08-30-2012 03:01 PM
Are there any least-squares regression procedures that can correct for heteroskedasticity and produce output that can be used with the PLM procedure to create prediction intervals?
My data is a long panel, but since it appears that output from the PANEL procedure is incompatible with PLM, I created dummy variables for cross-sections & time in order to force the data into a model like GLM or REG. Unfortunately, the data exhibits heteroskedasticity. I have found various corrections for this with the REG procedure (ACOV, HCC, SPEC, WHITE), but haven't found any that are avaialble for GLM, ORTHOREG, MIXED, etc.
Thank you in advance for any suggestions.
08-30-2012 10:36 PM
Sorry. I don't know . But you can try to include some polynomial (e.g. x^2 x^3 ......) into your REG GLM to see whether it can eliminate heteroskedasticity. Or you maybe should be at SAS Forecasting forum to find some help.
08-31-2012 07:52 AM
Within the linear mixed model procs (MIXED, GLIMMIX, HPMIXED), there are two methods for accommodating heteroskedasticity. If the lack of homogeneity is in your cross-section variables, you can add a group= option in the random/repeated statement that will result in separate estimates of the covariance parameters by group. If the lack of homogeneity is in your time variable, then assuming you are fitting time as a CLASS variable, you could try heterogeneous covariance structures such as arh(1), fa(1), toeph, or unstructured. The choice will depend on the spacing of the time variable.
I hope this applies to what you are attempting.