NET 2: Empirical Research on Network Effects
Gandal, N., "Hedonic Price Indices for Spreadsheets and an Empirical
Test for Network Externalities," Rand Journal of Economics (25), 1994,
160-70.
Gandal, N., Kende, M., and R. Rob, “The Dynamics of Technological
Adoption in Hardware/Software Systems: The Case of Compact Disc
Players,” Rand Journal of Economics (31), 2000, 43-61.
1
Gandal, Rand Journal of Economics, 1994
Did spreadsheet programs that were compatible with Lotus
command a premium?
Idea is that compatibility gives users of your product access
to other product’s network.
Results not due to switching costs because the spreadsheet
market was growing a lot b/w 1986 and 1991.
Direct effects because people want to share files and indirect
with compatible software e.g. database programs.
2
•Paper uses product level data on spreadsheet prices and
characteristics between 1986 and 1991
•Estimate hedonic pricing equations of form:
ln(pit) = Zit  + it
•Z are product attributes, one of which is a compatibility dummy,
others “network” variables are dummies for linking capabilities
between spreadsheets (LINKING), (GRAPHS). There is a
variable that measures capacity of the spreadsheet as well.
•By 1991, nearly all high end spreadsheets were compatible.
That was not the case in 1986, when many of the premium
brands were not compatible with Lotus format.
•Article examines consumers’ valuation of compatibility, yet
formally doesn’t address firms’ decisions to become compatible.
Would need a structural model for this.
3
Data
There is an unbalanced panel of 91 model-observations. In addition
to the basic editing functions common to all spreadsheets, the
DATAPRO report contains the following characteristics and
features. I briefly summarize the available data.
(1) The dummy variable LOCOMP takes on the value one if the
program is compatible with the Lotus (WKS, WK1) format.
Otherwise the variable takes on the value zero. If the program is
compatible with the Lotus format, it is capable of exchanging files
with Lotus and other spreadsheets that support the Lotus format.
(2) The dummy variable RECALC takes on the value one (zero) if
the program can (can not) automatically recalculate when new
entries are made.
(3) The dummy variable SORTING takes on the value one if the
4
program
can sort a group of data observations on at least two levels.
(4) The variable GRAPHS is a dummy variable that takes on the
value one if the program is capable of performing all of the following
basic graphs: pie, bar, and line. If the program cannot perform these
basic functions, the variable GRAPHS takes on the value zero.
These basic functions are bundled because the early DATAPRO
reports collected the data in this manner.
(5) The variable WINDOW takes on the value one if the maximum
number of windows on-screen simultaneously is between two to
fifteen and takes on the value two if this maximum is sixteen or more.
If this feature is not available, the variable WINDOW takes on the
value zero. This variable was defined in this manner because some
spreadsheets offer a maximum that is limited only by hardware
features.
(6) The dummy variable LINKING takes on the value one if values in
several worksheets can be updated at the same time.
5
(7) The dummy variable EXTDAT takes on the value one if the program
provides links to external data bases, and zero otherwise. This link can
be either proprietary or through SQL support. If this feature is
available, databases on mainframes can be downloaded directly into the
spreadsheet.
(8) Macros allow a user to automate repetitive tasks. The dummy
variable PROGRAM takes on the value one if macros can be written in
this manner.
(9) Macros can also be written in “learn mode.” In learn mode, the
keystrokes that are to be replicated are typed and the spreadsheet
converts these keystrokes into a Macro. The dummy variable LEARN
takes on the value one if the program enables the users to automate
repetitive tasks in this manner.
(10) The dummy variable LANCOM takes on the value one if the
6
program
has the capability of linking independent users through a LAN.
(11) The dummy variable PRINT takes on the value one if three or
more of the following five advanced print functions are possible:
Sideways printing, Background Printing, Preview Mode, PostScript
Support, and Printing of non-contiguous worksheet portions.
(12) The variable PRESENT takes on the value one if worksheets and
graphs can be printed on the same page OR if multiple printing fonts
(and character sizes) are available. If both features are available, this
variable takes on the value two, while if neither feature is available, the
variable takes on the value zero. Although it seemed natural to group
the two presentation features together, nothing in the analysis changes
if these features are entered as separate variables.
(13) The dummy variable LOTUS takes on the value one if the
program is produced by Lotus Development Corporation and zero
otherwise.
7
(14) The variable MINRC measures the power of the spreadsheet and
is defined to be the minimum of the maximum number of rows and
columns that the spreadsheet can handle.
