Choosing Software to Estimate Spatial Models
Step 3 Run a GLS model with a Cochrane Orcutt lag structure for the temporal autocorrelation. This is likely to be the best model if you have more time-series points than cross sections in your data, an unusual situation for most spatial models. If the LSDV model was not seen as appropriate during Step 2, you can run a more general Random Eﬀects Model (REM), in which the unique impact of each cross section is assumed to be properly modeled as a unit speciﬁc error term, the distribution of which is expected to be random. This situation is analgous to the corrections for heteroscedasticity that have become widespread in GLS modeling with non-spatial data (see Hanushek and Jackson, 1977). This is a more general model than the LSDV, and does not stress the estimation by the inclusion of all those dummy variables, one for each unit. Including those dummy variables may help with the unique contribution of each unit to the model, but including so many variables in the model may undermine your ability to get good estimates of the eﬀects of the other substantively important variables you have included.