elasticity, estimates, measurement error, OLS, Hassett, interval, variables, capital formation, firm, Emory University, panel data, parameter estimates, firms, output growth, Business Capital, Caballero, Steven M. Fazzari, productivity shock, investment, capital accumulation, capital stock, plausible results, estimation method, investment data, Solow growth model, Journal of Public Economics, University of Chicago Press, cost, Brookings Papers on Economic Activity, cost elasticity, John B. Taylor, estimate, Federal Reserve System, OLS estimates, industry effects, Lawrence Summers, European Central Bank, Kenneth West, Andrew P. Meyer, estimation strategy, standard error, constant returns to scale, Center for Economic Studies, Werner Roeger, Hashem Dezhbakhsh
Content:
That Elusive Elasticity: A LongPanel Approach To Estimating The Price Sensitivity Of Business Capital Robert S. Chirinko, Steven M. Fazzari, and Andrew P. Meyer* June 2002 * We thank Hashem Dezhbakhsh, Harry Huizinga,
Daniel Levy, Doug Meade, Werner Roeger, Lawrence Summers, Kenneth West, and seminar participants at the Center for Economic Studies (Munich), the Centro de Investigacion y Docencia Economicas (Mexico City),
Emory University, the European
central bank, the European Commission, the
University of Illinois (Chicago) and the University of Missouri for comments and suggestions, and Kate von Koss for preparing the figure. The views expressed here do not necessarily reflect those of the Federal Reserve Bank of St. Louis nor the
Federal Reserve System. All errors, omissions, and conclusions remain the sole responsibility of the authors.
2 That Elusive Elasticity: A LongPanel Approach To Estimating The Price Sensitivity Of Business Capital Abstract
The sensitivity of business
capital formation to its user cost plays a key role in the analysis of many economic issues. Although this elasticity has been the subject of an enormous number of studies, a consensus remains elusive. We develop an estimation strategy that filters panel data in an original way and avoids several pitfalls  difficulttospecify dynamics, transitory timeseries variation, and positively sloped supply schedules  inherent in investment equations that can bias the estimated elasticity. Results are based on an extensive panel containing 1,860 manufacturing and nonmanufacturing firms. Our model generates a precisely estimated user cost elasticity of approximately 0.40. The method developed here may prove useful in estimating other structural parameters from panel datasets.
JEL Nos. E22 and H32
Corresponding Author: Robert S. Chirinko Dept. of Economics Emory University Atlanta, Georgia USA 303222240
Steven M. Fazzari Dept. of Economics
Washington University St. Louis, Missouri USA 63130
PH: (404) 7276645 FX: (404) 7274639 EM:
[email protected]PH: 3149355693 FX: 3149354156 EM:
[email protected] wustl.edu
Andrew P. Meyer Federal Reserve Bank of St. Louis St. Louis, Missouri USA 631660442 PH: 3144444647 FX: 3144448740 EM
[email protected] frb.org

2
Table Of Contents
Abstract I. Introduction II. Estimation Strategy III. The Panel Dataset IV. Empirical Results A. OLS Estimates B. Measurement Error C. IV Estimates D. SubSample Estimates E. Comparison To Alternative Approaches V. Summary And Conclusions References Appendix A: Trending Variables Appendix B: The Replacement Value of Capital Appendix C: ShortRun Frictions And Measurement Error Tables Figure
That Elusive Elasticity: A LongPanel Approach To Estimating The Price Sensitivity Of Business Capital I. Introduction Understanding the links between capital formation and price incentives has been a prominent topic on the quantitative research agenda for decades. This elasticity, which we refer to as , features prominently in several areas of economic research. Policymakers frequently alter price incentives for capital accumulation, and is a key element determining policy effectiveness and the resulting welfare changes. The validity of various growth theories depends on the value of . Although the vast majority of
academic research on capital formation utilizes the user
cost of capital as the central price variable, elasticity estimates vary widely. The range of estimated 's corresponds to an equally wide range of policy implications in tax simulation models. In a simplified version of the Ballard, Fullerton, Shoven, and Whalley (1985) computational general equilibrium (CGE) model, the change in welfare from equalizing capital
tax rates across industries is 70 percent larger when the user cost elasticity rises from 0.50 to 1.00. Similarly, Engen, Gravelle, and Smetters (1997, Table 5) show thatё when the income tax is replaced by a
consumption tax, the increase in steadystate
net output is 79 percent higher when the elasticity of 0.50 is replaced by a value of unity. Results from the twocountry model of Roeger, Veld, and Woehrmann (2000) are also sensitive to whether the elasticity is 0.50 or 1.00; a one percentage point cut in one country's corporate tax rate leads to a 70 percent larger increase in combined consumption with the larger elasticity. Fox and Fullerton (1991) find that, in CGE models, estimated welfare gains from tax initiatives depend much more on this elasticity than on the complex features and detailed disaggregation found in many simulation models. Starting with the seminal analysis of Harberger (1959, 1962), the user cost

2
elasticity, equivalent to the substitution elasticity between capital and other inputs
in a CES production technology, is central to assessing policy impacts.
The substitution elasticity is also essential for understanding longrun growth.
Values of near or below unity casts doubt on the validity of the Solow growth
model as conventionally formulated. Mankiw (1995, p. 287) presents a formula for
computing the impact of on the difference in rates of return on capital between
rich and
poor countries implied by the Solow model. When = 4.0,the implied rate of return difference is only 3
percentage points.1 However, as is lowered to 1.0 or
0.5, the implied differences become implausibly large, rising to 32 and 100
percentage points, respectively. In the Solow growth model, Klump and Preissler
(2000) show that the substitution elasticity is negatively related to the speed of
convergence toward the steadystate (if the economy has overaccumulated capital) and positively related to steadystate per capita output.2 If exceeds unity, longrun
growth is possible even without any technological progress (Pitchford, 1960; Barro
and SalaiMartin, 1995). Acemoglu (2001) shows that the relative contributions of
technological change and factor accumulation in accounting for longrun growth
depend on , which has further implications for the importance of biased technological change and the movement of factor shares.3
Despite the substantial research energies devoted to estimating the user cost elasticity, a consensus value remains elusive.4 For example, in the
Joint Committee1 These computations are based on the assumptions of a 10 percent return in the rich country and a capital elasticity in production of 2/3. 2 However, the latter result is not robust; in a Diamond overlapping generations model, Miyagiwa and Papageorgiou (forthcoming) show that and steadystate per capita output are negatively related (provided is sufficiently large). 3 Furthermore, in the original article introducing the CES production function, Arrow, Chenery, Minhas, and Solow (1961) note that the impact of factor endowments on
International Trade and the variation of relative income shares depend on the value of this elasticity. 4 See Chirinko (1993), Hassett and Hubbard (1997), and Mairesse, Hall, and Mulkay (1999) for surveys of the empirical literature.

3
On Taxation's (1997, Table 6) study of nine different tax models, user cost
elasticities range from 0.20 to 1.00. The wide range of estimated elasticities
reported in the literature may be attributed to a common source. Most econometric
studies rely on quarterly or annual timeseries variation in investment data to
identify the user cost elasticity. Three biases may result that weigh more or less
heavily in different studies. First, the specification of an investment equation
requires assumptions about dynamics. While economic theory is highly informative
about the determinants of the demand for the stock of capital, it is relatively silent
about the demand for the flow of investment. Misspecified dynamics can bias
estimates of the user cost elasticity (Summers, 1988). Of particular importance are
the nature of adjustment costs and the role of financing constraints, which have
received a great deal of attention in recent years and whose effects on
investment spending remain controversial.5 Second, coefficient estimates from investment
regressions may be biased if the timeseries variation of investment spending largely
reflects adjustments to transitory shocks and firms respond less to transitory than
permanent variation because of adjustment costs. An elasticity estimated with time
series data at quarterly or annual frequencies will tend to be lower than the "true" longrun elasticity.6 Third, if the supply curve of investment is upward sloping, as
is more likely in the short to mediumrun, studies incorrectly maintaining a perfectly
elastic supply schedule will tend to understate demand elasticities (Goolsbee, 1998).7 While the misspecification of dynamics has an indeterminate effect, the
5 Regarding adjustment costs, see the surveys by Hamermesh and Pfann (1996) and Caballero (1999). Regarding financing constraints, see the survey by Hubbard (1998) and the controversy in Kaplan and Zingales (1997, 2000) and the reply by Fazzari, Hubbard, and Petersen (2000). Chirinko, Fazzari, and Meyer (1999) show that excluding cash flow (a variable typically included to capture financing constraints) from an investment equation using annual data biases upward the estimated user cost elasticity. 6 This point has been noted by, among others, Eisner (1967), Lucas (1969). Berndt (1976), Shapiro (1986b), and Kiyotaki and West (1996). 7 This conclusion has been challenged by Hassett and Hubbard (1998) and Whelan (1999).