(15) The variable PRICE is the list price for a single copy of the
program.
(16) The variable LPRICE is defined to be the natural log of the
price.
(17) The variable LMINRC is defined to be the natural log of
MINRC.
Variables (1) through (6) are considered to be basic spreadsheet
features, while variables (7) through (12) are more advanced features.
Some of these advanced features (EXTDAT, PRESENT, and PRINT)
were not available until 1989.
8
In addition, to the above information, the year in which the
observation was taken and the date of introduction are available. This
allows the construction of time, age and vintage dummy variables,
which are important for the analysis. The time dummy variables are
denoted TIME87, TIME88, TIME89, TIME90, and TIME91.
Similar to the personal computer (hardware) market, most products in
this market were less than two years old. In this sample, fully 54
percent of the spreadsheets were new, 26 percent had been available
for one year, 9 percent were two years old and 11 percent had been
available for three or more years.
Lotus was the dominant firm throughout the period. Other major
spreadsheets include Microsoft (first with Multiplan and then with
Excel), Computer Associates (various versions of SuperCalc),
Paperback Software (VP-Planner), WordPerfect (PlanPerfect) and
Borland (Quattro and Quattro Pro) in the latter part of the sample.
9
TABLE 2: REGRESSION RESULTS (DEPENDENT VARIABLE IS LPRICE)
Regression (#1)
Regression (#2)
Regression (#3)
All independent Variables
Significant Variables
Preferred equation
coeff.
VARIABLES
CONSTANT
TIME87
TIME88
TIME89
TIME90
TIME91
LMINRC
3.73
-0.076
-0.42
-0.64
-0.74
-0.82
0.11
0.41
LOTUS
0.59
0.44
GRAPHS
0.45
0.36
WINDOW
0.17
0.18
LOCOMP
0.76
0.70
EXTDAT
0.52
0.67
LANCOM
0.25
LINKING
0.18
0.43
LEARN
0.03
PROGRAM 0.13
PRESENT -0.08
PRINT
0.20
SORTING -0.21
TLANCOM
TLMINRC
10
TLINKING
(t-stat)
coeff.
(10.92)
(-0.45)
(-2.37)
(-3.50)
(-4.13)
(-4.48)
(1.31)
(3.42)
(3.46)
(1.97)
(2.94)
(2.03)
(2.16)
(1.71)
(5.30)
(3.02)
(3.10)
(3.16)
(1.62)
(1.51)
(2.22)
(0.18)
(0.70)
(-0.54)
(0.86)
(-1.27)
3.76
-0.062
-0.44
-0.70
-0.79
-0.85
(t-stat)
(12.31)
(-0.38)
(-2.67)
(-4.20)
(-4.90)
(-5.30)
0.11
coeff. (t-stat)
3.12
-0.07
-0.45
0.92
0.90
0.85
(1.59)
Regression (#4)
New Models Only
coeff.
(9.50)
(-0.43)
(-3.03)
(1.71)
(1.67)
(1.59)
0.26
(3.24)
-0.57
-0.94
0.56
(4.36)
0.46
(3.62)
0.46
(3.51)
0.52
(4.18)
0.17
(2.14)
0.14
(1.92)
0.72
(5.28)
0.66
(5.17)
0.55
(4.05)
0.57
(3.93)
0.21
0.21
(1.65)
(1.91)
0.61
-0.34
-0.31
(3.28)
(-3.07)
(-1.49)
(t-stat)
2.61
(-1.73)
(-2.91)
1.63
1.47
1.54
0.26
(2.00)
-0.52
-0.25
0.33
(-2.76)
(-0.76)
(4.91)
(1.65)
(1.51)
(1.56)
(1.20)
11
Price Indexes
Table 3 shows the quality-adjusted (hedonic) price indexes for regressions 2,3 and
4 from Table 2. The numbers in parentheses are the percentage price declines
from the previous year. The average yearly decline in quality adjusted prices was
15 percent for the full sample and 22 percent for new products in the sample. For
the second regression, price indexes are calculated by taking the exponentiated
estimated coefficients on the time dummy variables, with the coefficient on T86
normalized to zero. For regressions 3 and 4, the procedure is slightly more
complicated. See Berndt and Griliches (1993) for details.
