Conditional logistic regression is used to analyze matched observations in cohort and case-control epidemiologic studies, and other clustered-data situations where interest centers on within-cluster association of a dichotomy with predictor variables. Inference from conditional logistic regression models is based on the randomization likelihood from equiprobable reassigments of the observed predictor variable patterns within each cluster among the cluster members, independently across clusters. Both asymptotic and exact methods of inference are available.
When these predictor variable patterns are random rather than preselected, the frequentist properties of inferential methods using the conditional likelihood thus depend on the multivariate predictor variable distribution. Generally, the same issues arise as with analysis of variance or covariance of the predictor variables in groups defined by the dichotomy. For example, in a 1:1 matched case-control study with one predictor, the randomization likelihood depends on the matched case-control differences, as does the paired t-test. Distributional properties that degrade the power of the paired t-test, such as heavy tails, similarly affect tests based on conditional logistic regression. Conditional logistic regression analogues of standard rank tests may perform better under these circumstances, in this simple situation and in models involving adjustment by covariates.
After describing the nature of the above problem, we show that conditional logistic regression analogues of standard rank procedures are easily derived and implemented. Practical significance will be illustrated through an example from the occupational epidemiology of nasopharyngeal carcinoma, where adjusted sign tests from conditional logistic regression models grossly outperform competitors in detecting occupational associations.
Learning Objectives: At the conclusion of this session, the participant will be able to:
Keywords: Statistics, Biostatistics
Presenting author's disclosure statement:
Organization/institution whose products or services will be discussed: None
I do not have any significant financial interest/arrangement or affiliation with any organization/institution whose products or services are being discussed in this session.