Dominitz, Jeff: Identification and Estimation with Contaminated and Partially Verified Data
World Conference Econometric Society, 2000, Seattle

Jeff Dominitz, Carnegie Mellon University and Resolution Economics, LLC
Robert P. Sherman, California Institute of Technology
Identification and Estimation with Contaminated and Partially Verified Data
Session: C-8-16  Monday 14 August 2000  by Dominitz, Jeff
A sample is said to be contaminated if it is drawn from a mixture of a distribution of interest, F, and another distribution, G. Without parametric assumptions on F and G, characteristics of F such as moments and quantiles are not identified. However, given a lower bound on the probability of being drawn from F, Horowitz and Manski (1995) show that finite bounds on these characteristics are identified and easily estimated. In some applications, more information is available: a subset of the sample is known to be drawn from F. Often these verified observations are not a random sample from F and so point estimation based only on verified observations is biased. The Horowitz and Manski procedure applies but is inefficient because it makes no use of the extra information on verification. This paper adapts the method of Horowitz and Manski to develop tighter bounds using the extra information. In addition, the method is more direct and easier to generalize to other models than an alternative method proposed by Lambert and Tierney (1997). The estimator is shown to be consistent and is applied to measurements of organic pollutant concentrations.


File created by Jurgen Doornik with eswc2000.ox on 2-01-2001