The Expectation-Maximization (EM) algorithm is a well-established estimation procedure which is used in many domains of econometric analysis. Recent application in a discrete choice framework (Train, 2008) facilitated estimation of latent class models allowing for very exible treatment of unobserved heterogeneity. The high exibility of these models is however counterweighted by often excessively long computation times, due to the iterative nature of the EM algorithm. This paper proposes a simple adjustment to the estimation procedure which proves to achieve substantial gains in terms of convergence speed without compromising any of the advantages of the original routine. The enhanced algorithm caps the number of iterations computed by the inner EM loop near its minimum, thereby avoiding optimization over suboptimally populated classes. Performance of the algorithm is assessed on a series of simulations, with the adjusted algorithm being 3-5 times faster than the original routine.
|Place of Publication||Tilburg|
|Number of pages||12|
|Publication status||Published - 2013|
|Name||CentER Discussion Paper|
- EM algorithm
- discrete choice models
- latent class models