4
estimated user cost elasticity will be biased toward zero by transitory timeseries
variation and positively sloped supply schedules.
These potential problems all stem from a common source  the use of
investment data as the measure of capital formation. We avoid these problems by
developing an approach that relies directly on capital stock data and exploits in an
original way the substantial information available from panel data. We focus on the firstorder condition relating the longrun desired capital stock (K*) to the longrun desired values of output (Y*) and the user cost (C*). This specification underlies
virtually all investment studies since Jorgenson's (1963) pathbreaking work on the
neoclassical model of capital accumulation, and can be represented as follows:
(1) K* = G[Y*, C*]. + The difficulty with estimating (1) is that the desired values are not readily observable. We use panel data, long in the time dimension, to estimate the variables in (1) as timeaverages within firms. With empirical counterparts to K*, Y*, and C* defined, it is straightforward to estimate G[.] directly across firms. This relatively simple yet fully rigorous approach estimates technology parameters immune to the three biases discussed above. Our study proceeds as follows. Section II introduces the estimation strategy. The econometric equation is derived from the firm's profitmaximization problem, and longrun values of the variables entering the regression equation are measured as timeaverages. Our estimation strategy accounts for a variety of productivity shocks, omitted variables, and firm fixed effects, and uses panel data in a way that differs substantially from prior panel studies. Section III discusses the panel dataset, containing 1860 firms for the period 1972 to 1991, and the construction of the variables. Section IV presents our OLS and IV results. Both techniques yield

5
similar estimates of the user cost elasticity of approximately 0.40. This estimate is
higher than the elasticity of 0.25 reported by Chirinko, Fazzari, and Meyer (1999)
based on the same dataset but using an investment model. Thus, the three problems
affecting investment equations  difficulttospecify dynamics, transitory timeseries
variation, and positively sloped supply schedules  impart a discernible bias toward
zero. Nonetheless, the user cost elasticity remains far from unity, the value defining
the frequently used CobbDouglas production function and determining the cutoff
at which tax incentives become cost effective (in a static sense). Section IV also
assesses the importance of measurement error, examines the sensitivity of the
estimates to various subsets of the sample, and compares our approach to related
work with panel data. Section V offers a summary and conclusions.

6
II. Estimation Strategy
Our econometric model follows directly from the behavior of a firm that
maximizes its discounted flow of profits over an infinite horizon. We analyze the
firm's choices in longrun equilibrium, thus eliminating the need to model
adjustment costs, delivery lags, and vintage effects. Under these assumptions, the
firm always produces its longrun desired level of output with its longrun desired
mix of inputs. The critical consequence is that the firm's dynamic optimization
problem is transformed into a static problem. To determine the firm's demand for
capital, we need only calculate the marginal product of capital evaluated at the long
run levels of inputs and output.
We assume that production possibilities are described by the following CES
technology,
(2) Y*f,t = {(K*f,t[( 1)/]) + (1)(X*f,t[( 1)/])}[/( 1)]Uf,t ,
where Y*f,t is longrun desired real output for firm f at time t, K*f,t is the longrun desired real capital stock, X*f,t is the longrun desired level of all other factors of production, and Uf,t represents a stochastic productivity shock.8 An attractive feature of the CES technology is that it depends on only three parameters characterizing returns to scale (), the distribution of factor returns () and, of particular importance for this study, substitution possibilities between the factors of production (). The CES function is strongly separable, and can be expanded to include many additional factors of production (e.g., intangible capital) without affecting the estimating equation derived below. This feature gives the CES specification an important advantage relative to other technologies that allow for a
8 The limiting value of (2) as  > 1 is the CobbDouglas production function under the additional assumption that =1.

7
more general pattern of substitution possibilities (e.g., the translog, minflex
Laurent). Our approach does not require price and quantity data on the other factors
of production (with limited availability and reliability at the firm level) to recover the
key parameter of interest.
Differentiating (2) with respect to capital, we obtain the following relation for the marginal product of capital ( Y*f,t / K*f,t),
(3) Y*f,t / K*f,t = () Y*f,t[1+ (1 )/]K*f,t [1/]Uf,t[( 1)/].
Profitmaximization implies that this marginal product of capital equals the Jorgensonian user cost of capital (C*f,t), which combines interest, depreciation, and tax rates with relative prices (an exact specification of the user cost is deferred to Section III). Setting Y*f,t / K*f,t equal to C*f,t and rearranging (3), we obtain the following expression for the longrun desired capital stock, (4) K*f,t = C*f,t[]Y*f,t[]Uf,t[ ], = (), =  , = (+ 1 )/, = ( 1)/. Note that, with a CES production function, the user cost elasticity of capital is equivalent to the substitution elasticity between capital and other inputs (multiplied by minus one). The central difficulty with estimating (4) is the that the longrun values are not observed. Most previous research addresses this problem by differencing the log of (4) to obtain an equation for investment. As discussed in the Introduction, however, this approach relies on investment data, and may therefore generate biased

8
estimates. To avoid these potential problems, we measure the capital stock
directly, and then estimate the longrun desired levels of capital, output, and the user
cost as time averages over several years. We refer to the years over which an
average is computed as an interval. As shown in Figure 1 for a representative
variable Wf,t (Wf,t = {Kf,t, Yf,t, Cf,t}), we divide our sample into three intervals
indexed by a subscript, = 0,1,2. The intervals are 19741980 (=0), 19811986 (=1), and 19871992 (=2). We assume that W*f,t equals Wf,, where the latter is
the mean of Wf,t over an interval. As we will discuss below, the=1 and
=2 intervals are used for parameter estimation; the =0 interval is used only to form
instruments and define classifications that split the sample.
With the variables in (4) defined in terms of the=1 and =2 intervals, we
take logs, and obtain the following equation,
(5) kf, = cf, + yf, +  uf,, kf, = ln[Kf,] = ln[K*f,t], cf, = ln[Cf,] = ln[C*f,t], yf, = ln[Yf,] = ln[Y*f,t], = ln[ ],
= 1,2.
where uf, is an error term that follows directly from the technology and represents productivity shocks. We model productivity shocks as follows,
(6) uf, = [vf + vi + w + wi, + wf,].
The productivity shock is decomposed into firmspecific (vf) and industryspecific (vi) components, as well as components that vary over the intervals, w, wi, and wf,. With this error structure, estimates can be obtained by differencing (5) and (6) between the intervals,

9
(7) kf, = cf, + yf, +  uf,,
kf =  cf + yf   i  wf.
=  , = (1 )/( ), = w, i = wi,, wf = wf,.
Since there are only two intervals, firstdifferencing eliminates the temporal dimension to the model, and subscripts have been omitted in the final equation. Consequently, the parameters are estimated in a crosssection regression. Equation (7) is our estimating equation relating growth in the capital stock to growth in the user cost and output. Fixed firm and industry effects are eliminated by differencing, and fixed interval effects (notably, biased technical change) are captured by the constant (). Industry effects that vary across intervals are captured by industry dummies (i). This framework allows us to estimate the parameter of central interest in this study, the elasticity of the capital stock with respect to its user costs () Additionally, we can recover the returns to scale elasticity,, as a nonlinear combination of the estimated and parameters. Our estimation strategy filters panel data in a way that differs substantially from prior panel studies. By taking interval averages, we use very lowfrequency variation to estimate the longrun values of the regression variables. This approach avoids the difficult specification problems that necessarily arise with investment regressions based on data at quarterly or annual frequencies. Differencing equation (7) across intervals controls for firm fixed effects, as well as productivity shocks and omitted variables that vary across intervals. The remaining crosssection variation provides ample degrees of freedom for estimation.

10
Consistency of the OLS
parameter estimates depends on the relation
between the stochastic element,wf, and the regressors, especiallyyf. This
correlation is not likely to be a problem for two reasons. First, wf is that part of
the productivity shock that remains after accounting for all fixed and industry
effects. Major technological changes (e.g. telecommunications, computing, the
internet) are likely to have their largest effects on all firms (captured by) or on all
firms in an industry (captured byi) with only a small residual impact that is firm
specific. Second, only part of the productivity shock enters the error term. As
noted by Shapiro (1986a), including output in a factor demand equation can com
pletely absorb the productivity shock. When the elasticity of substitution is unity,
equals zero, and uf vanishes (cf. equations (4) and (6)). When deviates from
unity, the impact of the productivity shock is nonetheless diminished by . Despite
these arguments that OLS estimates will not be appreciably affected bywf, we
present two alternative estimates that are robust to simultaneity. First, we impose
constant returns to scale (=1 implying =1). Thus, yf, the variable most likely to
be correlated with wf, no longer appears as a regressor. Second, we present IV
estimates and Hausman tests using the variables in the=0 interval as instruments.
This econometric model is robust to four potentially important distortions.
First, the parameter estimates are robust to trending variables. See Appendix A for
formal consideration of this issue and the role played by differencing in eliminating
firmspecific trends. Second, the estimates are unlikely to be influenced by
additional factors that may affect the specification of the production function or the
firstorder conditions. For example, the estimating equation is robust to including
additional factors of production. Markups that vary across firms are captured by a
firmspecific fixed effect eliminated by differencing. Moreover, the information
processing revolution may have led to biased technical change over the past 20
years. In terms of the CES technology, biased technical change is represented by

11
temporal variation in and, like w, will be reflected in the constant. Third,
studies implementing the Jorgensonian framework have often been criticized for
failing to distinguish between desired output and actual output (e.g., Coen, 1969;
Hall, 1995). By using timeaverages in the econometric equation, we recognize this
important distinction. Fourth, the estimates are unlikely to be affected by measure
ment error in the capital stock. Classic measurement error will be part of the error
term, and hence innocuous. A plausible situation where measurement error may be
systematic arises when an increase in the pace of technological change effectively
increases the depreciation of fixed capital through obsolescence, an effect not
captured in our fixed depreciation rate assumption. However, an increase in depre
ciation rates would lead to a systematic overstatement of capital in=2, and would
be captured by the constant. If omitted variables or measurement error are both
firmspecific and intervalvarying, consistent estimation becomes an issue. In this
case, the IV estimates in Section IV.C, coupled with the measurement error analysis
in Section IV.B, provide a useful safeguard to check the parameter estimates.
In sum, the estimation strategy developed here collapses the time dimension
of firm panel data by defining three intervals and then timeaveraging the data within
an interval. The first interval is used to form instruments or sort variables into
contrasting classes. The second and third intervals are used for estimation. A
variety of productivity shocks, omitted variables, and fixed firm effects are
accounted for by estimated parameters or differencing. Production function
parameters are thus estimated in a crosssection of timeaveraged, differenced firm
data. This econometric model does not solve the estimation problems inherent with
investment models  difficulttospecify dynamics, transitory timeseries variation,
and positively sloped supply schedules  that may bias estimates of the user cost
elasticity. Instead, our approachavoids these problems by exploiting panel data and
estimating directly the firstorder condition for capital.

