topic Re: Mean difference test for multiple imputed dataset in Statistical Procedures

Mean difference test for multiple imputed dataset

Recep — Sun, 31 Jan 2021 19:53:42 GMT

Hello there,

I'm trying to calculate the mean difference of a variable among two independent groups. I know that the variable is NOT normally distributed so I use the Wilxocon test below:

proc npar1way data = sample_data wilcoxon;
class gender;
var test_score;
run;

The data "sample_data" is generated by multiple imputation with almost 40 imputations. The original data has about 10,000 records which brings the imputed data to 400,000 records (40x10,000). Since the sample size is artificially very big due to imputations the mean difference looks extremely statistically significant.

I was wondering how can I take into account that the data is constructed by 40 imputed data so that the significance of the mean difference can be calculated appropriately? Is there an option in the "proc npar1way" or do I need to use the estimators produced by PROC MIANALYZE in the next steps to calculate the mean difference of the two samples?

Thanks a lot in advance!

Re: Mean difference test for multiple imputed dataset

Ksharp — Mon, 01 Feb 2021 12:09:45 GMT

Calling @Rick_SAS

Re: Mean difference test for multiple imputed dataset

Rick_SAS — Mon, 01 Feb 2021 12:17:33 GMT

You have to include

BY _Imputation_;

in your analysis. That will give you 40 estimates. You can then combine those estimates into a single estimate by using PROC MIANALYZE or some other adjustment method.

Re: Mean difference test for multiple imputed dataset

SAS_Rob — Mon, 01 Feb 2021 12:27:47 GMT

If you want to combine the results from the Wilcoxon test then you will need to include BY _IMPUTATION_ in your NPAR1WAY code as was already mentioned. But you will also not be able to use MIANALYZE to combine the nonparametric test but instead will need to combine the actual Chi-Square test statistics that are generated by the Wilcoxon test. Dr. Paul Allison has a MACRO on his website called %COMBCHI that does this correctly.

https://www.sas.upenn.edu/~allison/combchi.sas