmtp
sjmz(Amr, a) = E
akmtr=t
d
-Pdakmt EkkmrakmTEkvkz
-
The left side is the discounted sum of current and future marginal revenues. Given our specification
of shares, we can plug in the specific derivatives that are relevant depending on whether we are
evaluating the same product being advertised or another one in the portfolio, and whether the other
product is in or out of the subcategory of the advertised product.
The right hand side is the discounted sum of current and future marginal costs associated with
advertising level akmt in the current period. With costs specified as linear in advertising, we know
that d akmt = 1. We are left to deal with the effect of current advertising on future advertising.
Current period advertising may effect future advertising by raising future shares which increases
60
the optimal level of future advertising. However, the effect of current advertising on future shares
is rather small, and the effect of small changes in shares on optimal advertising is even smaller.
Further, this will be discounted. The size of this distortion will be no larger than 2.9%, and the
computation of that bound is provided in Appendix F. As such, I will assume that current advertising has a negligible effect on future advertising. To the extent that this approximation is bad, it
will lead me to overstate the true costs by no more than 2.9%.
Since the marginal revenues take into account expectations of rival behavior, assuming that the
observed world is in MPE will give us that these equal to marginal costs.
Given this, the procedure is as follows:
1. Form expected shares from today, forward for T periods. For each periodt, product j, and
market m, regress observed shares {si, ... , sT}on current period advertising stock At, a current
period effect ao, and a DMA effect am. Fitted values to these regressions will make firm
expectations that are correct in expectation, though not necessarily correct in outcomes. Do
this for the inside option share, the share of the subcategory given the inside option and the
share of the product given the subcategory. Given that we observe market outcomes, these
share expectations will include expectations about rival behavior and about demand shocks.
The implementation in the paper sets T= 12, as remaining advertising carry-over after twelve
months is negligible.
2. Having expected shares, we have all of the elements necessary to compute marginal revenue,
or the left hand side of the first order condition above.
3. Set marginal revenue equal to marginal cost.
E.2 Counterfactual Computation
The counterfactual is a bit trickier to compute, as we do not have observed future shares from
which we can back out the co-operative's expectations about future shares. Instead, I will solve for
these expectations in the following way:
1. Make an initial guess as to the monopoly's optimal strategy Co.
2. Solve for category shares given that strategy
3. Use these shares as the co-operative's expectation of future shares
61
4. Solve for the optimal strategy a"" given those expectations
5. Check if Ilane" - a
<E
6. If yes, stop, solve for category shares and profits given anew
7. If no, set anew = co and go back to step 2.
Appendix F - Cost Computation Error From Simplification
Recall, we had the firm first order condition:
O0
FOC :
d
(pjr-MCjr)pMr
jc(b5 r=t
00
sjmr(Amra) = kkmt+ -
#tkkmrdak
d
Eakmz-
d~+1akmt
dakmt
From this FOC, I want to show that for r > t, this term is negligibly small. I will first proceed
by noting that current avertising may only affect expected future advertising through its affect on
tomorrow's advertising stock, which affects the optimal choice of tomorrow's advertising through
its affects on tomorrow's market shares. Over time, this effect must by construction get smaller,
as the effect of current advertising on future advertising stock depreciates exponentially with time.
As such, it will suffice to show that the effect of today's advertising on tomorrow's advertising is
very small and bound all future period effects above by this small number.
So, we have:
dEakm,t+1
dakmt
__
dEakm,t+1 dESkm,t+1 dAkm,t+1
dESkm,t+1 dAkm,t+1
dakmt
From here, I note that this situation will be a 'worst case scenario' when there is only one product
in the market acting as a monopoly. The intuition is as follows: in the oligopoly case, each product
advertising only realizes a fraction of its own return to advertising, whereas a monopoly will realize
all of it. Given the logit demand function, advertising has an s-shaped resonse, and since the
shares are all significantly less than 0.5 in our sample, optimal advertising will increase in shares
everywhere in our sample. So, each term of this product is larger in the monopoly case, so the
product must be larger in the monopoly case. As such, I will focus only on the monopoly case and
use it as an upper bound.
62
Now, note that
d Skm,t+1
Akm,t+1
E(
d~kt~ ddkm~~l=
Eskm,t+1E(I
-
dAkm,t+1
-Sk'
- skm,t+l)
dakmt
1
We now need to compute the first term in the product.
)-I
+ akmt
Plugging in functional forms for the
monopoly case of the first order condition above, we get:
)jlrt 'Y
00
z~t
(Pkr -
EWt
T
MCkr) Ymr -
t
I
-+ a
-
0t
Y.T
+akmt
d
00
E [skmr (I1 - skms ) ] =
r=t
pi" kkmT
dakmt
Eakms -
-t(Pkr
- mCkr)I-mi
t ylE[Skmr(1 - Skm-)]
Mk)m
Vt(k
kkmt+E
t+1 lPT-kkmrdm
Eakr
7
da*m
-
(Pkt - mckt)Mmt7l(1
dESkmt
kkmt + E
-
ESkmt) - (Pkt - mckt)IlmtylEskmt
3=t+
"-kkmr
Eakmr
d
(Pkt - mckt)Imtyl(l
-
2
ESkmt)
kkmt
where the inequality follows from the fact that more advertising today leads to more expected
advertising tomorrow, as shares are well below one half. Now, given all of this, we have that
dEakm,t+1
<
dakmt
-
(Pk,t+1 - mck,t+1)/m,t+I Y17(1 -
2
Eskm,t+1)ESkm,t+E(1 - Skm,t+1)
kkmt(l+akmt)
Given this expression, I can plug in the 'worst case scenario' values of the variables observed in
the data to obtain a number. Since a priori, I have not estimated kkmt, I wil assume that it is exactly
equal to 1.15, which is the pecuniary cost of running the advertisement plus the typical commission
to the advertising agency. In addition, worst case advertising is when there is no advertising, so I
will set akmt = 0- So,
dEakm,t+l
(84) (0.047)2(0.685) (1 - 2(0.130)) (0.130) (1 0.130) = 0.0106
dakmt 1(1+0)
As this is an upper bound on the bias contributed by one period ahead, we know that two periods
ahead is bounded by
63
dEakm,t+2 _
dEakmt+I
dakmt
eakmt
and so forth. So the infinite sum, setting / = 0.95 and X = 0.685 and assuming expected future
costs are the same in each period and equal to k:
0
1
t-d
f 't-dakmt
Eakm, -
P
-Ea±
-3
-- k-m
1 -PX
+' < 2.7201 * (0.0106)k = 0.0288k.
deakmt
So, in the absolute worst case scenario, I will overstate the true observed marginal cost of advertising in each period by 2.88% of the expectation of costs, and there is every indication that this
number is typically far smaller, given that in practice, the observed values of the above expression
are smaller than the maximum observed in the sample and that in reality there is oligopoly which
depresses these terms.
64
2
Estimating the Cost of Strategic Entry Delay in Pharmaceuticals: The Case of Ambien CR
Abstract
With the Hatch-Waxman Act of 1984, the FDA included an unchallengable exclusivity
period for new approved drugs, independent of patents. This generates an incentive for firms to
strategically delay the introduction of new versions of drugs until just before patent expiration
of the originals in order to take market share away from new generics rather than its own
original product in its time of FDA exclusivity. Using detailed prescribing and pricing data,
I document that delay and estimate cost to consumers in the prescription sleep aid market at
about $644 million over seven years.
2.1
Introduction
Intellectual property policy is motivated by a desire to encourage firms to innovate. Few industries
have as obvious a need for innovation incentives as pharmaceuticals. However, by granting temporary monopolies to innovators, markets can be distorted in many ways. In particular, distortions
may come about due to the more or less fixed lifecycle that a drug gets upon approval. After the
lifecycle ends, a drug's sales tend to plummet as cheaper generic versions of the drug take over the
market. Given the impending cliff, we might imagine that firms would try any and all strategies to
make their investments last longer.
One way that a firm can extend the return on the research investment is to introduce a new version
of the drug that is about to lose its exclusivity. Since the molecule has already been tested, the
research and development costs are certainly less for the new version than for the first. In addition,
it is not unreasonable to think that a new presentation of an existing drug might be preferable to
some people, and in some cases may even be better for most people. Indeed, new versions of
drugs are granted new periods of market exclusivity by both the FDA (following the successful
completion of clinical trials) and the Patent Office (beginning at date of patent application).
A common new version seen in pharmaceutical markets is the time release version of a drug. To
make this, the firm takes the same molecule from the orginal drug and puts it into a different release
mechanism which releases the medicine to the body more slowly. For some drugs, this reduces
the number of doses per day. For others it simply is a more continuous release of that product.
65
While the time release versions need to be verified as safe and efficacious through clinical trials,
the molecule need not be rediscovered.
When should a firm introduce its time release version of an existing drug? If it is introduced during the existing drug's lifecycle, it will be cannibalizing its own market. In a world with lifecycle
concerns, there is an incentive to delay the introduction of a new version such that its lifecylce has
limited overlap with that of the original. In that way, the time release version will be competing
with the generic version of the original drug rather than the original branded version. It can compete as a horizontally and perhaps vertically differentiated product with a rival instead of stealing
share from it's own original, IP protected, branded product.
In prescription pharmaceutical market, Huckfeldt and Knittel (2011) have documented a pattern of
delay for new versions of existing drugs. They note that reformulations tend to enter the market in
the two years prior to patent expiration of the orginals, while there is no scientific reason why they
should systematically enter in that window. If the products being delayed have significant value,
we should be worried that potential consumer surplus is being lost in the intervening years of the
delay.
In a perfect world, the regulator would like to see products that generate surplus introduced as soon
as they possibly can be. They would like to do this while granting as little exclusivity as they must
to induce the innovation of such products. Meanwhile, the regulated firm would like to have as
much exclusivity as possible so as to get as much profit from its initial research investment as it
can.
In this paper, I will propose a policy that will induce early entry of new versions of drugs for
which there is significant consumer value. In particular, I will propose that exclusivity be granted
to new versions immediately upon entry and last until the date it would otherwise end if it had been
delayed. That is, the length of exclusivity would vary based on entry date, and the expiration date
would be fixed relative to the date of expiration of the orginal drug's exclusivity.
I will evaluate this proposed policy empirically using an example from the prescription sleep aid
market. In 2005, two years before the patent for Ambien was set to expire, Sanofi-Aventis brought
to market Ambien CR, a time release version of the popular drug. I will estimate demand in the
prescription sleep aid market using a discrete choice framework and evaluate welfare under the
realized world and a counterfactual world where Ambien CR is introduced seven years earlier,
or five years after the introduction of the original Ambien. I bound below the potential welfare
gain from earlier entry of Ambien CR at $644 million over the seven year delay. That amounts to
around ten percent of the total market revenue over that period.
66
2.2
Regulatory Framework
Intellectual property protection of pharmaceuticals comes from two main sources: the U.S. Patent
Office and the Food and Drug Administration, henceforth FDA. These two agencies provide somewhat complementary exclusivity as will be explained below.
2.2.1
Patents
In the prescription pharmaceutical industry there are both regulations and incentives imposed on
firms from the government. Historically, new molecules discovered by firms have been awarded
patents, which run for twenty years from the date of application. In those twenty years, a rival firm
may not copy the product with the same molecule or use that molecule in the development of its
new product. Note that the patent life is not initiated from market entry, but from patent application
date.
While patents are typically granted somewhat liberally, they are challengeable in court. If it is
believed that a patent is not innovative or has already been done, a firm may challenge that patent
through the courts and ultimately get it overturned.
2.2.2
FDA Exclusivity
In order for a drug to enter the market, it must pass several stages of clinical trials. These trials
are both costly and time consuming. It may be several years between the granting of a patent and
the entry of a new molecular entity. Thus, the effective length of market exclusivity varies based
on how long it takes the drug to get approved. This might cause one to worry that the incentive to
innovate will be watered down due to the regulatory delay to market entry.
In the meantime, regulators also noticed that consumers reaped large benefits from the entry of
cheaper generic drugs. In an attempt to continue to encourage new innovation while also incentivizing generic entry, the Hatch-Waxman Act of 1984 was passed. The act provides a fairly comprehensive set of regulations for the pharmaceutical industry, but in it are two main countervailing
forces which generate the incentives that are examined in this paper: what is called the Paragraph
IV generic challenge and FDA granted exclusivity.
The Hatch-Waxman Act Paragraph IV provides incentive for potential generic entrants to challenge patents by giving 180 days of generic exclusivity to the first succesful patent challenger. To
67
balance this with the high cost of trials and the desire for innovations of new drugs, Hatch-Waxman
provided for five years of unchallengeable exclusivity granted by the FDA from the time of market
entry for new molecular entities and three years of unchallengeable exclusvitity for reformulatoins,
along with other lengths for other kinds of drugs. Table 2.1 shows exclusivity periods for various
types of new drugs. This paper will focus on an example of a reformulation.
What kinds of patents are most likely to be challenged through Paragraph IV challenges? Hemphill
and Sampat (2011) document that it is 'weaker' patents that are not on the molecule itself that are
most commonly and most successfully challenged. Among these types of patents are reformulations.
Unlike a patent, FDA exclusivity is not challengeable in court, so each new drug that passes through
trials is awarded FDA exclusivity, regardless of patent status. Given the regulatory structure, an
incentive to distort entry decisions to extend product lifecycles exists. A firm realizes that a reformulation's patent is more likely to be challenged, so it is more likely to have a short amount of
effective exclusivity.
Given this regulatory framework, it is not surprising that a firm might prefer to take its shorter
effective exclusivity attached to its reformulated drug in a time where it will not be cannibalizing
its own sales on the original branded product. That is, by waiting to release a new formulation of
an existing molecule to just before the date of patent expiration of the original drug, a firm is able
to effectively extend its exclusivity in that market by another several years and avoid competing
with itself.
2.2.3
Policy Proposal
Given that there may be welfare gains from earlier entry of the extended release version of Ambien
and assuming that the pattern for other drugs would be similar, it is instructive to think about
policy choices that would lead to earlier entry of extended release versions of drugs. We could,
for example, discontinue giving exclusivity to branded reformulations. While this seems attractive
on the surface, it is unclear the effect it would have on the firm. By revealed preference, it would
seemingly decrease profits, as they have chosen to delay product entry.
Since extended release versions tend to be challenged more via Hatch-Waxman Paragraph IV
generic challenges (Hemphill 2011), firms likely expect to be challenged as soon as their FDA
exclusivity expires. Given that they will have only three years of exclusivity on the extended release product, they prefer to have it later when it will not be cannibalizing sales from their original
formulation.
68
One way around this concern would be to give the firm three years exclusivity of the branded
reformulation after the patent expiration date of the original formulation, regardless of approval
date. That is, if Ambein CR entered in quarter one of 1998, it would still have exclusivity through
quarter three of 2010, just as it did in the realized world. The only difference now is that there is
no delay incentive and consumers would gain the benefit of expanded product variety. In addition,
the firm would be able to claim all of the profits from 1998 until 2005 that it did not in the realized
world.
2.3
The Prescription Sleep Aid Market
According to the Center for Disease Control, in 2008, roughly 35% of adults reported having
insomnia or sleep problems at least ten days out of a month. It has been emphasized in the medical
community, however, that insomnia is not a disease in and of itself and is often a sign or symptom
of more serious health problems. While insomnia is not a disease, per se, it has been treated
medically for many years. With the advent of new non-benzodiazapines such as Ambien in the
1990s and Lunesta in the 2000s, the number of people medically treated for their sleep problems
has drastically increased.
The treatment for sleep problems can be complicated. Ideally, doctors would like to isolate the
cause of the sleep insufficiency. Often, however, identifying the root cause is difficult, and a
temporary solution for sleep abnormalities is prescribed. Before 1992, the closest drugs to a prescription sleep aids were the anti-anxiety benzodiazepines. However, if a patient does not suffer
from anxiety problems, this drug class is less than ideal due to serious side effects. Another alternative at the time was over the counter drugs such as Unisom. These drugs are quite a bit less
powerful and less risky in terms of side effects. As a result, they report lower efficacy than the
more powerful prescriptions.
In 1992, Sanofi-Aventis introduced Ambien, the first drug in the non-benzodiazepine sleep aid
market. Since its introduction, non-benzodiazepines have become increasingly popular, and increasing proportions of people suffering from sleep abnormalities have obtained prescriptions for
these drugs. Their growth in popularity can be seen in Figure 2.1. While very effective, these drugs
do have many serious side effects, including hallucinations, depression and memory loss. Due to
their habit forming nature, they are, in general, not approved for long term use.