12
III. The Panel Dataset
Our estimation method requires a panel dataset that is long in both the cross
section and timeseries dimensions and that contains crosssection variation in the
user cost. We link data sources from the Compustat Industrial Database maintained
by Standard and Poors (containing financial statement data) and Data Resources,
Inc. (DRI, containing user cost and industry data). In this section, we discuss the
construction of the variables used for regression estimates of equation (7), for
instruments, and for sorting firms into contrasting classes.
For the user costs (C), we have data for 26 different capital assets (24 types
of equipment and two types of structures). The basis for these user costs, from Hall
and Jorgenson (1967) and modified by DRI, is:
(8)
Ci,j,t = [pIj,t / pYi,t] [(1  mj,t  zj,t) / (1txt)] [rt + j]
where pIj,t is the assetspecific purchase price for asset j at time t, pYi,t is the industry i output price at time t, j is the assetspecific economic depreciation rate, and txt is the income tax rate. The investment tax credit (mj,t) and the discounted value of tax depreciation allowances (zj,t) also vary across assets. The financial cost of capital (rt) is a weighted average of the cost of equity (the dividendprice ratio for Standard & Poor's Composite Stock Price Index plus an expected longrun growth rate of 2.4 percent, with a weight of 0.67) and the cost of debt (average yield on new issues of highgrade corporate bonds adjusted to a AAA basis, with a weight of 0.33). The nominal cost of debt is reduced by its tax deductibility and the expected inflation rate, defined as a weighted average of past GDP deflator
growth rates. Industryspecific user costs are a weighted average of the asset user costs. The weights are

13
the proportion of total capital in an industry accounted for by each of the 26
assets.9 This industry information is then merged with the firmlevel Compustat
data using each firm's S.I.C. code1.0
Measurement of the capital stock (K) is important for our study. Compustat
does not provide an acceptable measure of the capital stock because book values of
net plant and equipment likely understate current replacement values in periods of
inflation. In addition, accounting depreciation rules may not accurately reflect
economic depreciation.
We measure the current replacement value of capital with a threestep, iterative algorithm.11 First, choose a seed value. We use the book value of net plant
and equipment from the firm's first observation in Compustat. The nominal seed
value is deflated by a weighted average of investment goods price deflators, where the weights are determined by the specific capital asset mix of each industry12.
9 Note that these weights are from the Bureau of Economic Analysis capital flow tables and reflect asset usage by establishment. The Compustat data reflect ownership by company. The combination of industry aggregate data for the user cost and firm data for investment and other items may induce measurement error because some firms operate in a variety of industries. To the extent that such measurement error is constant within firms, however, it will be captured in firm fixed effects. 10 We average the quarterly DRI user cost data at the firm level to obtain an annual user cost that corresponds to the Compustat data. The averages account for differences in firms' fiscal years, and therefore introduces some firmlevel heterogeneity into the user cost data. 11 This conceptual approach has been used for firmlevel panel data at least since Salinger and Summers (1983). 12 Because the book value of net plant will usually be less than the replacement cost when there is inflation, the use of net plant as a seed in 1974 distorts the measurement of the replacement cost of capital. This distortion, however, is unlikely to affect the estimated parameters for three reasons. First, the distortion will disappear as new investment is added to the capital stock at current replacement value and old capital is depreciated. The early part(=0) of our sample is used only for instruments. The effect of the seed value on the regression data, therefore, is attenuated because the capital series consist largely of new investment expenditures by the=1 and =2 periods. Based on the average depreciation rate of 14.8 percent, only 32.6 percent of the 1974 seed value will remain at the beginning of the estimation period in 1981. Second, a proportionate distortion of the seed value relative to the "true" replacement cost across firms is eliminated by our econometric procedure that takes logs and then first differences the capital

14
These are the same weights employed in the user cost computation described
above. Second, subtract capital lost to (geometric) depreciation. The firm's
depreciation rate is the weighted average of the rates for individual assets from DRI,
again using industryasset proportions as weights. Thus, there is a consistency
between the depreciation rates used in constructing the capital stock and user cost
data. Third, add in new investment. In most cases, this step simply adds the
deflated value of the Compustat capital expenditures variable. The deflator is the
weighted average of each industry's investment goods price deflators. At the micro
level, however, we must take into account that a firm's capital stock may rise or
decline due to acquisitions or divestitures that are not included in the capital
expenditure variable. If the data indicate a significant acquisition or divestiture, we
use accounting identities to calculate the impact of this activity on the capital stock.
Details of the capital stock calculation appear in Appendix B.
Output (Y) is gross sales during the year reduced by cash discounts, trade
discounts, and returned sales or allowances to customers. Sales will differ from
output by the change in finished goods inventories. While this difference may be
nontrivial in the shortrun, it will have very little impact on the longrun averages
used in our estimation. Nominal sales figures from Compustat are deflated by
industryspecific output price indexes from DRI.
For some of the results that follow, we sort the data into contrasting sub
samples depending on whether a classifying variable averaged over the=0 pre
estimation period (19741980) is above or below its median. Three variables are
used as classifiers: the cash flowcapital ratio (CF/K), the size of the capital stock
(K), and the BrainardTobin Q. Cash flow is income after taxes plus noncash
stock data that enter the regressions. Third, any remaining random measurement error in the capital stock affects the dependent variable only and, therefore, it does not bias coefficient estimates, although it would raise standard errors.

15
expenses, primarily depreciation and amortization. The numerator of Q is the
market value of equity plus the book value of debt less the book value of
inventories. The denominator is the replacement value of the capital stock measure
discussed above.
To protect against results driven by a small number of extreme observations,
we exclude observations in the onepercent upper and lower tails from the distributions of the firmspecific variables.13 Firms included in the
data set must
have some observations for each variable in all three of the intervals. Our final
data set contains 1,860 firms from all sectors of the economy.
Growth in firm variables between the =1 and =2 intervals enter directly into
the econometric model, and hence are key for estimating the parameters in equation
(7). While this equation is estimated as a crosssection of firms, the value of each
firm observation is based on temporal variation between intervals. Statistics for the
interval growth rates appear in the top panel of table 1. It is clear from the large
coefficient of variation statistics for the growth in capital, output, and the user cost
(2.3, 2.4, and 2.7, respectively) that there are substantial differences across firms in
the regression variables. It is not surprising that capital and output growth differ
across firms, but the heterogeneity in the user cost growth is remarkable relative to
most other studies. We rely on this heterogeneity for identification of the elasticity of the capital stock with respect to the user cost.14
13 We checked the robustness of our results when we deleted both the onehalfpercent and twopercent tails. The effect on the results was negligible. Because the user cost is computed from stable industry and aggregate data, we did not delete data in the tails of the user cost variable distribution. 14 The bottom three panels of table 1 give information on the levels of capital, output, and the user cost within each interval to provide a sense of the data. The capital and output statistics are in millions of 1987 dollars.

16
IV. Empirical Results
A. OLS Estimates
Ordinary least squares estimates of the structural parameters from equation
(7) appear in table 2. The focus of our study is on the user cost elasticity of capital,
, which is also the substitution elasticity between capital and other factors of
production. In column 1, our benchmark estimate of is 0.367 with a standard
error of 0.067. The null hypothesis that the user cost elasticity is zero can be
strongly rejected at any conventional level of significance. It is also clear, however,
that our estimate of is much smaller than unity, the value implied by the Cobb
Douglas production function and often assumed in applied work.
As shown by equation (7), the estimated returns to scale elasticity ( ) is a
function of the
regression coefficients on the growth in both output () and the user
cost (). The OLS estimate of the returns to scale elasticity,, is 1.135 also with a small standard error.15 With our estimated parameter values, the primary reason that
the estimated returns to scale elasticity modestly exceeds one is that the coefficient
of output growth in our capital growth regression is somewhat less than unity
(=0.925). As shown in equation (7), an estimated in the neighborhood of unity
generates results for close to constant returns for any admissible value of . It is
interesting to note that the effect of output is much stronger here than in panel data
studies using investment data (cf. Chirinko, Fazzari, and Meyer, 1999). We believe
the reason for these more plausible results is that, unlike typical investment
equations, our estimation method captures longrun,
permanent changes in output,
15 The returns to scale elasticity is recovered from the estimated coefficients with the following formula: = (1 )/( ) when > . The variance of depends in a complicated way on the variances and covariances of the estimated and . We use an approximate formula based on a secondorder Taylor series expansion of about the estimated values of and : V[] = {V[] (1 )2 + V[] (1 )2  2 C[,] (1 ) (1 )} / ( )4, where V[.] and C[.] are the variance and covariance operators, respectively.

