Specifically, the production function is assumed to be quasi-concave and monotonic (Varian 1992). Furthermore, specific functional forms, such as the Cobb-Douglas production function, entail additional restrictions.
 It bears pointing out that total benefits of IT spending can still be large even if marginal benefits are zero or negative. In fact, a high marginal product may be a sign of underinvestment.
All future references to profitability will be in reference to equilibrium (long run) profitability.
 In particular, see Berndt (1991) for an excellent discussion of simultaneity in supply-demand systems and Gurbaxani and Mendelson (1990) on the role of technology diffusion.
The pursuit of value need not be zero sum. Some types of competitive tactics, such as raising rivals' costs (Salop and Scheffman 1983), actually lead to a loss of total value, even though these activities may be privately beneficial.
In fact, to facilitate a cumulative research tradition, the complete data set is available from the authors on request, along with a data appendix describing the variables in more detail.
This sample size refers to a complete set of productivity and profitability measures for which there is more than one observation over the five years in a 2-digit SIC industry. The sample may decrease further for analyses that require additional variables, but the base case analyses for productivity and profitability are conducted on exactly the same data points.
Interestingly, the average life age of 642 systems reported in Swanson and Beath (1989) is 6.6 years, which is quite close to our 7-year upper bound.
 While this approach is not the only method used for conducting productivity analyses, it is by far the most common in the context of calculating the elasticities and marginal products of inputs. The Cobb-Douglas equation has the virtues of simplicity and empirical validity. Variations of this approach include the use of more complex function forms such as the translog or using different estimation methods such as stochastic production frontiers or data envelopment analysis. These alternatives can be employed to address other issues related to the productivity question such as substitution between inputs or relative efficiencies of various firms (see e.g. (Banker and Maindratta 1988; Caves and Barton 1990; Lee and Barua 1994))
 For example, the instruments for the 1992 data points would be the 1991 values of IT Capital, Non-IT Capital and Labor Expenses, along with the sector and time dummy variables. These are far from perfect instruments.
 The marginal product is equal to the elasticity divided by the percentage of IT in Value-Added which is .0930. Therefore, the gross marginal benefit is: .0883/.0930 = 94.9%.
Note that because labor expense is an annual cost, a zero net return would require a gross return of 1.0. That is why the returns to labor appear much different than the returns to capital which represent an accumulation of spending over time and have an inflation adjusted cost on the order of 5-10%.
 This appears to be substantially higher than a "reasonable" estimate of the capital cost based on the Jorgensonian cost of capital (Christensen and Jorgenson 1969). The Jorgensonian cost is a function of the risk free rate, a risk premium, depreciation charges, and capital gains or losses. Following Hall (1993b) we use 6% as the risk free rate and assign a risk premium of 3%. The Bureau of Economic Analysis (1993) assumes computers depreciate over a period of 7 years while IS Labor stock depreciates over 3 years by assumption. Weighting the two over the sample average composition gives an average life of 4.5 years or a depreciation rate of 22% per year. Finally, holders of IT Stock face capital losses of approximately 4% per year because the quality-adjusted costs of new computers (and therefore the value of old computers) declines at 19% per year (Gordon 1987) and the costs of IS Labor increase by 4% per year. Accounting for the above factors yields a total cost of capital of 35% per year. However, it should be noted that other factors, such as taxes, the benefits of learning, the options value of investments and unmeasured costs and benefits can substantially affect the true costs of capital, although they are difficult to quantify.
 The actual choice of denominator for the IT measure does not affect the results substantially. However, there is a possible negative bias when sales is used as the denominator: irrespective of the contribution of IT, an unexpectedly good sales figure will increase profits but lower the IT ratio, inducing a negative coefficient. The use of other inputs such as employees or capital avoids this bias. We chose employees to most closely replicate Strassmann's (1990) analyses, which do not use sales as the denominator.
 In our empirical analysis the actual distinction is somewhat more complex because neither prices nor industry boundaries are measured exactly. If industry definitions were exact, the dummy variables included in our performance regression should control for differences in the rate of price change across industries and eliminate empirically the negative effect on profitability. However, because in practice the effective competitors of the firms in our sample do not map perfectly on to the 2-digit SIC code definitions we used and the industry dummy variables do not account for differences in price changes within industries over time, some of the residual price changes will still be reflected in our estimates.
 In principle, the dummy variables we included for each industry in the basic profitability regressions should have partially controlled for the effect of industry-wide IT spending on profits. As mentioned in an earlier footnote, imperfections in the match between true competitors and 2-digit SIC codes, and the fact that the dummy variables are invariant over time lead to less control for spurious industry effects. As a result, to the extent that a firm's IT budget is correlated with its competitors' spending, the coefficient on IT will in part reflect the indirect effects of higher overall industry spending on IT.
 By contrast, an R2 of 95% or more has been achieved for both production function analyses and consumer surplus analyses (e.g. (Brynjolfsson 1993b; Brynjolfsson and Hitt 1993)).