Since 1992, two additional molecules have entered the market: Zaleplon (Sonata) and Escopiazepine (Lunesta). The patents of both Sonata(2007) and Ambien(2006) have expired, and generic
69
entry has occurred. As can be seen in Figure 2.1 (and as predicted by the incentives), Ambien CR
was introduced less than two years prior to Ambien's patent expiration, which is consistent with
the Huckfeldt(201 1) findings that the window of 2 years before the patent expiration is the most
common time for reformulations to appear.
The different molecules in the class are differentiated horizontally: Sonata has a shorter halflife than Ambien while Lunesta has a much longer half-life, making their properties somewhat
different. With a longer half-life, Lunesta remains in the body longer and tends to be better for
staying asleep while Sonata packs most of its punch at the beginning, making it a natural choice
for falling asleep. Ambien, still the most popular sleep aid, is between the two. Ambien CR is a
time-release version of the original. While the half life of the molecule is the same, the delivery
mechanism of the drug releases the molecule more slowly into the system, making it commonly
prescribed for staying asleep.
2.4
Theoretical Framework
2.4.1
Random Utility Model
To analyze demand in the prescription sleep aid market, I will use a characteristics based nested
logit formulation. Consumer i's utility for product j, in market, m is specified as:
uijm
The mean utility, 3jm =
aPjm +
=
apjm+PXjm+
jm+ Eijm
(2.1)
Xjm + 4jm, and the outside option utility, ujom = 0. Here, Pjm
is the price of good j in market m, Xjm are product attributes, and gjm is a quality attribute that is
observed to market participants but unobserved by the econometrician. If Eijm ~ iid extreme value
type I, then we have the familiar multinomial logit with market shares given by:
Sjm =
3 im
+E2e
(2.2)
While this formulation has computational convenience, it has significant drawbacks. In particular,
substitution patterns are determined only by market shares. In particular, with generic drugs having
extremely similar if not identical properties to their branded counterparts, it is difficult to believe
that individual taste shocks are independent across all products. A simple way to address this
70
concern is to allow for correlation in taste shocks conditional on a pre-determined nesting structure,
or a nested logit. While this does not allow for individual heterogeneity in the taste parameters, it
does allow richer forms of substitution and correlation between taste shocks for similar products.
In particular, taste shocks cijm ~ F, where
F(E) - e- nEN(YjEneXP( Ia))l-
(2.3)
In this framework, n represents a nest of products out of the set of nests, N, and a is a measure of
within nest correlation. As a -+ 1, the within nest correlation of taste shocks becomes perfect, and
as a -+ 0, the taste shocks are independent and we have our familiar multinomial logit back. For
the model to be consistent with utility maximization, a E [0, 1].
This model gives way to market shares of the following form. First, the market share conditional
on product nest is:
(2.4)
Dm
sin=
where Dn is a within nest utility term,
Dn = Ejenebm-(l)
(2.5)
Nest market shares are given by,
D -n)(2.6)
Sn
EnENDn
This gives us product market shares equal to:
e3 Jm
Si = Siln *sn =
(2.7)
D9EnNDnIlG)
We normalize the utility of the outside good to 1, so its market share is
71
1
so =
(2.8)
E NDn
Following Berry(1994), we can take logs and rearrange the market shares of the product and the
outside good to get:
ln(sjm) - ln(som) = -apjm + fXj
+ aln(sjln) + gjm
(2.9)
This rearranged expression for market shares allow me to estimate the model using simple linear
instrumental variables methods.
In this setting, there is reason to believe that the nested logit framework is reasonable. We can think
of the idiosyncratic error term as dependent on diagnosis of the individual. When a patient goes
to the doctor and receives a diagnosis, the choice of prescription is dependent on that diagnosis.
Drugs often fall into very natural groups based on diagnosis, which we do not observe. In the
simplest case, a branded molecule and its generic version clearly have correlated idiosyncratic
shocks. In most states, the doctor just prescribes the molecule, and pharmacies have mandatory
substitution rules to generics if they are available, unless the doctor insists that the branded is
medically necessary.
In other cases, drug classes naturally separate into different diagnoses. In the prescription sleep
aid market, according to conversations with doctors, there are three main diagnoses which could
correspond to three nests, where demand within each nest can depend on a common idiosyncratic
shock: trouble falling asleep, trouble staying asleep and general trouble sleeping. The zaleplon
molecule, which includes Sonata and its generic belong in the first nest, the standard release zolpidem tartrate molecule fall into the second nest and Lunesta and Ambien CR fall into the third nest.
The diagnosis nesting structure will be employed in the preferred specification in this paper.
We might think of an alternative nesting structure that puts drugs into nests only by their molecule
designation. In the appendix, I will work through that nesting structure as well. My preferred
structure will be based on diagnosis due to discussions with doctors, though the molecule nesting
structure will yield both qualitatively and quantitatively similar results.
2.4.2
Alternative Demand Models
While Ellison et. al (1997) use a multi-stage budgeting approach to modeling demand in their
pharmaceutical demand framework, there are good reasons to use the characteristics based frame-
72
work as in Branstetter et. al (2011) and Stem (1996). First, the multi-stage budgeting demand
system gets unwieldy in terms of dimensionality as the number of products increases. While the
prescription sleep aid market has only six products, I would like to develop a framework here that
could be applied to any number of drug classes. For example, the antidepressant market has more
than fifty products.
In addition, the characteristics based approach allows for easily and quickly computed counterfactuals of different kinds, as we often know the characteristics of products even before they enter
the market. It is more difficult to imagine how to compute such counterfactuals in the multi-stage
budgeting framework.
2.4.3
Counterfactuals and Welfare
Given the demand estimates, I will compute welfare under a counterfactual world in which entry
of Ambien CR happens seven years sooner than observed and under the observed world to see
what the welfare gains would be in the counterfactual. When thinking about estimating the cost
of delay, we might think that welfare gains from an additional product in the market could come
from three sources: an additional product in the market would provide downward price pressure
on the existing products, consumers have a willingness to pay for having the additional options,
and consumers get welfare from earlier entry of the generic.
In this paper, I will focus only on the welfare from the additional product on the market for a
number of reasons. First, there is little evidence in my data that there is substantial downward
price pressure from non-generic entry by therapeutic substitutes. Second, the welfare gained from
generic entry has been addressed in the literature (Branstetter 2011) and found to be large. However, the policy implication I will propose will not include earlier generic entry. If the welfare
gain simply from expanded product variety is sufficiently high, it will highlight the significance of
strategic product delay.
Consumer Welfare Given the nested logit framework, expected compensating variation may be
computed following Train (2003) as:
E(CV)
-- PotentialMarket[ln(l+En(Ejcne3
a
(1a))1
ln(l +En (Ejene/(1
(2.10)
73
Given that I am taking the supply side as given, I will need to assume the counterfactual prices
and advertising levels in 3}m. My strategy will be to assume prices that are biased upward, so the
estimate of welfare gain will be biased downward. Given a low end estimate of welfare, if it is
still large, it will highlight the importance of strategic delay to consumer welfare. In addition, I
will assume that advertising provides no welfare benefits, making it unnecessary to guess as to
what advertising levels would be in the counterfactual. These assumptions should give a weakly
lower estimate of total welfare from an earlier introduction of a product. In the appendix, I will
run the analysis including advertising as a demand shifter in the welfare calculation. Continuing
work addresses the question of how to think about pharmaceutical advertising in the calculation of
welfare.
Producer Welfare In this paper, I will not estimate producer welfare differences from earlier
introduction of Ambien CR. However, I will propose a policy that fixes the date of exclusivity
expiration of the reformulation relative to the exclusivity expiration of the original drug. Then, by
revealed preference, the firm could always replicate the strategy in the observed world and get the
same profits as before. If the product is introduced sooner, it gets a longer period of exclusivity
and as long as there are positive profits to be made in the period before observed entry, firm profits
must be better in the counterfactual.
2.5
Data
I use data from three main sources: I use the prescriber data Xponenent from IMS Health to derive
market shares, national advertising in IMS Health's IPS dataset, and price data from medicaid
reimbursements. In addition, I use a survey from the Center for Disease Control to construct my
measure of market size.
2.5.1
The Prescription Sleep Aid Market, The Outside Good and Market Shares
The IMS Health Xponent data set follows specific prescribers across time, tracking their prescribing behavior on a monthly basis. Here, I have a ten percent random sample of all prescribers who
prescribed at least one prescription sleep aid in 1996. The random sample is refreshed with new
observations every year to reflect the overall growth in prescribers. In this paper, I use prescribing
data from 1998 through 2008, due to limitations in advertising data in 1996 and 1997.
74
To define the size of the market, I use the CDC Behavioral Risk Factor Surveillance System Annual
Survey. This is an annual survey of a representative subsample of the adult population conducted
at the state level that assesses many health questions. In 2008, the survey in all states asked the
question, "How many days did you not get enough sleep in the past 30 days?" If someone responds
with a number greater than ten, I call that person a potential consumer of prescription sleep aids. I
calculate the percentage of respondents who are potential consumers and multiply that by the state
adult population to get the potential market size. As it may be the case that respondents answer
that they have no sleep problems because they are on sleep aids, I add to that number the number
of consumers taking sleep aids. I assume that each potential consumer may get as many as three
prescriptions in a given quarter. In previous years, this question shows up only in a subset of states.
To obtain estimates for missing state quarter combinations, I regress the observed values on time
dummies and use imputed values where actual estimates are not available.
While I do not use the micro data on prescribing, each presciber does have an associated address.
I use these addresses to identify geographic markets. I construct market shares by aggregating the
prescriptions by geographic market and period, multiplying by ten and dividing by the potential
market size. Due to granularity and quality issues in the pricing data, I will use nine geographic
divisions, as defined by the United States Census Bureau. A market will be defined as a region of
the United States in a given quarter. Potential market sizes are aggregated to the region level, and
prices and shares are calculated at the quarter level. I define the market share of the outside option
to be the potential market size less the market shares of each of the products in the market.
We can think of the outside option being over the counter drugs, treatment of the underlying cause
of the sleeplessness or non-medical treatment such as lifestyle and diet adjustments.
Figure 2.2 depicts the regions used in the analysis.
2.5.2
Product Characteristics
Advertising The IMS Health IPA data set includes information on number of free samples given,
the 'retail value' of those free samples priced at Average Wholesale Price (AWP) and the number
of physician detail visits the pharmaceutical company made in a given month where the drug in
question was part of the portfolio of the advertiser.
In practice, free samples and doctor visits are highly collinear. From conversations with doctors,
the collinearity is hardly surprising given that most detail visits include the giving of free samples.
75
As such, in estimation I am able only to identify a coefficient on either samples or doctor visits. I
use one plus the log of the number of samples as a proxy for all of this type of advertising.
Unfortunately, this data is only available at the national level, so all variation is over time.
Prices While I am able to infer the Average Wholesale Price (AWP) from the two advertising
measures of number of samples and retail value of samples for those products which advertise in
this way, it is well known in the industry that AWP is not a good measure of the average price that
a drug sells for in the market. In some circles, AWP is referred to as "Ain't What's Paid" due to
the numerous (unobserved) discounts and rebates that nearly all insurers and pharmacies receive.
However, Duggan and Scott-Morton (2006) argue that the average price that medicaid pays per
perscription is a good measure of the average price of a drug on the market. This is convenient
because through CMS, medicaid quantities and reimbursements are publicly available at the state
and quarter level. As my measure of price, I use the total medicaid units dispensed divided by the
total medicaid reimbursements during a quarter for a particular product multiplied by thirty, which
is the standard amount dispensed per trip to the pharmacy for this class of drugs.
Some discussion of this measure is warranted, however. Some doctors write prescriptions to allow
for refills, so each prescription may not correspond to exactly thirty pills, while decisions are only
made at the point of prescription. It seems natural that the price that a physician would respond
to would be based on the per pill price rather than some expectation of the per prescription cost.
That being said, re-running the analysis with average price per prescription yields qualitatively and
quantitatively similar results.
Medicaid price information has an important advantage in addition to being publicly available
and free: the geographic variation in this price measure is exogenously set by states. Medicaid's
reimbursement formula has both a term set by the state and a term set by the federal government
authority. For more information on the specifics of the formula are available at CMS and in Duggan
et. al. (2006), but the important part is that the geographic variation is exogenous. The time series
and between product variation in these prices is still endogenous and will need to be instrumented.
Figure 2.3 depicts the average price over time of the products in the market in 2008 dollars. Figure
2.4 depicts the average price of Ambien across regions over time.
Of note in this picture is the time beyond 2007 when generic Zolpidem Tartrate entered the market.
Different states have different policies about reimbursements for branded drugs when generic drugs
were available. Some would reimburse only at the rate of the generic while others would simply
76
keep the same formula as before. Hence, we see some regions with distinct drops in prices while
others have some price increases.
Price increases upon generic entry have been documented before (Huckfeldt 2011) and have been
explained as a price discrimination story. We might also think that the firm is trying to phase out
the branded original molecule in favor of their patented extended release version.
Data was removed in several states due to suspected serious nonclassical measurement error issues.
For example, in one quarter, South Dakota reported more prescriptions of Ambien on medicaid
than there were total residents in the state. In addition, they reported those coming in at a per
prescription price of around a penny. There were similar issues in Tennessee and Maine. Finally,
Arizona did not report medicaid prescription drug claims for most of the sample period. All of
these states were removed from all data.
Other Product Attributes In the previous literature, product attributes have included number of
contraindications or number of presentations (dosage sizes) (Branstetter 2011) or time since market
entry, number of pills taken per day and generic or pioneer status (Stem 1996). Since in this market
there are only six products and only four with unique chemical properties, it is important to choose
attributes that vary across drugs. I use the half life, denoted HALFLIFE, of the molecule and
the time since entry, denoted TIMESINCEENTRY in addition to a brand dummy on Ambien CR,
denoted ER, region and time fixed effects. I define time since entry as the number of quarters since
the original branded molecule of a drug was introduced. In this way, we can view a drug with a
large time since entry as one that has successfully stayed on the market for a long time, possibly
due to safety and efficacy in the population.
To complete the nested logit formulation, a conditional share term is included in Equation 2.9.
I get this term by taking the prescriptions of the product and divide that by the total number of
prescriptions of products in the nest. This quantity will also be endogenous since as shown in
Equation 2.4, the conditional share is a function of 3 jm which makes it a nonlinear function of the
unobserved characteristics j,. It then must be instrumented.
2.5.3
Identification
As mentioned above, the price, advertising and conditional share term have endogeneity concerns
due to the presence of the unobserved characteristic 4jm. Prices and advertising are chosen by the
firm which sees gjm and takes it into account. The conditional share term also depends on jm. As
77
such, I will employ an instrumental variables approach in the spirit of Berry, Levinsohn and Pakes
(1995)- henceforth BLP.
The geographic component of variation in price is set exogenously by states as previously described, however much of the variation in the data is across time and between drugs. Changes in
prices happen with changes in demand conditions that are unobserved to the econometrician. In
the spirit of BLP, I will instrument price with own and rival product characterics. As new drugs
enter the market, the closeness of competition in each market depends on the characteristics of all
of the competitors. Prices and advertising are chosen with the closeness of competition in mind.
Entry is a supply side change that will help us trace out the demand curves.
For these instruments to be valid, we need only that the characteristics not be endogenously chosen.
Since the half life of a drug is a chemical property, it seems reasonable to suspect that it is not
manipulated to meet demand shocks. Also, conditional on being in the market, the time since
entry cannot be changed by a firm. Finally, entry is something that depends on FDA clinical trials
as well as random discoveries and so might reasonably be thought of as having exogenous timing.
The exogeneity of attributes and market structure seem to be reasonable assumptions.
Given the number of characteristics, there will be sufficient instruments to accommodate our three
endogenous variables.
In principle we could also think of using prices in other markets as instruments in the spirit of
Hausman (1994). The justification of these instruments is that prices in one region should be correlated to prices in another through their common correlation with marginal cost, while a demand
shock that is localized to California should not affect prices in New York. In my setting, however,
all prices are correlated across regions through their correlation with AWP, which seems likely to
be related to demand conditions and not simply costs. In the appendix, I use these instruments,
and the analysis will fail the test of overidentifying restrictions, though the estimates will be qualitatively similar.