17
and is not affected by the transitory variation that may unduly influence
investment regressions with annual or quarterly data.
The second column of table 2 presents results from including twodigit
industry dummies in the benchmark regression (thei terms in equation 7). These
dummies control for industrylevel productivity shocks between intervals =1 and
=2 or, more generally, any industryspecific effects. The structural parameter
estimates do not change much when the dummies are included. The estimate rises
from 0.367 to 0.440, and is virtually identical. The standard error of, however,
rises by a factor of more than four. With the twodigit dummies in the model, is
estimated very imprecisely. The structure of our user cost data accounts for this
increase in the standard error of. While there is some firmspecific variation in
the user cost within industries, the most important differences in the user cost occur
across industries. The estimate is therefore much less precise with industry
dummies in the model. For this reason and given the modest change in, the
remaining regressions in table 2 exclude the industry dummies.
As discussed in Section II, the most likely source of correlation between the
error term and the
independent variables in these OLS regressions comes from the
correlation between firmspecific productivity shocks embedded in the error term
and firm output growth. This potential simultaneity problem can be avoided by
imposing constant returns to scale (=1 implying =1), an assumption that removes
output growth as a regressor. The third column of table 2 presents a regression with
the output growth coefficient constrained to unity. The estimate changes only trivially when constant returns to scale are imposed (from 0.367 to 0.372).16 These
result supports our contention that the user cost elasticity is consistently estimated
16 While the R2 decreases trivially from 0.564 (column 1) to 0.560 (column 3), the constraint of constant returns to scale is rejected at the onepercent level, a result driven by the large number of observations used in estimation.

18
by OLS in our framework.
The final column of table 2 presents estimates based on the assumption of a
CobbDouglas production function, which is defined by a unitary elasticity of
substitution ( =1) and constant returns to scale (=1). Not surprisingly given the
prior results for and , the restrictions associated with the CobbDouglas are
easily rejected at the onepercent level relative to the unconstrained model in the
first column.
B. Measurement Error What role might measurement error play in biasing the estimated downward and away from a unitary elasticity (as emphasized recently for investment models by Goolsbee, 2000)? We consider three sources of measurement error. First, measurement error introduced in the construction of the capital stock will have a modest effect on the estimates because the capital stock enters as the dependent variable. In situations where measurement error in the dependent variable takes the classic form or is fixed for a given firm, industry, or interval, the elasticity estimates will be unaffected. Second, measurement error in the independent variables may arise for various reasons, and can have direct and indirect effects on the estimated. To assess the direct effects, assume that the true value of this elasticity is unity. If the OLS estimate is inconsistent because cf is afflicted with classic measurement error, the variance of this measurement error would have to account for at least 60 percent of the variance in the observedcf.17 This seems highly implausible, especially since the estimator accounts for measurement error arising from fixed firm, industry, and 17 The asymptotic bias on the estimated is given by the following formula: (#') = (VAR[f]/VAR[cf]) #, where 'and # are the estimated and true values of , respectively, and f is the measurement error. If # =1, then the variance ratio must be at least equal to 0.60 given an OLS estimate of '=0.40.

19
interval effects by differencing. An indirect effect could result from measurement
error in the other independent variable. Ifyf is measured with error, we can use
the formula proposed by Rao (1973) and an auxiliary regression of yf on cf to
assess the extent of bias on the user cost elasticity. Making the rather extreme
assumption that onehalf of the variance of yf is measurement error, we obtain the
somewhat surprising result that the estimated is biased upward toward unity. However, the bias is a trivial 0.043.18
Third, the assumption that the longrun values of K and Y appearing in the
model of Section II are measured as timeaverages over a interval may not be valid
because of various shortrun frictions. For example, irreversibility constraints or
asymmetric adjustment costs suggest the possibility that the average values of K and
Y might differ from their longrun values in a frictionless model. As shown in
Appendix C, these shortrun frictions introduce measurement error into K and Y.
The effects on have been analyzed above. Measurement error can adversely
affect the reported results but, with our estimation strategy, it does not appear to be
quantitatively important.
C. IV Estimates The OLS estimates of equation (7) are consistent under the assumption that the error term is independent of both output and user cost growth. As discussed in Section II and suggested by the results with the constant returns model in Section IV.A, these are reasonable assumptions with our estimation method. Nonetheless, 18 The bias on the estimated is given by the following formula, (#') = (( by,c)/(1R2y,c)) (VAR[f]/VAR[yf]), where 'and # are the estimated and true values of , respectively, is from equation (7), by,c is the coefficient on cf and R2 the correlation coefficient from the auxiliary regression of yf on cf and a constant, andf is the measurement error. We assume that onehalf of the variance in the output variable is measurement error; hence, VAR[f]/VAR[yf] = 0.50. We further assume that equals its estimated value under IV of 1.402. From the auxiliary regression, by,c = 0.061 and R2y,c = 0.0003; hence, (#') = 0.043.

20
we present instrumental variables estimates in table 3 to explore the robustness of
our OLS results. The instruments are constructed from data in the=0 interval.
Effectively, we employ data from 19741980 to predict growth in the user cost and
output between the 19811986 and the 19871992 intervals. The instrument list
includes the user cost (Ci,=0), capital stock (Ki,=0), the outputcapital ratio ((Y/K)i,=0), and the cash flowcapital ratio ((CF/K)i,=0). In addition, we included the annualized growth rates of capital, output, cash flow, accounts receivable, and
cash and cash equivalents defined over the=0 interval.
The benchmark IV estimate of in the first column of table 3, 0.390 is almost
identical to the benchmark OLS estimate from table 2 of 0.367. Not surprisingly,
the standard error rises with IV, but we can still strongly reject both the hypotheses
that equals zero or unity. Unfortunately, the IV estimates of are not as
reasonable. Because of the large coefficient on output growth (), the point
estimate of the returns to scale elasticity () is 0.603. The standard error of is
much larger with IV than with OLS, but the IV estimate still rejects constant returns
to scale in favor of decreasing returns. However, we do not consider this result
reliable because of our inability to find good instruments for output growth. The partial R2 statistic developed by Shea (1997) provides quantitative
confirmation of this interpretation. This statistic measures the relevance of
instruments for each estimated coefficient after removing the explanatory power used in instrumenting other regressors. The partial R2 for is 0.040, dramatically lower than the partial R2 of 0.515 for .19
To pursue this issue one step further, we reestimate the model with IV
imposing constant returns to scale (=1). Under this assumption, =1, and we no
longer need to instrument output growth. The results appear in the third column of
19 The partial R2 statistic is preferable to the more commonly used firststage R2 for reasons discussed by Shea (1997).

21
table 2. The IV estimate of is only modestly affected by imposing constant
returns. The user cost elasticity estimate rises to 0.434 from 0.390, a change well
less than one standard error. This result demonstrates that, even if the IV estimate
of returns to scale is unreliable due to the lack of relevant instruments for output
growth, this difficulty does not "contaminate" conclusions about , which is the
primary focus of our study.
The second column of table 3 presents IV estimates with twodigit industry
dummies. This specification accounts for industrylevel productivity shocks
between the =1 and =2 periods (cf. equation (7)). The point estimate of hardly
changes from the benchmark value (0.373 versus 0.390). As was the case for the
OLS estimates with industry dummies, however, the standard error of rises
dramatically, almost by a factor of three; we cannot reject the hypothesis that
equals zero in this regression. The problem here is again that most of the variation
in the user cost is across industries, with much less firm heterogeneity within
industries. The resulting collinearity between the industry dummies and user cost growth compromises the precision of the estimated .20
As a final test of the validity of the OLS estimates, we performed Hausman
tests on the parameters. The Hausman statistics are asymptotically distributed 2(1) under the null hypothesis that the OLS estimates are consistent. For the
benchmark model, the
test statistic is 0.07 and for the constant returns to scale model it is 0.92.21 Both test statistics are far below the 90 percent critical value for the 2(1) distribution of 2.71. These tests support the validity of the OLS estimates
of . Taken together, the unconstrained OLS and IV estimates strongly suggest that
20 Because of collinearity, it was not useful to include industry dummies in the model as both regressors and instruments. In the second column of table 2, the instrument set for output and user cost growth is the same as for the other IV regressions; the industry dummies are instrumented by themselves. 21 The Hausman test is not defined for the model that includes industry dummies because the

