An argument for preferring Firth bias-adjusted estimates in aggregate and individual-level discrete choice modeling

Using maximum likelihood estimation for discrete choice modeling of small datasets causes two problems. The first problem is that the data often exhibit separation, in which case the maximum likelihood estimates do not exist. Also, provided they exist, the maximum likelihood estimates are biased. In this paper, we show how to adapt Firth's bias-adjustment method for use in discrete choice modeling. This approach removes the first-order bias of the estimates, and it also deals with the separation issue. An additional advantage of the bias adjustment is that it is usually accompanied by a reduction in the variance. Using a large-scale simulation study, we identify the situations where Firth's bias-adjustment method is most useful in avoiding the problem of separation as well as removing the bias and reducing the variance. As a special case, we apply the bias-adjustment approach to discrete choice data from individuals, making it possible to construct an empirical distribution of the respondents' preferences without imposing any a priori population distribution. For both research purposes, we base our findings on data from a stated choice study on various forms of employee compensation.