10-10-2013 07:17 PM
I'm conducting a conditional logistic regression analysis with an interaction term between a continuous variable ('exposure') and an ordinal variable ('age,' with four levels); there are twoadditional covariates, one of which is categorical with three levels (education) and one of which is continuous (smoking)). The log regression coefficients correspond perfectly with the results of contrast statements I'm using to generate confidence intervals for the interaction term (e.g., looking at the effect of my exposure at different age quartiles). I'd like to generate a forest plot type graph, but the 'oddsratio' statement/graph is giving me very difference results from what I see in the logistic regression results/contrast results. My guess is that the 'oddsratio' statement is not evaluating the effect using the same reference category as the model statement but I've been unable to get this squared away. The code is below. Any ideas? Thanks in advance.
proc logistic data=set descending;
class education (ref='0')/param=ref;
model case_status=age smoking education exposure age*exposure /expb cl;
oddsratio 'dose' exposure / at (age= 0 1 2 3);
contrast 'age=0 exposure 5 vs 0' age_quart 0 exposure 1 age*exposure 0/e estimate=exp;
contrast 'age=1 exposure 5 vs 0' age_quart 1 exposure 1 age*exposure 1/e estimate=exp;
contrast 'age=2 exposure 5 vs 0' age_quart 2 exposure 1 age*exposure 2/e estimate=exp;
contrast 'age=3 exposure 5 vs 0' age_quart 3 exposure 1 age*exposure 3/e estimate=exp;
10-11-2013 07:57 AM
Probably not the base cause, but since age_quart is not in the model statement, the contrasts strike me as non-stimable, as written. Other things--you say smoking is continuous, but that means that the contrasts will be calculated at the mean value for smoking. Does that even make sense in terms of your research question?
10-11-2013 08:33 AM
Thanks for the feedback, very much. Age and age_quart are the same variable; I forgot to delete '_quart' when trying to simplify the code. Yes, that does make sense actually, thanks for the clarification. So, if I set cigarettes to zero in the oddsratio statement I should be able to duplicate the results is my guess? Is that feasible to your knowledge? Thanks again.
10-11-2013 10:45 AM
That rings a bell somewhere. Setting to zero should give some measure of concurrence in the results.