2.6
Results
2.6.1
Demand Estimates
Demand estimates are displayed in Table 2.2. For the sake of completeness, I have included both
the logit and nested logit estimated by OLS and IV. The instrumental variables estimates on price
are different from the OLS estimates in the way we would expect, giving us downward sloping
78
demand curves. Given the significance of the within nest conditional share term and its magnitude
strictly between zero and one, we can reject the logit model in favor of the nested logit model. It is
notable, however, that the OLS nested logit gives us a very similar coefficient on price as the IV. I
attribute this to the substantial exogenous variation in prices outlined above. Despite the similarity
in price coefficients, the OLS estimates will fail the Hausman specification test, leading me to take
column (4) as my preferred specification.
As to the magnitude of the coefficients, the price coefficient implies a mean own price elasticity of
around -2.5. This is significantly more elastic than estimates of own price elasticities by Branstetter(20 11) in the anti-hypertensive drug market. This may partially be explained by the fact that
anti-hypertensive drugs are almost always covered by insurance, while sleep aids are much less
frequently. Advertising has an important positive effect on demand, as we would expect. However,
that effect is smaller in magnitude than those found by Higgins (2011) in the anti-hypertensive
market. Again, we might be able to explain this through the insurance story. Suppose that free
samples have an effect on doctors in the following sense: a patient is willing to take whatever a
doctor gives him if insurance is going to cover it. Given that, a doctor finds it easiest to give the
patient whatever samples he has on hand, making that the drug of choice. However, now suppose
insurance will not cover the drug. The patient now cares more about the price and is less willing to
be swayed by the convenience of a free sample. Since free samples usually only come in pouches
of one or two doses, the final price of a treatment will not be drastically reduced by the one or two
free doses.
As for other product attributes, time since entry also comes up positive, signifying there is a benefit
to having a drug be on the market for awhile. Given that drugs are approved only by having an
average effect beat that of placebo, each year that a drug is on the market without being recalled
reduces the expected probability that the drug will have an adverse effect bad enough to trigger a
recall. Finally, half life and the extended release dummy are both positive, suggesting that there is
benefit to a longer lasting drug and that there is a willingness to pay for extended release.
2.6.2
Welfare Analysis
Counterfactual Setup
Given the above demand estimates, I can use Equation 2.10 to compute
counterfactuals. To do this, I must first construct 3counterfactual. For this calculation, I will as-
sume that the counterfactual world is one in which Ambien CR enters the prescription sleep aid
market five years after Ambien enters the market, or in 1998, quarter 1. The reason for this is
twofold. First, we might think it reasonable for a certain amount of time to pass between the orig79
inal molecule entry and a reformulation if for no other reason than getting through clinical trials.
Second, my sample begins in this time period, so it is a convenient place to draw a line.
In addition to a welfare estimate, I will include a graph of expected compensating variation by
quarter. In that way, we can imagine how much welfare is gained at different points in time.
Finally, I am only computing a welfare differential over the period that has Ambien CR in the
counterfactual world and not in the realized world. In practice, we might think that having Ambien
CR on the market sooner would have effects on the more recent market. I will ignore any of
these potential welfare effects. Given that the coefficient on Time Since Entry is positive, I would
postulate that I am ignoring further positive welfare gains rather than losses.
Prices As noted previously, the supply side of the market is not the focus of this paper, so assumptions must be made about prices in the counterfactual world. Looking back at Figure 2.3
suggests that prices do not seem to respond much to entry of a therapeutic substitute. As such, I
will assume that prices remain the same for all therapeutic substitutes to Ambien CR in the counterfactual world. If anything, we would expect those prices to decrease with another competitor.
Assuming that they do not will be giving conservatively low estimate of the welfare gain of adding
the new product.
The trickier question is what one should do about the prices of Ambien and Ambien CR in the
counterfactual world. A multiproduct firm does set prices differently than a single product firm.
Given that these products are substitutes, one would expect that they would not compete their
prices down. Furthermore, upon the actual entry of Ambien CR and prior to the entry of generic
Ambien, we can see in Figure 2.3 that Ambien does not move its price from trend for a few periods,
and then raises its price in what seems to be an attempt to phase itself out in favor of Ambien CR.
It might be doing that in anticipation of patent expiration of Ambien.
To again get a lower bound estimate on welfare, I will assume that Ambien prices remain the same
in the counterfactual world. To predict Ambien CR prices in the counterfactual, I predict Ambien
CR prices in the realized world using the BLP instruments, Ambien's price and a time trend. From
those estimates, I impute the value of Ambien CR's price in the counterfactual if that price is at
least five dollar's more than Ambien's price. If it is not, I replace the Ambien CR price in the
counterfactual with five plus the price of Ambien, as in the realized world, Ambien CR is never
more than five dollars more expensive than Ambien. These choices on prices should all amount to
a low estimate on the welfare gain from adding the new product.
80
Advertising Thinking about how advertising affects welfare is complicated. We might think that
advertising is simply a transfer from the pharmaceutical company to the doctor as a bribe to get
him to prescribe one drug over a substitute or the outside option. If this is the case, we need to
worry about value lost due to mismatches. This is very interesting and the subject of future work,
however in this setting I will assume that mismatches induced by advertising are not an issue.
In addition, whether advertising affects the information set or is a demand shifter (as I have modeled it) is important for welfare.
I will choose not to take a stand on the welfare effects of advertising and be satisfied with having
controlled for it in the choice decision as my main specification. An alternative specification
in the appendix describes how I choose counterfactual advertising levels and computes welfare
differences under the assumption that advertising is a demand shifter. Given that the coefficient on
advertising is positive, this specification will give larger welfare gains to sooner entry. I am again
choosing the more conservative welfare estimate route for the main specification. However, the
magnitude of the difference in welfare gains further accentuates the need for further research on
the welfare costs and benefits of pharmaceutical advertising.
Results Given the counterfactual scenario presented above, over the period from 1998 until the
actual time of Ambien CR entry in quarter three of 2005, consumers gain approximately $644
million from the addition of Ambien CR to the market. That number is not uniform over the length
of the period, given that there are welfare benefits to a drug being on the market for an extended
period of time.
The welfare gains plotted against time are presented in Figure 2.5.
Of particular note is that the welfare benefit is substantially decreased at the period right around
the beginning of 2005. This is due to the entry of the branded drug, Lunesta. Given the diagnosis
nests, Ambien CR and Lunesta are in the same nest, and the incremental benefit of a new nest is
larger than that of a new product within a nest.
To give an idea of the economic significance of the welfare gain numbers, Figure 2.6 plots the
welfare gain as a percentage of total revenue in the prescription sleep aid market in the realized
world. The welfare gains are above 5% of the size of the market for most of the period, and above
15% of the market in the years when counterfactual entry is occurring.
All of this is to say that as a fraction of the size of the market, the benefits that consumers are not
realizing due to strategic entry delay are not trivial. If we can come up with a simple policy fix to
remove the incentive to delay, there may well be huge potential benefits.
81
Discussion In the discussion of the welfare benefits to Hatch-Waxman Paragraph IV challenges,
Higgins points out that many of the potential benefits may be accruing to parties other than patients.
If the computed benefits are coming from price decreases, many of those benefits may be captured
by insurance companies or pharmacies. If computed benefits are coming from changes to the
amount of advertising attention doctors receive, the benefits may be going to doctors rather than
patients. However, in this study, I only compute the consumer welfare change due to expanding
product variety. Expanded product variety intuitively seems like something more likely to benefit
consumers.
Also worth considering is that prescription drugs are a unique industry in the amount of uncertainty
there is in whether or not a product will work. Patients, together with doctors, make decisions based
on their expectations of success, but evidence about how a particular drug will affect a particular
patient is scarce. Clinical trials need only show that a drug perform better than placebo on some
pre-determined population. If a patient is not similar to the average person in that population, it
is unclear how well it will work. Further, advertising may bias expectations. Computing welfare
taking into account these things is difficult. As such, these results are meant only to be suggestive
that there is a large willingness to pay for expanding product variety that we may be able to attain
with policy choices.
Given that the first date of potential entry is private information to the firm, it is important to
craft a policy that removes the incentive for delay rather than setting exclusivity from entry for
reformulations. The proposed policy here may not have given exactly $644 million in additional
consumer surplus- it could be more, as Ambien CR may have entered even sooner than the first
quarter of 1998 or less if it required longer in development. However, if the removal of the delay
incentive resulted in any earlier entry, Figures 5 and 6 demonstrate that there is the potential for
significant welfare gains.
2.7
Conclusion
When the Hatch-Waxman act was passed, there was tremendous concern about balancing the benefits of generic entry with the incentives for innovation. As a result, the FDA instituted an unchallengeable five year exclusivity period for all newly approved drugs. While the patent system has
built in protections (i.e. the opportunity to challenge in court) for weak, non-innovative patents,
the FDA exclusivity period does not. This exclusivity period generated an incentive to delay the
entry of branded reformulations of drugs. This incentive has been documented in the literature and
is broadly recognized in the market.
82
Strategic entry delay limits consumer welfare significantly through the withholding of product variety. In this paper, I conservatively estimate a consumer welfare gain of $644 million for Ambien
CR entering the market in 1998 rather than in 2005. This number is significant as a percentage of
the market size and is spread out over the seven year delay.
In reality, nearly all drug classes contain extended release formulations of important drugs. Given
the pattern that Huckfeldt (2011) shows, entry delay is a significant concern not only for sleep aids,
but also for all of these other drug classes. Sleep aids are a relatively small class compared with
say, anti-hypertensive drugs or antidepressants. Thinking carefully about the incentives generated
by our exclusivity rules in the FDA has the potential to create huge gains for consumers.
References
[] Branstetter, L., Chatterjee, C., & Higgins, M., (2011). "Regulation and welfare: evidence
from Paragraph IV generic entry in the pharmaceutical industry," NBER Working Paper No.
17188.
[2] Ellison, S., Cockburn, I., Griliches, Z., & Hausman, J., (1997). "Characteristics of demand
for pharmaceutical products: an examination of four cephalosporins," RAND Journalof Economics, 28(3), 426-446.
[3]
Duggan, M., & Scott Morton, F., (2006). "The Distortionary Effects of Government Procurement: Evidence from Medicaid Prescription Drug Purchasing," The QuarterlyJournalof
Economics, 121(1), 1-130.
[4]
Hausman, J., Leonard, G., & Zona, J., (1994). "Competitive analysis with differentiated products. Annales D 'Economie et de Statistique, 34, 159-80.
[5]
Hausman, J., (1997). "Valuation of New Goods Under Perfect and Imperfect Competition."
in Bresnahan, Timothy and Robert J. Gordon, eds. The Economics of New Goods: NBER
Studies in Income and Wealth, 58, 209-37.
[6]
Hemphill, S.C., & Sampat, B.N., [2012], "Evergreening, Patent Challenges, and Effective
Market Life in Pharmaceuticals", Journalof Health Economics, 31(2):327-39, March.
[7]
Huckfeldt, P., Knittel, C., (2011). "Pharmaceutical Use Following Generic Entry: Paying
Less and Buying Less," NBER Working Paper No. 17046.
83
[8]
Nevo, A., (2001). "Measuring Market Power in the Ready to Eat Cereal Industry," Econometrica, 69(2), 307-342.
[9]
Berry, S., Levinsohn, J., Pakes, A., (1995). "Automobile Prices in Market Equilibrium,
Econometrica, 63(4), 841-890.
[10] Petrin, A., (2002). "Quantifying the Benefits of New Products: the Case of the Minivan,"
Journalof PoliticalEconomy, 110(4) 705-729.
[11] Stem, S., (1996). "Market Definition and Returns to Innovation: Substitution Patterns in
Pharmaceutical Markets," Working paper, MIT Sloan School of Management.
[12] Train, K., (2003). Discrete Choice Methods with Simulation, Cambridge University Press,
Cambridge, UK.
Tables and Figures
Approval Type
New Chemical Entity
Reformulation
Biologic
Pediatric
New Indication
[
Requirements
Molecule not previously used passes clinical
trials
Molecule previously used, but altered passes
trials
New biologic product passes clincial trials
Clinical trials run (no approval necessary) on
pediatric population
Clincial Trials run for new indication
Exclusivity Granted
5 years
3 years
12 years
6 months
Table 2.1: Different FDA Exclusivity Lengths By New Drug Type
84
None
Figure 2.1: Quantities over Time
Quantity over Time
)
00
0
0.0(0
CO)
a.
(U 0
U) 0
0C%4
0
-
1998q1
20dOql
20d2q1
20d4q1
Period
20d6q1
Ambien
Sonata
Ambien CR
Lunesta
Generic Ambien
Generic Sonata
2008q1
Figure 2.2: Regions Used in the Analysis
Midwest
Northeast
West
85
Figure 2.3: Average Price over Time
0
0
T-
0
0
CDt
1948q1
2OdOql
200 12q1
20014q1
20d6ql
2008q1
Date
Ambien
Sonata
Generic Ambien
Lunesta
Ambien CR
Generic Sonata
Figure 2.4: Ambien Price Across Regions
0N
0
Cl)
CO
0o
1998ql
2Od0ql
20d2ql
20d4q1
Date
86
20d6ql
200 18q1
Table 2.2: Baseline Results
Results
PRICE
LOGSAMPLES
(1)
(2)
(3)
(4)
Logit
logdiff
Logit IV
logdiff
Nested Logit
logdiff
Nested Logit IV
logdiff
-0.0322***
(0.00216)
-0.0368***
(0.00212)
0.238***
(0.0117)
-0.0178***
(0.00105)
0.176***
(0.00659)
0.862***
(0.0329)
0.592***
(0.0125)
0.0732***
(0.00113)
2.297***
(0.0650)
-8.960***
(0.124)
Yes
Yes
-0.0253***
(0.00193)
0.190***
(0.0121)
0.495***
(0.0586)
0.578***
(0.0220)
0.0695***
(0.00116)
1,022
0.898
1,022
0.227***
(0.0237)
aHALFLIFE
TIMESINCEENTRY
ER
Constant
Time Fixed Effects
Region Fixed Effects
Observations
R-squared
Hansen's J
p-value
0.532***
(0.0243)
0.0668***
(0.00239)
1.657***
(0.123)
-8.310***
(0.177)
Yes
Yes
0.541***
(0.0214)
0.0661 ***
(0.00192)
1.639***
(0.109)
1,022
0.763
1,022
-7.941 ***
(0.222)
Yes
Yes
5.083
0.025
Hetroskedasticity robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
87
2.049***
(0.0773)
-8.505***
(0.140)
Yes
Yes
0.214
0.644
Figure 2.5: Welfare Gains in Each Time Period from Ambien CR Presence
UO
M
C~)
LO
C C3J
.2
O
oCDJ
V,
1997q3
i999q3
2001q3
Date
20d3q3
2005q3
Figure 2.6: Welfare Gain as a Pct of Total Revenue
UO
C
4)
V)
0
CU
1997q3
199 9q3
2001q3
Date
20d3q3
20d5q3
Appendix A- Hausman Instruments
Below I provide demand estimates using prices in other cities as instruments, as suggested by
Hausman (1997). While a compelling case can be made for the validity of the BLP instruments in
88
the case of pharmaceuticals, the prices in other regions instruments relies critically on correlations
between prices only being due to marginal cost commonalities. In my price measure, all prices are
correlated through their common correlation with the Average Wholesale Price, which is set by the
firm with demand shocks in mind.
Table 2.3: Results with Hausman Instruments
(1)
Nested Logit IV
logdiff
thirtyprice
-0.0339***
(0.00347)
0.298***
(0.0302)
0.625***
(0.0799)
0.534***
(0.0269)
logad
logconditional2
halflife
timesinceentry
0.0747***
(0.00145)
1.866***
(0.114)
-8.299***
(0.194)
pfe2
Constant
Time Fixed Effects
Region Fixed Effects
Yes
yes
1,022
Observations
0.857
R-squared
57.401
Hansen's J
<0.0001
p-value
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<O.I
While it is comforting that the estimates don't differ significantly apart from the coefficient on
advertising, the test of over identifying restrictions fails spectacularly, leading me to reject the
Hausman instruments in favor of using only the BLP instruments.
89
Appendix B- Alternative Nesting by Molecule
In this section, I re-run the entire analysis using the molecule nest. For this, I group together all of
the zolpidem tartrate molecule: Ambien, Ambien CR and generic Zolpidem. Sonata and generic
Zaleplon go in a nest together, and everything else gets its own nest. Estimates are below.
