topic Re: Pooling results after multiple imputations: For the proportional odds assumption tests in Statistical Procedures

Pooling results after multiple imputations: For the proportional odds assumption tests

ejay0503 — Mon, 18 Jun 2018 21:37:47 GMT

Hello, all

I used PROC LOGISTICS to run ordinal logistic models on 5 multiply imputed data (My outcome is ordinal with 3 categories).

I also included "the unequalslopes" function in order to check the proportional odds assumption. I was able to get pooled results on the 5 ordinal logit analyses but not on the 5 assumption test results. Is there a way to obtain pooled results on the proportional assumption tests (indicated in SAS output as "Linear Hypothesis Test Results").

I would greatly appreciate your help.

Here is my code:

proc logistic data=name; 
	by _imputation_; 
	model DV (order=data) = 	var1 	var2 	/  unequalslopes MAXITER =1000 ;
	var1:       test    var1_4 =	var1_3; 
	var2:       test    var2_4 =	var2_3; 
run;

data perm.name2print; set perm.nameprint; parameter = compress (parameter); run; 

proc mianalyze parms (classvar=classval)= perm.name2print;
 modeleffects intercept var1   var2 ;
 ods output parameterestimates = mianalyze_parms; 
run;

data OR;  
 set mianalyze_parms;
 OR=exp(estimate);
 LCL_OR=exp(LCLMean);
 UCL_OR=exp(UCLMean);

proc print; var parm OR LCL_OR UCL_OR;
run;
ods rtf close;

Re: Pooling results after multiple imputations: For the proportional odds assumption tests

pau13rown — Mon, 18 Jun 2018 21:52:35 GMT

it's not a good way to determine whether you believe proportional odds is a reasonable assumption in this case. That's quite a separate matter from the analysis and estimation

Re: Pooling results after multiple imputations: For the proportional odds assumption tests

ejay0503 — Mon, 18 Jun 2018 21:54:43 GMT

Thanks for the reply, Paul. Then what would you suggest to make sure the data meet the proportional odds assumption?

Re: Pooling results after multiple imputations: For the proportional odds assumption tests

pau13rown — Mon, 18 Jun 2018 22:30:52 GMT

it's more a matter of whether your audience will believe it and biological reasons or a priori reasons are more cogent than p-values, especially when simultaneously confessing missing data. A visual assessment is much more persuasive i think: http://support.sas.com/kb/37/944.html. Otherwise i think the stokes book (https://www.amazon.com/Categorical-Data-Analysis-Using-System/dp/0471224243) describes how to restructure your dataset to fit a partial proportional odds model

Re: Pooling results after multiple imputations: For the proportional odds assumption tests

ejay0503 — Mon, 18 Jun 2018 23:58:50 GMT

Thank you for the helpful links. I personally agree that a visual assessment is much more persuasive, but a proportional odds test is more popularly used in my field, because I think it provides more objective results. I very appreciate your prompt replies, but perhaps I should stick to the assumption tests though.

All the best,

Re: Pooling results after multiple imputations: For the proportional odds assumption tests

pau13rown — Tue, 19 Jun 2018 00:14:55 GMT

a final note regarding the significance test: i assume the null hypothesis is "Ho: everything is fine" ie "the proportional odds assumption is reasonable", thus the test is anti-conservative - all you need is a lot of missing data and you'll show every time that the proportional odds assumption cannot be rejected (using a significance test). Better to change your colleagues' default thinking on the matter