Read on Twitter
So I was waiting for @nberpubs to tweet this out last week (which happened Friday pm). Anyway, here is my paper with @msinkinson and Matt Backus, where we test the common ownership hypothesis in RTE Cereal. Methodology Thread for #econtwitter: 1/
https://chrisconlon.github.io/site/bcs_cereal.pdf
https://chrisconlon.github.io/site/bcs_cereal.pdf
Classic IO question (back to Bresnahan's "rotations of demand") is to look at data on prices and quantitiy and figure out if firms are setting prices consistent w/ Cournot, Bertrand, Perfect Competition, Monopoly, Common Owners. This is harder than it looks to get right 2/
Recent work by @steventberry and @philhaile shows that we can nonparametrically identify conduct using exclusion restrictions (essentially variables that should NOT be correlated with costs - ie: ones that move marginal revenue instead). 3/ https://onlinelibrary.wiley.com/doi/abs/10.3982/ECTA9027
In practice we compare two different models of competition implying different markups and look at violations of moment conditions. We show the key input is the markup difference between two models (monopoly and Bertrand, etc.) 4/
We develop a simple (semi-parametric test). First predict mc=price-markup using determinants of costs for each product, and get the residual. This can be done flexibly h(*) and we use random forests. This avoids "is this log? linear? or exponential?"" 5/
Second we predict the markup difference between any pair of models using all features of all products in the market. (Why? Markups depend on supply and demand shocks for everyone's products). Here, nonparametrics/random forests matter a lot. 6/
Then we can just test to see if our predicted markup difference is correlated with our cost residual. When it is, it means we probably have the wrong model of price setting (bc your cost is correlated with things it shouldn't be!) 7/
We then try this for different sets of instruments: commodity price of corn for Rice Krispies, and rice for Corn Flakes, consumer demographics (which affect demand but not cost), "BLP instruments", and the Chamberlain 98 demand optimal IV and different markups. 8/
The punchline is that the usual linear specification of MC is sensitive to the choice of IV restrictions while the semiparametric (random forest) version is not. The demand optimal IV also perform well with or without the semiparametrics which is neat. 9/
If you want to know why things work you'd probably have to read the paper (or invite us for a seminar). 10/10 https://chrisconlon.github.io/site/bcs_cereal.pdf