Table 2.4: Results with Molecule Nesting
(2)
(1)
Nested Logit OLS Nested Logit IV
logdiff
logdiff
thirtyprice
logad
logconditionall
halflife
timesinceentry
pfe2
Constant
-0.0179***
(0.00100)
0.161***
(0.00638)
0.825***
(0.0213)
0.521***
(0.0118)
0.0802***
(0.00112)
3.345***
(0.0739)
-9.021 ***
Region Fixed Effects
Time Fixed Effects
(0.119)
Yes
Yes
-0.0283***
(0.00173)
0.200***
(0.00921)
0.359***
(0.0930)
0.535***
(0.0196)
0.0720***
(0.00190)
2.393***
(0.190)
-8.406***
(0.127)
Yes
Yes
Observatio ns
R-squared
1,022
1,022
0.907
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<O.I
In some ways this seems natural and is the approach used both by Stem (1996) and Branstetter,
Chatterjee and Higgins (2011). However, I do not use it for only one main reason: conversations
with doctors suggest and the fit of the model confirm that the diagnosis nesting structure is how
decisions typically work in the sleep aid market. I run the full analysis here, and it gives a slightly
higher, but similar estimate of welfare gains. The overall expected gain over the seven years in this
specification, assuming that advertising is worthless (depicted in figures below) is $691 million.
The overall expected gain in this specification assuming that advertising enters as a demand shifter
is $3.44 billion.
90
Figure 2.7: Welfare Over Time With Molecule Nesting
Welfare Gains Over Time - Molecule Nests
LO
C)
0
0
AR
0
1997q3
1999q3
2001q3
Date
20d3q3
20d5q3
Figure 2.8: Welfare By Pct of Total Revenue with Molecule Nesting
Welfare Gains By Fraction of Market Revenue
4) tO
C)
19917q3
1999q3
200d11q3
20d3q3
20d q3
Date
Appendix C- Welfare Analysis Including Advertising
Here, I will include advertising in the welfare analysis, assuming that advertising is a demand
shifter that benefits someone materially. However, I am not considering the possibility that adver91
tising could bias expectations or create patient drug mismatches. This analysis will be expected to
produce a bigger welfare gain than the preferred specification because advertising has a positive
coefficient, and including a new product which will likely advertise will bring further benefits.
Given that Ambien CR and Ambien are produced by the same firm, one must consider that advertising for Ambien in the counterfactual world could be different than it is in the realized world. I
have not done a formal analysis of the supply side, so I cannot give a precise estimate of counterfactual values of advertising for either Ambien CR or Ambien. Instead, I will predict advertising
using market structure and time fixed effects. The result is decreased advertising for Ambien in
the counterfactual, but significant advertising for Ambien CR, with higher total advertising.
Below are the analogous graphs for both level of welfare gain per quarter and welfare gain as a
percentage of realized total revenue in the market. Over the seven year counterfactual span, the
welfare gain in the counterfactual is $4.79 billion.
Figure 2.9: Welfare Over Time Counting Advertising
Welfare Gain With Advertising
0
C
0
C:
C)
V0
C
Dat
Lo
1997q3
1999q3
20di q3
Date
92
20d3q3
2005q3
Figure 2. 10: Welfare By Pct. of Total Revenue Counting Advertising
Welfare Gain as Fraction of Market Revenue
0
CN
J0
CD
1947q3
199 19q3
200d11q3
Date
93
20d3q3
20d5q3
3
The Regulation of Prescription Drug Competition and Market Responses: Patterns in Prices and Sales Following Loss
of Exclusivity 6
Abstract
We examine six molecules facing initial loss of US exclusivity (LOE, from patent expiration or challenges) between June 2009 and May 2013 that were among the 50 most prescribed
molecules in May 2013. We examine prices per day of therapy (from the perspective of average revenue received by retail pharmacy per day of therapy) and utilization separately for
four payer types (cash, Medicare Part D, Medicaid, and other third party payer - TPP) and
age under vs. 65 and older. We find that quantity substitutions away from the brand are much
larger proportionately and more rapid than average price reductions during the first six months
following initial LOE. Brands continue to raise prices after generics enter. Expansion of total
molecule sales (brand plus generic) following LOE is an increasingly common phenomenon
compared with earlier eras. The number of days of therapy in a prescription has generally
increased over time. Generic penetration rates are typically highest and most rapid for TPPs,
and lowest and slowest for Medicaid. Cash customers and seniors generally pay the highest
prices for brands and generics, third party payers and those under 65 pay the lowest prices,
with Medicaid and Medicare Part D in between. The presence of an authorized generic during
the 180-day exclusivity period has a significant impact on prices and volumes of prescriptions,
but its impact varies across molecules.
3.1
Introduction and Background
Since the early 2000s many brand name prescription drugs have lost the exclusive right to sell
their products in the U.S. due to patent expiration and challenges. This loss of exclusivity (LOE)
resulted in substantially lower prices for payers and consumers and reduced revenues for brand
name prescription drug manufacturers. Moreover, since the 1980s, the speed with which generics
have gained market share from brands following LOE has accelerated 7 . Research has investigated
6
This chapter is jointly written by Murray L. Aitken (IMS Institute for Healthcare Informatics), Ernst R. Berndt
(Sloan School of Management, Massachusetts Institute of Technology, and NBER), Barry Bosworth (The Brookings
Institution), lain M. Cockburn (School of Management, Boston University, and NBER), Richard G. Frank (Department of Health Care Policy, Harvard Medical School, and NBER), Michael Kleinrock (IMS Institute for Healthcare
Informatics), and Bradley T. Shapiro (Department of Economics, Massachusetts Institute of Technology)
7
See, for example, Aitken, Berndt and Cutler [2008], Aitken and Berndt [2011], Berndt and Aitken [2011], and
Generic Pharmaceutical Association [2012, 2013].
94
market responses to LOE by examining the rate at which generics are substituted for brands 8 , the
path of prices (generic, brand and molecule average) paid 9 , total (brand plus generic) molecule
utilization following LOE10 , the relationship between number of generic manufacturers entering
markets and prices, and trends in the duration of market exclusivity prior to initial LOE" , among
other issues12
In recent years a set of policy debates related to LOE in the U.S. has centered on provisions in
the Hatch-Waxman Act that govern entry of generics. Most recently these provisions have figured prominently in the 2013 Supreme Court ruling on "pay for delay". The Hatch-Waxman rules
have increased the rate at which generic drug companies file so-called Paragraph IV challenges to
the patents held by originator companies, either contesting the validity of the patents that protect
brand name products or asserting that their version of the drug does not infringe them. The growth
in these challenges is in large measure due to the strong incentives created by the provision of
the Hatch-Waxman Act that grants a 180-day period of generic sales exclusivity to the successful
challenger. In these circumstances, the generic may be able to win substantial market share from
the brand with only a modest discount off the brand price, though it is important to note that the
substantial profits that can accrue to the generic during this period may be reduced by additional
competition from the brand that has the right to contract with a generic manufacturer to market
a so-called authorized generic product13 . In fact launching an authorized generic in the context
of Paragraph IV challenges from a generic manufacturer has become a widespread industry practice14 . Following the 180-day period of limited competition (duopoly with brand and exclusive
generic, or triopoly with the addition of an authorized generic), unfettered competition emerges.
As a result, the generic penetration and price reductions during the first 180 days following LOE
8
Berndt, Cockburn and Griliches [1996], Caves, Whinston and Hurwitz [1991], Cook [1998], Ellison, Cockbum, Griliches and Hausman [1997], Ellison and Ellison [2011], Frank and Salkever [1992, 1997], Grabowski and
Kyle [2007], Grabowski, Kyle, Mortimer, Long and Kirson [2011], Grabowski, Long and Mortimer [2011, 2013],
Grabowski and Vernon [1992, 1996],Griliches and Cockburn [2004], Hurwitz and Caves [1998], Reiffen and Ward
[2005], Saha, Grabowski, Birnbaum, Greenberg and Bizan [2006], Scott Morton [1999, 2000] and Wiggins and
Maness [2004].
9
Cook [1998], Frank and Salkever [1992, 1997], Regan [2008], and Wiggins and Maness [2004].
10
Aitken, Berndt and Cutler [2008], and Caves, Whinston and Hurwitz [1991].
"Grabowski, Kyle, Mortimer, Long and Kirson [2011], Grabowski and Kyle [2007], and Hemphill and Sampat
[2012].
12
Other aspects studied include factors determining the extent and composition of generic manufacturer entry (Scott
Morton [1999, 2000]), characteristics of molecules that impede or accelerate generic penetration (Grabowski, Long
and Mortimer [2011,2013]), differential therapeutic class composition between retail and mail order generic drug
dispensing (Wosinska and Huckman [2004]), and variation among states with large public funded programs such as
Medicaid in exploiting cost savings opportunities following the brand's LOE (Avalere Health LLC [2010], Brill [2010]
and Kelton, Chang and Kreling [2013]).
13Appelt [[2013]; Berndt, Mortimer, Bhattacharjya, Parece and Tuttle [2007], Branstetter, Chatterjee and Higgins [2011]; Federal Trade Commission [2002, 2009, 2011], Grabowski, Kyle, Mortimer, Long and Kirson [2011],
Hemphill and Sampat [2012], Olson and Wendling [2013], Panattoni [2011], and Wang [2012].
14
Federal Trade Commission [2011].
95
can differ substantially from subsequent generic price and quantity movements when generic entry
is unfettered 15
We extend this literature by comparing the magnitude of quantity movements with the size of price
reductions during and after the 180-day exclusivity period. We do so by carefully examining six
molecules facing initial LOE between June 2009 and May 2013 that were among the fifty most
prescribed molecules in the US in May 2013. We focus on retail selling prices, and are able
to disaggregate buyers in the overall retail market and separately examine the relative price and
quantity movements before and following LOE for four classes of payers: Medicaid, Medicare
Part D, commercial and other third party payers (TPPs), and cash customers. Relatively little is
known about retail price and quantity movements during the 180 day exclusivity period16 or about
the magnitude of any differences by payer type in prices paid, and in the speed and extent of shifts
from brands to generics. Finally, since information is available on the age of customers receiving
medications, we examine relative prices and generic substitution rates for those under age 65 with
those age 65 and older.
Our analyses yield the following main findings. First, quantity substitutions away from brands
are much larger proportionately and more rapid than average molecule price reductions during
the first six months following LOE. Second, brands continue to raise prices after generics enter.
Third, expansion of total molecule sales (brand plus generic) following LOE is an increasingly
common phenomenon compared with prior observations. Fourth, generic penetration rates are
generally highest for third party payers and lowest for Medicaid. Fifth, cash (and seniors over age
65) generally pay the highest prices for brands and generics, third party payers (and those under
age 65) pay the lowest prices, with Medicaid and Medicare Part D prices being in between those
of cash and third party payers. Sixth, the presence of an authorized generic during the 180-day
exclusivity period has a significant impact on prices and volume of prescriptions, but its impact
varies across molecules. In two of the cases studied, the brand and its licensee collectively retained
an almost two-thirds share of the market by volume, in the others they captured less than half.
Price discounts off the brand prevailing fee during the "triopoly" period also showed substantial
variation. In some cases, the price of the authorized generic product was between the brand and
the independent generic, in others it was significantly below the independent generic.
15
Appelt [2013], Berndt, Mortimer, Bhattacharjya, Parece and Tuttle [2007], Olson and Wendling [2013] and Reiffen and Ward [2007]. For a discussion of the implications of Paragraph IV challenges on consumers' and producers'
welfare, see Branstetter, Chatterjee and Higgins [2011].
16
See, however, Federal Trade Commission [2011], ch. 3 and Alpert, Duggan and Hellerstein [2013].
96
3.2
Data and Methods
The IMS Health Incorporated National Prescription Audit (NPA) data base tracks prescriptions
dispensed at a nationally representative sample of retail, mail order and long term care pharmacies
and is projected to an estimate of total national prescriptions dispensed through these pharmacies
on a monthly basis. We limit our analysis to prescriptions dispensed at retail pharmacies, and
focus on the 50 most prescribed molecules (measured by number of prescriptions dispensed during
May 2013). For each of these molecules, data are available on the distribution by payer type:
Medicaid, Medicare Part D, commercial or other third party payer (TPP), and cash; for about half
the transactions, information is also available on the age of the patient dispensed the prescription:
65 and over, or under age 65.
These NPA data reflect the perspective of the retail pharmacy, and prices measured at this point
in the distribution chain correspond to the retail prices that the U.S. Bureau of Labor Statistics
attempts to measure in constructing its monthly Consumer Price Index for Prescription Drugs17 .
The total revenue received by the dispensing pharmacy is the sum of the customer's co-payment
or coinsurance contribution, plus the amount (if any) paid to the dispensing pharmacy by the third
party payer - Medicare Part D insurer, Medicaid, or commercial or other third party payer. This
pharmacy price therefore includes reimbursement for the active pharmaceutical ingredient, a dispensing fee, and any customer cost sharing contribution 18 . For our purposes, however, it is important to note that this pharmacy price already includes margins realized by wholesalers, pharmacies
and any other distributors, and is therefore generally larger than the price (net of rebates and discounts) received by the brand and generic manufacturers. These rebates and discounts can be
substantial, in some cases larger than the amount that consumers actually pay 19 .
Starting from the 50 most prescribed molecules in May 2013 we used information from the FDA's
Orange Book to identify the six of these molecules that faced initial LOE during the June 2009 May 2013 time period 20 . For each of the six molecules, we identified the National Drug Codes
17
Berndt, Cutler, Frank, Griliches, Newhouse and Triplett [2000], and U.S. Department of Labor, Bureau of Labor
Statistics [2011].
181f the customer presents the pharmacy with a coupon, the value of that coupon is attributed to the patient copayment or coinsurance contribution; how well the NPA is able to capture coupon transactions is not publicly known.
19
Federal Trade Commission [2011] and Centers for Medicare & Medicaid Services [2012]. According to the
Takeda 2011-2012 Prescription Drug Benefit Cost and Plan Design Report (Table 25, p. 31), the 2011 survey of
employers receiving rebates from their pharmaceutical benefit management firms revealed that the average (lowest,
highest) rebate received for a 30 day retail prescription was $10.06 ($2.64, $21.71) and for a mail order script was
$34.74 ($2.70, $72.59). The trend in average rebates was increasing in time, with the Report stating "...the growth
in rebate amount may be due in part to the large number of brands going generic in the next year or two. Increasing
rebates to capture market share before losing patent is a common strategy."
20
One brand product, Plavix (generic name clopidogrel), faced an at-risk launch by Apotex in August 2006; later
97
(NDCs) for the brand's strengths/formulations, and the monthly number of prescriptions dispensed
by payer type (Medicare Part D, Medicaid, third party commercial payer, and cash customer), customer age (number under age 65, age 65 and over, and age unknown), average customer cost
sharing contributions, and mean reimbursement to the pharmacy by TPPs. Similar monthly data
were obtained for generic and authorized generic versions for the strengths/formulations of the
molecule. The NDCs for authorized generics were identified, and were alternatively treated separately or combined with NDC codes of abbreviated new drug application (ANDA) holders. Data
on the mean number of extended units (EUs, e.g., tablets, capsules) per prescription along with recommended daily dosing data from the Drugs@FDA website were used to convert average price per
prescription to average price per day of therapy (only Augmentin XR differed from once-daily dosing, with its recommended dosing being twice daily for 7-10 days). These NDC-specific average
data were then aggregated up to the molecule level separately for brands, generics and authorized
generics, by payer type and customer age group, using relative number of prescriptions by NDC
code as weights.
3.3
Findings: Characteristics of Six Initial Generic launches, June 2009May 2013
The most salient characteristics of the six molecules are summarized in Table 3.1, in chronological
order by date of LOE from left to right. When originally approved by the FDA as New Drug
Applications (NDAs), three of the six were new molecular entities, whereas the other three were
new formulations of a previously approved molecular entity. Four of the six were designated for
standard review by the FDA, and two were tagged for priority review. The six molecules are in
six different therapeutic classes, and represent five different original NDA holders (Sanofi Aventis
is the only multiple NDA holder at two). Effective market exclusivity (time between initial NDA
approval and first ANDA entry) varies substantially among the six molecules-from about 5.5
years for Ambien CR to about 15.0 years for both Cozaar and Lipitor. Although all six derive from
the 50 most prescribed molecules, because several are new formulations of an older molecule,
the market size (measured by total number of monthly prescriptions) varies dramatically across
that month Bristol Myers Squibb (the firm marketing Plavix in the US) obtained an injunction preventing further
production and distribution by Apotex, and subsequently won a patent infringement case against Apotex. As the
single non-brand entrant in the market for the three-week period, Apotex charged about 87% of the brand's price, but
stuffed inventory channels with massive sales. Although Apotex had been awarded tentative Paragraph IV exclusivity,
this exclusivity was forfeited by Apotex. By late 2007 Apotex's clopidogrel inventory had essentially disappeared
from the US retail marketplace. We therefore treat the May 2012 launch of generic clopidogrel as the initial LOE for
Plavix, and not the August 2006 at-risk launch by Apotex. Additional details are given in Berndt-Aitken [2011]; for a
journalist's account of the Plavix - clopidogrel episode, see Smith [2007].
