Sociology Hosted 2019 Clogg Lecture

“Identification of Treatment Effects in Fixed Effects Models for Longitudinal and Clustered Data: Problems, with Illustrations from Demography”
When Apr 25, 2019
from 11:00 AM to 12:00 PM
Where 406 Oswald Tower
Add event to calendar vCal

In observational studies where the goal is to estimate the effects of a causal variable, it is typically necessary to assume all confounders are measured.  In many situations where this assumption is made, it is not credible, leading to biased estimates of effects.  For longitudinal data, and for clustered data, where respondents are observed within a hierarchical structure, fixed effects models are often used to adjust for unobserved confounders, associated with the observed confounders, that are constant over time (the longitudinal case) or common to all units in a cluster.   Using such models, coefficients associated with the causal variable are then deemed effects.  But the causal estimand is never defined, and in the longitudinal case, there are many causal estimands of potential interest. Further, the ignorability conditions that would need to be met in order to impart a causal interpretation to estimated coefficients (or functions of these) in fixed effects panel data models will almost certainly not be met in observational studies.  I illustrate these points using the literature from economics on the effect of marriage on men's earnings.  In the clustered case, units within a cluster are modeled under the assumption of no interference between units.  Despite talk of spillovers between units, this is not reflected in the modeling. When the no interference assumption is relaxed, many possible estimands can be defined, and, depending on the estimand, it may become necessary to deal with the timing of events in different units.  New identification conditions are also needed in this case.  These points are illustrated using the literature on teenage pregnancy.