We specify and estimate a dynamic game to study the equilibrium relationship between market structure and innovation in the automobile industry. The quality of each firm’s product for the average consumer, the key state variable, is modeled as stochastically increasing in innovation, the dynamic control, which is proxied by patent applications. Equilibrium innovation is a function of market structure, the vector of quality levels of all active firms, and the cost of R&D. Our main findings are as follows: (a) optimal innovation has an inverted-U shape in own quality; (b) holding own quality constant, innovation is declining in average rival quality but increasing in quality dispersion; and (c) following entry, each incumbent’s innovation declines, but aggregate innovation increases in most market structures. These findings are broadly consistent with the Schumpeterian hypothesis that market power leads to more innovation.

I reexamine the inverted-U relationship between competition and innovation (modeled and tested by Aghion et al. (2005)) by using data from publicly traded manufacturing firms in the United States. I control for the possible endogeneity of competition by using a trade-weighted average of industry exchange rates as an instrument. I find a mildly negative relationship between competition (as measured by the inverse of markups) and innovation (as measured by citation-weighted patents). The negative relationship is robust to many alternative assumptions and specifications. To reconcile the mildly negative relationship in the U.S. data with the inverted-U relationship that Aghion et al. (2005) find in the U.K. data, I modify their theoretical model and show that the modified model can explain both negative and inverted-U relationships. The key theoretical assumption is that the U.K. manufacturing industries are technologically more neck-and-neck than their counterparts in the United States. I find support for this assumption in the data. The different empirical results between the two countries may also arise because of differences in data and samples.