Hello,
I am using Stata 16's ITSA package to estimate the effect of a court decision on state level crime rate. I am estimating a multiple group ITSA with one treatment and one control group, with 65 pre-intervention observations and 43 post-intervention observations for each panel.


The actest command identifies autocorrelation at multiple lags, so I chose to estimate an AR-1 model using the prais option and tscorr for the autocorrelation type. The models contain DW statistics close to 2 (ranging between 1.96 and 2.06 depending on the type of crime considered), however the rho values remain high (0.39 to 0.58). Does this suggest that autocorrelation still biases the results after the prais correction? I also estimated the same models with the Corchrane-Orcutt transformation, and the values are very similar. I notice that when the DW stat approaches 2 and reported rho value decrease, the f-test value for the model also decreases, suggesting poorer model fit, which seems contradictory.


I would also like to estimate DIckey-Fuller statistics for each outcome variable to check for non-stationarity, though this does not seem possible with panel data.


I would appreciate any help interpreting these diagnostic to ensure I have appropriate model fit. Thank you very much.