Pre- and Post- measurements across time when time is not of importance

[Melk]

I have pre-post intervention data on a very small number of subjects (n=7).

The coutcome is continuous; measurements were taken across time at pre- and post-intervention for each patient. Time was not measured consistently, so 1 patient may have measurements at 3, 5, 7, and 9 while another may have 1, 5, 6, and 8.

My research question is: Is the intervention effective in lowering the outcome measurement? Although time was measured, outcome measurements over time is actually not part of our hypothesis.

What statistical test would I use to best assess this data? Would time be a time-varying covariate here?

[Melk]

I have pre-post intervention data on a very small number of subjects (n=7).

The coutcome is continuous; measurements were taken across time at pre- and post-intervention for each patient. Time was not measured consistently, so 1 patient may have measurements at 3, 5, 7, and 9 while another may have 1, 5, 6, and 8.

My research question is: Is the intervention effective in lowering the outcome measurement? Although time was measured, outcome measurements over time is actually not part of our hypothesis.

What statistical test would I use to best assess this data? Would time be a time-varying covariate here?

[StatsMan]

Yes, TIME would most likely be a within-subject covariate in your model.  PROC GLIMMIX might be a good choice for the model if you feel that your sample size is large enough.  You can use the RANDOM statement to model the correlations between the observations on the same subject on the r-side and use one of the spatial structures (perhaps TYPE=SP(POW)) to account for the unequally spaced time points you observe.  

If all of the subjects received the same treatment, then perhaps comparing their baseline response to a response near the max observed time would be of interest?  You could model change from baselineas your response or write a CONTRAST statement to compare the response at the times of interest to you.  If you have a control group that received no treatment or more than one treatment group, then you could compare the results across the treatments or the treatment vs control using the LSMEANS statement with the AT option to set the TIME value of interest.  

The EFFECT statement could be used to set up a spline effect on TIME.  That might be especially useful given the small number of subjects you have.  

[Melk]

Thank you so much for your response. Is GLIMMIX generally appropriate for such a smalle sample of n=7?

If so, would this tell me if my post was significantly lower in y than my pre?

proc glimmix data=dat;
  class Trial subject;
      model y = Trial | time / solution;
      random intercept / subject=subject;
      random intercept time / type=sp(pow) subject=subject*Trial;
run;
quit;

where trial = pre or post.