Reduced-Form Behavioral Public Finance Josh Schwartzstein Based on joint work with Sendhil Mullainathan, Kate Baicker, and Bill Congdon 1 / 37 Background Growing recognition that behavioral tendancies matter for thinking about a variety of policy problems I commodity taxation – e.g., sin taxes (e.g., Gruber and Koszegi 2001, O’Donoghue and Rabin 2006, Chetty et al. 2009) I savings – e.g., new levers like defaults available to policy-maker (e.g., Madrian and Shea 2001, Carroll et al. 2009, Beshears et al. 2008) I I I health insurance design – no longer clear incentives/insurance trade-off (e.g., Baicker, Mullainathan, and Schwartzstein 2013) unemployment insurance (e.g., DellaVigna and Paserman 2005, Spinnewijn 2014) energy efficiency – may want to subsidize energy-efficient behaviors by more than suggested by externalities alone (e.g., Allcott, Mullainathan and Taubinsky 2011) 2 / 37 Background To make progress, the literature has largely proceeded by drawing out the implications of particular biases in these settings I I commodity taxation – present-bias (e.g., Gruber and Koszegi 2001, O’Donoghue and Rabin 2006), inattention (e.g., Chetty et al. 2009, Goldin 2014) savings – present-bias (e.g., Carroll et al. 2009), inattention, anchoring (Bernheim et al. 2011) I health insurance design – present-bias (e.g., Newhouse 2006), false beliefs (e.g., Pauly and Blavin 2008) I unemployment insurance – present-bias (e.g., DellaVigna and Paserman 2005), overoptimism (e.g., Spinnewijn 2014), reference-dependent preferences (e.g., DellaVigna et al. 2014 ) I energy efficiency – inattention (e.g., Allcott, Mullainathan and Taubinsky 2011; Allcott and Taubinsky 2014), present-bias (e.g., Heutel 2011) 3 / 37 Observation This literature can feel a bit scattered I Biases vary from paper to paper I It is difficult to easily assess which results extend and which depend on the specific psychologies considered Compare with traditional public finance: provides a much more integrated treatment for how different policy levers affect total welfare Goal: develop a reduced-form framework that highlights lessons that are robust across specific psychologies and contexts 4 / 37 Roadmap 1. Illustrate features of framework through a specific example on health insurance design I Based on Baicker, Mullainathan, and Schwartzstein (2014, first draft 2010) 2. Give a flavor for broader framework and implications I Based on Mullainathan, Schwartzstein and Congdon (2012) 5 / 37 Keep in Mind Some things to keep in mind I I The reduced-form approach is a complement, not a substitute for drawing out the implications of specific behavioral assumptions The focus will be on normative issues, but not so much on how to identify welfare I I I Others have focused on the identification question, e.g., Bernheim and Rangel (2009), Hojman and Green (2008), Rubinstein and Salant (2011) The focus will primarily be on analyzing the impact of standard price levers rather than nudges Reduced-form approaches have a long tradition in public finance I E.g., in the spirit of the “sufficient statistics literature” (e.g., Chetty 2009) 6 / 37 Health Insurance Problem of health insurance design: trade-off between financial protection and moral hazard Example: I Josh gets a headache of severity s I Treatment provides benefit b(s) and costs c I It is socially efficient for Josh to get treated only if headache is sufficiently severe: b(s) > c The problem of health insurance I I I Insurance means Josh only pays p < c Josh seeks treatment if b(s) > p 7 / 37 Standard Model Standard model: The concern is overutilization due to moral hazard 8 / 37 Example: Type-II Diabetes Diabetes: I Can have many serious complications (e.g., blindness, lost limbs, coma) I Broad consensus of how to treat I Adherence to treatment recommendations has been shown to reduce the probability of adverse events But many do not adhere I I Recent study: almost half of diabetic patients did not consistently