MIANALYZE w/ LOGISTIC: how to get global null hypothesis test as output?

[EricVanceMartin]

Using multiply imputed data with nominal categorical variables that have been recoded as k-1 dummy variables.

I want to know: Does the predictor variable overall have a (bivariate) relationship with the outcome?

My strategy has been to do a series of separate logistic regressions with each variable (or group of dummies), and I want a global null hypothesis for each regression. (I am doing this in lieu of bivariate correlations, since this does not seem to make sense with dummy coded categorical variables. If you have other thoughts, let me know.)

MIANALYZE gives me the parameter estimates, but not the global test, which does appear for the individual regression for each MI dataset. I want the global test as an output of MIANALYZE. I see the TEST statement / MULTI option, but I'm not sure this is what I want.

Ideas?

[jf]

I don't really understand your question. But, when you do Proc Logistic, you will see this --"testing global null hypothesis : beta = 0" in output. the statistics for the test are likelyhood ratio, score and wald. either one will give you a global test result.

[EricVanceMartin]

Thank you for replying. Sorry if it wasn't clear.

I'm using a multiply imputed dataset (m=20). The results of PROC LOGISTIC (20 separate regressions) are passed to PROC MIANALYZE to summarize into a single result. My syntax takes the parameter estimates and covb as inputs and produces a single set of estimates with significance tests.

Of course PROC LOGISTIC also produces the global null hypothesis tests for each regression. I am asking how (or whether) I can get MIANALYZE to summarize the results of the global null hypothesis tests in the same way.

For example, maybe some way of averaging the ∆-2LL and df for a test. Also, as I mentioned, there is some kind of "MULTI" test option, but I don't understand what it does.

My discussion about the data (dummy coded categorical) was just to show why I couldn't simply do (and run through MIANALYZE) bivariate correlations, as one would ordinarily do to select variables for multiple regression (which is what I'm doing).

Message was edited by: Eric Martin

[SteveDenham]

Eric,

Check out the documentation for the TEST statement and in particular where it leads to Figure 58.6.  In that example, test3: a1=a2=a3 is a global test with two degrees of freedom.  So for your example, try

test: beta1=beta2=...=betan;

where the beta's should be replaced with the terms as presented in the MODELEFFECTS statement.

Steve Denham