re: st: rmanova or anova with repeated command, what to use?

["Airey, David C" <david.airey@vanderbilt.edu>]

.

Sounds like a great experiment! One thing I would do is to start with graphs if you have not. For example, make a 3 panel graph (baseline, placebo, active) where each panel plots each patient as a connected line of 7 points (t0-t6), so you have 13 lines per panel. You can also make 13 separate plots, one per patient, with 3 connected time lines for each treatment, to get a detailed sense of the effects per patient over the 3 conditions. Plotting all 13 x 3 = 39 lines in one plot may be too busy, but you could color by treatment. You can simplify this last plot by collapsing across patients to means for each time point and plot 3 average time profiles (baseline, placebo, active). I think this will help you "see" what's going on before you try to model the data, and you'll have a sense of whether the model matches what might be obvious in the plots. The plots will also show outliers, problems with variances, etc. that you might want to consider.


        Baseline Placebo Active
        t0-t6    t0-t6   t0-t6
patient1
patient2
...
patient3

If I'm understanding the design, you have two repeated measures variables, treatment condition (baseline, placebo, active; 3 levels) and time (pre, 0, 2, 4, 8, 12, 22 hours after exposure), where every patient sees every treatment by time combination giving you a total of 13 x 7 x 3 observations. Presumably you balanced the order of treatments. You could also have an order factor in the model if you wanted to check that. Leaving it in may soak up some of the error variance. Anyway, if you wanted to model both group and time, you probably would want to treat time as at least an ordered factor, if you did not want to give each person's treatment by time profile just a slope and intercept. The t0 time point "pre" may be used directly or it may be used to form differences with t1-t6, and or could be used as a baseline covariate. The Stata xtmixed command is something you should consider using instead of anova or manova, but it's complicated enough that a statistician's help is n!
 eeded, unless you have an example where your design was analyzed and you can emulate.

It sounded like you plan to do repeated measures ANOVAs at each time point. I think this is going to miss seeing what is going on.



> Dear statalisters,
> 
> Some time ago we performed a randomized crossover study in which we
> monitored the lung function of a group of volunteers  during three days. Day
> 1 was used to determine baseline lung function whereas day 2 and 3 were used
> to monitor lung function after inhaling either placebo (air) or active gas
> (oxygen) under hyperbaric  condition. During each measurement day lung
> function was measured 6 times. All variables and observations were put in a
> dataset which initially had the following format:
> 
> ID           Group      Time                     Result1 etc
> 1             0          0                             6.19
> 2             0          0                             5.97
> .             .          .                              .
> .             .          .                              .
> 13            2          22                           5.33
> Etc
> 
> ID exist of 13 persons (nr 1-13)
> Group: 0 (baseline), 1 (placebo), 2 (active)
> Time: pre, 0, 2, 4, 8, 12, 22 hours after exposure
> 
> We want to perform a repeated measures anova as all subject performed all
> three test days which makes the groups not independent. The format we had in
> mind was to search for differences between the groups at a specific time
> point and to look for a correlation between Result1, etc and variable time.
> 



*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/