Calculating intraclass correation coefficient in SAS

[Abishekaa]

This is my sample dataset:

PTID _NAME_ COL1

1           M1        11     

1           M2        20

1           M3         . 

2           M1        15 

2           M2        25

2           M3        12

3           M1        10

3           M2        9

3           M3        15

I want to calculate the intraclass correlation coefficient to assess the correlation of (m1-m3) within each patient..so basically an icc for each patient.

I have tried to use the %intracc macro:

 %intracc(data=trans,target=PtID,
               depvar=COL1,rater=_NAME_,print=3);

But this only outputs one overall icc values, while I want one icc for each patient..Any help is appreciated!

[Rick_SAS · Accepted solution]

I don't think the ICC gives you a statistic for each patient. The ICC is a scalar value that combines information across all groups. You get one statistic that describes how closely the measurements within a group agree with each other. In your example, you have specified that the patients are the groups. If all the patients have ratings that are close together (like {10, 9, 11}) you will conclude that the correlation within classes is strong.

Think of the analogy with the usual (Pearson correlation). You don't get a correlation for each observation, you get one correlation for the whole data set.

[StatDave]

Actually, it is possible to obtain within-subjects measures if your data represent the ratings from a set of raters on a set of subjects (or items) and your interest is the agreement between raters or within subjects. The generalized linear mixed models (GLMM) approach available with the MAGREE macro can provide estimates of the random effects for subjects as well as raters. For each subject or rater, the macro can provide measures of the departure from agreement. Example 2 in the Results tab of the macro documentation illustrates how rater- and subject-specific agreement statistics can be obtained and interpreted. If your PTID variable represents the subjects and your _NAME_ variable represents the raters, then the following macro call requests the GLMM-based statistics. However, to avoid problems in fitting the model required for this method, you will probably need more than three subjects. But more subjects does not guarantee that model fitting problems won't occur.

%magree(data=trans, items=ptid, raters=_name_, response=col1,
        stat=glmm, options=ratereffects itemeffects )