fill prescriptions, doubling their risk of hospitalization (Sokol et al. 2005) 9 / 37 Examples of Underuse of High Value Care Table 1a: Examples of Underuse of High Value Care Estimates of return to care Usage rates of clinically relevant population Statins Reduce all cause mortality (Relative Risk .90), cardiovascular disease mortality (RR .8), fatal myocardial infarction (RR.82), non-fatal MI (RR.74), and strokes (RR .86) Adherence < 70% Beta-blockers Reduce mortality by 25% post heart attack Adherence < 70% Anti-asthmatics Reduced Hospital Admissions (OR .58). Improvement Adherence < 50% in airflow obstruction (OR .43) Anti-diabetics Decrease of cardiovascular mortality (OR .74); risk of hospitalization halved Adherence < 65% Immunosuppresants Reduction in the risk of organ rejection seven-fold Adherence < 66-75% Recommended Preventive Care Care of known efficacy including immunizations, disease management, follow-up care post surgery Pre-natal care Inadequate prenatal care increases infant mortality (RR 2.13) <55% care delivered (36% diabetics receiving semi-annual blood tests, 40% colorectal cancer 45.8 % received adequate or better care 10 / 37 What’s Going On? Mis-measurement? I Perhaps unobserved side effects drive underuse I Given clinical evidence, hard to argue for many of these examples Heterogeneity? I I Perhaps there is a lot of heterogeneity and people properly self sort Evidence tends to be inconsistent with this hypothesis I e.g., studies examining heterogeneous demand responses to co-pay changes by clinical status (Goldman et al. 2006) Dynamic moral hazard? I Perhaps people underuse preventive care because they are insured in the future I We focus on underuse where benefits seem to outweigh costs to the consumer ; i.e., uninsurable costs of non-adherence are large 11 / 37 What’s Going On? Present-bias? I In deciding whether to get recommended preventive care, a present-biased agent underweighs the delayed benefits (Newhouse 2006) I Should get treated if b(s) = v (s) − k(s) > p, v (s) represents delayed benefits and k(s) represents immediate costs I Gets treated if βv (s) − k(s) > p ⇐⇒ b(s) − (1 − β)v (s) > p β≤1 12 / 37 What’s Going On? Inattention or symptom salience? I An inattentive agent underweighs the benefits of filling an antidiabetic prescription since diabetes is “silent” much of the time (Rubin 2005, Osterberg and Blaschke 2005) I Should get treated if b(s) = b(v + n + o) > p where v represents salient symptoms, n opaque or non-painful symptoms, and o other symptoms I Instead gets treated if b(αv + βn + o) > p ⇐⇒ b(s) + [b(αv + βn + o) − b(v + n + o)] > p α ≥ 1, β ≤ 1 13 / 37 What’s Going On? False beliefs? I Once symptoms have abated, an agent with false beliefs may believe that continuing a TB drug regimin is no longer necessary (Pauly and Blavin 2008) I Should get treated if b(s) > p I Instead gets treated if b̂(s) > p ⇐⇒ b(s) + (b̂(s) − b(s)) > p 14 / 37 Resolution Observation: Agents with present-bias, inattention, and false beliefs all make choices according to whether b(s) + ε(s) > p In these cases, there is a clear wedge between “decision” and “hedonic” utility (Kahneman et al. 1997), given by ε(s) I Not true for all “behavioral factors”, e.g., anxiety (Koszegi 2003) Call misbehavior resulting from ε 6= 0 behavioral hazard I suboptimal choices resulting from mistakes or behavioral biases I in contrast to privately optimal but socially suboptimal choices resulting from misaligned incentives 15 / 37 Simple Insight Simple Insight: We can generally draw out the implications of “behavioral hazard” for welfare calculations and optimal co-pay formulas I We don’t need to separately consider present-bias, symptom salience, and false beliefs to see how BH changes central insights of the “standard” model I The feature of a behavioral bias that matters most is how it affects who is marginal with respect to the copay I But, as we saw above, we can re-write ε in terms of primitives