sas gee repeated measure

[AmyJuly]

I have two groups of people. One group of people are healthy and the other group of people is sick. For people in both groups, I let them do exercise for 10 mins and then let them rest for 3 mins. I measure their blood pressure over these 13 mins. I am wondering whether I can use the blood pressure over these 13 mins to tell whether a person is sick or not? In other words, whether the blood pressure varies between the two groups of people. Can we use the blood pressures, 13 times, as the biomarker to differentiate the health and sickness.
My intuition is to use logistic model to establish the outcome healthness with the predictor blood presure. But the problem is the outcome is not the repeated measure, while the predictor is. The SAS GEE procedure handles the outcome repeated, not the opporsite direction. How should I tackle this problem? Thanks!

[sbxkoenk]

Hello @AmyJuly ,

I think I would do this with the GLIMMIX procedure, but if you want to use the GEE procedure, you can turn to ALR.

If the responses are binary (that is, they take only two values), then there is an alternative method to account for the association among the measurements. The alternating logistic regressions (ALR) algorithm of Carey, Zeger, and Diggle (1993) models the association between pairs of responses by using log odds ratios instead of using correlations, as ordinary GEEs do.

There are other alternatives, but let's first see if you can sort it out using the info above,

Koen

[AmyJuly]

Thanks! I don’t have to restrict to GEE. How should I do this in Glimmix to account for the correlation of predictors instead of response? Thanks again!

[sbxkoenk]

Hello,

I am sorry. On closer inspection, I do not think that alternating logistic regressions (ALR) in the GEE procedure will help you out. It's not doing "the opposite direction" you are talking about.

I am absolutely sure that PROC GLIMMIX can do what you want, but I do not have example code readily available.

I guess some PROC GLIMMIX specialists will chime in here.

I am doing a lot of TSDM (Time Series Data Mining) and there it's common to have time series as predictors for a binary outcome (in manufacturing predictive maintenance for example). But "my" time series are always longer than 13 equi-distant measurements. 13 is maybe not enough to apply the techniques I usually apply.

In Time Series Data Mining the idea is that you reduce (compress) the (auto-correlated) time series to get some useful (more or less independent / uncorrelated) characteristics out of it. These characteristics (i.e. candidate predictors) can then be used for (binary target) prediction.
What I like to use is the Time Series Dimension Reduction (TSDR) Package in PROC TSMODEL. I mostly apply TSDR by Singular Value Decomposition (SVD) but you could also realize TSDR by Random Projection (RP) or by Fourier transformations or by ...!
I am not sure the above will work with only 13 measurements in your time series and with a limited amount of patients (low power to reliably estimate many coefficients and to reliably select a winning model from many challenger models). I always have massive amounts of subjects (objects) and my time series are long.

I could also use PROC GLIMMIX for TSDM but that would require lots of memory and lots of CPU time. Maybe more than available.
So I always apply PROC TSMODEL TSDR followed by a logistic regression or any other data mining algorithm for binary target prediction. The nice thing with TSDM is that you can also predict time-to-event (on top of just event) and also recurrent events.

Kind regards and good luck with your analysis,

Koen

[AmyJuly]

Thanks very much for your quick reply and that is very interesting!

[StatDave]

Your response is blood pressure, not sickness. The approach you could use is to fit a model on BP that includes a group (sick, healthy) and time as predictors. So, one possible model using PROC GEE would be the following that allows the groups to have different intercepts and different slopes on time. In this model, the test of the group parameter is a test of whether the group intercepts differ. The test of the group*time parameter is a test of whether the group slopes differ. If either is significant, then that would suggest that an individual's BP profile over time can predict sickness. 

proc gee;
class group id;
model bp=group|time;
repeated subject=id*group;
run;

A similar approach could be taken with a random effects model using either PROC MIXED or PROC GLIMMIX.

[SteveDenham]

If you are interested in using bp to predict illness, just invert what @StatDave  gave you.

proc gee;
class group id;
model group=bp|time/dist=binomial;
repeated subject=id*group;
run;

In this case, bp and time are used to predict group membership.  There should be some diagnostics that tell how good of a predictor bp is, but we are outside my comfort zone.

SteveDenham