If a CLASS or STRATA variable has a missing value, then PROC MULTTEST removes that observation from the analysis.

When there are missing values for test variables, the within-group-and-stratum sample sizes can differ from variable to variable. In most cases this is not a problem; however, it is possible for all data to be missing for a particular group within a particular stratum. For continuous variables and Freeman-Tukey tests, PROC MULTTEST re-centers the contrast trend coefficients within strata where all data for a particular group are missing. Re-centering the MEAN tests could redefine your t tests in an undesirable fashion; for example, if you specify coefficients to contrast the first and third groups (contrast -1 0 1) but the third group is missing, then the re-centered coefficients become –0.5 and 0.5, thus contrasting the first and second groups. If this is the case, you can run your t tests in separate PROC MULTTEST invocations, then combine the data and adjust the p-values by using the INPVALUES= option. However, you might find this re-centering acceptable for the Freeman-Tukey trend tests, since the contrast still tests for an increasing trend. The Cochran-Armitage and Peto tests are unaffected by this situation.

PROC MULTTEST uses missing values for resampling if they exist in the original data set. If all variables have missing values for any observation, then PROC MULTTEST removes the observation prior to resampling. Otherwise, PROC MULTTEST treats all missing values as ordinary observations in the resampling. This means that different resampled data sets can have different group sizes. In some cases it means that a resampled data set can have all missing values for a particular variable in a particular group/stratum combination, even when values exist for that combination in the original data. For this reason, PROC MULTTEST recomputes all quantities within each pseudo-data set, including such items as centered scoring coefficients and degrees of freedom for p-values.

While PROC MULTTEST does provide analyses in missing value cases, you should not feel that it completely solves the missing-value problem. If you are concerned about the adverse effects of missing data on a particular analysis, you should consider using imputation and sensitivity analyses to assess the effects of the missing data.