Estimating multi-level discrete-time hazard models using cross-sectional data: Neighborhood effects on the onset of adolescent cigarette use

Investigating the effects of social context (e.g., neighborhood or school context) on the timing of behaviors (such as cigarette use initiation) requires both multi-level modeling and event-history analysis, and often requires the construction of a retrospective person-period data set from cross-sectional data. In this article we describe procedures for constructing such a data set and discuss modeling strategies for estimating multi-level discrete-time event history models. We show that the estimation of two-level discrete-time models involves three distinct modeling assumptions (the assumptions that individual- and neighborhood-level covariates have the, same effect at all time points and the assumption that the baseline logit-hazard curves in each neighborhood are parallel) and discuss methods of relaxing and empirically testing each of these assumptions. Estimation can be simplified in some cases if we additionally assume that the shape of the baseline logit-hazard curve in each neighborhood can be approximated by a simple functional form. The methods described here are applicable to a wide variety of questions where the dependent variable of interest is either onset or cessation. Here we apply these methods to the analysis of cigarette use initiation in a sample of 1,979 11- to 18-year-olds drawn from 79 neighborhoods of Chicago. We find that the racial composition of a neighborhood accounts for roughly half of the difference in age of smoking initiation between Black and White teenagers. Specifically, we find that living in a neighborhood with a large percentage of Black residents is associated with a lower hazard of adolescent cigarette use initiation than is living in neighborhoods with few Black residents.