Resources
This page collects notes and code that have been helpful over the years but that do not appear in my research papers.
Nested Logit Demand
Materials prepared in support of the Miller and Weinberg (2017) research project on the U.S. beer industry.
 Notes that derive market shares and the first and second derivatives of demand for the onelevel and twolevel nested logit demand systems.
 R Code with functions that provide the market shares and derivatives (nlfunctions.R) and a short script that demonstrates how the functions can be called (nlstarter.R).
Perry Porter (1985 AER) Model of Cournot Competition
Materials prepared in support of the Miller and Podwol research project. The Perry and Porter (1985) model provides a way to analyze some markets in which there isn't much product differentiation.
 See McAfee and Williams (1992 JIE) for the mathematics.
 R Code with functions that allow for calibration and merger simulation (pp_functions.R) and a short script that shows how to use the functions (pp_starter.R). Mergers often are profitable in the Perry and Porter (1985) version of the Cournot model.
BLP Contraction Mapping
C Code to implement the contraction mapping for random coefficients logit models of demand. There are three versions:
 contrMap_blp_BHRW uses the speedenhancing mathematical simplification suggested in the Brunner, Heiss, Romahn, and Weiser (2017) working paper, hereafter BHRW, for standard BLPstyle models.
 contrMap_rcnl_BHRW modifies the BHRW simplification for a random coefficients nested logit (RCNL) model in which the outside good is in its own nest. The speed of convergence is damped as shown to be necessary by Grigolon and Verboven (2014 ReStat).
 contrMap_rcnl also is for the RCNL model but does not use the BHRW simplification. Thus it is about 3x slower. With very small datasets, such as the pseudo cereals data of Nevo (2000), the speed of the BHRW algorithms might lead to computational instability on some computers.
The code must be compiled before use with Matlab. The RCNL functions can be used with more standard (nonnested) BLP specifications if the nesting parameter is set to zero. However, more computational steps are required to accommodate the RCNL, so if the model does not have a nesting structure then the BLP function will be faster. A useful reference for all things BLP is the Conlon and Gortmaker (2019) working paper titled "Best Practices for Differentiated Products Demand Estimation." It proposes using an accelerated fixed point algorithm such as SQUAREM. Based on various comparisons in the literature, the approach of using compiled C code with the BHRW simplification appears to deliver speed that is comparable to SQUAREM, though I'm not aware of any explicit comparisons.