98
molecules. Specifically, for the three months prior to initial LOE, the market size ("Mean TRX
Before LOE" in Table 3.1) varied from largest (Lipitor - atorvastatin) to smallest (Augmentin XR
-amoxicillin/clavulanate potassium) by a factor of 1251:1; market size of the second largest (Plavix
- clopidogrel) relative to second smallest (Ambien CR - zolpidem tartrate) varied by a factor of
4.66:1, while the market size of the third largest (Lexapro - escitalopram oxalate) relative to the
third smallest (Cozaar - losartan) was 2.33:1. Hence, the sample of six molecules represents great
variation in market size of the molecular formulations.
All six brands faced successful Paragraph IV challengers, although because of its infringing prepatent expiration entry, Apotex forfeited its 180-day exclusivity for Plavix (clopidogrel). As a
result, Plavix was the only brand to face unrestricted generic entry at the time of its LOE. Of the
remaining five molecules, following LOE four of the brands (Cozaar, Ambien CR, Lipitor and
Lexapro) launched authorized generics thereby creating a triopoly market for 180 days, while the
remaining brand (Augmentin XR) did not launch an authorized generic and thereby faced duopoly
competition for 180 days. By seven (twelve) months after initial generic entry, the molecule with
the smallest pre-patent expiration market size that also involved complex manufacturing processes,
(Augmentin XR) had just two (two) competitors, a reformulated extended release molecule with
modest pre-patent expiration market size (Ambien CR) had four (five) competitors, while the
molecule with the largest pre-patent expiration market size (Lipitor) had only six (seven) competitors. Three other molecules (Cozaar, Lexapro and Plavix) each had 13 or 14 competitors at
both seven and twelve months following initial ANDA entry.
A more detailed discussion of each of the six molecules is provided in the Appendix.
3.4
Results
We now discuss results, in separate sub-sections for generic penetration rates by payer type and
age; quantities post-LOE relative to pre-LOE; number extended units per prescription; prices by
payer type pre-LOE, during any 180-day exclusivity periods, and post- 180-day exclusivity; as well
as prescription shares and prices for brands, Paragraph IV challengers and authorized generics
(AGs) during the 180-day exclusivity period.
3.4.1
Generic Penetration Rates, Over All and By Payer Type and Age
We computed the generic penetration rate as the proportion of all prescriptions for a given molecule
dispensed as generics (or authorized generic). With six molecules, five payer types and two age
99
groups, the number of quantitative findings is voluminous. As an overview, we first ask, how long
in number of months does it take for a molecule to reach certain specified generic penetration
thresholds? Results for 60% and 90% generic penetration thresholds are presented in Table 3.2. A
number of findings are striking.
First, looking at the six molecules over all payer types (the top row in each molecule panel), we
observe that for all six drugs, the time required to reach a 60% generic penetration threshold is
three months or less. A 60% generic penetration threshold was reached in one month or less by
Lexapro, Plavix, Cozaar and Lipitor (<1 month is interpreted as the average generic penetration
in the first month of entry exceeding 0.6); for the two smallest market drugs, Augmentin XR and
Ambien CR, the 60% threshold over all payer types was two and three months, respectively. The
90% threshold for all payer types is attained within two months for Plavix (that faced unfettered
generic entry throughout), four months for Cozaar and Augmentin XR, six months for Lexapro,
and nine months for Lipitor; only for Ambien CR is the 90% threshold more than a year (13
months). This very rapid shift from brand to generic following LOE is much greater than has been
reported in earlier US studies2 1
A second set of findings in Table 3.2 reflects differences in the speed of generic penetration by
payer type. Looking at each of the four payer types and at all payers within each molecule, we see
that for all six molecules, third party payers (TPPs) are always the most (or tied for the most) rapid
in reaching the 60% generic penetration threshold, whereas Medicaid is the slowest. To reach the
90% generic penetration threshold, in all cases but one (Lipitor, Medicare), TPPs take the shortest
amount of time, followed by Medicare, cash customers, and finally, Medicaid. The relative speed
with which TPPs reached high generic penetration thresholds could reflect in part aggressive formulary management by pharmaceutical benefit managers (PBMs) working on behalf of TPPs and
Medicare Part D prescription drug plans. Other researchers have noted the relatively slow generic
take-up by Medicaid and have leveled criticism at the program for failing to exploit available cost
savings (e.g., Brill [2010].) It is worth noting, however, that manufacturer rebates to Medicaid from
brand manufacturers are several times larger than those from generic manufacturers (currently on
average about 30+% for brands having recent price increases vs. 13% for generics) 22 , and during
21
See, for example, Grabowski and Vernon [1992,1996], Griliches and Cockburn [1994], Ellison, Cockburn,
Griliches and Hausman [1997], Cook [1998], Aitken, Berndt and Cutler [2008], and Berndt and Newhouse [2012].
22
For the purpose of calculating Average Manufacturer's Price (AMP) and Best Prices that trigger Medicaid rebates,
since 2007 the prices of authorized generics have been treated as brand prices, not generic (Federal Trade Commission
[2011], p. 13 and Appendix J). For brands, in addition to the statutory 23.1% rebate, the extent to which the brand's
current quarter AMP has exceeded growth in the Consumer Price Index-Urban since product launch is applied to the
rebate, implying that for many brands, the total Medicaid rebate is above 30%. For non-innovator multisource drugs
(independent generics), the rebate is 13% of AMP, with no adjustment for excess price inflation. See Medicaid Drug
Rebate Program [2013].
100
the first few months following LOE in which there is only limited competition (say, at the beginning of the 180-day exclusivity period), it is possible that net of rebates, prices to Medicaid may
be lower for brands than generics, thereby rationalizing a slower speed of generic substitution by
the Medicaid programs. Recall that the prices we measure here are the total consumer plus TPP
payments to retail pharmacies, not prices net of rebates paid by payers or received by generic manufacturers. The lower brand than generic price net of rebates to Medicaid is not just a theoretical
possibility. In their evaluation of state Medicaid program responses to Prozac's (fluoxetine) LOE in
August 2001, for example, Kelton, Chang and Kreling [2013, p. 1207] report that during 2001Q3
- the first full quarter in which generic fluoxetine was available - net of estimated Medicaid rebates
the average price of a 20 mg tablet/capsule of branded Prozac at $1.91 was slightly less than the
average price of a 20 mg tablet/capsule of generic fluoxetine at $1.95.
A third set of results in Table 3.2 is a negative finding: Evaluated at either the time required to
reach a 60% or 90% generic penetration threshold, over all payer types those in the under age 65
group take on average about the same length of time to reach the 60% generic penetration as do
older adults (age 65 and over); for the 90% threshold over all payers, for three of the six drugs
seniors substitute more rapidly, and for one more slowly, than do those under age 65; for the other
two drugs the switching speed is the same. Moreover, when one examines time to reach thresholds
across payer types, there does not appear to be any dominant pattern for older adults vs. those
under age 65.
The patterns observed in Table 3.2 suggest that consumers take FDA judgments about interchangeability at face value and benefits managers make polices independent of clinical or demographic
circumstances. We note in passing because TPPs often manage prescription drug benefits for both
firms' employees and their retirees (via retiree wraparounds), and because Medicare beneficiaries
include some individuals under age 65 (e.g., End-Stage Renal Disease beneficiaries and Social
Security Disability insured beneficiaries), there is considerable overlap between prescriptions paid
for by Medicare and those dispensed to customers age 65 and over.
Finally, the extent and speed of generic penetration in the single case with 180-day exclusivity but
no authorized generic - Augmentin XR - appears to be a bit less aggressive and rapid than for
Cozaar, Lexapro, and Lipitor (but not for Ambien CR), each of which had an authorized generic
during its 180-day exclusivity period; Cozaar, Lexapro and Lipitor reached the 60% threshold in
one month or less, whereas Augmentin XR and Ambien CR needed two-three months. With no
180-day exclusivity and unfettered generic entry at patent expiry, Plavix reached the 90% generic
penetration threshold in the shortest amount of time - two months.
101
3.4.2
Quantities Post-LOE Relative to Pre-LOE
Typically we expect demand curves to slope downward - as price declines, other things equal,
quantity demanded increases. For quite some time, however, it has been conventional wisdom that
total brand plus generic utilization of a molecule declines following patent expiration as average
molecule price falls 23 . This has been attributed to the fact that brands reduce their marketing as
the date of patent expiration and initial loss of exclusivity (LOE) approaches, i.e., other things are
not equal. Brands then tend to terminate almost all marketing efforts immediately following LOE.
As reported by Aitken, Berndt and Cutler [2008], however, this is not always the case. When the
statin Zocor went off-patent in 2006, payers and PBMs aggressively switched individuals taking a
relatively costly statin brand drug Lipitor (still under patent protection) to generic versions of Zocor
(simvastatin), and also initiated new patients on simvastatin instead of Lipitor. We now examine
whether the Zocor experience is unique or has become more common. Results are presented in
Table 3.3, alternatively for average utilization six to nine months post-LOE relative to the three
months pre-LOE, and for 10-12 months post-LOE relative to the three months pre-LOE, over all
payer (buyer) types and both age groups. A number of results are worth noting.
First, for four of the six molecules, across all payer types the total molecule utilization at both
six-nine and ten-twelve months post-LOE is greater than during the three complete months prior
to LOE. For one of the molecules (Plavix-clopidogrel) total molecule utilization is essentially the
same, whereas for one other molecule (Ambien CR - zolpidem tartrate) there is a very slight
reduction in total molecule utilization post-LOE. The reason for post-LOE total molecule growth
in utilization being so large for Augmentin XR is unclear to us, but we note from Table 3.1 that
the number of monthly prescriptions in the three complete months pre-LOE was very small for
Augmentin XR24
Second, considerable variation occurs among payer types. Notably, in the four cases when postLOE total molecule utilization increases, the increase is driven primarily by TPP payers (although
for Lipitor cash and Medicare customers are also large drivers of increased molecule utilization).
This is consistent with the large shift to generics and the substantial reduction in average molecule
price that results. The smallest utilization increase or largest decrease occurs for Medicaid payers.
In the one case where total molecule utilization decreases slightly post-LOE (Ambien CR), it is
Medicare and Medicaid prescriptions that decline most post-LOE. Since we observed that Medi23
See, for example, Berndt, Cockburn and Griliches [1996], Caves, Whinston and Hurwitz [1991], Cook [1998],
Ellison and Ellison [2011], Grabowski and Vernon [1996], Hurwitz and Caves [1988], Huskamp, Donohue, Koss,
Berndt and Frank [2008] and Regan [2008].
24
The number of Augmentin XR prescriptions in the three months pre-LOE displayed no growth pattern; the increase post-LOE represents a break in trend (results not shown).
102
caid is generally the slowest payer to switch from brand to generic, the relatively large reductions
post-LOE by Medicaid payers suggests they are not staying with the brand, but rather are either
discontinuing treatment with that molecule altogether or are instead switching to another molecule,
either brand or generic. This issue merits further examination.
Table 3.3 shows that in all four cases in which 10-12 months post-LOE total molecule utilization
increased, total molecule utilization by cash customers increased. Note that because certain retail chains such as WalMart and Target have recently introduced $4 prescriptions for 30 days and
$9.99 prescriptions for 90 days, customer out-of-pocket cash payments for these prescriptions were
likely less than the typical co-payment or coinsurance customer contribution to the pharmacy for
a first-tier generic drug under a private or public insurance plan formulary benefit design. While
this shift to box merchandiser pharmacies might explain some of the growth in cash payer total
molecule utilization, because Medicaid beneficiaries typically have very low if any co-payment
for first tier generic drugs, the shift to box merchandiser pharmacies is unlikely to be the source
of the post-LOE decline in Medicaid total molecule utilization. Alternatively, price declines to
the uninsured may generate more demand response for the molecule than do price declines to an
insured population. The extent to which the increase in price reduction-induced utilization can
be decomposed into increased demand from existing patients (through better compliance, i.e., the
resumption of temporarily discontinued therapy) versus demand from new patients, for example
those that switch in to a molecule experiencing generic entry from another branded molecule or
are now receiving initial drug treatment with a generic, is unclear, potentially has important implications for consumers' welfare, and merits further analysis25 .
Third, there is considerable heterogeneity among the five molecules having a 180-day exclusivity period. For Augmentin XR having no authorized generic entry, and for Ambien CR - both
extended release reformulations - the post-LOE utilization experiences are dramatically different,
with Augmentin XR showing a very substantial increase and Ambien CR a slight decline. While
the other three molecules having an authorized generic present during the 180-day exclusivity each
experienced a post-LOE total molecule utilization increase, the extent of this increased utilization
varied considerably - being largest for Cozaar, followed by Lipitor and then Lexapro.
Finally, regarding total molecule utilization post- vs. pre-LOE by age group, as seen in the final
columns of Table 3.3, there is no striking pattern. In most but certainly not all cases, those under
age 65 increase more or decrease less in their total molecule utilization than do those age 65 and
25
Professor David Kreling of the University of Wisconsin School of Pharmacy has communicated to us informally
that the composition of generic drugs offered by the box mass merchandisers at $4 (30 day) or $9.99 (90 day) prescription prices tends to be dominated by mature generic drugs, rather than by molecules more recently having experienced
initial generic entry.
103
over six to nine months after LOE, but even this modest trend is mitigated at ten to twelve months
after LOE.
3.4.3
Number Extended Units Per Prescription
One of the strategies employed by PBMs has been to encourage beneficiaries to switch from filling
30-day prescriptions at "brick and mortar" retail pharmacies to ordering 90-day prescriptions via
mail order. Such a switch has been accomplished in part by 90-day co-payments being only twice
as large as 30-day co-payments, thereby reducing beneficiaries' per diem co-payment amount.
This strategy has been particularly attractive for maintenance medications that treat chronic diseases (i.e., are taken each day indefinitely), but obviously is less practical for medicines needed
immediately to treat acute conditions or episodes26 . In response to seeing reduced foot traffic in
their brick and mortar stores from this shift to 90-day mail order prescriptions, several retail pharmacy chains have begun offering co-payment incentive schemes similar to those by the PBM mail
order firms, calling them "channel neutral" co-payments27
One implication of this shift from 30- to 90-day prescriptions is that the number of extended
units (EUs, e.g., tablets, capsules) per prescription is likely to have increased somewhat during
our 48-month sample time period (June 2009 - May 2013). Since our data reflect only retail
pharmacy dispensing and excludes mail order, we expect this increase in number of EUs per Rx
to be modest. However, we also expect that the extent to which prescriptions contain more EUs
per prescription will vary considerably among the six molecules in our sample since they are in six
distinct therapeutic classes of medicines treating either chronic or acute conditions. A consequence
is that the average price per prescription not taking into account shifts in the number of EUs per Rx
could give a misleading picture of prescription drug per diem price changes over time. Therefore,
before presenting results on price differences among payer types and age groups, we digress and
first examine trends among our six molecules on number of EUs per Rx.
Due to space limitations, we do not present detailed results on EUs per Rx here; detailed figures
can be accessed at the IMS Institute for Healthcare Informatics website (www.theimsinstitute.org)
. Our principal findings on number of EUs per Rx can be summarized as follows. For all molecules
except Augmentin XR, the number of EUs per Rx increased over time, with the largest increasing
being from about 39 to 45 for Cozaar. For Augmentin XR with twice daily dosing recommended
26
Wosinska and Huckman [2004] discuss variations in the therapeutic class composition between prescriptions
dispensed by retail versus mail order.
27
Trends in mail order vs. retail prescription co-payment levels and coinsurance rates are discussed in Berndt and
Newhouse [2012], pp. 241-48.
