Dear R:
I am trying to fit a doubly multivariate LME (DM) where I have two response
variables measured on two occasions per person. Specifically, reading and math
scores measured at the beginning and ending of a school year. The response
variables have a correlation of r = .85.
The response variables in the data matrix are stacked in a vector with a dummy
code flagging each outcome and with time variables for each outcome.
The model was fit by removing the overall intercept, but creating fixed effects
and random effects for each using the following:
mult2.lme<-lme(fixed=score~-1+read+math+time.m+time.r, data=mult.samp,
random=~-1+read+math+time.r+time.m|childid)
This worked and seemed to produce reasonable estimates. I then ran a model using
only a single outcome (reading) and found that the estimates are very similar,
so I am relatively confident in the results of the model.
Now, I have a couple of questions regarding the DM LME:
1) If I wanted to explore model assumptions, i.e., homoscedasticity and
dependence among the residuals, how might I do this. For example, how might I
specify an AR1 structure? I have explored the assumptions in the single response
model, but am having trouble now.
2) I presume it is safe to explore the intercepts and slopes using lmList () one
variable at a time. Is this correct?
3) The AIC statistic for the single response model is half the size of the DM
model. It is my understanding that this statistic is more appropriate than LL
test in this scenario. Is this correct? If so, is there a reason that may
explain the larger AIC in the DM model? Or, is this indicating that the single
response model is a better fit?
4) Last, is there any other issue that I am missing? I may not have asked the
question, but would appreciate any suggestions related to the model.
Regards,
------
Harold C. Doran
Director of Research and Evaluation
New American Schools
675 N. Washington Street, Suite 220
Alexandria, Virginia 22314
703.647.1628
<http://www.edperform.net/>
[[alternative HTML version deleted]]