of specific behavioral models 16 / 37 Further Details on the Basic Setup Setup: I Individual has wealth y I Insurance costs premium P I When healthy, the individual has utility U(y − P) I With probability q, she falls sick with varying degrees of severity s ∼ F (s) I Absent treatment, the sick individual receives utility U(y − P − s) Treatment costs society c and the patient co-pay p; its benefit b(s) depends on severity: b 0 (s) ∈ [0, 1] I I I To simplify certain statements, will assume in the talk that b(s) = s With treatment, the sick individual receives utility U(y − P − s + b(s) − p) 17 / 37 Choice to Receive Treatment Without behavioral hazard: Choose to get treated if b(s) > p With behavioral hazard: Choose to get treated if b(s) + ε(s) > p I Shouldn’t think of ε(s) as fixed, but as systematically varying across diseases and treatments I I I drugs treating chronic conditions: ε < 0 antibiotics for children’s ear infections or treatment of back pain: ε > 0 To simplify formulas today, ignore heterogeneity in behavioral hazard across people I I Does not matter for what I’ll discuss Paper considers more general case With BH misutilization is not solely a consequence of health insurance 18 / 37 Re-Thinking The Welfare Impact of Co-pay Changes What are some implications of BH for optimal co-pays? Planner’s problem: Demand Pr(b(s)+ε(s)>p) max W = E [U] subject to P = p z }| { M(p) ·(c − p) Raising the co-pay has two effects: 1. It reduces insurance value - doesn’t matter whether there is behavioral hazard 2. It affects utilization - interpretation of demand response depends on whether there is behavioral hazard 19 / 37 Welfare Effects of Changing the Co-pay: No BH Without BH: Differentiate W and normalize: Demand Response Sufficient Statistic E [U 0 (C )|m=1]−E [U 0 (C )] E [U 0 (C )] z }| { dW dW / = −M 0 (p)(c − p) − | dp dy | {z } Welfare Impact z }| { I · M(p) {z ·M(p) Insurance Value } FOC for optimal co-pay c − pS = pS I η |{z} Demand Elasticity I More elastic demand implies higher co-pay I Insurance optimally partial: p S ∈ (0, c) if η > 0, I > 0 20 / 37 Optimal Copay: Behavioral Hazard Taking BH into Account: Differentiate W and normalize: Demand Response Not Sufficient Statistic z }| { dW dW / ≈ −M 0 (p)(c − p + ε) −I · M(p) {z } | dp dy | {z } Sign Could Become (-) Welfare Impact FOC c − pB I ε ≈ − B η p pB I More elastic demand does not necessarily mean higher co-pay I E.g., the fact that elasticities for beta blockers and cold remedies are similar does not necessarily imply similar co-pays 21 / 37 Re-thinking Optimal Co-pays (Continued) I c − pB ε ≈ − B B η p p I Optimal co-pays create incentives for more efficient treatment decisions, not just insurance value: when I = 0, p B = c + ε(s(p B )) I I I optimal co-pays price “internalities” health insurance can actually lead to more efficient utilization It can be optimal to fully cover effective treatments, even if they are ineffective for some insurees: can optimally have p B ≤ 0 even when M 0 (0) < 0 22 / 37 Implication Modal evidence for moral hazard has been the demand response This can be very misleading I Being marginal does not necessarily imply indifference Implications can even be the wrong sign! More than an abstract concern 23 / 37 Example Example: Choudhry et al. (NEJM, 2011) I I Study impact of eliminating copays for recent heart attack victims Randomly assigned patients discharged after heart attacks to a control group with usual coverage I I copayments in $12 to $20 range Or a treatment group with no co-payments I For statins, beta blockers and ACE inhibitors 24 / 37 /,))*.'%"($G"* 3"2&.$E'0*$2#"("0."* !"#$%&&&* ;7,$)*.'%"($G"* A=5JbAT!7* !"4$P-)'.M"(7* !_\[]! !^\^]! 14$E07* ![\8]! A))*:*.)$77"7* !_\^]! +a9\99`*<'(*$))*.'6B$(&7'07* *!:`]* !:8]! !37%! !