Hi,
Background:
Consider a simple ANOVA model to test for differences between treatments on some continuous PD-outcome variable. For this simple example, Levene's test can be applied to test homoscedasticity between treatments on the residual term. Using proc mixed, it is not possible to get Levene's test automatically from the model, but extracting the absolute values of the studentized residuals and then running the model again with the absolute residuals as the dependent variable, e.g. like this:
proc mixed data=residuals;
class treatment;
model abs_resid = treatment;
run;
and then looking at the p-value for the F-statistics would be the equivalent for Levene's test.
Current case:
However, I have a more complicated linnear mixed model, LMM, with treatment as a categorical explanatory variable, a continuous covariate (the baseline measurements) and subject within treatment (parallel study). Also, measurements are taken at two post-baseline timepoints, so there is a repeated measurements situation. I use all post-baseline measurements and include the (post-baseline) timepoint and the interaction treatment*timepoint as categorical effects in the model. This is this model definition:
proc mixed data=example_data;
class subject treatment timepoint;
model change_from_baseline = treatment|timepoint baseline / ddfm=kenwardroger2 outp=residuals RESIDUAL;
repeated timepoint / subject=subject (treatment );
run;
Questions:
Given my more complicated model, compared to the ANOVA situation:
- Would Levene's test still be applicable using the simple code stated under background above (i.e. taking the absolute values of the residuals from the LMM as the dependent variable in a simple ANOVA)?
- Or would I need to run the full LMM model code using absolute residuals as the dependent variable? If I do that, I get a p-value for the all the effects treatment, timepoint, treatment*timepoint and baseline. Should I then check that all F-test p-values are non-significant?
- Or, if Levene's test is not applicable at all for more complicated models, is there another more appropriate test that can be used? If not, I guess graphical investigation of predicted values vs. residuals would have to suffice to check the homoscedasticity assumption.
Many thank for any useful input in this situation!
Kind Regards,
Joakim