22
is approximately 0.40.
D. SubSample Estimates
Table 4 explores our results further by assessing the stability of structural
parameter estimates in several subsamples chosen to address issues that have arisen
with empirical investment models. In each subsample, we expect to remain
similar to its value in the full sample results. All estimates are with the benchmark
model. The first panel presents results with the sample split by the ratio of cash
flow to the capital stock. In investment regressions using annual data, Chirinko,
Fazzari, and Meyer (1999) found that including cash flow had a significant effect on
the estimated . We interpreted that finding in the context of the extensive literature
on finance constraints and firms' investment spending. The approach here, however,
emphasizes the longrun impact of the user cost on the capital stock. We therefore
expect financial constraints to be less important. The first panel of table 3 presents
results from data split according to the presample median cash flowcapital ratio. If
financial constraints were important at the horizon relevant for our estimation, we
would expect the estimated to be significantly different across high and low cash
flow firms that differ by their inadequate access to finance. There is little evidence
of such an effect in our data. The OLS point estimate of is somewhat larger for
the high cash flow firms than for low cash flow firms, but the difference is less than
two standard errors. Similar results hold for the IV regressions except that is
relatively larger for the low cash flow firms. The formal test statistics ('s) for the
equality of the 's from the two subsamples have pvalues greater than 0.35, easily sustaining the null hypothesis of equal 's.22
standard error of the IV estimate is slightly smaller than the standard error of the OLS estimate. 22 The null hypothesis that '= " (where the 'and " refer to estimates based on the low and high subsamples, respectively) is evaluated by in the following auxiliary equation based on equation (7): kf =  cf  cf * If + 'yf * If + " yf * (1If)  '* If  " * (1If)  wf., where If is an indicator variable equal to 1 for the low subsample and 0 for the high subsample and

23
Our second sort is by size, defined by the median average capital stock in
the presample period (=0). The technologies utilized by firms may vary
systematically by size, and the technology parameters estimated here may change
accordingly. Moreover, size is frequently used to identify firms that may be finance
constrained. External finance may be relatively costly for smaller firms because
they are not be able to bear the substantial fixed costs of obtaining external funding
or they lack visibility in external capital markets. Relative to the results in table 2,
the OLS point estimates of are higher for small firms and lower for large firms.
With IV, the point estimates have the reverse pattern, and both are lower than the
comparable estimate of 0.390 based on the full sample (table 3). None of these
differences are statistically significant.
Finally, we split the data at the median value of the BrainardTobin Q
variable to address how sensitive our estimation strategy is to investment dynamics.
Firms with high values of Q are presumably further from their longrun equilibrium
capital stock. Therefore, if our estimation method did not adequately account for
investment dynamics, we might expect a difference in the estimated's across the
highQ and lowQ subsamples. In table 4, the user cost elasticities are virtually
identical in the OLS results across the Q subsamples. The lowQ firms have a
modestly higher user cost elasticity than the highQ firms in the IV regression, but
the difference is not statistically significant. This result provides additional support
for the way our estimation method addresses the problems with complicated
investment dynamics, avoiding these difficult specification issues by focusing
directly on the longrun growth of the capital stock.
= ' " and is distributed asymptotic t under the null hypothesis that'= ". In the IV regressions, each individual instrument, zf, appears twice in the instrument list as follows, zf * If and zf * (1If).

24
E. Comparison To Alternative Approaches
Prior studies estimating (the user cost or substitution elasticity) can be set
into three categories. Most previous research has been based on timeseries data at
the aggregate or industry levels. Prominent examples of this work are the exchanges
between Hall and Jorgenson (1967, 1969, 1971) and Eisner and Nadiri (1968,
1970), Eisner (1969, 1970), and Coen (1969). Hall and Jorgenson's initial work
was based on a CobbDouglas production function, and hence equals 1.00 by
assumption. Eisner and Nadiri estimated freely, and reported that the
responsiveness of capital to its user cost was 0.16. This gap has not been closed by
subsequent research. Several important concerns, however, have been raised about
elasticities estimated from aggregate data suggesting that such estimates may be
biased downward due to problems with firm heterogeneity, simultaneity,
measurement error, and capital market frictions.
These issues were difficult to address with aggregate data because of the
limited amount of variation, and a more recent set of studies has exploited the
substantial information in panel data. While some of these concerns can be
addressed, these studies usually remove firm effects by differencing; thus, transitory
timeseries variation heavily influences the estimated user cost elasticity. A recent
example is Chirinko, Fazzari, and Meyer (1999), who find a user cost elasticity of
0.25 for a panel of firms. A similar elasticity is reported by Goolsbee (2000), who
analyzes a panel of equipment assets. Cummins and Hassett (1992) and Cummins,
Hassett, and Hubbard (1994, 1996) develop a novel approach, focusing on those
years in which there are sizeable tax policy changes to mitigate concerns about
endogeneity and measurement error. In these studies, crosssection variation is key.
Nonetheless, based on some auxiliary assumptions, the implied user cost elasticity
for U.S. firm data in Cummins, Hassett, and Hubbard (1994) is somewhat lower

25
than that obtained by Chirinko, Fazzari, and Meyer2.3 These studies use
investment data, and the biases associated with investment models mentioned above
may be important.
A third class of studies focuses on longrun relations between the capital
stock and its determinants. To mitigate the distorting effects of complex dynamics,
Caballero (1994) exploits the innovative idea that the user costs elasticity can be
estimated in a cointegrating equation that includes the capital/output ratio and the
user cost. Because cointegration is an asymptotic property, this estimate can be
biased downward in finite samples. Using aggregate quarterly data for equipment
spending and the StockWatson (1993) correction to adjust the estimates for the
effects of transitory variation, Caballero obtains a range of elasticity estimates, from
0.40 to 0.93, depending on the number of lags used in the correction. Also
exploiting cointegration properties, Mairesse, Hall, and Mulkay (1999) and Harhoff
and Ramb (2001) estimate error correction models (ECM) containing the longrun
relation between the capital stock and its determinants and the percentage changes
in these variables to capture shortrun dynamics. Firmlevel data and fixed effects
are used in both studies. The parameters are imprecisely estimated, a result that
may be due to estimating both longrun and shortrun parameters in the ECM with
transitory timeseries variation. Kiyotaki and West (1996) specify a model that
includes deviations of the desired from the actual capital stock, and estimate desired
capital in terms of a future projection from a twostep VAR procedure. With
quarterly aggregate data for Japan, they find that the shortrun and longrun user
cost elasticities are 0.05 and 0.07, respectively. The authors attribute these very
small responses to transitory variation in the user cost series as represented by a
pronounced tendency for mean reversion. Caballero, Engel, and Haltiwanger (1995)
estimate a model similar to Caballero (1994) with plantlevel equipment spending.
23 See Chirinko, Fazzari, and Meyer (1999, section 5) for further details.

26
They obtain a range of elasticities across twodigit industries from 0.01 to 2.00,
with an unweighted average of approximately unity. If we assume that the
structures elasticity is onethird as large as that for equipment (per the results of
Cummins and Hassett, 1992), then the overall user cost elasticity is approximately
0.70.
The elasticity estimates of Caballero, Engel, and Haltiwanger and those
presented in this paper are both based on a panel, but are not directly comparable
for a variety of reasons, including the use of plantlevel vs. firmlevel data, the
specification of the longrun determinants of the capital stock, and the manner in
which the problem of capital stock dynamics is addressed. The Caballero, Engel,
and Haltiwanger estimates are based on a cointegrating relation that emphasizes the
time dimension of the panel, and deviations from longrun values are accounted for
with the StockWatson correction. By contrast, our approach uses the time
dimension of panel data to measure longrun variables in each interval, and then
estimates the user cost elasticity from the crosssection dimension of the panel.
Given these differences, it is not surprising that we obtain different results.

27
V. Summary And Conclusions
The elasticity of business capital to its user cost has been the focus of much
research attention over the past 40 years. Among other issues, this parameter is
central in translating the effects of tax policy into real outcomes and assessing the
validity of alternative models of longrun growth. Prior work has relied in almost all
cases on timeseries variation in investment data at the aggregate, industry, or firm
level to estimate this elasticity. This paper offers a different approach. The
estimation strategy developed here classifies the time periods into three intervals,
and then averages the firmlevel panel data within each interval. The data are
differenced across intervals, and production function parameters are estimated in a
crosssection of timeaveraged, differenced firm data. Our approach accounts for a
variety of productivity shocks, omitted variables, and firm effects. This econometric
model does not solve the estimation problems inherent with investment models 
difficulttospecify dynamics, transitory timeseries variation, and positively sloped
supply schedules  that may bias estimates of the user cost elasticity. Instead, our
approach avoids these problems by exploiting panel data in an original way and
estimating directly the firstorder condition for capital.
We find that the user cost elasticity can be consistently and precisely
estimated by OLS, and is approximately 0.40. Relative to a comparable investment
study (Chirinko, Fazzari, and Meyer, 1999), the results here suggest that investment
models impart a discernible bias toward zero in estimates of the user cost elasticity.
To the central question of whether the CobbDouglas assumption is valid, our
results offer a strikingly negative answer. This robust finding raises questions about
the frequent use of the CobbDouglas production function in theoretical and
empirical models and about the costeffectiveness of various tax proposals for
stimulating capital formation.

28
Apart from our immediate objective, the method developed here may prove
useful in estimating other structural parameters from longpanel datasets. Our
approach, which uses interval averages to estimate longrun desired values of
regression variables, could be applied to other problems where shortrun dynamics
may obscure longrun structural relations. There are likely to be a number of
applications in, for example, labor and industrial organization, where the availability
of longpanels and interest in structural parameters may make this method feasible
and informative.