^`]! Implication Modal evidence for moral hazard has been the demand response This can be very misleading I Being marginal does not necessarily imply indifference Implications can even be the wrong sign! More than an abstract concern How can we tell? 25 / 37 Understanding the Marginal Internality Use health responses I I Define H(p) = E [m(p; s)b(s) − s] to equal the aggregate level of health Standard model: H 0 (p) = M 0 (p)p I I Can infer the health response directly from the demand response and the co-pay With behavioral hazard: − I H 0 (p) ε = −1 p pM 0 (p) Can infer the degree of behavioral hazard from the return to the last private dollar spent on treatment I Can only equate H 0 (p) = M 0 (p) · p when we are confident there is no marginal behavioral hazard 26 / 37 Understanding the Marginal Internality Use health responses I I Define H(p) = E [m(p; s)b(s) − s] to equal the aggregate level of health Standard model: H 0 (p) = M 0 (p)p I I Can infer the health response directly from the demand response and the co-pay With behavioral hazard: − I ε H 0 (p) = −1 p pM 0 (p) Can infer the degree of behavioral hazard from the return to the last private dollar spent on treatment I Can only equate H 0 (p) = M 0 (p) · p when we are confident there is no marginal behavioral hazard 26 / 37 Re-Thinking Optimal Co-pays Re-Thinking Optimal Co-pays I Optimal co-pays also depend on marginal health value, unlike in standard model ! 0 B H (p ) I c − pB ≈ + −1 η pB p B |M 0 (p B )| I Helps rationalize VBID (Chernew et al. 2007): lower cost-sharing for treatments with greater health benefits I Fixing the demand response, co-pays should be lower when this response translates into relatively worse health 27 / 37 3$e'(*%$7.,)$(*"%"047*f/$4$)*'(*0'0<$4$)*3Jg*,074$-)"*$0G&0$g*=I/g*74('M"h* !"#$%&&&* T$4"*B"(*`99* B"(7'0*K"$(7* /,))* .'%"($G"* ;7,$)* .'%"($G"* ``\9* `8\d* NO-6)#4*K8=6M8# "`^]* $-))#4*K8=6M8# ;6<6=.#=6>*#?@3A#B"CD#:EFG#?:EHIJ:E@@C# 'JK6)-8D#:E:L# P*E#6+#%(OQ# ;7,$)*.'%"($G"* /,))*.'%"($G"* :9`9* 8d^_* 8:[`* 88i_* `[_8* `_j8* `9ii* `9`:* [[8* [8_* :ji* :^9* `:`* `:_* The Pitfalls of Ignoring Behavioral Hazard The pitfalls of ignoring behavioral hazard Taking behavioral hazard into account can dramatically influence the inferences we draw from data Let’s compare the optimal co-pay under behavioral hazard, p B , to the co-pay a “neo-classical” analyst who ignores behavioral hazard would think is optimal I.e., ignoring corners and multiplicity the neo-classical analyst believes p N is optimal, where p N solves: −M 0 (p) · (c − p) − I (p) · M(p) = 0 28 / 37 Numerical Example Consider a numerical example, where we make the following assumptions: I Quadratic utility: U(C ) = αC − βC 2 I Uniformly distributed disease severity: s ∼ U[0, s̄] I Specific values for income, etc.: y = 2500, α = 7000, β = 1, q = .1, s̄ = 200 I Coinsurance rates perfectly coincide with co-pay levels: cost c = 100 I Consider the deviation between p B and p N as a function of ε(s) ≡ ε̃: ε̃ ∈ {−99, −50, 0, 50, 99} 29 / 37 Numerical Illustration ε̃ = −99 ε̃ = −50 ε̃ = 0 ε̃ = 50 ε̃ = 99 Neo-classical Co-Pay (p N ) 99.90 95.25 90.92 87.13 0 Optimal Co-pay (p B ) -2.85 42.78 90.92 141.08 192.86 1. p B < p N if ε̃ < 0; p B > p N if ε̃ > 0 2. p B increasing in ε̃ 3. p N instead decreasing in ε̃ 4. Deviation between p B and p N can thus be huge 30 / 37 Important Point Illustrates: When behavioral hazard is really positive or really negative, situations where the neo-classical analyst believes co-pays should be low are precisely instances where co-pays should be high and vice-versa! Intuition: First take case of really positive behavioral hazard. 1. Neo-classical optimal co-pay: Almost everybody gets treated when p=c ⇒ looks like there is approximately no benefit to controlling moral hazard ⇒ low co-pay 2. Optimal co-pay: Many people who demand treatment at p = c are inefficiently doing so ⇒ big benefit to raising the co-pay above cost ⇒ very high co-pay (provided people are not extremely risk averse) 31 / 37 Intuition Continued Now take the case of really negative behavioral hazard. 