topic Re: Mixed models for correlated repeated measurement in Statistical Procedures

Mixed models for correlated repeated measurement

Yasu — Thu, 05 Jan 2023 08:07:04 GMT

Hi,

I'm trying to analyze correlated repeated measurements (e.g. systolic blood pressure and diastolic blood pressure) with proc MIXED.

I want to analyze both outcomes simultaneously and consider the correlation between variables as that between random effects (random intercepts for each variable).

How should I specify the repeated statement (especially subject= option)?

Below is an example of the data and code that I want to implement .

data bp;

do id=1 to 100;

do bp=1 to 2;

do visit=0 to 4;

input aval @@;

output;

end;

cards;

180 165 155 150 140

110 95 90 85 80

185 175 160 150 135

110 90 85 85 80

...

;

int1=(bp=1);

int2=(bp=2);

run;

proc mixed data=bp;

class id bp visit;

model aval = bp*visit;

random int1 int2 / type=un subject=id G;

repeated visit / type=un subject=id R;

run;

Re: Mixed models for correlated repeated measurement

PaigeMiller — Thu, 05 Jan 2023 11:34:30 GMT

https://communities.sas.com/t5/Statistical-Procedures/Repeated-and-Random-effect-in-Proc-Mixed/m-p/413708#M21713

https://support.sas.com/resources/papers/proceedings14/SAS026-2014.pdf

Re: Mixed models for correlated repeated measurement

Yasu — Fri, 06 Jan 2023 02:12:05 GMT

Sorry, there is a typo in the description of repeated statement. I specified the repeated statement as follows:

repeated bp*visit / type=un subject=id R;

Re: Mixed models for correlated repeated measurement

StatsMan — Fri, 06 Jan 2023 14:00:23 GMT

The Kronecker product structures on the REPEATED statement might be helpful here. Look into TYPE=UN@AR(1) or UN@UN. These structures set up an unstructured correlation across the two BP series and then either AR(1) or UN within a series. Alternatively, another common approach is to use the REPEATED statement to model correlation within one of the BP measures (using TYPE=AR(1) or UN or whatever you would like with SUBJECT=BP*PATIENT) and then adding correlation for all observations from the same subject through a RANDOM statement (random int / subject=patient).