Date: Fri, 29 Sep 2006 08:33:10 -0400 Peter Flom "SAS(r) Discussion" Peter Flom Re: interactions (effect measure modification) To: Kevin Roland Viel text/plain; charset=US-ASCII

>>> Kevin Roland Viel <kviel@EMORY.EDU> 9/28/2006 5:58 pm >>> wrote <<< One published report performed separate analyses for male and females. A reviewer suggested that we also perform such an analyses. I disagree based on current knowledge, but I performed the analyses anyway. We have approximately 360 people in this study.

In the combined analysis using a interaction term the results were:

Score (-1,0,1): 22.1 (p = 0.17) Score X SNP: -6.1 (p = 0.59)

Females only

Score (0,1): 6.6 (p = 0.52)

Males only

Score(-1,1): 19.0 (p < 0.01)

In each analyses, score is an additive effect, i.e. one variable, up to three levels. Males cannot (usually, see PS) have a score of 0, but no female had a score of -1.

Under such scoring we might expect:

Score Mean -1 x 0 2x 1 3x

In other words, the means of the group with score = 0 is the average of the means of the groups with score = -1 and score = 1.

Translating the male results to the female results would mean the comparison at hand, 0/1, would have an effect size of 9.5. Translating the female results to the male results would mean the comparison at hand, -1/1, would have an effect size of 13.2.

I would like to reply that there is not enough evidence to suggest that sex is an effect measure modifier. Nominally, 6.6 is different from 9.5 (or 13.2 is different from 19.0), but I am not willing to say that a large sample size (with -1 females) would support the contention that an interaction term or separate analyses should be the main results.

Would anyone like to comment or suggest another type of analyses? >>>>

I may be missing something, but I don't understand what you are doing here. What was the DV, and how was it scored? Clearly there must be some IV besides sex, but what was it? Was there more than one? How was the IV that was in the interaction scored?

What statistical methods did you use?

In general, though, the question of whether to look at interactions or do stratified analysis should be answered based on the nature of the research question. Stratified analyses are easier to interpret, but do not give all the infomration that the analysis with interaction gives - in particular, they do not allow about the size and significance of the difference in effect size

HTH

Peter

Back to: Top of message | Previous page | Main SAS-L page