Empirical Industrial Organization: An Introduction Holger Sieg August 24, 2015
Overview Some of the core areas of empirical IO are the following: Demand and Supply of Differentiated Products. Productivity Analysis and Production Function Estimation. Static Models of Entry, Exit, and Mergers. Dynamic Models of Demand and Supply of Durable Goods. Dynamic Models of Investment and Innovation. Static and Dynamic Models of Auctions. Models of Network Formation.
Theory-Based Estimation in IO The prevalent approach in modern empirical IO is based on structural or theory-based estimation: Data Collection. Model Selection and Specification. Identification and Estimation of the Model. Parametric versus Non-parametric Identification. Endogeneity: Instruments & Control Functions. Goodness of Fit Analysis. Validation of the Model and Specification Testing. Hypothesis Testing. Policy Analysis and Predictions.
Alternative Quantitative / Empirical Approaches Descriptive Approach: If we have access to an interesting, often confidential, data set, we can gain insights into problems that have not been studied before. In that case, a purely descriptive approach that summarizes the main empirical regularities is typically the first step of the analysis. Analyzing data without a well-defined modeling framework is, however, only of limited usefulness. Experimental Approach: If a firm allows us to conduct a controlled experiment, we can estimate causal treatment effects. In practice, It is difficult to convince firms to participate in controlled experiments. Hence, there a few compelling experimental studies in empirical IO. Calibration Approach: Sometimes we can solve an interesting new model and obtain quantitative insights based on carefully calibrated versions of a model. This approach is more widely accepted in macro economics.
An Example: Crawford & Yurukoglu (2012) To help fix ideas, I m going to briefly summarize a recent paper that showcases the power of the structural empirical approach in IO. Crawford & Yurukoglu (2012, AER), The Welfare Effects of Bundling in Multichannel Television Markets. They consider the Multichannel TV = cable, satellite, or telco video. For convenience, I ll just call this the cable industry.
Some Institutional Background Producers: Disney (ABC, ESPN, Disney,..), Viacom(CBS, Showtime,..), Time Warner (CNN, HBO,...), Comcast (NBC,..) Distributors (Cable Operators): Comcast, Cablevision (NY), DirectTV,... Bundles: Comcast offers Limited Basic for $14 (local programming), Digital Economics for $30 (add CNN, Discovery, etc), Digital Starter (add MTV, ESPN, CNN, etc) for $50, etc. There is an upstream market, in which producers negotiate fees with distributors, and a downstream market in which consumers buy bundles of channels from local cable operators.
Context It is widely believed that bundling can be profitable for firms by allowing them to price discriminate: People differ in their WTP for bundle components. Bundling implicitly allows firms to charge different prices to different people for components. Example: Sports and Gardening TV channels. This rationale is regularly taught in economics and strategy despite no compelling empirical evidence of its effects in practice. The cable and satellite television industry is the canonical (conjectured) example of the theory. In the US, policymakers have long advocated à la carte channel offerings. If this were mandated, what would happen to consumer welfare?
Is there an experimental alternative? Could one predict what would happen under a la carte using experimental methods? How would you do that? Convince one cable operator to offer channels a la carte and see what happens to quantities and prices in that/those markets? It wouldn t work. For two reasons: 1. An operator s contracts with channels wouldn t let them do it. Thus would need to convince not only the cable operator, but all its channels. Unlikely! 2. One firm wouldn t be enough. Outcomes in the cable industry depend on equilibrium outcomes that depend on the choices of all the firms in the market.
Is there a quasi-experimental alternative? Could one predict what would happen under a la carte using quasi-experimental methods? How would you do that? Compare markets that do and do not offer bundles and predict what quantities and prices might be? It wouldn t work. Bundling in TV markets is ubiquitous across the globe The only option is to build a structural model of the industry as it exists now and predict what would happen in a world without bundles.
