Hello,
I have individual time series data from 32 EEG channels. In the data, time runs from -200ms to 800ms. Zero is where a stimulus is presented. Two groups of subjects (10 patients v 20 controls) carried out four experimental conditions. My data have been averaged across trials, separately for each group, participant and condition. In other words, for each participant I have a separate time series for each condition, which has been averaged across trials.
For individual EEG channels I want to overplot the average EEG signal for both groups (on the y-axis) as a function of time (from -200 to 800ms, in 10ms increments), and indicate where in time the two time series are statistically significantly different from each other - i.e. effectively a set of multiple pairwise comparisons, one for each time point.
So far, to deal with the fact that we have a repeated measures design, I have used PROC MIXED as follows:
ods output diffs = diff;
PROC MIXED DATA=mlm2 COVTEST METHOD=ML;
CLASS sub group time;
MODEL EEGdata = group time group*time / SOLUTION
OUTP=pred;
RANDOM INTERCEPT / SUBJECT=sub TYPE= un;
LSMEANS group*time / diff;
run;
data subset; set diff;
if group = _group then delete;
if time ^= _time then delete;
timepoint_id = group||_group||time1;
raw_p = Probt;
keep timepoint_id raw_p;
run;
proc multtest inpvalues=subset holm hoc fdr;
run;
The general idea was to use LSmeans to get a set of pairwise comparisons which would give me a complete set of uncorrected t-tests. Then select the ones I want and run those p-values through proc multtest to do the statistical corrections. I would have liked to use the slice option, but could not find a way to obtain adjustments for the p-values, directly, so I settled on this rather clumsy route. I know you can use the adjust option in the lsmeans statement, but the only way I could think of using this to prevent computing all pairwise comparisons (thereby obtaining a very conservative adjustment as a result), was to use the lsmestimate route - but coding that looked horrific with 100 time points.
In summary:
1) Does anyone see an objection to the use of proc mixed to run the basic model in this way?
2) I am aware that taking account of autocorrelation in the data would be ideal, but including a random effect like this:
random intercept time / subject=sub type=ar(1)
caused the models to run forever, even when I resampled the timeseries to a much coarser scale. This also made me think that I might be doing something wrong.
3) Are there better / smarter / more elegant ways of obtaining the pairwise comparisons I am seeking.
Bottom line, help needed please ....
Thanks
Piers C