29
References
Abel, Andrew B., and, Eberly, Janice E., "The Effects Of Irrevesribility And Uncertainty On Capital Accumulation,"Journal Of Monetary Economics 44 (December 1999), 339378.
Acemoglu, Daron, "Labor And CapitalAugmenting Technical Change," M.I.T. (July 2001).
Arrow, Kenneth J., Chenery, Hollis B., Minhas, BagichaS., and Solow, Robert M., "CapitalLabor Substitution and Economic Efficiency,"Review Of Economics And Statistics 43 (1961), 225250; reprinted in Production and Capital: Collected Papers of Kenneth J. Arrow Vol. 5 (Cambridge:
Harvard University Press, 1985), 50103.
Ballard, Charles, L., Fullerton, Don, Shoven, John B., and Whalley, John,A General Equilibrium Model for Tax Policy Evaluation (Chicago: University of Chicago Press (for the NBER), 1985).
Barro, Robert J., and SalaiMartin, Xavier,Economic Growth (New York: McGrawHill, 1995).
Berndt, Ernst R., "Reconciling Alternative Estimates of the Elasticity of Substitution," The Review of Economics and Statistics 63 (February 1976), 5968.
Caballero, Ricardo J., "Small Sample Bias And Adjustment Costs," The Review Of Economics And Statistics 76 (February 1994), 5258.
Caballero, Ricardo J., "Aggregate Investment," in John B. Taylor and Michael Woodford (eds.), Handbook Of Macroeconomics, Volume 1B (Amsterdam: Elsevier NorthHolland), 1999), 813862.
Caballero, Ricardo J., Engel, Eduardo M.R.A., and Haltiwanger, John C., "PlantLevel Adjustment and Aggregate Investment Dynamics,"Brookings Papers on Economic Activity (1995:2), 154.
Chirinko, Robert S, "Business Fixed Investment: A Critical Survey of Modeling Strategies, Empirical Results, and Policy Implications,"
Journal of Economic Literature 31 (December 1993), 18751911.

30
Chirinko, Robert S., Fazzari, Steven M., and Meyer, Andrew P., "How
Responsive Is Business Capital Formation To Its User Cost?: An Exploration With
Micro Data," Journal of Public Economics 74 (October 1999), 5380.
Coen, Robert M., "Tax Policy and Investment Behavior: Comment," American Economic Review 59 (June 1969), 370379.
Cummins, Jason G. and Hassett, KevinA., "The Effects of Taxation on Investment: New Evidence From Firm Level Panel Data,"National Tax Journal 45 (September 1992), 243252.
Cummins, Jason G., Hassett, Kevin A., and Hubbard, R. Glenn, "A Reconsideration of Investment Behavior Using Tax Reforms as Natural Experiments," Brookings Papers on Economic Activity (1994:2), 160.
Cummins, Jason G., Hassett, Kevin A., and Hubbard, R. Glenn, "Tax Reforms and Investment: A CrossCountry Comparison,"Journal of Public Economics 62 (1996), 237273.
Eisner, Robert, "A Permanent Income Theory for Investment: Some Empirical Explorations," American Economic Review 57 (June 1967), 363390.
Eisner, Robert, "Tax Policy and Investment Behavior: Comment,"American Economic Review 59 (June 1969), 379388.
Eisner, Robert, "Tax Policy and Investment Behavior: Further Comment," American Economic Review 60 (September 1970), 746752.
Eisner, Robert, and Nadiri, M. Ishaq, "Investment Behavior and NeoClassical Theory," The Review of Economics and Statistics 50 (August 1968), 369382.
Eisner, Robert, and Nadiri, M. Ishaq, "Neoclassical Theory of Investment Behavior: A Comment," The Review of Economics and Statistics 52 (May 1970), 216222.
Engen, Eric, Gravelle, Jane, and Smetters, Kent, "Dynamic Tax Model:s Why They Do The Things They Do,"National Tax Journal 50 (September 1997), 657682.

31
Fazzari, Steven M., Hubbard, R. Glenn, and Petersen, Bruce C.,
"InvestmentCash Flow Sensitivities Are Useful: A Comment On Kaplan and
Zingales,"
Quarterly Journal of Economics 115 (May 2000), 695705.
Fox, Tyler, and Fullerton, Don, "The Irrelevance of Detail in a Computable General Equilibrium Model," Economic Letters 36 (March 1991), 6770.
Garber, Steven, and Klepper, Steven, "Extending the Classical Normal ErrorsinVariables Model," Econometrica 48 (September 1980), 15411546.
Goolsbee, Austan, "Investment Tax Incentives, Prices, and the Supply of Capital Goods," Quarterly Journal of Economics 93 (1998), 121148 .
Goolsbee, Austan, "The Importance of Measurement Erron in the Cost of Capital," National Tax Journal 53 (June 2000), 215228 .
Hall, Robert, E., "Comment On `PlantLevel Adjustment and Aggregate Investment Dynamics,'"Brookings Papers on Economic Activity (1995:2), 4751.
Hall, Robert E., and Jorgenson, Dale W., "Tax Policy and Investment Behavior," American Economic Review 57 (June 1967), 391414.
Hall, Robert E., and Jorgenson, Dale W., "Tax Policy and Investment Behavior: Reply and Further Results," American Economic Review 59 (June 1969), 388401.
Hall, Robert E., and Jorgenson, Dale W., "Application of the Theory of Optimum Capital Accumulation," in Gary Fromm (ed.), Tax Incentives and Capital Spending (Washington: Brookings Institution, 1971), 960.
Hamermesh, Daniel S., and Pfann,Gerard A., "Adjustment Costs In Factor Demand,"Journal Of Economic Literature 34 (September 1996), 12641292.
Harberger, Arnold C., "The Corporation Income Tax: An Empirical Assessment," in Tax Revision Compendium, Vol. 1 (Washington: House Committee on Ways and Means, 86th Congress, First Session, 1959), 231250.
Harberger, Arnold C., "The Incidence of the Corporation Income Tax,"
Journal of Political Economy 70 (June 1962), 215240.

32
Harhoff, Dietmar, and Ramb, Fred, "Investment And Taxation In Germany
 Evidence From FirmLevel Panel Data ," in Deutsche Bundesbank (ed.),Investing
Today For The World Of Tommorrow: Studies On The Investment Process In
Europe (Berlin: SpringerVerlag, 2001), 4773.
Hassett, Kevin A., and Hubbard, R. Glenn, "Tax Policy and Investment," in Alan J. Auerbach, ed.,Fiscal Policy: Lessons from Economic Research (Cambridge: MIT Press, 1997), 339385.
Hassett, Kevin A., and Hubbard, R. Glenn, "Are Investment Incentives Blunted By Changes In Prices Of Capital Goods?,"International Finance (October 1998), 103125.
Hubbard, R. Glenn, "CapitalMarket Imperfections and Investment,"Journal of Economic Literature 36 (March 1998), 193225.
Joint Committee on Taxation, Joint Committee On Taxation Tax Modeling Project And 1997 Tax Symposium Papers (Washington: U.S. Government Printing Office, 1997).
Jorgenson, Dale W., "Capital Theory and Investment Behavior,"American Economic Review 53 (May 1963), 247259.
Kaplan, Steven N., and Zingales, Luigi, "Do InvestmentCash Flow Sensitivities Provide Useful Measures Of Finance Constraints?," Quarterly Journal Of Economics 112 (February 1997), 169215.
Kaplan, Steven N., and Zingales, Luigi, "InvestmentCash Flow Sensitivities Are Not Valid Measures Of Financing Constraints,"Quarterly Journal Of Economics 115 (May 2000), 707712.
Kiyotaki, Nobuhiro and West, Kenneth D, "Business Fixed Investment and the Recent Business Cycle in Japan," in Ben S. Bernanke and Julio J. Rotemberg (eds.), NBER Macroeconomics Annual 1996 (Cambridge: MIT Press, 1996), 277344.
Klump, Rainer, and Preissler, Harald, "CES Production Functions and Economic Growth," Scandinavian Journal of Economics 102 (March 2000), 4156.

33
Lucas, Robert E., Jr., "LaborCapital Substitution in U.S. Manufacturing",
in Arnold C. Harberger (ed.),The Taxation of Income from Capital (Washington:
Brookings Institution, 1969), 223274.
Mairesse, Jacques, Hall, Bronwyn H., and Mulkay, Benoit, "FirmLevel Investment In France And The United States: An Exploration Of What We Have Learned In Twenty Years,"Annales d'Economie et Statistiques 5556 (1999), 2767.
Mankiw, N. Gregory, "The Growth Of Nations,"Brookings Papers On Economic Activity (1995:1), 275310.
Miyagiwa, Kaz and Papageorgiou, Chris, "Elasticity Of Substitution And Growth: Normalized CES In The Diamond Model,"Economic Theory (forthcoming).
Pitchford, John D., "Growth And The Elasticity Of Substitution,"Economic Record 36 (December 1960), 491 504.
Rao, Potluri, "Some Notes On The ErrorsInVariables Model,"The American Statistician 27 (December 1973), 217218.
Roeger, Werner, Veld, Jan in 't, and Woehrmann, Don I. Asoka, "Some Equity and Efficiency Considerations of International Tax Competition," European Commission (August 2000).
Salinger, Michael A., and Summers, Lawrence H., "Tax Reform and Corporate Investment: A Microeconomic
simulation study," in Martin S. Feldstein (ed.), Behavioral
simulation methods in Tax Policy Analysis (Chicago: University of Chicago Press (for the NBER), 1983), 247273.
Shapiro, Matthew D., "The Dynamic Demand For Labor and Capital," Quarterly Journal of Economics 101 (August 1986a), 513542.
Shapiro, Matthew D., "Investment, Output, and the Cost of Capital," Brookings Papers on Economic Activity (1986:1b), 111152.
Shea, John, "Instrument Relevance in Multivariate Linear Models: A Simple Measure,"Review of Economics and Statistics 79 (May 1997), 348352.

