This paper proposes a methodology that accounts for the selection effect due to non‐random entry in duration models using latent‐class models. A mixed proportional hazard model with continuous finite mixture unobserved heterogeneity (MPH‐CFM) is introduced to correct for the potential bias induced by the selection effect. Conditions for identification, consistency, and asymptotic normality of the MPH‐CFM are provided. The estimator is used to investigate the duration of new entrant Canadian manufacturing firms. For the current application, the MPH‐CFM is compared with alternative duration models and found to be superior. Empirically, the results indicate that there are two classes of firms. Class I starts with high hazard and decreases non‐monotonically while Class II has a negligible hazard. These empirical results can be used to understand alternative models of firm dynamics. Copyright © 2017 John Wiley & Sons, Ltd.