1. Neo-classical optimal co-pay: Almost nobody gets treated at p = c ⇒ looks like huge benefit to controlling moral hazard ⇒ high co-pay I (appears optimal so long as people are not extremely risk averse) 2. Optimal co-pay: Even at a co-pay of zero, people at the margin of getting treated have a benefit above cost ⇒ no benefit to controlling behavior by raising the co-pay above zero, but there (may be) an insurance value cost ⇒ the co-pay should be at most zero. 32 / 37 Using Choudhry Data If we assume moral hazard (don’t look at health outcomes) I Full coverage ⇒ $106 increase in spending (per patient) I Average patient share 25% ⇒ extra care consumed has monetized health value of at most $.25 on the dollar, or $26.50 overall I Marginal (social) dollar has -$.75 net return I Eliminating copayments bad policy (abstracting from insurance value) Taking behavioral hazard into account I Take a small part of the health impact: mortality reduction I .3 percentage point reduction ⇒ $3000 value using common value of statistical life ($1 million of death averted) I Marginal (social) dollar has $27 net return I Eliminating copayments very good policy 33 / 37 Using Choudhry Data If we assume moral hazard (don’t look at health outcomes) I Full coverage ⇒ $106 increase in spending (per patient) I Average patient share 25% ⇒ extra care consumed has monetized health value of at most $.25 on the dollar, or $26.50 overall I Marginal (social) dollar has -$.75 net return I Eliminating copayments bad policy (abstracting from insurance value) Taking behavioral hazard into account I Take a small part of the health impact: mortality reduction I .3 percentage point reduction ⇒ $3000 value using common value of statistical life ($1 million of death averted) I Marginal (social) dollar has $27 net return I Eliminating copayments very good policy 33 / 37 Summarizing a Few Points I With behavioral hazard, health insurance can provide more than just financial protection: it can lead to more efficient health delivery I I Demand responses no longer form a sufficient statistic for setting co-pays I I Optimal insurance may be full or nothing We derive a formula for the optimal-copay level that can be empirically implemented based on demand and health responses Situations where the neo-classical analyst believes co-pays should be low can be precisely instances where co-pays should be high and vice-versa 34 / 37 Mullainathan, Schwartzstein and Congdon (2012) Broader insight: Significant progress can be made without relying on specific psychological assumptions about why behavior may deviate from the optimum I What matters for much of the analysis is the wedge between the marginal private benefit and demand curves Mullainathan, Schwartzstein and Congdon (2012) review how this plays out in other policy problems, e.g., I Taxing to finance a public good I Taxing to correct an externality I Taxing to provide social insurance We also consider nudges 35 / 37 Welfare Impact of Policy Change The analysis proceeds much like it did with health insurance because the welfare impact of a marginal increase in the tax to finance a transfer can typically be broken into the same two terms: Impact of Changing Behavior + Impact of Transferring Resources I Commodity taxation - distorts behavior; benefits stem from increasing revenue I Taxing to correct an externality - changing behavior is beneficial; need not have other benefits Again, behavioral hazard matters because it changes who is marginal Impact of Changing Behavior = M 0 (t)[t + ME + |{z} MI ] =−ε 36 / 37 More Broadly I Behavioral public finance often seems messy I I I Many different biases Lessons seem context specific However, there are some common features across biases and policy contexts I I There is a wedge between marginal private value and demand Nudges influence demand (next talk) I By building a framework based on these features, the goal is to highlight broader lessons I Behavioral public finance is not as messy as it seems 37 / 37