topic Proc mixed: Using 2 or more Subject IDs in Statistical Procedures

Proc mixed: Using 2 or more Subject IDs

alothary11 — Sat, 30 May 2020 11:20:15 GMT

I am using a dataset with both household and person identifiers. For example the household ID may have 5 or so numeric values (00001) and then there would be a personal identifier in the household (1). How can I load both identifiers into my models? I have tried to add the variables or simply list both but could not figure out how to properly use both in my model.

proc mixed data=hrs noclprint covtest;
class HHID PN;
model age=
/solution ddfm=bw;
random intercept / subject= HHID PN;
run;

Re: Proc mixed: Using 2 or more Subject IDs

ChrisNZ — Sat, 30 May 2020 11:21:22 GMT

Re-titled and moved to stat procedures.

Re: Proc mixed: Using 2 or more Subject IDs

PaigeMiller — Sat, 30 May 2020 11:29:29 GMT

@alothary11 wrote:

I am using a dataset with both household and person identifiers. For example the household ID may have 5 or so numeric values (00001) and then there would be a personal identifier in the household (1). How can I load both identifiers into my models? I have tried to add the variables or simply list both but could not figure out how to properly use both in my model.

proc mixed data=hrs noclprint covtest;
class HHID PN;
model age=
/solution ddfm=bw;
random intercept / subject= HHID PN;
run;

What happens when you do it this way? Do you get an error? What error?

You could modify the input data set to concatenate HHID and PN into a single variable.

Re: Proc mixed: Using 2 or more Subject IDs

alothary11 — Sat, 30 May 2020 13:38:39 GMT

Thank you, that is what I ended up doing. I had to concatenate the variables and then use the input function to create a numeric value.

data ex;set ex.exall;
ID_all = cats(HHID,PN);
run;
data ex2; set ex.exall;
ID= input(ID_all, 8.);
run;

Just posting what I found so others might be able to search and find it. Thanks for the replies!

Re: Proc mixed: Using 2 or more Subject IDs

SteveDenham — Mon, 01 Jun 2020 12:23:37 GMT

That would enable you to model a single variance component based on the composite variable. However, if you wish to model a 2 level design, where person is nested within household, you might try:

proc mixed data=hrs noclprint ;
class HHID PN;
model age=
/solution ;
random intercept HHID/ subject= PN;
run;

I removed the covtest option from the PROC MIXED statement, as it depends on a Wald test that is really not a good test given the distribution of variances, and also removed the ddfm=bw, as you now want a containment hierarchy for the variance components.

If you truly need to test whether the variance components are non-zero I would recommend changing to PROC GLIMMIX and using the options under the COVTEST statement, which give likelihood ratio based tests that are more appropriate.

SteveDenham