Handling heteroskedasticity in Glimmix regression

[dsuchoff1]

Greetings,

I am having a hard time finding out how to specify an error heteroskedasticity model using Glimmix to run a mixed regression. 

I found this resource from UCLA, which shows how it can be done in Mixed; however, the authors use the local option in the repeated statement to specify that variance depends on the level of the dependent variable. I do not see the local option in Glimmix.

The authors give the following code as an example:

data nys1; set nys1;
  if time = 1 then ind2 = 1; else ind2 = 0;
  if time = 2 then ind3 = 1; else ind3 = 0;
  if time = 3 then ind4 = 1; else ind4 = 0;
  if time = 4 then ind5 = 1; else ind5 = 0;
run;

ods output  CovParms =t;
proc mixed data = nys1 method=ml;
  class id timecat;
  model attit = x time /solution;
  repeated timecat /subject=id type=cs r local=exp(ind2 ind3 ind4 ind5);
run;

How might this be conducted using Glimmix? 

Thanks!

[jiltao · Accepted solution]

the LOCAL= option is not available in the RANDOM statement in PROC GLIMMIX, so you cannot fit the same model in PROC GLIMMIX, unfortunately.

I personally think the LOCAL=EXP() option might be a too strong assumption in some cases, and you might consider using the type=csh residual option in the RANDOM statement in PROC GLIMMIX to fit something similar.

[SteveDenham]

@jiltao 's recommendation to use type=csh is spot on for the type of heterogeneity of variance between time points.  If you wish to model correlated heterogeneity, type=arh(1) works well for data that has timepoints equally spaced in time.  If they are not equally spaced, the csh structure or the  heterogeneous Toeplitz (toeph) come to mind.  If you have sufficient data, an unstructured covariance matrix will readily accommodate heterogeneity and non-constant correlation.

EDIT:

If you want a linear exponential autoregressive structure, SAS/STAT15.1 also has sp(lear)(time) {not timecat].

SteveDenham

[dsuchoff1]

Thank you kindly!