34
Stock, James H., and Watson, Mark W., "A Simple MLE of Cointegrating
Vectors in Higher Order Integrated Systems,"Econometrica 61 (July 1993),
783820.
Summers, Lawrence H., "Comment: 'Will 'The' Neoclassical Model of Investment Please Rise?: The General Structure of Investment Models and Their Implications for Tax Policy,'" in Jack M. Mintz and Douglas D. Purvis (eds.),The Impact of Taxation on Business Investment (Kingston, Ontario: John Deutsch Institute for the Study of Economic Policy, Queen's University, 1988), 171173.
Whelan, Karl, "Tax Incentives, Material Inputs, And The Supply CurveFor Capital Equipment," Board of Governors (May 1999).

35
Appendix A: Trending Variables
This appendix considers the effects of trending variables on the specification
of the model. We begin with the following decomposition for variable Wf,,t into nongrowth (ng) and growth (g) components, where W corresponds to any of the
model variables, K, Y, or C (note that, unlike in the text, we explicitly include an
index for the interval even when it is redundant),
(A1) Wf,,t Wngf,,t + Wgf,,t ,
(A2)
Wngf,,t µngf, + ngf,,t,
T
µngf, =
Wng f,,t / T,
t=1
(A3)
T ngf,,t / T = 0, t=1 Wgf,,t µngf, [(1+gf)t 1].
In (A2), the nongrowth component equals the mean over the interval (µngf,) and a deviation from the mean value (ngf,,t) that averages to zero. These summations are over all T time periods that are in the interval. In (A3), the growth component is proportional to the mean, and increases at a firmspecific rate of gf. As in Section II, we measure the longrun value of W (W*f,t) as a timeaverage over a interval (Wf,),

36
(A4) W*f,t = Wf,
T = Wf,,t / T , t=1
T
=
Wngf,,t / T +
t=1
T Wgf,,t / T, t=1
= µngf, +
T ngf,,t / T + t=1
T µngf, [(1+gf)t 1] / T, t=1
T = µngf, + 0 + µngf, [(1+gf)t 1] / T, t=1
= µngf, * H[gf], T H[gf] 1 + [(1+gf)t 1] / T, t=1
Our estimator (7) uses the difference between the=2 and =1 intervals in the logarithms of Wf,, (A5) Ln{Wf,=2}  Ln{Wf,=1} = Ln{µngf,=2} + Ln{H[gf]}  Ln{µngf,=1}  Ln{H[gf]}, = Ln{µngf,=2 / µngf,=1}, which is the percentage change in the nongrowth component of Wf,. Thus, the variables entering the estimating equation are not distorted by firmspecific growth.

37
Appendix B: The Replacement Value of Capital
The capital stock is a key variable in this study, and this appendix provides
details about how we overcome several significant problems in measuring the
capital stock from accounting data. The obvious proxies for the capital stock in the
Compustat data, book values of gross or net property, plant, and equipment, are not
acceptable measures of the economic value of the capital stock for two reasons.
First, they value assets at the historical cost prevailing when the assets were
acquired and therefore contain a mix of historical price levels that cannot be easily
adjusted for inflation. Second, accounting depreciation rules likely do not capture
economic depreciation correctly. The iterative "perpetual inventory" algorithm
described here addresses these problems.
The first step in our procedure is to choose a seed value for the iteration. We
use the nominal book value of net property, plant, and equipment for firm f from its
first observation in the data set (NPLANTf,0). To convert this value to real terms we
employ data on the share of different kinds of capital assets (indexed by j) in the
firm's twodigit SIC industry i. Denote this share asi,j. The amount of capital
(i,j NPLANTf,0) should be deflated by the assetspecific price index pj,0. Then the
real seed value of the capital stock (Kf,0) is defined as:
(B1)
K
f
,0
=
j
i,
j NPLANT f p j,0
,0
.
Starting from this seed value, the remainder of the capital stock for firm
k is constructed iteratively from:
(B2)
K f ,t + 1 = K f ,t i, j (1  j
j ) +
KCHG
f
,t
j
i,
j
(1  p j,t
j
)
.
The first term in equation (B2) is the depreciated value of the period t capital stock
that remains in period t+1. The depreciation ratej for each asset j is determined by
DRI from the "double declining balance" formula:

38
(B3)
j = 1  e 2 / LIFE j ,
where LIFEj represents the estimated average service life for capital asset j. The second term in equation (B2) represents the addition (or deletion) to the period t+1
capital stock accounted for by new investment, acquisitions, or divestitures in period
t. The variable KCHGf,t (discussed in detail below) represents the nominal addition (or subtraction) of new capital goods for firm f in period t prices. The deflation
method for KCHGf,t is the same as for the seed value in equation (B1). We assume
that new capital is acquired at the beginning of period t and depreciates one full year
before entering the period t+1 capital stock. (We also constructed capital stock
series using a half year's depreciation for KCHG and found that it had only a trivial
impact on the results.) If a firm adds to its capital stock in period t only through
conventional capital spending, the KCHGf,t variable in equation (B2) would equal the firm's investment (If,t), that we obtain from Compustat's capital expenditure data in the sources and uses of funds statement. In practice, acquisitions and divestitures
can augment and deplete the capital stock independent of reported investment.
Many panel studies delete firms with substantial acquisitions or divestitures.
However, there are a large number of observations with acquisitions and divestitures
in the Compustat data. Deleting these observations reduces the sample size and
could induce a selection bias. We therefore develop a method to account for
acquisitions and divestitures when constructing the capital stock data. (To the
extent that acquisitions or divestitures create outliers in the data, these should be
captured by our outlier detection algorithm described in Section III.)
The capital change variable (KCHGf,t) in equation (B2) is defined in a way that accounts for large acquisitions and divestitures. We appeal to the following
accounting identities to derive a formula for KCHGf,t,:
(B4)
GPLANT f ,t = I f ,t + ACQUIS f ,t  RETIRE f ,t

39
(B5)
NPLANT f ,t = I f ,t + ACQUIS f ,t  DEPR f ,t
GPLANT f ,t = NPLANT f ,t = ACQUIS f ,t = RETIRE f ,t = DEPR f ,t =
the change in gross plant and equipment from the end of year t1 to the end of year t; the change in net plant and equipment from the end of year t1 to the end of year t; acquisitions in year t; retirements in year t,24 and accounting depreciation in year t.
In the event of an acquisition, KCHGf,t equals If,t + ACQUISf,t. Because Compustat
does not have reliable figures for ACQUISf,t, we rearrange equation (B4) to obtain:
(B6)
I f ,t + ACQUIS f ,t = GPLANT f ,t + RETIRE f ,t or
KCHG f ,t = GPLANT f ,t + RETIRE f ,t
In the event of a divestiture, we want to decrease the capital stock by the
depreciated value of the capital sold. In this case:
(B7)
KCHG f ,t = NPLANT f ,t
If there is no major acquisition or divestiture, then we retain the basic formula:
(B8)
KCHG f ,t = I f ,t
We now need an empirical test to determine whether afirm has undergone an
acquisition or divestiture in a given year. There are two
rules of thumb that aid us in
this search. First, GPLANTf,t is normally less than If,t because of retirements. Therefore, if GPLANTf,t > If,t by a substantial amount, it signals an acquisition with a high probability. Second,GPLANTf,t is normally greater than
24 Compustat defines retirements as "a deduction from a company's property, plant, and equipment account resulting from the retirement of obsolete or damaged goods and/or
Physical structures."

40
RETIREf,t because retirements are the only way to reduce gross plant and
equipment in the absence of a divestiture. Therefore, ifGPLANTf,t < RETIREf,t
by a substantial amount it signals a divestiture.
We define a "substantial" amount as a discrepancy of ten percent or more.
The point of imposing the ten percent limit is to make acquisition and divestiture
adjustments conservative. That is, we only deviate from the standard formula when
there is clear evidence that this formula is misleading. In this case, if
(B9)
GPLANT f ,t  I f ,t > 0.1, GPLANT f ,t 1
then we assume an acquisition and define KCHGf,t from equation (B6). In contrast,
if
(B10)
GPLANT f ,t + RETIRE f ,t <  0.1, GPLANT f ,t 1
then we assume a divestiture and define KCHGf,t from equation (B7). If neither
rule holds, we simply define KCHGf,t as investment spending, as in equation (B8).

