Using the glimmix radial smoother

[SteveDenham]

Hi Susan,

It seems like I am hijacking Samir's thread, but it still all boils down to using splines in a mixed model setting.

As far as selecting which to go with--the missing data means the two approaches give very different results.  The group that is early right censored gives a continually increasing estimate under the "only random effect spline" which is probably closer to the biology.  The fixed-effect spline/random spline model yields values that decrease once the data starts to be missing en masse--which is not unexpected, as the least squares estimates will regress toward the mean under the MAR assumption that usually applies to REML.

Now I am waiting for some NLMIXED code that incorporates a general log likelihood function for a 3 parameter Gompertz distribution/growth model.

In the end, I'll probably end up selecting the model that minimizes the standard error of predicted values across the range of non-missing data.

Steve

[Sam28041977]

Dear SLD,

The first random statement allow me to do within subjects comparison of treatments.

Every subject in the study receive all the treatments.

The second random statement capture the correlation from repeated measurements within the same treatment.

I hope this is clear.

Samir

[sld]

Hello Samir,

Yes, that clears up my concern. Thank you. SUJET(TRT) is read as "sujet nested within trt" so my guess was that SUJET was the experimental unit assigned to a single level of TRT, rather than SUJET being a block containing smaller experimental units to which levels of TRT were assigned (in your words, each subject receives all treatments). Because both SUJET and TRT were defined in your model, I think that GLIMMIX would interpret SUJET(TRT) as SUJET*TRT which makes sense.

Susan