Table 3. Price Indexes for Spreadsheets (1986=1.00)
Year
Regression (#2
Regression (#3)
Regression (#4)
1986
1.00
1.00
1.00
1987
.94 (6.0)
.93 (7.0)
.57(43.0)
1988
.64 (31.7)
.64 (30.9)
.39 (31.6)
Percentage decline from previous year in parentheses
12
1989
.49 (22.6)
.50 (21.9)
.31 (20.5)
1990
.45 (8.6)
.48 (4.0)
.27 (12.9)
1991
.42 (5.5)
.46 (4.2)
.28 (-3.7)
Further Tests of the Significance of Lotus Compatibility
In the first subsample, 15 of the 40 model observations (38 percent) were not compatible with the
Lotus platform and some of these "incompatible" spreadsheets were relatively expensive. This is not
the case with the second subsample, in which 42 of the 51 model observations (82 percent) were
compatible with the Lotus platform. Further, the spreadsheets not compatible with the Lotus platform
in the second subsample ranged in price from $39.00 to $60.00, while the spreadsheets that were
compatible with the Lotus platform ranged in price from $60.00 to $695.00.
It is thus encouraging that the variable TLOCOMP (which is LOCOMP interacted with a dummy
variable for the second (1989-1991) sample period) is insignificant in both regressions 3 and 4. Adding
the variable TLOCOMP to regression 3, one obtains a coefficient (t-stat) of -0.12 (-0.53), while adding
the same variable to regression 4 yields a coefficient (t-stat) of 0.02 (0.055). Thus, the Lotus
compatibility parameter is essentially the same for both subsamples.
Despite the fact that some of the incompatible spreadsheets were relatively expensive in the (19861988) subsample, the average price of a lotus compatible spreadsheet in this subsample ($365.00) is
much higher than the average price of an incompatible spreadsheet ($80.00). Thus this subsample was
further restricted to spreadsheets that cost less than $200.00. In this restricted sample of non-premium
spreadsheets, there were 24 observations, of which 9 were compatible with the Lotus platform. The
mean price of lotus compatible spreadsheets in this subsample was $151.00 versus $80.00 for the 15
incompatible spreadsheets. Gandal (1992) shows that although the lotus compatibility effect declines
slightly in magnitude, it is still highly significant. Thus, the effect of Lotus compatibility continues to
hold for non-premium spreadsheets.
13
Comments:
•Common critique of all hedonic price equations is question of
demand or supply effects (they are very reduced form). Is this
really a premium or does it just cost more to make Lotus
compatible spreadsheets?
•Finds 93% premium for Lotus compatibility. (e.66 -1)
•At first glance, the premium seems quite large, but once one
thinks about the importance of compatibility…
•Results are robust to all sectors of the market (premium,
generics, etc.)
•Gandal (1995) extends the results to Database Management
Software and multiple standards. He finds similar results, but
the only “important” standard is the Lotus (WK) standard.
14
Gandal, Kende, Rob (Rand Journal of Economics, 2000)
The dynamics of technological adoption in hardware/software systems:
the case of compact disc players
First Structural Model used in Empirical Network Literature. See also
Rysman (2001).
This paper examines the diffusion of a hardware/software system. For
such systems there is interdependence between the hardware-adoption
decisions of consumers and the supply decisions of software
manufacturers.
Paper considers the CD-industry and estimates the (direct) elasticity of
adoption with respect to CD-player prices and (the cross) elasticity with
respect to the variety of CD-titles.
Results show that the cross elasticity is significant.
Model can be used to quantify the effect of various policies aimed at
15
speeding up the diffusion of a system.
Methodology useful for public-policy analysis regarding
the benefits of backward compatibility for other systems. (Example
HDTV).
Firm strategy: it is claimed that high-tech firms can enhance their
profits by subsidizing the adoption of their technology. (Elasticities
of hardware sales with respect to hardware prices and with respect to
software availability.)
Dynamic model for estimating demand (technology adoption) is
applicable even when there is no complementary software industry.
Consumers explicitly trade-off the lower prices which will result
from waiting one period with the loss of one period's services from
the durable product.
16
Step 1: Learn about the CD industry. (Helpful for modeling purposes)
Step 2: Informal reduced form regressions -- get feel for data
Step 3: Develop a structural model.
Step 4: Econometric Specification and Estimation of Structural Model.
Step 5: Discuss, Interpret, and Check Robustness of the Results
Step 6: Perform counterfactuals
17
Step 1: Key institutional details:
(i) in the case of CD-players, hardware prices are essentially
exogenous.
Philips licensed the technology to more than 30 firms. Hardware
market for compact disc players quite competitive.
(ii) There are cost instruments for CD-variety.
These cost instruments are the fixed cost of installing capacity for
pressing compact discs.