What do we need in our model? Consumers: Demand for individual channels. As this is what people will be buying in an unbundled world. Firms: Costs, Pricing, Bundling. As we need to understand the structure of costs and nature of competition to predict unbundled outcomes Bargaining between TV channels and TV distributors As unbundling will change existing pricing structures upstream (ESPN gets $4/month for everyone in a bundle - will want more if some people opt out)
Features in a structural industry model of the multichannel TV industry CY (2012) incorporate the following features: Estimate demand for 50+ individual TV channels Estimate marginal costs for these channels; Estimate bargaining parameters for pairs of channel and distributor conglomerates e.g. Disney-Comcast, Viacom-DirecTV Simulate a counterfactual world where firms compete with a fixed fee for access and then individual prices for channels.
Findings 1. If marginal costs didn t change, consumers would be 19% better off, confirming the predictions of the discriminatory theory. 2. But costs do change: channels and distributors re-negotiate Channels would be an average of 103% more expensive to downstream firms without bundles. These cost increases are passed on to consumers, reducing the benefit of greater freedom of choice. 3. After renegotiation, consumers are no better off and would likely be worse off, e.g. if there were any equipment, marketing, and/or admin costs.
Discussion Even these conclusions come with important caveats. CY (2012) assume: 1. Channels that are watched more are valued more. Maybe not quite right for sports channels? 2. Channels currently on TV stay there. 3. Quality of those channels stays the same. 4. More generally, don t model how consumers trade off price and channel quality.
The Seven Skills of a Structural Econometric Modeler There are many steps involved in the production of a structural research paper It s therefore important to have a variety of skills.
1. Institutional Detail One very important insight of structural research is that details matter! In an IO context, markets are different. In order to understand the effects of competition in a market and thus to evaluate policies in that market, one must understand the institutional aspects that characterize the market environment. A fringe benefit: one may become an expert about certain markets! You will not succeed in IO unless you know your Industry and its Institutions!
2. Data Discovery To answer any empirical question requires good data. Finding new data sources is often the job of the researcher, particularly for the young scholars. In practice, this is often the highest hurdle to addressing relevant and pressing policy questions. Some data sources are proprietary and can be very expensive to obtain. For example, Crawford and Yurukoglu spent more than $100K on their data. You need to start looking for data immediately!
3. Economic Theory How do the economic agents behave? What drives consumer demand? What role do product characteristics/advertising/dynamics play? What drives firms costs? What is the appropriate model of competition? You need to be able to specify a tractable yet realistic model of the institution being studied.
4. The Econometric Model Given the institution, the available data, an economic theory of agent behavior in the market, and the feasibility and performance of alternative econometric methods, you need to blend these different elements together without losing internal consistency. Are all the salient features captured? Are you capable of measuring the effects of interest? The answers to these questions are application-specific.
5. Econometric Theory You need to be able to incorporate a reasonable error structure into your economic model to generate meaningful orthogonality conditions. Given the data and econometric model what are the sources of error underlying observed behavior? Unobserved variables? Measurement error? Optimization error? What identifies the key effects of interest? What are the appropriate estimation techniques?
6. Computational Skills Can you manage the data? Can you manage the estimation? Can you manage the policy simulations? Can you code in Stata? Matlab/Gauss? Fortran/C? Can you figure out why your estimation won t converge? Why your standard errors don t make sense? Can you do it in weeks instead of months? Do you have the energy and the perseverance to do that?
7. Presentation and Writing Last but definitely not least: can you present your work? Can you describe it? Perhaps it isn t surprising, but many of those who are good at (3)-(6), often have trouble at making their insights understandable to others
Course Outline Demand, Supply, and Competition. Demand Estimation: dominant application in IO, but increasingly in other fields. Productivity Estimation which is common in IO, even more popular in Trade, Development, and Growth. Static Entry Games that endogenize the market structure. Single-Agent Dynamic Decision Models. Competitive Dynamic Models of Industry Structure. Dynamic Games of Oligopolistic Behavior. Topics.