topic Re: PROC SURVEYLOGISTIC and interaction: 95%CI for OR of X(=0,1) at Time=t for X*Time in New SAS User

PROC SURVEYLOGISTIC and interaction: 95%CI for OR of X(=0,1) at Time=t for X*Time

slegleye — Sun, 11 Nov 2018 19:42:09 GMT

Hello everyone.

I try to compute OR and 95%CI for OR for a binary variable X interacting with a continuous variable TIME (in the product X*Time) for some values of Time=t in PROC SURVEYLOGISTIC. I have SAS 9.4 TS, Level 1M5.

Here is the code :

proc surveylogistic data=work;
cluster username ;
class X /param=glm;
model attrition(ref='0')=X TIME X*TIME / varadjust=morel expb rsq ;
run ;

There is an option in PROC LOGISTIC, but I cannot find any way for doing the same in PROC SURVEYLOGISTIC.

Any suggestion is welcome !

Re: PROC SURVEYLOGISTIC and interaction: 95%CI for OR of X(=0,1) at Time=t for X*Time

Rick_SAS — Mon, 12 Nov 2018 01:14:16 GMT

I think you are looking for the CLODDS option on the MODEL statement.

Formulas/explanation

Re: PROC SURVEYLOGISTIC and interaction: 95%CI for OR of X(=0,1) at Time=t for X*Time

slegleye — Mon, 12 Nov 2018 09:59:19 GMT

Thank you, Rick.

In fact I want to compute OR for a time-varying variable, at certain time points: I do time-discrete logistic regression, TIME is continuous and X is binary and time-varying.

The database WORK is in person-year format : each individual in my panel (experiment) is observed (ATTRITION=0) until he/she is excluded (ATTRITION=1). For an individual who is observed n-1 times, ATTRITION has value 0 the first n-1 lines and ATTRITION has value 1 the last line at time=n when he/she is excluded. PERSONYEAR has value n-1 for this individual.

If I use simply PROC LOGISTIC, I can use the ODDSRATIO statement :

PROC LOGISTIC data=work ;
class X/param=glm ;
Model attrition(ref='0')=TIME X TIME*X PERSONYEAR ;
ODDSRATIO X /AT (PERSONYEAR=1 18 36) ;
RUN ;

The problem is that using proc logistic provides underestimated standard errors of the OR. Using PROC SURVEYLOGISTIC with the statement CLUSTER ID (ID being the identification number of the individuals) provides correct standard errors. But it seems there is no similar statement in PROC SURVEYLOGISTIC.

In PROC SURVEYLOGISTIC, it seems that the option CLODDS cannot be used with AT suboption. Or maybe I am wrong ?

Best,

Re: PROC SURVEYLOGISTIC and interaction: 95%CI for OR of X(=0,1) at Time=t for X*Time

Rick_SAS — Mon, 12 Nov 2018 20:01:37 GMT

I see. In that case, I think you want to use the LSMEANS and SLICE statements, as described in the first example in @SAS_Rob Agnelli (2014), which seems similar to your design.

Be sure to use the GLM parameterization, as you are already doing. The LSMEANS statement supports the ODDSRATIO and CL option, and the SLICE statement enables you to specify the X values at which you want to estimate the odds ratio.

Re: PROC SURVEYLOGISTIC and interaction: 95%CI for OR of X(=0,1) at Time=t for X*Time

slegleye — Mon, 12 Nov 2018 19:38:24 GMT

Hello Rick, hello everybody.

Thank you again.

You picked up the right example in the paper by Agnelli (2014): I also read it. It concerns categorical variables and unfortunately, it does not work for continuous variables...

In my case, with X*TIME, TIME being continuous and X binary, and the statement "lsmeans X*TIME/oddsratio cl diff;", I get an error message : "ERROR: Only CLASS variables allowed in this effect".

Any other suggestion ?

Re: PROC SURVEYLOGISTIC and interaction: 95%CI for OR of X(=0,1) at Time=t for X*Time

ChuksManuel — Thu, 30 Jul 2020 05:32:06 GMT

Hello,

DId you finally solve this problem?

I have a similar problem of interaction between a categorical and a continuous variable using proc surveylogistic and i could not figure it out.Let me know how you solved it.