Some Computational Aspects of Feasible GLS Estimation of Large Panels in R

Econometric estimation of panel data models by feasible generalized least squares (FGLS) provides an example of how conceptually simple problems may run into computational bottlenecks. I address the main computational tasks of FGLS within the R system for statistical computing, comparing different tools from the point of view of computational efficiency. I concentrate on estimating two models: the popular “random effects” with two error components and the less restrictive “general GLS” specification, which does not fit into the standard computational framework usually employed for the former. I compare the standard solution (partial time demeaning) with two alternative strategies, based respectively on algebraic properties and on object-oriented programming. I show how, while naive implementations become infeasible with large datasets, both list operators and object-oriented matrix routines available in the R environment make the problem tractable for most practically relevant sample sizes on any machine. I conclude by briefly discussing the parallelization of critical tasks.