Hi all,

I have some gridded climate data that I'm trying to fit a regression model for; 265 grid cells x 153 time steps. So far I've managed to fit a fairly decent model using a pooled OLS (i.e. no time or grid cell - specific effects), but I found the residuals to be heteroskedastic, so I've used an HC3 covariance matrix in the fit to account for this in the error calculation.

I want to investigate whether temporal or spatial autocorrelation needs to be accounted for as well. Running a Durbin-Watson test on the pooled residuals doesn't appear to show any autocorrelation, but I want to analyse the temporal and spatial effects separately to make sure nothing has been masked by pooling all the data together.

To start with, I fitted a panel regression fit using grid cell location & time as my dimensions (I was thinking that if panel effects exist, then I could load my data into R and use splm somehow), but an F-test suggested that neither fixed nor random effects models offered any advantage over the pooled OLS model. Looking at the tutorials online, I can't seem to find an examples involving spatiotemporal data (the only results I've found are where the variables are time-independent).

I'm not sure what to do next. How is autocorrelation detected when using gridded time series data, and what is done in OLS to account for these effects? In light of the null result of my F-test and the inconclusive Durbin-Watson test I did earlier, is this even worth doing? Also, if possible I would like to know if R and/or Python libraries exist to handle these analyses, as those are the languages I'm most comfortable with.

Thanks!