Announcement

Collapse
No announcement yet.
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • Using weights in linear mixed models

    Hi!
    I have run a linear mixed model with repeated measures, but have some trouble with heteroscedasticity. The code and the stdres vs fitted values plot looks like this:
    Code:
    mixed depvar c.agesp*##i.gender cov1 cov2 cov3..........  || ID: age, mle cov(un) residuals(exp, t(time))
    Click image for larger version

Name:	stdfit.png
Views:	1
Size:	15.1 KB
ID:	1375326
    and the fitted vs observed plot looked like:
    Click image for larger version

Name:	fittedobserved.png
Views:	2
Size:	12.9 KB
ID:	1375332

    I feel I have tried almost everything to reduce heteroscedasticity, but nothing seems to help much. However, when including a frequency weight in the model:
    Code:
    mixed depvar c.agesp*##i.gender cov1 cov2 cov3.......... [fw=time]  || ID: age, mle cov(un)
    the stdres vs fitted plot became:

    Click image for larger version

Name:	stdfitfreqweight.png
Views:	2
Size:	138.7 KB
ID:	1375328
    and the fitted vs observed plot looked like:
    Click image for larger version

Name:	fittedobservedfreqw.png
Views:	2
Size:	128.7 KB
ID:	1375330
    which seems to indicate a much bettwe fit than the fitted vs observed plot above.

    However, I have never used frequency weights before, and I am not sure whether it is appropriate to use it here (I used it since the earlier time points contain much more observations than the later ones).

    Any comments would be greatly appreciated!


    Kjell Weyde
    Attached Files
Working...
X