topic Re: PROC FREQ McNemar test in Statistical Procedures

PROC FREQ McNemar test

cminard — Thu, 21 Jan 2021 18:22:26 GMT

Why can't SAS compute McNemar's test when zeros are in a column? I think that it is because SAS doesn't "know" that there are zeros.

For example, consider the following data set for a hypothetical case-control study.

data one;
	input caseevent controlevent number;
	cards;
	0 0 0
	0 1 9
	1 0 0
	1 1 15
	;
run;

In this data set, controls are matched to cases. We have 0 case-control pairs where both the case and the control do NOT have the event.

There are 15 pairs where they both have the case.

For discordant pairs, we have 9 instances where the control had an event, and the case did not.

Also, we have 0 instances where the case had an event, and the case did not.

Running PROC FREQ and requesting McNemar test:

proc freq data=one;
	tables caseevent*controlevent / agree;
	weight number;
	exact mcnemar;
run;

We get the appropriate table

But, there is no McNemar test and SAS log gives a NOTE:

So, apparently SAS thinks we are missing data.

I can calculate McNemar test (with or without continuity correction) manually. In this case X2=(|9-0|^2)/9=9 with one degree of freedom.

Or continuity corrected X2=((|9-0|-1)^2)/9=7.1.

But, it there some way to get sas to do this without having to program it? Further, I would prefer the Exact McNemar test. But, maybe I just need to program conditional Binomial tests for this.

Re: PROC FREQ McNemar test

SAS_Rob — Thu, 21 Jan 2021 18:45:26 GMT

You will need to add the ZEROS option to the WEIGHT statement.

proc freq data=one;
tables caseevent*controlevent / agree;
weight number/zeros;
exact mcnemar;
run;

Re: PROC FREQ McNemar test

cminard — Thu, 21 Jan 2021 20:46:12 GMT

Thanks, did not know about the zero option. I suppose that if I have raw data, then this is not going to work. I would have to count the observations in each cell first and create a data set with these summary statistics (frequencies). Then, I could use the zero option with the weighted statement.

Re: PROC FREQ McNemar test

SAS_Rob — Thu, 21 Jan 2021 21:05:10 GMT

Yes, with raw data you could assign a count of 1 to each observation and then add an observation with the missing row/column combination and a count of 0.