statistical methods for assessing drug-cancer associations have changed. In
the original smaller database we computed standardized incidence ratios
(SIRs) and evaluated statistical significance by applying confidence limits
to the observed number of cases using the Poisson distribution. In our
studies employing the newer larger database we continued our retrospective
cohort study design with SIRs34,
but soon switched to Cox proportional hazards analysis35,36,
and then to a nested case-control design in the study cohort, using
conditional logistic regression analysis37-39.
found evidence (next two slides) that the switch from 10 to 50 controls per
case was worthwhile when there are relatively few exposed subjects.