B0760
Title: Generalised measurement error models
Authors: Thomas Augustin - LMU Munich (Germany) [presenting]
Abstract: From social sciences and econometrics to biometrics and technometrics, measurement error is an omnipresent problem: many variables of interest are latent and prone to non-negligible measurement error. A bundle of methods has been developed in the literature to adjust inference for error-proneness of data. However, these methods typically rely on the so-called classical error model/testing theory, making rather strong, and often untestable, assumptions about the eo ipso unobservable error process. However, in the spirit of Manski's Law of Decreasing Credibility, untenable assumptions undermine the practical relevance of the results. Against this background, generalised measurement error models are proposed. The measurement error distribution is flexibly described by a credal set of probability distributions, for instance, by a neighbourhood model or by a parametrically constructed model based on a set of parameter values. Nakamura's method of corrected score functions will be extended to deal with set-valued expectations, enabling the derivation of set-valued estimators that satisfy some generalised criteria of unbiasedness. Finally, some ideas are discussed on how to overcome in addition some doubtful independence assumptions in the measurement error modelling process.