104
for 7-10 days, while the number of EUs per Rx for the branded version increased from about 32
to 38 (16 to 19 days' supply), for the generic version it declined from 36 to 32 (18 to 16 days'
supply). The smallest number of EUs per Rx occurred with Ambien CR prescriptions, for whom
the increase during the sample time period was from almost 29 to just under 31 EUs per Rx. We
conclude that while there is heterogeneity in EUs per Rx across the six molecules, the additional
variability over time implies that it is preferable to measure trends in price per day of therapy rather
than price per prescription.
3.4.4
Price Per Day of Therapy By Payer Type
Prices per day of therapy by payer type for the six molecules over the June 2009 - May 2013 time
period are graphed in the six panels of Figure 3.1; the solid red vertical line denotes the date of
initial LOE, while the dotted red vertical line represents 180 days later, corresponding with the
expiry of any 180- day exclusivity (except for Plavix). Several results in Figure 3.1 merit special
attention.
First, although the decline in average molecule price per day of therapy at the time of initial LOE is
evident for all molecules, the price drop is most dramatic for Plavix (clopidogrel) - panel in lower
right corner, where generic entry at the time of LOE was unfettered, resulting in the average price
per day of therapy falling from about $6.50 to about $3.50 (46%) a month after initial LOE.
Second, we see important differences in prices across different classes of payers. We treat these
cautiously, given the potential for differences in the level of unobserved rebates received by different classes of payers. (Recall that our price measure is from the perspective of average revenue
per day of therapy received by the pharmacy, which does not include manufacturers' rebates to
Medicaid and other payers; net of such rebates, Medicaid price premia might be much smaller,
and perhaps may even be non-existent.) Nonetheless, it is interesting to note that both pre-LOE
and during the 180-day exclusivity window, cash payers paid the highest prices, TPPs the lowest
prices, with Medicaid and then Medicare Part D in between. Price differences among payer types
generally tended to decline over time, and by the end of the sample time period (May 2013) cash
payers paid the highest prices for four of the six molecules (Ambien CR, Augmentin XR, Lexapro
and Plavix), while Medicaid paid the highest prices for Cozaar and Lipitor. For Lipitor, the average
price per day of therapy for Medicaid at just under $3 was about twice the amount paid by TPPs.
Third, an intriguing phenomenon we observe in Figure 3.1 is that while at the time of initial LOE
and during the 180-day exclusivity, there is a noticeable immediate price decline, for Cozaar,
Ambien CR, Lipitor and Lexapro - each of which was in a triopoly market structure- the average
105
price per day of therapy is relatively flat during the remainder of the 180-day exclusivity window,
and then (except for Ambien CR) falls sharply immediately following expiry of the exclusivity
window. Somewhat counter-intuitively, the post-LOE price decline is larger and more sustained
for Augmentin XR - in a duopoly market structures with no authorized generic entrant during
the 180-day window. Why a three-competitor market structure generates higher and more stable
prices than does a duopoly runs counter to basic economic intuition and merits further analysis 28.
In particular, future research might focus on the role played by the brand and its authorized generic
agent in maintaining price discipline during the 180-day exclusivity window.
Finally, although we do not present price trends per day of therapy separately for those under age
65 and age 65 and over, we find there do not appear to be systematic differences paid per day of
therapy by age group, and that any differences are relatively small in magnitude.
What is clear, however, in comparing quantity data in Tables 3.2 and 3.3 with price trends in
Figure 3.1 is that even though the price reductions in the 180 days immediately following initial
LOE are significant, they are much smaller and slower than the very dramatic increases in generic
penetration rates, i.e., for these prescription drugs, following LOE, quantities move much more
quickly and proportionately than do prices.
3.4.5 Prices and Prescription Shares During the 180-Day Triopoly
In Table 3.4 we present means (and standard deviations) for price per day of therapy and prescription shares during the four 180-day triopolies we observe in our data set (Cozaar, Ambien CR,
Lipitor and Lexapro). Although our sample size includes only four molecules, and any results
should therefore be viewed as tentative, several preliminary results are worth noting.
First, in terms of maintaining market share in the face of LOE, Sanofi's experience with Ambien
CR and Pfizer's experience with Lipitor stand out. As seen in the bottom three panels of Table
3.4, while for both Cozaar and Lexapro the mean brand share during the triopoly falls to about
15%, at 44% and 40%, respectively, the Ambien CR and Lipitor brand shares were much larger;
in addition, while the AG share for both Cozaar and Lexapro was about 35%, for Ambien CR
and Lipitor it was just under 25%. An implication is that when one sums the prescription share
for the brand plus that of its authorized generic, for both Cozaar and Lexapro this comes to about
28
For a preliminary theoretical analysis and empirical implementation based on data from the 1980s and 1990s, see
Reiffen and Ward [2007].
106
50%, whereas for Ambien CR and Lipitor it soars to 64-68%29. For the successful Paragraph IV
challenger, therefore, while being able to capture approximately 50% of prescriptions during the
triopolies involving Cozaar and Lexapro, during Ambien CR's and Lipitor's180-day exclusivity
period, Actavis Elizabeth and Ranbaxy - the independent non-authorized generic entrants - only
garnered about 35% of prescriptions. Thus the rewards during 180-day exclusivity to being the
successful Paragraph IV challenger are mixed. Industry analysts have suggested that Pfizer was
able to secure its 40 per cent substantial share by giving payers, mail order firms and pharmaceutical benefit managers (PBMs) large rebates, aggressively marketing $4 copay coupons in the print
media, and maintaining and perhaps even increasing direct-to-consumer marketing efforts before
LOE and during the 180-day exclusivity period 30 . We have no information regarding whether
Sanofi undertook similar actions to protect its brand and authorized generic shares during the Ambien CR exclusivity period (although as noted below during the exclusivity period the brand over
generic price premium was only 22 per cent). In addition, whether the more recent Lipitor experience remains historically unique or instead ushers in a new form of triopoly competition remains
to be seen; the only other major brand facing initial LOE since the 2011-2012 Lipitor LOE was
Plavix, but as noted earlier, because Apotex forfeited its Paragraph IV exclusivity, Plavix faced
unfettered generic entry at the time of its initial LOE later in 2012 rather than a triopoly market
structure.
A second set of findings (in the top panels of Table 3.4) involves revenues or average prices (any
patient co-payment plus third party payer reimbursements) received by retail pharmacies. Here
the outlier drugs are Cozaar and Ambien CR, not Lipitor. For both Lipitor and Lexapro, the
AG average price is in between that of the brand and the generic (the successful Paragraph IV
challenger), whereas for Ambien CR and Cozaar the AG average price is even lower than that of
the generic; this latter phenomenon of AG retail prices being lowest has also been documented
by the Federal Trade Commission [2011, ch. 3], but our finding based on more recent data of the
independent generic average price being the lowest is novel. In all four cases, however, the brand
has the highest price, being 25-30% higher than the generic except in the cases of Ambien CR and
Cozaar where the brand-generic premium ranges between about 15-22%. It is worth emphasizing
again that the prices measured here are those recouped by the retail pharmacy, and not the prices
at which they acquire drugs from generic manufacturers (which are typically much lower) 31.
29
When the authorized generic is launched by the subsidiary of the brand (e.g., Winthrop for Sanofi Aventis), the
brand franchise captures all non-independent generic sales dollars. However, when the brand licenses out authorized
generic marketing rights to an independent generic manufacturer, the brand franchise still benefits from royalties it
receives from the independent generic manufacturer. In recent years, according to the Federal Trade Commission
[2011, p. 85], this royalty rate has been 90% and above.
30
See, for example, Drug Channels [2012].
31
Ibid. Also see Federal Trade Commission [2011, chs. 3 and 6] and Olson and Wendling [2013].
107
3.4.6
Cash vs. Full Sample Average Price Levels and Growth Rates
Our final set of analyses involves examining relative price levels and growth rates of prescriptions
paid for by cash in comparison to the full sample of retail dispensed prescriptions. Several outcomes are plausible. One line of reasoning is that because cash customers must pay the full price
of the drug, cash purchasers must value the prescription at a relatively high level, and knowing
this, pharmacies can exploit this fact by charging cash customers higher prescription prices. A
related view is that cash customers are less informed and cannot shift market share across products, whereas benefit managers for third party payer insurers are more knowledgeable concerning
alternative treatments for various conditions, and can use this knowledge and bargaining power
to obtain lower prescription prices from pharmacies. An alternative view is that cash customers
will be on average more price sensitive, and therefore will seek out those pharmacies advertising
discounted prescriptions, such as the mass merchandiser pharmacies, thereby paying lower prices
than those with insurance. While these views result in diverse predictions for relative price levels
paid by insured vs. the uninsured, without additional assumptions they make no predictions on
relative growth rates of prescription prices for cash vs. insured customers.
To measure mean price growth rates, for each molecule we compute the average of log [P(t)/P(t- 1)]
over the selected time interval, which for relatively small price changes such as that observed with
our monthly data, yield results that can be interpreted as approximating the mean monthly per cent
growth rate in price.
In the top row of each of the six drug molecule panels in Table 3.5, we present mean cash/full
sample prices (standard deviations in parentheses) over the full 48-month June 2009 - May 2013
sample time period, and then for three sub-periods: pre-LOE, during the 180-day exclusivity, and
post- 180 day exclusivity. In the case of Plavix, there was no 180-day exclusivity and we therefore
present relative cash/full sample prices for only the pre-LOE and post-LOE time periods where in
the latter generic entry is unfettered. Several results are worth highlighting.
First, for all six molecules prices paid by cash customers are generally greater than those for the full
sample. Over the entire 48-month time period and six molecules, the average cash price premium
is about 17%, ranging from 11% for Lipitor to 24% for Augmentin XR. In the pre-LOE time
periods, cash prices for brands are on average about 16% greater than for the full sample, and this
cash price premium grows slightly to about 18% during and following 180-day exclusivity. There
is remarkable little variability in the cash price premium during the 180-day exclusivity window
(small standard deviation), and generally (except for Augmentin XR) considerably more variability
following unfettered generic entry in the post- 180 exclusivity day time frame.
108
While results on relative price levels are clear and quite robust, for relative growth rates our findings are more nuanced. Looking at the pre-LOE column in Table 3.5, when we compare the mean
log [P(t)/P(t-1)] over the six molecules, we observe that cash prices increase on average about
0.0078 percent per month, very slightly greater than the full-sample prices that increase on average
0.0073 percent monthly, which when accumulated over 12 months results in cash prices annually
growing at 0.6% more rapidly than full sample prices. During any 180-day exclusivity period, on
average cash prices fall slightly less rapidly (-0.0268 vs. -0.0294) than do the full sample prices,
but this inequality is reversed following any 180-day exclusivity or when unfettered generic entry
occurs, during which time cash prices fall about 1.2% more rapidly (-0.0345 vs. -0.0228) than do
full sample prices. This last result is consistent with the pricing strategies of the mass merchandisers such as WalMart that offer $4 30-day or $9.99 90-day prescriptions to their customers once
unfettered generic entry occurs, typically lower prices than those available from chain and independent retail pharmacies. Finally, averaged over all six molecules and the entire 48-month time
period, cash prices fall very slightly more rapidly (-0.0117 vs.-0.0097, a difference of -0.0020 per
month, or about 2.4% annually) than do full sample prices. Hence, cash vs. full-sample differences
in price levels are quite substantial though stable over time, but differences in their growth rates
vary during exclusivity and LOE sub-periods, although in general these growth rate differences are
relatively small, i.e. the cash price level premia are proportionately stable over time.
3.5
Summary and Concluding Remarks
The extent and rate at which generic drugs capture market share in U.S. retail drug markets as
brands lose market exclusivity has increased sharply over the last decade. This heightened generic
penetration has been particularly evident for third party payers (TPPs), and likely reflects the increased bargaining power derived from formulary design by pharmaceutical benefit management
(PBM) firms serving TPPs' drug benefit plans. One implication of this phenomenon is that to
the extent "reverse payment" or "pay for delay" settlements result in delayed generic entry, this
phenomenon may have important implications for calculating the impact on purchasers of slower
access to lower cost medicines. The relatively large number of top-selling drugs facing initial loss
of exclusivity (LOE) in the U.S. in 2012 and 2013 has been unprecedented, with the resulting
patent cliff revenue losses for brands in 2012 alone approaching $29 billion and contributing to an
overall 1% decline in U.S. nominal pharmaceutical spending 32 , but providing a temporary windfall for the profit margins of wholesalers, retail and mail order pharmacies. Whether these recent
impacts on pharmaceutical spending will be repeated is unclear, particularly since the total dollar
32
1MS
Institute for Healthcare Informatics [2013], pp. 8,12; FiercePharma [2012].
109
revenues of brand drugs facing initial LOE in the next few years is expected to be considerably
smaller than in 2012 and 201333.
For four of the six molecules experiencing initial LOE in our 2009-2013 time frame, total monthly
molecule utilization post-LOE exceeded that prior to patent expiration, reflecting the combined
effects of cross-molecule substitution, new patients gaining access to lower-cost medicines, and
non-adherent patients resuming drug treatment 34 . This post-LOE increase in molecule utilization
runs counter to prior understanding of this market, and also likely reflects the increased aggressive
implementation of formulary design policies by PBMs. An implication is that not only do patent
protected brands need to worry about their own patents expiring, but their brand's revenues can also
be adversely affected if a competitive brand faces initial generic entry, for the newly competitive
generic molecule can steal market share. More generally, these post-LOE utilization increases
are creating novel complexities in defining drug markets for antitrust and other litigation-related
damage assessments.
Our data also document that the probability of a brand's patent being challenged by a potential
generic entrant is now very high, continuing an increasingly combative litigation trend reported
by the Federal Trade Commission [2011], which often results in the Paragraph IV first-filer being awarded 180-days of exclusivity. With their expected revenues being threatened, brands have
responded by launching their own authorized generic (AG), thereby creating a 180-day triopoly
calm before the patent cliff storm, with competition among the brand, its authorized generic and
the successful Paragraph IV challenger. In spite of only modest average molecule price reductions
during this 180-day exclusivity period (much less than after it expires), the substitution of prescription quantities away from the brand is already very large. Since through its combined sales of the
brand and its AG the brand franchise can moderate the revenue loss from LOE, the financial lure
of generics being awarded 180-day exclusivity is decreased. However, the evidence we find, as
has the Federal Trade Commission [2013, chapter 7], suggests that the existence of an AG during
the 180-day exclusivity period does not dampen the extent of generic entry post-exclusivity: on
the 181st day following initial LOE, the number of generics competing with the brand and with
each other tends consistently to be large, in our sample between seven and 14. Whether Pfizer's
relatively successful defense of Lipitor during the 180-day exclusivity period through the use of
coupons, rebates, other discounts and aggressive direct-to-consumer advertising is an historical
33
Drug Channels [2011, 2012].
1f one thinks of time before LOE with only the brand on the market as a monopolist implementing a uniform
price policy, and the post-LOE period as one in which there is price discrimination between brand-loyal and more
price-sensitive consumers, the finding here that post-LOE utilization is greater than pre-LOE suggests that a necessary
condition for economic welfare to have increased post-LOE is satisfied; see Varian [1989, pp. 617-623 and 632-636]
for a discussion of necessary conditions for price discrimination to increase economic welfare relative to a uniform
pricing policy.
34
110
quirk or instead is a harbinger of future more aggressive attempts by brands to protect brand revenues as patents expire, remains to be seen.
A novel set of findings we have reported here involves identifying separate retail prices by payer
type - cash, Medicare Part D, Medicaid and other commercial TPP. Our results suggest that declines in retail prices reflect the incentives to each payer to pursue policies that result in price
declines. In particular, the slower decline in Medicaid prices following LOE may be more rational
than previously thought. Three caveats are worth noting, however. First, the prices we calculate
represent the average total revenue received by a retail pharmacy for a dispensed prescription,
converted to price per day of therapy. This average revenue is the sum of any reimbursement the
pharmacy receives from a private or public payer for a dispensed prescription, plus any co-payment
or coinsurance amount contributed directly by the patient at the point of sale. This average revenue,
viewed from the perspective of the retail pharmacy, differs from and is likely considerably larger
than the average acquisition cost of the drug the retail pharmacy pays wholesalers or manufacturers35 . It also differs from the average revenue net of rebates and other discounts that is received
by the manufacturer selling to wholesalers or providers, and the average amount per prescription
contracted among payers, PBMs and manufacturers - in both these latter cases, rebates from the
manufacturer to payers and PBMs make it likely that these other prices are lower than the average
total revenue received by the retail pharmacy. Second, because the retail pharmacy average revenue price is not directly affected by rebates from manufacturers to public and private payers, the
relative brand-generic prices that payers provide retail pharmacies might differ considerably once
rebates from manufacturers to payers are taken into account. In particular, as noted earlier, provisions of the Affordable Care Act of 2010 mandate that Medicaid receive a rebate of 13% off the
average manufacturers' price (AMP) for a generic, and on average approximately 30+% discount
off AMP for a branded patent-protected drug. Third, our sample is very small - six molecules each
for 48 months, and thus these small sample findings cannot at this point be generalized to the entire
U.S. retail pharmaceutical market (a similar analysis of a much larger sample is on our research
agenda).
