topic Re: Identifying individual-level outliers in a repeated measures study in Statistical Procedures

Identifying individual-level outliers in a repeated measures study

pronabesh — Fri, 03 Jan 2014 09:35:41 GMT

Hi SASusers,

I am looking to identify outliers in a repeated measures study, where study participants were measures at baseline, 1, 2, 6, 12 and 18 months. The measures of interest are neuro psych test scores which seem to have improved with time due to learning and re-test effect.

I was trying to adapt an approach by Welch et al (paper attached) to identify individual-level outliers from longitudinal data by fitting a random-effects model with subject-specific slopes and examining the standardized residual . However, the standardized residual from my model is the residual deviation from the sample parameter estimates and not within individual. Is there any way to obtain standardized residual at individual-level?

The code I used is below:

proc mixed data=outliers noclprint covtest;
    class studyid time;
    model outcome=time/ddfm=bw outp=predicted;
    random intercept time/subject=studyid type=ar(1) gcorr;
    lsmeans timefr/diff cl alpha=0.2 slice=time;
    ods output diffs=diff1 lsmeans=lsm1;
run;

Re: Identifying individual-level outliers in a repeated measures study

AnalytX — Fri, 03 Jan 2014 12:22:19 GMT

Hi,

You can have a look to this paper named "Mixed Model Inﬂuence Diagnostics".

http://www2.sas.com/proceedings/sugi29/189-29.pdf

Best,

Re: Identifying individual-level outliers in a repeated measures study

SteveDenham — Fri, 03 Jan 2014 20:10:53 GMT

Use 's reference, and look at conditional residuals.

I want to address some other issues. With the time points being unequally spaced, an ar(1) structure isn't appropriate. Here is what I might consider using:

proc mixed data=outliers noclprint covtest plots=all;
    class studyid time;
    model outcome=time/ddfm=bw outp=predicted;
    random intercept /subject=studyid gcorr;

    repeated time/subject=studyid type=<I would suggest SP(POW). Other possibilities include CSH or UN, but for those I would drop the random statement above>;
    lsmeans time/diff cl alpha=0.2;* slice=time; /* No interactions so slice option can be dropped */
    ods output diffs=diff1 lsmeans=lsm1;
run;

Steve Denham

Re: Identifying individual-level outliers in a repeated measures study

pronabesh — Fri, 03 Jan 2014 22:00:39 GMT

Thank you Steve and AnalytX. As always, much appreciated!