Process Versus Product Innovation: Do Consumption Data Contain Any Information?

• Author: Thompson, Peter

Peter Thompson [*]

Doug Waldo [+]

In a model of endogenous growth in which product and process innovations are the joint outputs of an unspecified research program, we show that if quality growth is not captured by official price indices the usual isomorphism of product and process innovations breaks down. We derive and estimate a Euler equation for a representative consumer under the assumption of measurement error. Unobserved quality improvements account for at least half of growth, and real productivity growth in postwar United States was two to five times greater than measured total factor productivity (TFP) growth. We also find that at least 15% of the measured slowdown in TFP growth can be attributed to unobserved increases in the relative importance of product innovations.

1. Introduction

   Models of quality growth in consumer goods have become a prominent feature of growth theory in recent years. However, to date there have been relatively few attempts to estimate these models and little is known about the magnitude of quality growth. It is, of course, easy to explain the slow progress in empirical work: Analysis of the rate of quality growth is clouded by measurement problems. One component of technological advance, process innovation, results in reduced production costs and is readily measured by existing techniques of data collection. The other component, product innovation, results in new or improved products and is only partially captured by existing techniques. As Zvi Griliches (1994) stated in his presidential address to the American Economic Association, "Quality change is the bane of price and output measurement." Yet quality change is probably a large component of technological advance, and its absence from official statistics raises potentially serious problems in productivity measu rement.

   The impact of the BEA's hedonic regressions on the computer price index has become a cause c[acute{e}]l[grave{e}]bre of the measurement problem. Prior to 1986, the Bureau of Economic Analysis (BEA) had measured prices per computer with no adjustment for quality changes and the price index remained more or less constant from 1972 to 1986. In 1986, when the BEA changed its index to one that essentially measures the cost of computer calculations and backdated its results to 1972, the new price index showed an average annual decline of 14%. Raff and Trajtenberg (1995) have recently shown that the rate of change in quality-adjusted prices during the early years of the automobile industry was similar in magnitude.

   Numerous additional examples, albeit less dramatic, are reported in Baily and Gordon's (1988) exhaustive study of the measurement problem. For example, the BEA has assumed no quality improvement in construction since 1929; measured productivity in the banking industry has been almost constant since 1948, despite an enormous number of product innovations; and measured airline productivity growth since 1972 has fallen by 0.2% per year, although passenger miles per employee rose at an annual rate of 3.6% over the same period. Hausman (1994) has shown that similar problems of measurement arise if one ignores the introduction of new varieties. He measured the effect that new brands of cereal have on consumer surplus and suggested that the price index for cereals is overestimated by about 25% as a result of ignoring a decade's worth of new brands. This magnitude of error is remarkable in an industry that does not strike the casual observer as particularly innovative, and it suggests that annual productivity growth in cereals is underestimated by some 2% per year.

   The solution to the measurement problem is, of course, to adjust price indexes for quality changes measured at the product level. But we cannot expect that official statistics will adopt hedonic analyses any time soon, as hedonic regressions raise considerable methodological problems and are extremely costly. [1] Not surprisingly, the fraction of products for which estimates of measurement errors have been made remains very small. [2] For now then, we continue to look at real productivity growth through an empirical fog.

   In an attempt to dissipate some of this fog, a number of researchers have tried to infer the importance of quality growth from aggregate data. Some of these studies [3] have inferred rates of quality growth from estimated parameters in structural Schumpeterian models. Other studies have employed data on patents, trademarks, and research and development (R and D) inputs. [4] These studies have, without exception, tackled the measurement problem from the perspective of the producer. However, although researchers have noted and exploited the fact that producer behavior can provide some insight into the importance of unmeasured quality growth, it has not generally been realized that measurement errors have similarly observable implications for consumer behavior. In view of the considerable difficulties that have arisen with the construction of estimates of quality growth from production data, it seems sensible to inquire whether supporting evidence can be obtained from consumption data.

   In this paper, we therefore investigate whether we can extract from consumption data information on the rate of unmeasured quality growth. One might not expect this to be possible; generally, in the new growth models quality and productivity growth are isomorphic. However, if measurement error is introduced into such a model, this isomorphism is likely to break down, and there will be observable differences between the effects of product and process innovations. The breakdown of isomorphism might let us extract information from observable aggregate data about the rate of quality growth.

   To evaluate our conjecture, we construct a model of growth, in which economic advance is secured by discovery of better ways to produce existing goods, as well as by discovery of better goods. In our model, process and product innovations are the joint outputs of R and D activities. This results in an economy in which the mix of technological innovations between process and product innovations does not have any effect on welfare or the real aspects of equilibrium. However, the mix will affect measured real income growth under the assumption that process innovations are measurable while product innovations are not. In this setting, technological progress has empirically verifiable implications for aggregate consumption. We derive a Euler equation for the consumer's welfare-maximization problem in the presence of uneven technological growth. The resulting equation is an extension of the well-known Euler equation for iso-elastic utility, to which Hansen and Singleton (1982) first applied the generalized method of moments (GMM). When nominal variables are deflated by a price index that omits quality changes, the Euler equation includes a parameter measuring the fraction of growth caused by quality changes. In principle, this parameter can be estimated.

   Using postwar data for the United States, we conclude that as much as three-quarters of the real rate of productivity growth is accounted for by unmeasured product innovations. As measured total factor productivity (TFP) growth over the sample period averaged 1% per annum, our results yield an estimate for the rate of quality growth of up to 4% per annum, and hence a real rate of growth of up to 5% per annum. These figures are in concordance with other studies that used very different techniques. Applying patent count data to a quality ladders model of growth, Caballero and Jaffe (1993) report a growth rate of some 8% over the same period; Thompson (1996) relates firm-level stock market values to R and D expenditures and obtains a growth rate of 5% over the period 1973-1991. Most recently, and using a methodology closest in spirit to ours, Arroyo and Dinopoulos (1996) have used aggregate U.S. data to obtain a range of values for quality growth between 2.3 and 5.3% during postwar years.

   We would not suggest that estimates of measurement error obtained from aggregate data are a substitute for the careful empirical work that must be undertaken at the disaggregated level. But we do argue that as long as microlevel studies remain rare, aggregate studies do have practical applications. We illustrate their utility by asking whether our approach can shed light on the causes of the productivity slowdown. The basis of our investigation of the slowdown is the widely held suspicion that at least part of the slowdown is an illusion: Output expansion (i.e., measured GNP growth) may have decelerated, but the slack may have been taken up by increases in product innovation.

   Of course, mismeasurement, or nonmeasurement, of the rate of product innovation cannot by itself explain the slowdown. There has always been product innovation, and it has never been adequately measured. To construct a case for attributing at least part of the slowdown to measurement error, one must also show that measurement problems became more pronounced during the period 1966-1973. We separately estimate the share of product innovations in aggregate growth before and after the dates of the onset of the productivity slowdown. We conclude that quality's share of growth did indeed increase; with an elasticity of intertemporal substitution of 0.5, quality's share of growth is estimated to have risen from about 0.75 to 0.9. These estimates, in conjunction with the sample means of TFP growth pre- and postslowdown, let us infer the fraction of the productivity slowdown that can be attributed to measurement error. Our estimates suggest that 0.3 percentage points of the vanished growth can be accounted for by mea surement error. Put another way, about 15% of the productivity slowdown is attributable to measurement error. Our finding is consistent with conclusions drawn by Baily and Gordon (1988) from their detailed study of measurement problems.

   The layout of this...