difference-in-differences

[sks521]

Hi folks,

I am doing difference-in-differences analysis by using Proc Mixed as follow;

PROC MIXED DATA = diff1;
CLASS POST EXPOSED;
MODEL VISIT=POST|EXPOSED / SOLUTION ;
LSMEANS POST|EXPOSED / DIFF;
ESTIMATE 'D-I-D' EXPOSED*POST 1 -1 -1 1;
REPEATED /SUBJECT=PKID TYPE=UN ;
FORMAT EXPOSED STATUS. POST PERIOD.;
TITLE2 "RANDOM INTERCEPT MODEL FOR PRE/POST GP Visits by EXPOSURE GROUP";
TITLE3 "UNADJUSTED: INCLUDING LEAST-SQUARES MEANS ESTIMATES";
RUN;

But I am getting the following log;

{334 PROC MIXED DATA = diff1;
335 CLASS POST EXPOSED;
336 MODEL VISIT=POST|EXPOSED / SOLUTION;
337 LSMEANS POST|EXPOSED / DIFF;
338 ESTIMATE 'D-I-D' EXPOSED*POST 1 -1 -1 1;
339 RANDOM INT/SUBJECT=PKID TYPE=UN ;
340 FORMAT EXPOSED STATUS. POST PERIOD.;
341 TITLE2 "RANDOM INTERCEPT MODEL FOR PRE/POST ADHERENCE by EXPOSURE GROUP";
342 TITLE3 "UNADJUSTED: INCLUDING LEAST-SQUARES MEANS ESTIMATES";
343 RUN;

NOTE: Convergence criteria met but final Hessian is not positive definite.
NOTE: PROCEDURE MIXED used (Total process time):
real time 0.07 seconds
cpu time 0.06 seconds

}

And it's not giving any p-values in the output tables.

Can someone have a look and advise please?

Thanks

[PaigeMiller]

Go to your favorite search engine and type in

Hessian not positive definite

[sks521]

Actually I did that before posting the query on here.

It brings me to this forum where someone has posted a reply saying that either the model doesn't fit or the sample size is too small.

I am assuming my sample size is not too small (depending on what 'too small' means).

Not sure how to assess model fit?

So, I thought to post it on here.

Thanks.

[Rick_SAS]

For an overview of this issue, see

"Convergence in mixed models: When the estimated G matrix is not positive definite"

which includes tips and links to papers and SAS Notes that you can read for more information.