01-25-2013 12:08 PM
I am researching if pathogens are more or less infectious on the hosts from which they have been isolated.
So there is data on, say 4 pathogen lines (path), tested for %infection success (number diseased (d)/total inoculated(tot))) on all 4 hosts (host) from which they were originally isolated. In the infection matrix, the values on the leading diagonal are pathogens on their own hosts (home=1), while off diagonal are pathogens on novel hosts (home=0).
What model statements in GENMOD would test for a leading diagonal (home vs. away) effect?
Two approaches I tried using GENMOD give very different siginificance values
Approach 1 -
Get residuals from
Model d/tot = path host /link=logit dist=binom p r ;
Test home vs. away raw residuals - gives significant leading diagonal effect (P<0.02).
Visual inspection and plotting shows that the 4 leading diagonal resisduals are larger than all except one of the 12 off diagonal values, and the effect "looks real".
Approach 2 -
Compare AIC and likelihood for the above model, with the following more specified model
d/tot = path host home/link=logi dist=binom
Home has a non-siginificant (P=0.23) effect, and comaprison of model with home included gives an AIC improvement of only 0.5, and a likelihood improvement of ca. 0.7.
So approach 2 gives no evidence of a significant leading diagonal, home vs. away effect.
01-25-2013 09:41 PM
Hi, I assumed that Tot=D+H. After checking for interactions (home*path) and overdispersion, I arrived at the same result and conclusion as your second approach. I used proc logistic. - PG