With these caveats regarding rebates and small sample size in mind, we find that in general the
price levels paid retail pharmacies per day of therapy are highest for cash and Medicaid payers,
and are lowest for other TPPs, with the cash price premium over TPP prices being on average just
under 20%. In terms of growth rates, however, cash prices of patent protected molecules grow at
virtually the same rate as overall market prices (the annualized difference being +0.06%), during
180-day exclusivity the cash prices fall slightly less rapidly than overall market prices, and post
any exclusivity cash prices fall about 1% more rapidly per month than do overall market prices.
35
For details, see Centers for Medicare & Medicaid Services [2012]; also see Drug Channels [2011, 2012].
III
Averaged over all six molecules and all 48 months, cash prices fall slightly more rapidly (about
2.4% annually) than do overall market prices. An implication of this is that if for administrative and
logistical reasons statistical agencies such as the U.S. Bureau of Labor Statistics find themselves
disproportionately reliant on cash transaction price quotes, the potential consequences for price
mis-measurement are likely to be relatively minor 36 . Establishing this last conclusion will be the
focus of our immediate future research program.
References
[1] Aitken, M.L., & Berndt, E.R. (2011), "Medicare Part D at Age Five: What Has Happened
to Seniors' PrescriptionDrug Prices?", Report by the IMS Institutefor HealthcareBioinformatics, July.
[2] Aitken, M.L., Berndt, E.R., & Cutler, D.M. (2008), "Prescription Drug Spending Trends
in the United States: Looking Beyond The Turning Point", Health Affairs Web Exclusive
28(1):W151-60, 2009. Published online 16 December 2008, 10.1377/hlthaff.28. 1.w151.
[3] Alpert, A., Duggan, M., & Hellerstein, J.K. (2013), "Perverse Reverse Price Competition:
Average Wholesale Prices and Medicaid Pharmaceutical Spending", August. Available from
lead author at aealpert@uci.edu. Cambridge, MA: National Bureau of Economic Research,
Working Paper No. 19367, August 2013. Forthcoming, Journalof Public Economics.
[4] Appelt, S. (2013), "Authorized Generic Entry Prior to Patent Expiry: Reassessing Incentives
for Independent Generic Entry", Working Paper, University of Munich, May. Available from
the author at silvia.appelt@lrz.uni-muenchen.de.
[5] Avalere
Health
LLC
(2010),
"State
Policies
Regarding
Generic
Substitution,
2010".
Powerpoint
presentation.
Available
online
at
www.avalerehealth.net/research/docs/GenericSubstitution.pdf.
[6] Berndt, E.R., & Aitken, M.L. (2011), "Brand Loyalty, Generic Entry and Price Competition in Pharmaceuticals in the Quarter Century After the 1984 Waxman-Hatch Legislation",
InternationalJournal of the Economics of Business, 18(2):177-201, July 2011.
[7] Berndt, E.R., Cockburn, I.M., & Griliches, Z. (1996), "Pharmaceutical Innovation and Market Dynamics: Tracking Effects on Price Indexes for Antidepressant Drugs", Brookings Papers on Economic Activity: Microeconomics, 1996(2):1133-88.
36
Note that while a 2.4% annual mis-measurement may seem substantial, the expenditure weight of all generic
drugs experiencing initial LOE in any given year is typically less than five percent.
112
[8] Berndt, E.R., Cutler, D.M., Frank, R.G., Griliches, Z, Newhouse, J.P., & Triplett, J.E. (2000),
"Medical Care Prices and Output", ch. 3 in Anthony Culyer and Joseph P. Newhouse, eds.,
Handbook of Health Economics, Amsterdam and New York: Elsevier Sciences B.V., Vol.
1A:117-80.
[9] Berndt, E. R., Mortimer, R, Bhattacharjya, A.M, Parece, A., & Tuttle, E. (2007), "Authorized
Generic Drugs, Price Competition, and Consumers' Welfare", Health Affairs, 26(3):790-799.
[10] Berndt, E. R., & Newhouse, J.P. (2012), "Pricing and Reimbursement in US Pharmaceutical
Markets", ch. 8 in Patricia M. Danzon and Sean Nicholson, eds., The Oxford Handbook of the
Economics of the BiopharmaceuticalIndustry, New York: Oxford University Press, 201-265.
[11] Branstetter, L.G., Chatterjee, C., & Higgins, M.J. (2011), "Regulation and Welfare: Evidence
from Paragraph IV Generic Entry in the Pharmaceutical Industry", Cambridge, MA: National
Bureau of Economic Research, Working Paper 17188, June.
[12] Brill, A. (2010), "Overspending on Multi-Source Drugs in Medicaid", Washington DC:
American Enterprise Institute, AEI Health Policy Studies Working Paper @2010-01, July
21. Available online at http://www.aei.org/paper/100127.
[13] Caves, R. E., Whinston, M.D., & Hurwitz, M.A. (1991), "Patent Expiration, Entry, and Competition in the U.S. Pharmaceutical Industry", Brookings Papers on Economic Activity: Microeconomics, 1991, 1-66.
[14] Centers for Medicare & Medicaid Services (2012). "Survey of Retail Prices",
May 31. Available online at http://medicaid.gov/Medicaid-CHIP-Program-Information/ByTopics/Benefits/Prescription-Drugs/Survey-of-Retail-Prices.html.
[15] Cook, A. (1998), How Increased Competitionfrom Generic Drugs Has Affected Prices and
Returns in the PharmaceuticalIndustry, Washington DC: The Congress of the United States,
Congressional Budget Office. Available online at http://www.cbo.gov/.
[16] Drug Channels (2012), "Pfizer's Lipitor Strategy and the 2012 Generic Monster", March
15. Available online at http://www.drugchannels.net/201 2/03/pfizers-lipitor-strategy-and2012.html.
[17] Drug Channels (2011), "Ranbaxy Makes Three: The Battle for Generic Lipitor Profits", December 1. Available online at http://www.drugchannels.net/2011//I 2/ranbaxy-makes-threebattle-for-generic.html.
113
[18] Drug Facts and Comparisons 2011 Edition (2011), St. Louis, MO: Wolters Kluwer Health,
Inc.
[19] Ellison, G., & Ellison, S.F. (2011), "Strategic Entry Deterrence and the Behavior of Pharmaceutical Incumbents Prior to Patent Expiration", American Economic Journal: Microeconomics, 3(1):1-36.
[20] Ellison, S.F., Cockburn, I.M., Griliches, Z., & Hausman, J.A. (1997), "Characteristics of Demand for Pharmaceutical Products: An Examination of Four Cephalosporins", RAND Journal
of Economics, 28(3):426-46.
[21] Federal Trade Commission (2011), Authorized Generic Drugs: Short-Term Effects and LongTerm Impact,August. Available online at http://www.ftc.gov/opa/2011/08/genericdrugs.shtm.
[22] FiercePharma (2012), "Top 15 drug patent losses for 2013", November 1. Available online at
http://www.fiercepharma.com/special-reports/top- 1 5-patent-expirations-2013.
[23] Frank, R.G., & Salkever, D.S. (1997), "Generic Entry and the Pricing of Pharmaceuticals",
Journalof Economics and ManagementStrategy, 6(1):75-90.
[24] Frank, R.G., & Salkever, D.S. (1992), "Pricing, Patent Loss and the Market for Pharmaceuticals", Southern Economic Journal,59:165-79.
[25] Generic Pharmaceutical Association (2013), "Generic Industry by the Numbers", Generic
PharmaceuticalAssociation 2012 Annual Report. Available online at www.gphaonline.org.
[26] Generic Pharmaceutical Association (2012), Generic Drug Savings in the U.S.: Savings $1 Trillion Over 10 Years, Fourth Annual Edition, August 2. Available online at
www.gphaonline.org.
[27] Grabowski, H. G., Kyle, M.K, Mortimer, R., Long, G., & Kirson, N. (2011), "Evolving Brand-Name And Generic Drug Competition May Warrant A Revision Of The HatchWaxman Act", Health Affairs, 30(11):2157-66, November.
[28] Grabowski, H. G., & Kyle, M.K. (2007), "Generic Competition and Market Exclusivity Periods in Pharmaceuticals", ManagerialDecision and Economics, 28:491-502.
[29] Grabowski, H. G., Long, G., & Mortimer, R. (2013), "Biosimilars", unpublished manuscript,
Duke University, August. Forthcoming as ch. 12.8 in Patricia Danzon, ed., Encyclopedia of
Health Economics, New York: Elsevier.
114
[30] Grabowski, H. G., Long, G., & Mortimer, R. (2011), "Implementation of the Biosimilar
Pathway: Economic and Policy Issues", Seton Hall Law Review, 41(2):511-557
[31] Grabowski, H. G., & Vernon, J.M. (1996), "Longer Patents for Increased Generic Competition in the US: The Waxman-Hatch Act After One Decade", PharmacoEconomics, 10
Suppl2:110-123.
[32] Grabowski, H. G., & Vernon, J.M., "Brand Loyalty, Entry, and Price Competition in Pharmaceuticals After the 1984 Drug Act", Journalof Law and Economics, 35:331-350.
[33] Griliches, Z., & Cockburn, I.M. (1994), "Generics and New Goods in Pharmaceutical Price
Indexes", American Economic Review, 84(5):1213-32.
[34] Hemphill, S.C., & Sampat, B.N. (2012), "Evergreening, Patent Challenges, and Effective
Market Life in Pharmaceuticals", Journal of Health Economics, 31(2):327-39, March.
[35] Hurwitz, M.A., & Caves, R.E. (1988), "Persuasion or Information? Promotion and the Shares
of Brand Name and Generic Pharmaceuticals", Journal of Law and Economics, 31:299-320,
October.
[36] Huskamp, H. A., Donohue, J.M., Koss, C., Berndt, E.R., & Frank, R.G. (2008), "Generic
Entry, Reformulations, and Promotion of SSRIs", PharmacoEconomics,26(7):603-16.
[37] IMS Institute for Healthcare Informatics (2013), Declining Medicine Use and Costs: For
Better or Worse?, May. Available online at www.theimsinstitute.org.
[38] Kelton, C.M. L., Chang, L.V., & Kreling, D.H. (2013), "State Medicaid Programs Missed
$220 Million in Uncaptured Savings As Generic Fluoxetine Came to Market, 2001-05",
Health Affairs, 32(7):1204-11, July.
[39] Medicaid Drug Rebate Program (2013), last updated August 15, 2013. Available online at
http://www.medicaid.gov.
[40] Olson, L.M., & Wendling, B.W. (2013), "The Effect of Generic Drug Competition on Generic
Drug Prices During the Hatch-Waxman 180-Day Exclusivity Period", Washington DC: Federal Trade Commission, Bureau of Economics, Working Paper No. 317, April. Available
online at bwendling@ftc.gov.
[41] Panattoni, L. E. (2011), "The Effect of Paragraph IV Decisions and Generic Entry Before
Patent Expiration on Brand Pharmaceutical Firms", Journal of Health Economics, 30(1):12645.
115
[42] Regan, T. L. (2008), "Generic Entry, Price Competition, and Market Segmentation in the
Prescription Drug Market", InternationalJournalof Industrial Organization,26(4):930-48.
[43] Reiffen, D. E., & Ward, M.E. (2007), "'Branded Generics' as a Strategy to Limit Cannibalization of Pharmaceutical Markets", Managerialand Decision Economics, 28:251.267.
[44] Reiffen, D. E., & Ward, M.E. (2005), "Generic Drug Industry Dynamics", Review of Economics and Statistics, 87(1):37-49.
[45] Saha, A., Grabowski, H.G., Birnbaum, H.M.,Greenberg, P.E., & Bizan, 0. (2006), "Generic
Competition in the U.S. Pharmaceutical Industry", InternationalJournalof the Economics of
Business, 13(1):15-38, February.
[46] Scott Morton, F.M. (2000), "Barriers to Entry, Brand Advertising and Generic Entry in the
U.S. Pharmaceutical Industry", InternationalJournal of Industrial Organization, 18:1085104.
[47] Scott Morton, F.M. (1999), "Entry Decisions in the Generic Pharmaceutical Industry", The
RAND Journalof Economics, 30:421-440.
[48] Smith, A. (2007), "Federal judge whacks generic Plavix", CNN Money.com, June 10, 2007.
Available online at http://money.cnn.com/2007/06/19/news/companies/plavix/index.htm.
[49] Somers, J. (2010), Effects of Using Generic Drugs on Medicare's PrescriptionDrug Spending, Washington DC: The Congress of the United States, Congressional Budget Office,
September. Available online at http://www.cbo.gov/.
[50] Takeda (2011), 2011-2012 PrescriptionDrug Benefit Cost and Plan Design Report, Plano,
TX: Pharmacy Benefit Management Institute. Available online at http://www.pbmi.com.
[51] U.S. Department of Labor, Bureau of Labor Statistics (2011), "The Pharmaceutical Industry:
An Overview of CPI, PPI, and IPP Methodology", Office of Prices & Living Conditions,
October. Available online at http://www.bls.gov/ppi/pharmpricescomparison.pdf.
[52] Varian, H. R. (1989), "Price Discrimination", ch. 10 in Richard L. Schmalensee and Robert D.
Willig, eds., Handbook of IndustrialOrganization,Amsterdam: Elsevier Science Publishers,
Volume 1, pp. 597-654.
[53] Wang, X. (2012), "Understanding Current Trends and Outcomes in Generic Drug Patent
Litigation: An Empirical Investigation", unpublished Masters' honors thesis, Stanford University: Public Policy Program, May.
116
[54] Wiggins, S.N., & Maness, R. (2004), "Price Competition in Pharmaceuticals: The Case of
Anti-Infectives", Economic Inquiry, 42(2):247-63.
[55] Wosinska, M., & Huckman, R.S. (2004), "Generic Dispensing and Substitution in Mail and
Retail Pharmacies", Health Affairs - Web Exclusive, W4-409 to W4-416, posted 28 July
2004. Available online at www.healthaffairs.org. An abstract of the article was published in
the hardcopy edition of Health Affairs, 23(5):284.
Tables and Figures
Table 3.1: Characteristics of Six Initial Generic Launches, June 2009-May 2013
(IN CHRONOLOGICAL ORDER - ANDA ENTRY DATE, LEFT TO RIGHT)
Trade Name
Generic
Cozaar
losartan
Name
NDA
Approval
4/14/1995
Ambien CR
Augmentin XR
amoxicillin/clavu- zolpidem
lanate potassium tartrate
4/2/2005
9/25/2002
Lexapro
Lipitor
atorvastatin escitalopram
oxalate
8/14/02
11/17/1996
Plavix
clopidogrel
11/17/97
NME
NF
NF
NME
NF
NME
Standard
Standard
Standard
Priority
Standard
Priority
AUIRA Antihypertensive
Aminopenicillin
antibiotic
Insomnia
Antihyperlipidemic
SSRI antidepressant
NDA Sponsor
Merck
Dr. Reddys Labs
Pfizer
Forest
Mean TRX
Before LOE
Paragraph IV
Challenger
Initial ANDA
Entry Date
Authorized
634,846
2,281
Sanofi
Aventis
US
371,580
Antiplatelet
aggregation
inhibitor
Sanofi
Aventis US
2,854,162
1,485,203
1,730,464
Teva
Sandoz
4/6/2010
NME or New
Formulation
Review
Status
Therapeutic
Class
Ranbaxy
Ivax
4/21/2010
Actavis
Elizabeth
10/13/2010
11/30/2011
3/14/2012
Apotex
(forfeited)
5/17/2012
Sandoz
None
Winthrop
Watson
Mylan
None
3
2
3
3
3
12
14
2
4
6
14
13
14
2
5
7
14
13
Generic
Max. No.
