I've been asked to produce significance statistics for a number of surveys with very small subgroups. Some of the tables have 0 cells, and some have cells with less than 5 observations in them. Normally I produce CHISQs out of SURVEYFREQ for categorical variables, weighted TTESTS for continuous variables, and output the stats via ODS and apply data-driven stars in my reports where appropriate. In this case however, CHISQs are not appropriate for the small cells and just plain won't work for the 0 cells. Does anyone have a work-around for this? Any suggestions greatly appreciated. (I tried McNemars for the 2 x 2 tables which work, no idea if they are appropriate - but McNemars doesn't work for the 2 x >2 tables - these are all comparisons of two groups. Fishers won't work at all. Wald Chis (like the other Chis) also won't work.)
Hi.I am not familiar with proc surveyfreq.
But For your situation,It looks like the best way is to use exact analysis or drop/merge these problemed cells. Maybe also you can use Corresponding Analysis to see whether it is significant( But it is only referenced method).
For McNemars ,it is only suited for paired data ,but it is also not suited for your situation.
Chisquare Distribution is only used for enough freq ( > 5 for each cell ).
So I highly recommend you to calculate exact probability distribution by exact analysis.