Comparing procedures for estimating random coefficient logit demand models with a special focus on obtaining global optima

Abstract

We compare several nested fixed point and optimization procedures for computing the estimator of the widely-used empirical market demand model developed by Berry et al. (1995). It is well-known that the optimization may often lead to multiple local optima, which, if ignored, can lead to erroneous policy conclusions. By combining the frequencies of finding the global minima and the computing times, we propose a new indicator that provides the computing time needed for obtaining the global minima. Using this indicator, we find that the Spectral and Squarem methods (Reynaerts et al., 2012) outperform the benchmark contraction iterations method and the MPEC (Dubé et al., 2012) and ABLP (Lee and Seo, 2015) methods. Moreover, when the share of the outside alternative is relatively large, two derivative-free optimization algorithms, which require less calculations and coding than derivative-based algorithms, outperform the best derivative-based methods. A simple argument suggests that the latter statement is likely to be true for other versions of the model as well.