Interpreting Statistical Evidence by using Imperfect Models: Robust Adjusted Likelihood Functions

Abstract

The strength of statistical evidence is measured by the likelihood ratio. Two key performance properties of this measure are the probability of observing strong misleading evidence and the probability of observing weak evidence. For the likelihood function associated with a parametric statistical model, these probabilities have a simple large sample structure when the model is correct. Here we examine how that structure changes when the model fails. This leads to criteria for determining whether a given likelihood function is robust (continuing to perform satisfactorily when the model fails), and to a simple technique for adjusting both likelihoods and profile likelihoods to make them robust. We prove that the expected information in the robust adjusted likelihood cannot exceed the expected information in the likelihood function from a true model. We note that the robust adjusted likelihood is asymptotically fully efficient when the working model is correct, and we show that in some important examples this efficiency is retained even when the working model fails. In such cases the Bayes posterior probability distribution based on the adjusted likelihood is robust, remaining correct asymptotically even when the model for the observable random variable does not include the true distribution. Finally we note a link to standard frequentist methodology-in large samples the adjusted likelihood functions provide robust likelihood-based confidence intervals. Copyright 2003 Royal Statistical Society.

A simple and exploratory way to determine the mean–variance relationship in generalized linear models

“…Royall and Tsou [5] referred to the preserved consistency property as the first condition of robustness. They showed that likelihood functions on the basis of such working models could be adjusted to become robust if the condition is fulfilled.…”

Universal surrogate likelihood functions for nonnegative continuous data

“…In order to make inference about θ in the presence of φ, we can consider the profile likelihood function L(θ,φ(θ)), whereφ(θ) is the ML estimate of φ given θ (Kalbfleisch and Sprott 1970). Royall (1997) has elaborated that the profile likelihood is a proper representation of the evidence about θ . Now, let θ be the population mean and use the inverse Gaussian with the density function…”

“…If one applies the adjustment introduced by Royall and Tsou (2003) to L P (θ ), the resulting rescaled adjusted likelihood, denoted by L AP (θ ), is…”

“…situations one can modify the two likelihood functions for counts to become asymptotically identical to the profile likelihood for the mean parameter for general nonnegative continuous data. The idea of robust likelihood introduced by Royall and Tsou (2003) suffices to accomplish this task.…”

Measuring and illustrating statistical evidence in a cost-effectiveness analysis

“…Robust likelihood functions can be made insensitive to model misspecification (Royall and Tsou, 2003) in a regression setting (Blume et al, 2007). This is accomplished by adjusting the likelihood function, so that the adjusted likelihood behaves as if the model were correctly specified.…”