A Comparative Analysis of Alternative Econometric Packages for the Unbalanced Two-way Error Component Model

Notwithstanding it was originally proposed to estimate Error Component Models (ECM) using balanced panels, most applications use unbalanced panels. When unbalanced panels include both time and individual random effects, special computational problems arise. Wansbeek and Kapteyn (1989) analyze algorithms for estimating unbalanced two-way ECM; Baltagi et al. (2002) compares three classes of estimators for the unbalanced two-way error model. Here I show some differences between theoretical findings and empirical applications by investigating how various proposed algorithms are implemented in the most widely used econometric packages and by providing a comparative appraisal of the different methods the various packages used to estimate the unbalanced two-way ECM. An illustration examines the determinants of the bank deposit yield in Italy, comparing the outcomes of six popular econometric packages available for the analysis of panel data: E-Views, LIMDEP, RATS, SAS, STATA, and TSP. The packages give strikingly different numerical results. While the relevant documentation is often elusive as to algorithmic details, my findings suggest one reason for the differences lies in the means used for computing the variance of the two idiosyncratic error component terms. Finally, I examine the small-sample properties of the various algorithms by means of Monte Carlo experiments. According to preliminary results, these algorithms show approximately the same performance in terms of bias and variability.