41
Appendix C: ShortRun Frictions And Measurement Error
This appendix relaxes the assumption that the longrun values of capital and
output appearing in the model of Section II can be measured without error as time
averages over a interval  t f,,t/T Kf, = K*f,t and t Yf,,t/T Yf, = Y*f,t for
t=1,T. (Note that, unlike in the text but as in Appendix A, we explicitly include an
index for the interval even when it is redundant.) In particular, these equalities
might be disrupted if shortrun frictions affect longrun values. For example, a firm
facing irreversibility constraints will exhibit a reluctance to invest that lowers its
optimal capital stock. In this case, a negative, firmspecific measurement error
drives a wedge between Kf, and K*f,t and, through the production function, a wedge
between Yf, and Y*f,t. However, the same set of constraints create a "hangover effect," as a firm occasionally finds itself with capital that it would like to dispose of
but can not due to irreversibility constraints. Thus, the longrun capital stock for a
firm facing irreversibility frictions can be greater than or less than the longrun
capital stock for a frictionless firm (Abel and Eberly, 1999). An additional friction
is introduced by asymmetric adjustment costs implying that the average values of
capital and output will differ in general from their longrun values.
These considerations introduce measurement error (MEKf,) that drives a
wedge between Kf, and K*f,t (a similar analysis applies to Yf, and Y*f,t),
(C1)
Kf, = K*f,t + f, .
We assume that the measurement error between Kf, and K*f,t increases with the size of the "true" longrun capital stock and that MEKf, = f, K*f,t. Furthermore, f, consists of two components, mf and mf,, representing measurement error that varies by firm and by firm and interval, respectively. Combined with (C1), these assumptions generate the following model of measurement error,

42
(C2a) (C2b)
Kf, = K*f,t f, , f, = 1 + mf + mf,,
To understand the impact of measurement error on our estimating equation, we analyze (C2) in conjunction with the trending variables analyzed in Appendix A. We begin with K*f,t = Kf, / f,, and thus divide the rightside of (A4) by f,. Since f, is indexed by, it passes through the summation sign, and the version of (A4) accounting for measurement error and trending variables is as follows,
(C3)
K*f,t = µngf, * H[gf] / f, ,
where µngf, is the mean of the nongrowth component of f,,t and H[gf] is a function of the firmspecific growth rate. See Appendix A for details. Our estimator replaces the logarithm of K*f,t in (4) with the logarithm of the expression in (C3), and differences the resulting equation between the=2 and =1 intervals,
(C4) Ln{Kf,=2 / f,=2}  Ln{Kf,=1 / f,=1} = Ln{µngf,=2} + Ln{H[gf]}  Ln{µngf,=1}  Ln{H[gf]} + (mf + mf,=2)  (mf + mf,=1), = Ln{µngf,=2 / µngf,=1}  mf,, which is identical to (A5) in Appendix A with the addition of mf,. Note the firmspecific component of the measurement error (mf) cancels in the first difference.

43
The additional term, mf,, can be incorporated straightforwardly into the
econometric model. The error term in (7) is composed of a productivity shock,
wf,,that is firmspecific and intervalspecific and, to reflect the effects of
measurement error, it can be replaced by the composite error, ( wf, + mf,). Note
that the two components both vary by firm and interval. We expect the impact of
mf, to be relatively small because firm and industry effects have already been removed.
In the event that there is some correlation betweenmf, and the regressors
induced by measurement error, the estimated will be attenuated, that is, closer to
zero than the "true". Since the measurement error considered in this Appendix
does not affect the user cost variable directly, the effect on is indirect, and equals
the product of the attenuation of the coefficient on output and a coefficient
representing the correlation between "true" sales growth and "true" user cost
growth (Garber and Klepper, 1980, equation (2.6)). The quantitative impact of this
type of measurement error is assessed in Section IV.B (the Rao test). Furthermore,
since this type of measurement error applies only to capital and output, it will not
impact the estimate of in the regression in the third column of Table 2 ( = 1)
with the growth rate of the capitaloutput ratio as the dependent variable. This
estimate of and the Rao test both suggest that the type of measurement error
considered in this Appendix is quantitatively unimportant.
Figure 1: Defining the Intervals
Wf,t
Wf, = 0
Wf, = 1
Wf, = 2
t
· 1974
=0
·
·
·
1980/1981
1986/1987
1992
=1
=2
1980
Wf, = 0 =
Wf,t / 7
t = 1974
1986
Wf, = 1 =
Wf,t / 6
t = 1981
1992
Wf, = 2 =
Wf,t / 6
t = 1987
Table 1 Summary Statistics
1. Percentage Change Between Interval = 1 and = 2
Statistic Mean Median Standard Deviation Coefficient of Variation
Capital 36.4 16.9 81.9 2.3
Output 27.5 14.6 66.2 2.4
2. Levels in Interval = 0 Statistic Mean Median Standard Deviation Coefficient of Variation
Capital 320.8 33.1 848.7 2.6
Output 948.1 161.7 2562.0 2.7
3. Levels in Interval = 1 Statistic Mean Median Standard Deviation Coefficient of Variation
Capital 434.2 50.7 1141.3 2.6
Output 1169.5 211.9 3180.5 2.7
User Cost 6.9 11.8 18.7 2.7 User Cost 0.282 0.291 0.056 0.2 User Cost 0.242 0.246 0.046 0.2
4. Levels in Interval = 2 Statistic Mean Median Standard Deviation Coefficient of Variation
Capital 529.2 62.3 1410.1 2.7
Output 1404.2 253.7 4237.4 3.0
User Cost 0.219 0.218 0.028 0.1
Note: The statistics are derived from a sample of 1,860 firms constructed from Compustat and DRI sources as d of the text. The standard deviations represent crosssectional differences arising from firm heterogeneity in perce the =1 and =2 intervals (panel 1) and from firm heterogeneity in levels within an interval (panels 2, 3, and 4). capital and output are in millions of 1987 dollars
Table 2 Ordinary Least Squares Estimates
Unconstrained Regressions
Benchmark Model
Model with TwoDigit SIC Dummies
Constrained Regressions
=1
= 1 and = 1
0.367 (0.067)
0.440 (0.293)
0.372 (0.067)
1.0
1.135 (0.042)
1.152 (0.102)
1.0
1.0
0.925 (0.019)
0.926 (0.019)
1.0
1.0
0.084 (0.014)
R2
0.564
0.055 (0.114) 0.593
0.063 (0.013) 0.560
0.020 (0.013) 0.540
Note: Estimates of equation (7) with firmlevel panel data as described in section III.. Standard errors appear in parameters are (the user cost elasticity), (the returns to scale elasticity), (the regression coefficient on outpu intercept). See section IV.A for the formula used to compute and its standard error.
Table 3 Instrumental Variables Estimates
Unconstrained Regressions
Benchmark Model
Model with TwoDigit SIC Dummies
Constrained Regressions
=1
= 1 and = 1
0.390 (0.108)
0.373 (0.286)
0.434 (0.093)
1.0
0.603 (0.074)
0.633 (0.120)
1.0
1.0
1.402 (0.110)
1.364 (0.118)
1.0
1.0
0.049 (0.034)
0.324 (0.134)
0.059 (0.014)
0.020 (0.013)
Note: Estimates of equation (7) with firmlevel panel data as described in section III. Standard errors appear in p parameters are (the user cost elasticity), (the returns to scale elasticity), (the regression coefficient on outpu intercept). The instrument list is defined in section IV.C. In the second column, the industry dummies are instrum themselves. See section IV.A for the formula used to compute and its standard error.
Table 4 Ordinary Least Squares And Instrumental Variable Estimates: Various Sample Splits
Split by Cash FlowCapital Ratio
OLS Low High CF/K CF/K
IV Low High CF/Kl CF/K
Split by Capital Stock Size
OLS
IV
Low High Low High
Capital Capital Capital Capital
Split by Tobin
OLS
Low High
Q
Q
0.278 0.407 0.364 0.317 0.435 0.294 0.226 0.363 0.320 0.349 (0.075) (0.127) (0.102) (0.198) (0.139) (0.066) (0.226) (0.094) (0.076) (0.114)
1.308 1.019 0.989 0.673 1.042 1.284 0.646 0.782 1.214 1.046 (0.071) (0.049) (0.197) (0.128) (0.056) (0.064) (0.104) (0.120) (0.061) (0.054)
0.830 0.989 1.007 1.331 0.977 0.844 1.424 1.177 0.880 0.972 (0.025) (0.028) (0.128) (0.178) (0.029) (0.024) (0.172) (0.120) (0.025) (0.032)
0.040 0.125 0.000 0.014 0.105 0.064 0.016 0.019 0.057 0.107 (0.017) (0.024) (0.031) (0.066) (0.026) (0.016) (0.058) (0.033) (0.016) (0.027)
0.129 (0.144)
R2 0.541 0.575
0.047 (0.214)
0.141 (0.144) 0.556 0.582
0.137 (0.226)
0.029 (0.137) 0.631 0.562
Note: Estimates of equation (7) with firmlevel panel data as described in section III. Standard errors appear in p parameters are (the user cost elasticity), (the returns to scale elasticity), (the regression coefficient on outpu intercept). The instrument list is defined in section IV.C. See section IV.A for the formula used to compute Sample splits are based on the median value of the classifying variable in the =0 (19741980) interval. is the the difference between the 's for the contrasting classes, and is distributed asymptotic t under the null hypothes section IV.D for further details about this statistic.