Mechanism-Based Models with Nonparametric Bounds: A General Strategy for Estimating Age-period-cohort Effects

• In Event: 3269 - Statistical Advances

Abstract

In this paper we outline and extend the mechanism-based approach to identifying age-period-cohort (APC) models. The mechanism-based approach provides identification using Pearl's front-door criterion, but only under the assumption that at least one of the pathways between the temporal variables and the outcome of interest is fully specified. This can be a potentially strong requirement, especially when a limited set of variables are available to the researcher. However, we show that this assumption can be weakened since it is only the linear components of an APC model that are unidentified; accordingly, it is only the linear components that require mechanisms for identification. Furthermore, we provide a sensitivity analysis of the mechanism-based approach which can be used to assess the robustness of estimated APC effects in the presence of unobserved mechanisms. We illustrate our approach by examining the temporal effects of verbal ability, smoking, and subjective well-being. We conclude with a discussion of the advantages of mechanism-based models even when such models are known to be incomplete or partially misspecified.

Author

• Ethan Fosse, Princeton University