My lecture will focus on how to evaluate the statistical properties of a complex, structural model. In the main paper, I show through computational experiments that the Berry Levinsohn, & Pakes (1995) has pronounced finite sample bias for any realistic number of markets and products. This work builds on the Su & Judd (2012), Dubé, Fox, & Su (2012), and Skrainka & Judd (2012). My talk will show you the computational challenges of executing structural projects correctly.
If you do not work in Applied IO, the BLP model with not make a lot of sense. You can find a quick introduction in Chapter 13 Train’s book Discrete Choice Methods with Simulation. Skip his section on computation. Nevo’s A Research Assistant’s Guide to Random Coefficients Discrete Choice Models of Demand provides further description of the model and economics, but definitely ignore his section on computational methods.
My talk will cover the material in Skrainka (2012).
Here are the paper and slides for my talk:
To understand how approximating integrals incorrectly — i.e., with Monte Carlo methods — will compromise your results, see: Skrainka & Judd (2012): High Performance Quadrature Rules: How Numerical Integration Affects a Popular Model of Product Differentiation
Lastly, it is very important to get the right answer so that you don’t destroy your reputation — or at least think about whether your model is likely to be correct. Verification, Validation, & Uncertainty Quantification (VV&UQ), provides an epistemological framework to think about correctness of scientific models. I provide a quick introduction here:
- Short talk on “Correctness in Data Science” which went viral: http://youtu.be/kex-UXZTGU4
- Slides: https://www.slideshare.net/mobile/dominodatalab/data-science-popup-seattle-correctness-in-data-science