- Mark as New
- Bookmark
- Subscribe
- Mute
- RSS Feed
- Permalink
- Report Inappropriate Content
Multinom SAS/ IML module (Berry and Hurtado (1994) creates “Large-Sample Confidence Limits for Pi-Pj” in the results. (That is the heading in the results.) So, what is the minimum requirement of the size for large sample to use this module? Does it refer to the total sample or count for each category? Can this module or Catmod be used for small sample size (<15)?
Thanks for your help.
Accepted Solutions
- Mark as New
- Bookmark
- Subscribe
- Mute
- RSS Feed
- Permalink
- Report Inappropriate Content
That language typically means that the results are based on the asymptotic distribution of a statistic. There is no minimum size requirement, but the coverage probability of the confidence intervals approach their nominal values (eg, 95% coverage) for large samples. Typically in these types of computations, between-group proportions that are nearly equal (p1=p2=...=pn) have coverage probabilities that are better than for extreme proportions (p1=0.9, p2=0.05, p3=0.05).
If you forced me to give a number, I would guess n=100 if you have fairly even proportions and five or fewer groups. But that is just a guess.
If you are interested in CIs for multinomial proportions, you might want to look at the article "Simultaneous confidence intervals for multinomial proportions,"
which implements the methods in May and Johnson (1997) (Computer Methods and Programs in Biomedicine, p. 153–162).
- Mark as New
- Bookmark
- Subscribe
- Mute
- RSS Feed
- Permalink
- Report Inappropriate Content
Calling @Rick_SAS
- Mark as New
- Bookmark
- Subscribe
- Mute
- RSS Feed
- Permalink
- Report Inappropriate Content
That language typically means that the results are based on the asymptotic distribution of a statistic. There is no minimum size requirement, but the coverage probability of the confidence intervals approach their nominal values (eg, 95% coverage) for large samples. Typically in these types of computations, between-group proportions that are nearly equal (p1=p2=...=pn) have coverage probabilities that are better than for extreme proportions (p1=0.9, p2=0.05, p3=0.05).
If you forced me to give a number, I would guess n=100 if you have fairly even proportions and five or fewer groups. But that is just a guess.
If you are interested in CIs for multinomial proportions, you might want to look at the article "Simultaneous confidence intervals for multinomial proportions,"
which implements the methods in May and Johnson (1997) (Computer Methods and Programs in Biomedicine, p. 153–162).
- Mark as New
- Bookmark
- Subscribe
- Mute
- RSS Feed
- Permalink
- Report Inappropriate Content
Thanks. I appreciate your information and the SAS/IML function. I tried the function for different sample sizes to check the effect on the results.