Ordered response models and non-random personality traits: Monte Carlo simulations and a practical guide

The paper compares different estimation strategies of ordered response models in the presence of non-random unobserved heterogeneity. By running Monte Carlo simulations with a range of randomly generated panel data of differing cross-sectional and longitudinal dimension sizes, we assess the consistency and efficiency of standard models such as linear fixed effects, ordered and conditional logit, and several different binary recoding procedures. Among the binary recoding procedures analyzed are the conditional ordered logit estimator proposed by Ferrer-i-Carbonell and Frijters (2004) that recently has gained some popularity in the analysis of individual well-being, as well as the new developed "Blow-Up and Cluster" (BUC) estimator of Baetschmann et al. (2011). The Ferrer-i-Carbonell and Frijters estimator (FCF) performs best if the number of observations is large and the number of categories on the ordered scale is small. However, the BUC method performs similarly well and even outperforms the FCF estimator if the number of categories on the ordered scale is large. If the researcher is only interested in the relative size of coefficients with respect to a baseline, however, the easy-to-compute linear fixed effects model delivers essentially the same results as the more elaborate binary recoding schemes.