Mfrs. during
1-6 Months
No. Mfrs. At
7 Months
No. Mfrs. At
12 Months
Notes: NF is new formulation; AlIRA is Angiotensin 11 Receptor Antagonists; SSRI is selective serotonin
reuptake inhibitor; NDA Sponsor is patent holder at time of patent expiration; Mean TRX Before LOE is
the mean monthly number of prescriptions dispensed in the three full months preceding loss of
exclusivity.
117
Table 3.2: Months to Generic Penetration Rate Thresholds, June 2009 - May 2013
(IN CHRONOLOGICAL ORDER - ANDA ENTRY DATE, TOP TO BOTTOM)
Trade Name
Generic Name
Buyer
Type
Cozaar
losartan
All
TPP
Medicare
Cash
Medicaid
Threshold
All Buyers
60%
90%
1
4
1
2
1
3
1
9
8
21
Threshold
Under 65
60%
90%
1
5
1
2
1
2
1
7
8
17
Threshold
Over 65
60%
90%
1
3
1
3
1
3
1
9
8
24
Augmentin XR
All
2
4
2
4
2
4
amoxicillin/clavulanate potassium
TPP
Medicare
Cash
Medicaid
All
TPP
Medicare
2
2
2
5
3
3
3
3
3
3
9
9
13
10
20
2
2
2
5
3
3
3
3
3
9
9
12
10
20
4
3
5
4
17
14
23
22
2
2
4
4
3
3
3
3
16
1
1
2
1
Ambien CR
zolpidem tartrate
Lipitor
atorvastatin
Lexapro
escitalopram
oxalate
Plavix
clopidogref
Cash
Medicaid
10
22
-
3
10
All
TPP
Medicare
Cash
1
1
2
1
9
9
8
10
1
1
2
1
31
9
9
8
10
Medicaid
8
11
8
12
8
All
TPP
Medicare
Cash
Medicaid
All
TPP
Medicare
Cash
Medicaid
<1
<1
<1
7
1
6
4
4
8
11
2
<1
<1
<1
<1
7
1
6
4
3
8
11
2
<1
<1
<1
1
7
1
1
1
I
1
1
1
1
1
1
2
3
4
1
1
1
2
2
4
1
1
1
2
3
3
<1
118
27
8
8
8
11
11
5
4
5
9
13
2
Table 3.3: Mean Quantities Post-LOE Six to Nine Months, and Post-LOE Ten to Twelve Months
Relative to Pre-LOE Quantities, June 2009 - May 2013
(IN CHRONOLOGICAL ORDER - ANDA ENTRY DATE, TOP TO BOTTOM)
Trade Name
Generic Name
Cozaar
losartan
Buyer
Type
All Buyers
POST/PRE POST/PRE
6-9
10-12
All
160
1.86
TPP
1.61
1.84
Medicare
1.30
1.68
Cash
1.40
1.59
Under 65
POST/PRE POST/PRE
Over 65
POST/PRE POST/PRE
6-9
10-12
6-9
10-12
1.67
1.72
1.78
1.57
1.54
1.43
1.66
1.23
1.81
1.58
1.03
1.09
4.59
4.85
5.03
1.78
5.17
5.25
5.99
2.06
Medicaid
1.11
1.20
1.13
Augmentin XR
All
amoxicillin/clavu- TPP
lanate potassium Medicare
Cash
4.52
5.76
4.83
2.82
4.68
5.93
5.46
2.94
4.50
5.87
4.46
3.09
1.89
1.97
1.91
1.80
1.22
4.55
5.95
4.53
3.09
2.04
1.37
Medicaid
0.64
0.61
0.64
0.62
0.43
0.41
Ambien CR
All
zolpidem tartrate TPP
Medicare
Cash
0.96
1.00
0.77
0.86
0.96
1.00
0.78
0.84
0.97
1.00
0.85
0.80
0.97
1.00
0.87
0.80
0.92
1.00
0.71
1.12
0.89
0.97
0.70
1.04
Medicaid
0.80
0.73
0.81
0.74
0.62
0.52
All
TPP
Medicare
Cash
Medicaid
All
TPP
Medicare
Cash
All
TPP
Medicare
Cash
1.24
1.24
1.29
1.40
0.63
1.09
1.13
1.07
0.92
0.83
1.00
1.00
1.00
1.60
1.31
1.29
1.36
1.89
0.75
1.14
1.18
1.10
1.12
0.81
0.98
0.95
10
1.45
1.25
1.29
1.23
1.34
0.63
1.11
1.15
1.07
0.93
0.83
1.05
1.05
1.08
1.62
135
1.37
1.38
1.88
0.75
1.15
1.20
1.02
1.12
0.81
0.99
1.00
0.97
1.56
1.22
1.11
1.31
1.49
0.63
1.01
0.94
1.06
0.90
0.70
0.98
0.93
0.99
1.56
1.24
1.10
1.35
1.90
0.73
1.08
0.97
1.15
1.16
0.72
0.97
0.88
1.00
1.35
Medicaid
0.86
0.80
0.85
0.78
0.96
0.92
Lipitor
atorvastatin
Lexapro
escitalopram
oxalate
Medicaid
Plavix
clopidogrel
119
Table 3.4: Prices and Prescription Shares During the 180-Day Exclusivity Period for Branded,
suup
Generic and Authorized Generic Molec
Variable/Molecule
Cozaar
Ambien CR
Lipitor
Lexapro
Mean (s.d.) Branded
2.55 (0.06)
6.27 (0.04)
5.13 (0.12)
4.49 (0.20)
2.19 (0.07)
5.15 (0.35)
3.82 (0.44)
3.06 (0.14)
1.86(0.10)
4.46 (0.07)
4.46 (0.20)
3.99 (0.10)
0.16 (0.14)
0.44 (0.36)
0.40 (0.27)
0.15 (0.07)
0.50(0.09)
0.32 (0.23)
0.36 (0.18)
0.48 (0.14)
0.34 (0.05)
0.24 (0.16)
0.24 (0.12)
0.37 (0.07)
Price
Mean (s.d.) Generic
Price
Mean (s.d.) Authorized
Generic Price
Mean (s.d.) Branded
Share
Mean (s.d.).Generic
Share
Mean (s.d.) Authorized
Generic Share
120
Table 3.5: Cash Versus Full Sample Mean Price Levels and Monthly Growth Rates
(STANDARD DEVIATIONS IN PARENTHESES)
Trade Name
Generic Name
Cozaar
losartan
Measure
Mean Cash/Full
Log[P(t)/P(t-1)] Full
Augmentin XR
amoxicillin/clavu
-lanate
potassium
Arnbien CR
zolpidem
tartrate
Lipitor
atorvastatin
Lexapro
escitalopram
oxalate
Plavix
clopidogrel
Log[P(t)/P(t-1)]
Cash
Mean Cash/Full
LogfP(t)/P(t-1)]
Full
Log[P(t)/P(t1)]
Cash
Mean Cash/Full
Log[P(t)/P(t-1)
Full
Log[P(t)/P(t-1)]
Cash
Mean Cash/Full
Log[P(t)/P(t-1)j
Full
Log[P(t)/P(t1)
Cash
Mean Cash/Full
Log[P(t)/P(t-1))
Full
Log[P(t)/P(t-1)]
Cash
Mean Cash/Full
Log[P(t)/P(t-1)]
Full
Log [P(t)/P(t-1)
Cash
Full 48
Months
1.19
(0.107)
-0.017
(0.046)
-0.019
(0.048)
1.24
(0.027)
-0.006
(0.019)
-0.005
(0.029)
1.21
(0.041)
-0.005
(0.022)
-0.003
(0.020)
1.11
(0.157)
-0.013
(0.061)
-0.024
(0.092)
1.16
(0.039)
-0.005
(0.044)
-0.007
(0.036)
1.11
(0.053)
-0.012
(0.079)
-0.012
(0.063)
121
Pre LOE
1.17
(0.014)
0.007
(0.013)
0.010
(0.011)
1.23
(0.015)
-0.002
(0.007)
-0.002
(0.019)
1i16
(0.013)
0.011
(0.021)
0.012
(0.018)
1.15
(0.009)
0.009
(0.021)
0.009
(0.024)
1.16
(0.007)
0.010
(0.020)
0.010
(0.018)
1.09
(0.013)
0.009
(0.019)
0.008
(0.017)
During
180 Day
1.20
(0.005)
-0.031
(0.039)
-0.032
(0.043)
1.25
(0.035)
-0.035
(0.039)
-0.031
(0.056)
1.18
(0.021)
-0.033
(0.035)
-0.026
(0.031)
1.14
(0.011)
-0.023
(0.041)
-0.019
(0.032)
1.13
(0.012)
-0.025
(0.034)
-0.026
(0.032)
Post 180
Day
120
(0.133)
-0.011
(0.026)
-0t018
(0.047)
1.23
(0.034)
-0.003
(0.010)
-0.0002
(0.026)
1.24
(0.022)
-0.008
(0.010)
-0.008
(0.010)
1.02
(0.303)
-0.039
(0.064)
-0.093
(0.156)
1.21
(0.078)
-0.016
(0.031)
-0.032
(0.028)
1.17
(0.078)
-0.060
(0.140)
-0.056
(0.010)
E
0.
oE~
e
-
CO
200m7
2009m7
Medicaid
Medicare
Cash
2012m7
Third Party
2011m7
monthlydate
2010m7
Third Party
CashMdiad
-----......-
2011m7
monthlydate
.
. .......
Medicare
2012m7
I"
I'
Lipitor Unit Prices By Payer Type
2010m7
Cozaar Unit Prices By Payer Type
201m7
2()13m7
6
e
aL
CN-
0 E0L
2009m7
2009m7
1
Third Party
Cash
2011m7
monthlydate
Medicare
Medicaid
2012m7
2010m7
Cash
Third Party
Medicaid
2012m7
......... Medicare
2011m7
monthlydate
Lexapro Unit Prices By Payer Type
2010m7
Ambien CR Unit Prices By Payer Type
Figure 3.1: Price Per Day of Therapy By Payer Type
2013m7
2( 1m7
k
2009m7
2009m7
N* -
W
e
*L
ei
N
a
Cash
Third Party
.
.
2011m7
monthly date
Medicare
Medicaid
201m7
2010m7
Third Party
Cash
Medicaid
2012m7
..-.----.--..-------- Medicare
2011m7
monthlydate
Plavix Unit Prices By Payer Type
i
2010m7
Augmentin XR Unit Prices By Payer Type
--
r-
2013m7
2013m7
Appendix: Brief Description of the Six Molecules in the Data
Sample
Cozaar (losartan):
Two weeks after Cozaar, an antihypertensive, was approved by the FDA, Merck obtained FDA
approval for Hyzaar, a combination product that contained Cozaar (losartan potassium) as one
component and a very old off-patent beta blocker called hydrochlorothiazide as the other component. Teva successfully challenged a Merck patent underyling both Cozaar and Hyzaar, and was
granted 180-day exclusivity effective April 5, 2010 and launched its product one day later. Merck
contracted with Sandoz to launch an authorized generic version of Cozaar coinciding with the
launch of Teva's generic Cozaar. On October 6, 2010 when 180-day exclusivity expired, massive
generic entry occurred at all three Cozaar dosages, with 12 new ANDAs entering.
Augmentin XR (amoxicillin/clavulanate potassium):
Although GlaxoSmithKline (GSK) was the original NDA holder for Augmentin XR, a combination antibiotic product with amoxicillin - an old penicillin) and clavulate potassium as components, at the time of patent expiration in 2010, the NDA holder was Dr. Reddys Labs, Inc., a
manufacturer best known as a generic manufacturer. According to Drugs@FDA, Sandoz is the
only FDA-recognized ANDA entrant. Currently there are only two manufacturers (Dr. Reddy's
Labs, Inc. and Sandoz).This may be explained by the relatively small market size (2,281 mean
monthly prescriptions) and the complexity of manufacturing processes of the extended release version. Notably, the recommended use of this antibiotic is a twice-daily dosing for 7-10 days 37 ,
implying that the recommended number of extended units in a typical single episode prescription
(14-20) is smaller than for 30 or 90-day prescriptions of once daily maintenance medications (in
effect an even smaller market than is implied by the prescription count).
Ambien CR (zolpidem tartrate):
Ambien CR is an extended release version of Ambien immediate release (Ambien CR has a pharmacokinetic activity that provides an immediate dose release to facilitate getting to sleep, but then
provides a sustained release that facilitates a longer duration of sleeping). Sanofi Aventis US
obtained FDA approval for Ambien CR about 18 months before the Ambien immediate release
patent expiry. Actavis was granted 180-day exclusivity for the 6.25 mg formulation on October
13, 2010, while Anchen was awarded exclusivity for the 12.5 mg formulation on December 3,
37
Drug Facts and Comparisons [2011], p. 2088.
123
2010. Winthrop, a subsidiary of Sanofi Aventis US, launched an authorized generic for each of
the dosage forms within days of each exclusivity taking effect. Sublingual (under the tongue) and
oral spray formulations at varying dosages are also available having different brand names Edluar,
Intermezzo and Zolpimist, but they are not rated as therapeutic equivalents to Ambien CR by the
FDA.
Lipitor (atorvastatin):
Perhaps the most highly publicized patent expiry in the last decade was that for Pfizer's Lipitor
(atorvastatin), a drug controlling cholesterol lipids. Patents on all formulations were successfully
challenged by Ranbaxy, who was awarded 180-day exclusivity on its ANDA on November 30,
2011 that expired May 28, 2012. However, Pfizer contracted with Watson (later Actavis) 38 to market an authorized generic version of all dosages, and also initiated an aggressive coupon program
that reduced considerably the customer co-payment typically required for branded drugs on the
second or higher tier of a formulary 39 . During the exclusivity period, therefore, there were three
competitors (Pfizer's brand, its authorized generic through Watson, and Ranbaxy) and on the day
exclusivity expired, three additional ANDA holders entered at all four dosages (Apotex, Mylan
and Sandoz). A notable feature of the Lipitor-atorvastatin patent expiry is that even though it had
the largest pre-LOE market size, more than a year after the May 2012 unfettered ANDA entry,
there are only five ANDA holders competing at each dosage strength; as seen in Table 3.1, for
other molecules the number of ANDA holder entrants is much larger, at 12-15 entrants. The small
number of competitors in this market is curious - it may reflect the fact that the small number
of approved ANDA holders are each known to have substantial manufacturing capacity, or that
Pfizer's highly publicized aggressive protection of its brand reduced the perceived payoff to entry
by generic manufacturers 40 .
Lexapro (escitalopram oxalate):
Lexapro (escitalopram oxalate) is an antidepressant marketed in the US by Forest Labs; its NDA
was approved as a new formulation (it is an isomer of the earlier Forest antidepressant drug Celexa
- citalopram - both of which were licensed into Forest after being on the European market for a
number of years). Ivax was awarded 180-day exclusivity effective March 14, 2012 for all three
38
Pfizer entered into an authorized generic arrangement with a generic challenger (Arrow) on a separate patent
dispute, and Arrow was subsequently acquired by Watson, which merged with Actavis.
39
The coupon program was only available to cash or 3rd party patients in states that did not preclude use of coupons.
Though highly visible and heavily promoted in mass media, the number of patients that participated in the program
appears to have been quite small, likely well below 10% of those eligible.
40
0n the market's perception of the reputation of Pfizer and its generic subsidiary, Greenstone, see Federal Trade
Commission [2011], p. 82.
124
tablet versions. On the day after exclusivity expiration - September 11, 2012 - a total of nine
additional generic entrants came to market, each offering tablets at all three dosages, and the next
day two additional entrants were launched, for a total of 12 generic manufacturers of escitalopram
immediately following the exclusivity period.
Plavix (clopidogrel):
The launch of Plavix (clopidogrel) represented the culmination of contentious legal skirmishes
among Sanofi, Bristol Myers Squibb (the US marketer of Plavix), Apotex and various states' Attorneys General. Initial proposed settlements between BMS and Apotex allowing Apotex to have
several months of exclusivity were rejected by the Attorneys General, and eventually Apotex acceded to forfeiting any rights to exclusivity. The Plavix patent finally expired on May 17, 2012,
on which date with unfettered generic entry seven ANDAs launched at 75 mg and four at 300 mg;
as of June 10, 2013, there appear to be 13 ANDA entrants at the 75 mg formulation, and six at the
300 mg dosage.
125