Measuring consumer welfare in the CPU market: an application of the pure-characteristics demand model.

In this article, I estimate demand for the personal computer central processing unit and measure consumer welfare using the pure characteristics demand model. The model is based on a quasilinear utility function with multiplicative random variables and does not have the idiosyncratic logit error term, so that consumer welfare directly reflects consumers' valuation of product characteristics. Welfare calculations show that consumer surplus comprises approximately 90% of total social surplus and that large welfare gains have resulted from the introduction of new products.

1. Introduction

The past decade has seen a dramatic improvement in performance of the personal computer central processing unit (CPU). (1) Rapid advances in manufacturing technology have allowed a new generation of products to emerge every two to three years. The best known of the CPUs, the Intel-made Pentium, was introduced in 1993 with 60-megahertz (MHz) processing speed. Five new generations of processors have entered the market since then, with the Pentium III processor reaching 1-gigahertz (GHz) processing speed in 2000.

Although performance continued to improve over the 1990s, the average price of CPUs fell. The first Pentium processor was introduced at $878 but the Pentium II processor, faster than the Pentium processor by a factor of four, was marketed at $636 four years later. Figure 1 shows both the maximum and the quantity-weighted average prices of CPUs with the maximum processor speed (in MHz) from 1993 to 2000. It is clear that, despite frequent introductions of new products with higher speed, the average price shows a downward trend. The figure also shows that the maximum price does not necessarily increase as the maximum processing speed increases.

While consumers benefit from the drastic improvement in product quality, accompanied by price decline, firms incur a considerable cost from a high product turnover. Because the first mover gains transitory market power, firms are engaged in an "innovation race," which requires a few billion dollars for R&D activities and capital expenditures every quarter. (2) The Semiconductor Industry Association reports that R&D expenditures have grown at an annual rate of 15% over the last decade and that the ratio of R&D spending to sales reached 13.8% in 1999, which surpasses other high-technology industries. (3) Capital expenditures have also grown at an annual rate of over 15% over the same period, and the ratio of capital expenditures to sales fluctuates between 15% and 20%. (4)

[FIGURE 1 OMITTED]

Do benefits to consumers exceed the cost of innovation and firms' profits? In this article, I attempt to answer this question by estimating consumer welfare using my product-level data from two major firms in the CPU market: Intel and Advanced Micro Devices (AMD). My dataset provides quarterly prices and quantities sold from 1993 to 2000. Welfare gains from innovation have been studied in other industries, including Hausman (1999), who calculates gains in consumer welfare from the introduction of the cell phone; Petrin (2002), who estimates welfare gains from the development of the Minivan in the auto industry, and Nevo (2003), who estimates consumer benefit from new brand introductions in the ready-to-eat cereal market.

However, I use a new model of consumer demand, the so-called pure-characteristics demand model, to estimate consumer demand. The model is based on a quasilinear utility function with multiplicative random coefficients, but it does not have an additive idiosyncratic error term such as a logit error. Most of the recent empirical industrial organization literature relies on the conditional multinomial logit demand model in which the utility function has an (additive) logit error term to estimate demand and measure consumer welfare (McFadden, 1981; Berry, Levinsohn, and Pakes, 1995; Nevo, 2001; Petrin, 2002.) The logit error term facilitates demand estimation by insuring that all the purchase probabilities are nonzero at every value of the parameter vector and that the market-share equation has smooth derivatives.

The conditional multinomial logit model, however, imposes a restriction that consumers have idiosyncratic tastes for products that are independent of product characteristics. It is well known that this restriction can be problematic in measuring consumer welfare, especially when new products are introduced. For example, Petrin (2002) shows that, in a demand model with the idiosyncratic error term, a large part of welfare changes can come from the error term. Unless he interacts this error term with consumer demographic data, the model implies that consumers strongly dislike the observed characteristics of their choice relative to their alternative choice.

The pure-characteristics model provides an alternative way of estimating demand and measuring consumer welfare by eliminating the logit error term from the utility function. Because the model does not have any non-characteristics-related taste terms, consumer preferences are only related to product characteristics and consumer welfare directly reflects consumers' valuation of product characteristics.

Assuming that all relevant characteristics are observable, whether tastes play a crucial role in consumption independent of observed characteristics depends on the market. For example, in markets for paintings or fashionable clothes, consumers' tastes should matter more than in markets for computers or digital cameras. If these tastes are important in consumer preferences, the pure-characteristics model could be too "stringent" to reflect consumer heterogeneity. Nevertheless, it is reasonable to assume that, in the CPU market, consumers only care about product characteristics like the processing speed and the memory capacity inside the CPU (the level-2 cache) and that they do not have idiosyncratic tastes for products, which makes this market an appropriate place to apply the pure characteristics model.

The simplest case of pure-characteristics demand is the vertical demand model in which a utility function has one multiplicative random coefficient. Bresnahan (1987) used this model to analyze the American auto industry, and it has been widely used. However, in the vertical model, consumers agree on the product-quality order, which must coincide with price ranking, and only products in the adjacent "neighborhood" are substitutes for each other. These features confine the model to vertically differentiated products. In contrast, in the conditional multinomial logit demand model, all products are substitutes for one another.

The general pure-characteristics model, in which the utility function has multiple random coefficients, allows a more flexible substitution pattern, as consumers do not necessarily agree on the ranking of products. At the same time, the absence of the logit error term still guarantees that consumer welfare directly reflects consumers' valuation of product characteristics. Nevertheless, there has been no application of this model because an estimation strategy has not been available. Berry and Pakes (forthcoming) provide an estimation strategy, and my article is the first to estimate the model with real data.

My results show that consumer welfare surpasses producer surplus, measured by the variable profit, as well as firms' R&D spending. Consumer welfare is calculated in two ways in order to deal with changes in the value of the outside option over time. I first assume that the value of the outside option does not change over time, and all welfare changes come from products in the sample. My second method identifies the value of the outside option from the time dummy variables. Although there are considerable differences in the level of consumer welfare, both approaches indicate that consumer welfare increases when the rate of innovation goes up and that consumers gain far more than firms.

Much of the welfare gains come from the introduction of new products. A large part of consumer welfare comes from products at high-end, often new, products. Further, consumers who buy new products receive most of the surplus.

I also calculate consumer welfare using the conditional multinomial logit demand model. Consumer welfare is about 13 times higher on average. When the same demand estimates are used as in the pure characteristics model, it is 10 times higher on average. This suggests that one should pay close attention to the role of the logit error term in welfare calculations and be cautious when choosing a model.

The rest of the article is organized as follows. Section 2 provides an industry description with details of the dataset. Section 3 describes the demand model, followed by an outline of the estimation strategy in Section 4. Section 5 reports estimation results and Section 6 measures of consumer welfare. Section 7 examines effects of different market sizes on demand estimates and consumer welfare. Section 8 presents welfare results using the conditional multinomial logit demand. Section 9 concludes and discusses extensions.

2. Industry and data