(iii) Record companies were integrated into the production of
compact discs. The first CD (disc) pressing plant in the U.S was
opened by Sony/CBS Records (of Japan) in 1984.
The second plant was opened by Phillips/Polygram Records. CD
production/recording industry was an integrated “software”
industry
18
2.
19
Informal Regressions
Step 3: Structural Model
Entry decision of software firms:
If a software firm enters at time t, lifetime profits are
-Ft + t+1 + 2t+2 +… , where t+1 are operating profits
from sales of software in period t+1.
If a software firm waits and enters at time t+1, lifetime
profits (calculated at time t) are
-Ft+1 + 2t+2 +… (Enters at time t+1)
Free entry equilibrium condition requires that benefit of
waiting (cost savings) equal benefit from early entry
Ft -Ft+1 = t+1 = Ytf(nt),
where n = #of software firms, Y=installed base.
20
(1) log f(nt) = -log  - log Yt + log (Ft -Ft+1)
Consumer Adoption Decision
 is an individual characteristic measuring _
intensity of preference for the system, 0<< . The c.d.f. of  is F().
_
The total number of potential buyers is M=F()
If consumer  purchases in period t, net benefit is:
-Pt + [CS(pt+1 ) + 2CS(pt+2 ) +…], where P=hardware price,
p=software price, CS=Consumer Surplus.
If consumer purchases in period T+1, net benefit (evaluated at t) is:
- Pt+1 + [2CS(pt+2 ) + 3CS(pt+3 ) +…].
Pt - Pt+1 = tCS(pt+1 ) = tg(nt). (Indifferent consumer)
(2) log(t)=log(Pt - Pt+1) - log  - log g(nt).
21
4.
Econometric Specification & Estimation
F()=1
g(n)=anr
f(n)=bn
Comment: First assumption is somewhat different from what is
usually done. People typically assume a bell-shaped distribution
over . However, the assumptions give us tractable functional
forms to estimate and the insight of the model -- that one should use
cumulative variables -- is not dependent on these functional forms.
(1*) log (Nt) = 0 + 1 log Yt + 2 log (Ft -Ft+1) + t ,
(2*) log (M-Yt)= 0 + 1 log (Pt -Pt+1) + 2 log Nt + nt ,
where N=mn (m number of titles per firm); model implies 1 = -2
From the theoretical model, 1 >0, 2 <0, 1 >0, and 2 <0.
22
OLS Results with AR(1) terms are in Table 3
Instrumental Variables Estimation: Table 4
Cost Shifters (for Nt ): log (Ft -Ft+1) , (Ft -Ft+1)2
Cost Shifters (for Yt ): log (Pt -Pt+1) , (Pt -Pt+1)2
Theoretical direction of OLS bias: Estimates of cross
coefficients biased away from zero; estimates of own
coefficients biased towards zero.
Results in Tables (3) and (4) consistent with theoretical
direction of OLS bias.
23
5.
Discussion/Interpretation of Results
Elasticity of hardware sales with respect to variety
of software: - 2(M-Yt)/Yt
Elasticity of hardware demand with respect to a
permanent percentage price cut: - 1(M-Yt)/Yt
Ratio - 2/1 is independent of time.
From table (4), estimate of the ratio is 0.54. A 10
percent increase in CD titles would have had as
large an effect on sales as a 5 percent price cut.
24
Robustness of the Results
Potential Market (Index) M=300,000 (rather than 500,000).
Both the price and variety effects nearly double in absolute value.
But the estimate of the ratio (- 2/1) remains essentially
unchanged at 0.52.
We also examined an alternative set of instruments. In the
alternative case, we just employed log (Ft -Ft+1), as an
instrument. In this case, the consumer adoption equation is
exactly identified. The ratio (- 2/1) remains essentially
unchanged at 0.56.
We also estimated the consumer adoption equation using first
differences and estimated the consumer adoption equation using
an alternative specification. See the paper for discussion.
25
6.
Counterfactual: The Effect of Compatibility
Assume that it had been possible to make CD-players compatible
with LPs, and that the IV parameter estimates of table (4) describe
the true diffusion process.
Using simulations, we examine how compatibility could have
accelerated the adoption process.
We find that if the amount of variety had grown by 100 percent
between the first and the second quarter of 1985 the diffusion
process would have been shortened by 1.5 years.
While this counterfactual is purely a ``thought experiment'' for the
CD-system, it is of great relevance for other systems such
as high-definition television (HDTV).
26