My non-expert sense is that a lot of the empirical work I see in IO focuses on one very narrowly defined industry (e.g. mini-vans, video rental stores, concrete manufacturers) in which the production process can be well understood and data more easily comparable across firms. However, when I think about many developing countries, there just aren’t that many bigger firms – there may only be 200 firms with 10 or more workers in manufacturing in the entire country for example. How should we think about doing IO in such contexts?
[KS] That in many ways would be a help rather than a hinderance in that it facilitates collecting data and understanding firm interactions (unless it makes the data sample so small that reliable inference is difficult). One reason why a good number of IO studies focus on narrow industries is because the research question concerns competition in settings with particular features (if I am interested in measuring the welfare effects of new product introductions in an oligopolistic industry such as automobiles, I don’t necessarily need to worry about the market for trucks), just like studies of new product adoptions in developing country contexts focus on a relatively narrow product category (such as Conley and Udry’s work  on fertilizer use in pineapple production). I don’t think that the focus on a narrow setting is thus unique to IO or would render the techniques useless in developing country contexts, with the necessary adjustments to reflect differences in the institutional and market contexts.
[DK] I don’t see this as a major theoretical impediment to the study of IO topics in developing countries. The basic IO questions—competition between firms, regulation, firm entry and exit—are just as applicable to small firms as they are to large ones. Empirically this may generate a problem of small sample sizes for researchers seeking to investigate a very specific industry in a small country, but this would apply in the IO of developed countries as well.
The major difference I see in doing IO in these contexts is that we can no longer make certain assumptions that would seem reasonable in developed countries, for instance that the number of firms in a market/sector is a function of the demand for the products they sell, and not other factors such as the availability of credit in that market. Another example would be that while individual industries in developed countries usually have firms using relatively similar technologies, at relatively similar scales, in developed countries one often sees a large variety of technologies and a range of scales in the same industry. This both changes the set of estimation techniques that are available (since those that depend on firms having the same return to capital are not applicable) and also opens up a new set of questions.
Many of the firm policies used in developing countries to try and improve productivity are predicated on an assumption of multiple missing markets and legal and infrastructure constraints. How should we think about measuring productivity in such contexts, if firms are capital constrained, there are labor regulations which limit labor adjustment, many firms don’t use electricity or are subject to blackouts, and multiple other constraints which mean that the capital and labor firms are using are at best constrained optimal decisions. It is not clear to me whether methods like Levinsohn-Petrin which assume that firms can readily adjust intermediate inputs when faced with a productivity shock can be applied in cases where all these constraints limit the ability of firms to adjust, and if not, what we should do instead?
[DK] These concerns aren’t limited to the field of development—there several papers arguing that methods like Levinsohn-Petrin  may not be applicable even in some developed contexts as well (Blundell, Bond; Ackerberg, Caves, Frazer ).
The right approach depends upon the ultimate goal of the research. In papers that estimate the effect of some policy change—deregulation, opening to trade, etc—on productivity, I would argue that simpler approaches are probably better if only for transparency. For instance, much of the US productivity approach (e.g. work by Syverson, Haltiwanger and others) estimates productivity as the residual of a Cobb-Douglas where the parameters are calibrated as industry-level factor shares. If exit is a major feature of the data, then a structural correction as in Levinsohn-Petrin may be necessary.
For research that attempts to learn more about the exact mechanisms of productivity changes—for example an evaluation of a training program for small entrepreneurs—I would argue that a careful focus on a single industry is the right approach. This would allow researchers to customize their productivity estimation approach to the characteristics of the industry. In many cases, (say rice milling) the physical production function with respect to some inputs may be Leontief, and by incorporating this and other industry-specific details it may be possible to isolate the specific dimensions along which firms are differentiated, rather than estimate an overall distribution of TFP residuals.
[KS] The literature on estimating production functions has seen a resurgence of interest in recent years, in part because some of the assumptions made by earlier work appear strong beyond the developing country context. Work by Ackerberg et al. , for example, points to the problem of using lagged input decisions as an instrument in the presence of flexible inputs, and more recent work by Gandhi et al.  proposes a possible solution based on the firm’s optimal choice of flexible inputs. Such an approach might have merit in the contexts you describe as well. Since it relies crucially on identifying which inputs firms can adjust and which they can’t, however, it might be most appropriate in the context of a well-defined, relatively narrow industry, where firms’ production decisions can be well understood and credibly modeled. For questions where the research question requires measuring adjustments across larger portions of the economy – such as the trade liberalization questions mentioned previously – I would advocate for a mixed approach, where TFP residuals are recovered using alternative approaches including simpler ones such as traditional fixed effect estimators.
Are there important development questions that you think people are trying to address the wrong way by ignoring insights from IO?
[DK] To be fair, I don’t think that development is so entirely insulated from IO that development economists would be able to entirely ignore insights from IO. However, I do feel that the literature on demand for health products in developing countries might make better use of the IO literature on demand. The enormous literature on demand from the marketing field is also rarely cited in development, even though it seems naturally applicable.
The literature on microcredit evaluations might also explore more of the competitive impacts of microfinance. This is addressed briefly in a few papers – Banerjee, Duflo, Glennerster, Kinnan find few spillover effects onto nearby firms – but a more IO based approach that would focus on prices, total demand, and ultimately total welfare could also provide interesting other insights. For instance, we have little understanding of who new micro-entrepreneurs are competing against—other micro-entrepreneurs, or larger businesses.
What do you think, are IO methods applicable to development? Are we missing out by not using them more? Stay tuned for tomorrow’s final post in this series where Katja and Dan offer their thoughts on productive areas for using IO more in development.
- IO meets development