topic Re: Marginal effects with multiple imputed and bootstrapped data in Statistical Procedures

Marginal effects with multiple imputed and bootstrapped data

Alice86 — Tue, 22 Aug 2017 18:21:31 GMT

Dear all,
I need to calculate the average marginal effects for some logistic regression models im running. I use proc surveylogistic because my data must be bootstrapped and I was able to run marginal effects for some models. In other models where I have to impute data, I use proc surveylogistic followed by proc analyze.. I'm at a lost as to how I can calculate the marginal effects of the final estimates.

If anyone has any ideas please let me know. Thanks in advance!

Re: Marginal effects with multiple imputed and bootstrapped data

SAS_Rob — Wed, 23 Aug 2017 13:15:36 GMT

In order to use Proc MIANALYZE to combine the estimates, you would also need to have a standard error of the marginal effect (or some measure of variation). I suppose this could be done using the Delta Method, but I do not know of any published examples where someone has done this.

Re: Marginal effects with multiple imputed and bootstrapped data

Alice86 — Wed, 30 Aug 2017 16:09:29 GMT

Thanks a lot for your response Rob!

I essentially followed this step from SAS support (http://support.sas.com/kb/22/604.html😞

Marginal effects using results from PROC LOGISTIC

To compute the marginal effects using results from a model fit with PROC LOGISTIC (or other modeling procedure), include an ODS OUTPUT statement to write the ParameterEstimates table to a data set. Also specify the OUTPUT statement to save the predicted probabilities for a logistic model (use the P= option), or the x'b values for a probit model (use the XBETA= option) to a data set.

      proc logistic data=remiss;
        model remiss(event="1")=blast smear;
        ods output parameterestimates=logparms;
        output out=outlog p=p;
        run;

Use PROC TRANSPOSE to arrange the parameter estimates as a row and rename them so as not to conflict with the original variable names.

      proc transpose data=logparms out=tlog (rename=(blast=tblast smear=tsmear));
        var estimate;
        id variable;
        run;

Then use a DATA step to combine the parameter estimates and OUT= data sets and compute the marginal effects for each observation in the original data. Only the marginal effects for the response level representing the event of interest (REMISS=1) are computed below. The marginal effect for REMISS=0 could be similarly computed.

      data outlog;
        if _n_=1 then set tlog;
        set outlog;
        MEffBlast = p*(1-p)*tblast;
        MEffSmear = p*(1-p)*tsmear;
        run;
      proc print noobs;
        var smear blast MEff:;
        run;

I used mianalyze to get the appropriate standard errors but used the imputed runs to calculate the average marginal effects 🙂

Thanks again!