How do the social returns to R&D differ from the private returns? We must believe there is a positive gap between the two given the widespread policies of subsidizing R&D investment. The problem is measuring the gap: theory gives us a number of reasons why firms may do more R&D than the social optimum. Most intuitively, a lot of R&D contains “business stealing” effects, where some of the profit you earn from your new computer chip comes from taking sales away from me, even if you chip is only slightly better than mine. Business stealing must be weighed against the fact that some of the benefits of knowledge a firm creates is captured by other firms working on similar problems, and the fact that consumers get surplus from new inventions as well.
My read of the literature is that we don’t have know much about how aggregate social returns to research differ from private returns. The very best work is at the industry level, such as Trajtenberg’s fantastic paper on CAT scans, where he formally writes down a discrete choice demand system for new innovations in that product and compares R&D costs to social benefits. The problem with industry-level studies is that, almost by definition, they are studying the social return to R&D in ex-post successful new industries. At an aggregate level, you might think, well, just include the industry stock of R&D in a standard firm production regression. This will control for within-industry spillovers, and we can make some assumption about the steepness of the demand curve to translate private returns given spillovers into returns inclusive of consumer surplus.
There are two problems with that method. First, what is an “industry” anyway? Bloom et al point out in the present paper that even though Apple and Intel do very similar research, as measured by the technology classes they patent in, they don’t actually compete in the product market. This means that we want to include “within-similar-technology-space stock of knowledge” in the firm production function regression, not “within-product-space stock of knowledge”. Second, and more seriously, if we care about social returns, we want to subtract out from the private return to R&D any increase in firm revenue that just comes from business stealing with slightly-improved versions of existing products.
Bloom et al do both in a very interesting way. First, they write down a model where firms get spillovers from research in similar technology classes, then compete with product market rivals; technology space and product market space are correlated but not perfectly so, as in the Apple/Intel example. They estimate spillovers in technology space using measures of closeness in terms of patent classes, and measure closeness in product space based on the SIC industries that firms jointly compete in. The model overidentifies the existence of spillovers: if technological spillovers exist, then you can find evidence conditional on the model in terms of firm market value, firm R&D totals, firm productivity and firm patent activity. No big surprises, given your intuition: technological spillovers to other firms can be seen in every estimated equation, and business stealing R&D, though small in magnitude, is a real phenomenon.
The really important estimate, though, is the level of aggregate social returns compared to private returns. The calculation is non-obvious, and shuttled to an online appendix, but essentially we want to know how increasing R&D by one dollar increases total output (the marginal social return) and how increasing R&D by one dollar increases firm revenue (marginal private return). The former may exceed the latter if the benefits of R&D spill over to other firms, but the latter may exceed the former is lots of R&D just leads to business stealing. Note that any benefits in terms of consumer surplus are omitted. Bloom et al find aggregate marginal private returns on the order of 20%, and social returns on the order of 60% (a gap referred to as “29.2%” instead of “39.2%” in the paper; come on, referees, this is a pretty important thing to not notice!). If it wasn’t for business stealing, the gap between social and private returns would be ten percentage points higher. I confess a little bit of skepticism here; do we really believe that for the average R&D performing firm, the marginal private return on R&D is 20%? Nonetheless, the estimate that social returns exceed private returns is important. Even more important is the insight that the gap between social and private returns depends on the size of the technology spillover. In Bloom et al’s data, large firms tend to do work in technology spaces with more spillovers, while small firms tend to work on fairly idiosyncratic R&D; to greatly simplify what is going on, large firms are doing more general R&D than the very product-specific R&D small firms do. This means that the gap between private and social return is larger for large firms, and hence the justification for subsidizing R&D might be highest for very large firms. Government policy in the U.S. used to implicitly recognize this intuition, shuttling R&D funds to the likes of Bell Labs.
All in all an important contribution, though this is by no means the last word on spillovers; I would love to see a paper asking why firms don’t do more R&D given the large private returns we see here (and in many other papers, for that matter). I am also curious how R&D spillovers compare to spillovers from other types of investments. For instance, an investment increasing demand for product X also increases demand for any complementary products, leads to increased revenue that is partially captured by suppliers with some degree of market power, etc. Is R&D really that special compared to other forms of investment? Not clear to me, especially if we are restricting to more applied, or more process-oriented, R&D. At the very least, I don’t know of any good evidence one way or the other.
Final version, Econometrica 2013 (RePEc IDEAS version); the paper essentially requires reading the Appendix in order to understand what is going on.