For genetic epidemiological studies with binary outcomes, the case-only (CO) approach has been shown to be powerful for examining statistical interactions, in particular gene-environment interactions. For time-to-event outcomes, the CO approach has been extended in the context of randomized clinical trials (RCT), but has not yet been investigated in prospective observational data. We explore the CO approach for time-to-event outcomes in scenarios with main effects of different strength (small, moderate) and compare its results with classical Cox proportional hazard and logistic regression models. We use only the earliest observed events (as ‘cases’) in the CO approach and also consider censored events (as ‘controls’ in logistic regression) by a restricted follow-up scheme in a cohort design or a random subsample of these in a case-cohort design. In our simulation study, the CO approach was consistently valid in the cohort settings and had a similar power as the benchmark analyses. In contrast, in the case-cohort design, the CO approach was valid and more powerful only in the scenario with just one main effect. However, in the presence of two moderate main effects, estimators may be biased, with a moderately inflated type I error rate. In a real-world example of a cohort study, the CO design represents an efficient approach that can be applied at an early follow-up time. Under a variety of circumstances, the CO approach is as powerful as the standard models for time-to-event data in the cohort framework, but can be biased in the presence of two main effects in the case-cohort framework.