Identification and Estimation of Panel Data Models with Attrition Using Refreshment Samples

Abstract

This thesis deals with attrition in panel data. The problem associated with attrition is that it can lead to estimation results that suffer from selection bias. This can be avoided by using attrition models that are sufficiently unrestrictive to allow for a wide range of potential selection. In chapter 2, I propose the Sequential Additively Nonignorable (SAN) attrition model. This model combines an Additive Nonignorability assumption with the Sequential Attrition assumption, to just-identify the joint population distribution in Panel data with any number of waves. The identification requires the availability of refreshment samples. Just-identification means that the SAN model has no testable implications. In other words, less restrictive identified models do not exist. To estimate SAN models, I propose a weighted Generalized Method of Moments estimator, and derive its repeated sampling behaviour in large samples. This estimator is applied to the Dutch Transportation Panel and the English Longitudinal Study of Ageing. In chapter 4, a likelihood-based alternative estimation approach is proposed, by means of an EM algorithm. Maximum Likelihood estimates can be useful if it is hard to obtain an explicit expression for the score function implied by the likelihood. In that case, the weighted GMM approach is not applicable.

Publication
University College London Ph.D Thesis