p-value for CMH

[HappySASUE]

Is there a SAS procedure to calculate p-value for Mantel-Haenszel Stratum Weighted Method for risk difference in stratified samples? It provided the CI but I can't find p-value for this test. 

SAS code I used is

ods output CommonPdiff=CommonPdiff   ;

proc freq data  = resp order=data;

       tables stratum1*stratum2*trt01pn*resp/riskdiff(common)  ;

run;

I can't find option from

https://documentation.sas.com/doc/en/pgmsascdc/9.4_3.4/statug/statug_freq_details.htm

Many thanks

[StatDave]

You can do this by taking a model-based approach. Fit a logistic model that includes your treatment variable and your stratifying variables and include an LSMEANS statement with the DIFF option to compare the treatments. For example, assuming that your binary response variable is Y with values 0 and 1, where 1 represents the event of interest:

  proc logistic;
         class trt strat1 strat2 / param=glm;
         model y(event="1") = trt strat1 strat2;
         lsmeans trt / ilink diff cl;
         run;

You might also want to look at the discussion in this note which is more important when you want to estimate as well as test the risk difference. That note shows how it can be estimated as either the marginal effect of treatment or as a difference of LS-means.

[Ksharp]

Try add CMH option in TABLE.

[HappySASUE]

Thanks. 

There are 3 p-values, which one I should use? 

[Ksharp]

This three P value correspond to three Hypothesis .
It look like trt01pn and resp have correlation/association after adjusting MUTCAT and HGBCAT .
But these three Hypothesis has nothing to do with RISK . the second statistic is saying the two treat(trt01pn) has different response(resp) . But is more like ANOVA not like RISK .

Or could you also try RELRISK option ?

[HappySASUE]

Hi Ksharp, 

Thank you for your patience and help. I added the option, but still no p-value. It is very strange no p-value but CI for MH test. 

ods output CommonPdiff=CommonPdiff CMH=cmhpval ;
proc freq data = resp order=data;
tables mutcat*hgbcat*trt01pn*resp/riskdiff(common) cmh RELRISK ;
run;

[StatDave]

If you don't want to take the model-based approach I mentioned earlier, then you can use the COMMONRISKDIFF(TEST) option in PROC FREQ. If you want to test the risk difference, then you don't want the RELRISK option which gives the ratio of risks, not the difference.

[HappySASUE]

Thanks. 

Does proc logistic also create the risk difference p value? 

Best, 

[StatDave]

Yes, the DIFF option in the LSMEANS statement gives a p-value for the difference. The downside with the model-based approach is that you might encounter problems fitting the model particularly if the data are made sparse by the model specification.

[Ksharp]

Or you could check CI ?
For RISKDIFF ,check if CI contains 0 .
For RELRISK ,check if CI contains 1 .

[HappySASUE]

I had already used logistic regression, need this more sensitivity analysis.