Divergence-based estimation and testing with misclassified data

The well-known chi-squared goodness-of-fit test for a multinomial distribution is generally biased when the observations are subject to misclassification. In Pardo and Zografos (2000) the problem was considered using a double sampling scheme and ø-divergence test statistics. A new problem appears if the null hypothesis is not simple because it is necessary to give estimators for the unknown parameters. In this paper the minimum ø-divergence estimators are considered and some of their properties are established. The proposed ø-divergence test statistics are obtained by calculating ø-divergences between probability density functions and by replacing parameters by their minimum ø-divergence estimators in the derived expressions. Asymptotic distributions of the new test statistics are also obtained. The testing procedure is illustrated with an example.     

Comparison of causal effect estimators under exposure misclassification

Over the past decades, various principles for causal effect estimation have been proposed, all differing in terms of how they adjust for measured confounders: either via traditional regression adjustment, by adjusting for the expected exposure given those confounders (e.g., the propensity score), or by inversely weighting each subject's data by the likelihood of the observed exposure, given those confounders. When the exposure is measured with error, this raises the question whether these different estimation strategies might be differently affected and whether one of them is to be preferred for that reason. In this article, we investigate this by comparing inverse probability of treatment weighted (IPTW) estimators and doubly robust estimators for the exposure effect in linear marginal structural mean models (MSM) with G-estimators, propensity score (PS) adjusted estimators and ordinary least squares (OLS) estimators for the exposure effect in linear regression models. We find analytically that these estimators are equally affected when exposure misclassification is independent of the confounders, but not otherwise. Simulation studies reveal similar results for time-varying exposures and when the model of interest includes a logistic link.     

Robustness of the linear discriminant function to nonnormality: Johnson's system

The asymptotic distribution of the Errors of Misclassification in using the Linear Discriminant Function is investigated here. The purpose is to study the effects of nonnormality on these errors. The class of distributions considered is the Johnson's system. Each of the three random variables can be transformed to normality. In one particular case numerical evaluations are made, based on which it is possible to recommend whether or not it is necessary to make the transformation prior to classification. In a parallel study, we present similar results for the Edgeworth Series distribution, where the random variables cannot be transformed to normality.     

Analysis of ordered probit model with surrogate response data and measurement error in covariates

ABSTRACT

Often in data arising out of epidemiologic studies, covariates are subject to measurement error. In addition ordinal responses may be misclassified into a category that does not reflect the true state of the respondents. The goal of the present work is to develop an ordered probit model that corrects for the classification errors in ordinal responses and/or measurement error in covariates. Maximum likelihood method of estimation is used. Simulation study reveals the effect of ignoring measurement error and/or classification errors on the estimates of the regression coefficients. The methodology developed is illustrated through a numerical example.     

The Effect of Misclassification Costs on Neural Network Classifiers

The potential of neural networks for classification problems has been established by numerous successful applications reported in the literature. One of the major assumptions used in almost all studies is the equal cost consequence of misclassification. With this assumption, minimizing the total number of misclassification errors is the sole objective in developing a neural network classifier. Often this is done simply to ease model development and the selection of classification decision points. However, it is not appropriate for many real situations such as quality assurance, direct marketing, bankruptcy prediction, and medical diagnosis where misclassification costs have unequal consequences for different categories. In this paper, we investigate the issue of unequal misclassification costs in neural network classifiers. Through an application in thyroid disease diagnosis, we find that different cost considerations have significant effects on the classification performance and that appropriate use of cost information can aid in optimal decision making. A cross-validation technique is employed to alleviate the problem of bias in the training set and to examine the robustness of neural network classifiers with regard to sampling variations and cost differences.     

Bayesian misclassification and propensity score methods for clustered observational studies

Bayesian propensity score regression analysis with misclassified binary responses is proposed to analyse clustered observational data. This approach utilizes multilevel models and corrects for misclassification in the responses. Using the deviance information criterion (DIC), the performance of the approach is compared with approaches without correcting for misclassification, multilevel structure specification, or both in the study of the impact of female employment on the likelihood of physical violence. The smallest DIC confirms that our proposed model best fits the data. We conclude that female employment has an insignificant impact on the likelihood of physical spousal violence towards women. In addition, a simulation study confirms that the proposed approach performed best in terms of bias and coverage rate. Ignoring misclassification in response or multilevel structure of data would yield biased estimation of the exposure effect.     

Discriminant analysis of survey data

We consider the problem of the effect of sample designs on discriminant analysis. The selection of the learning sample is assumed to depend on the population values of auxiliary variables. Under a superpopulation model with a multivariate normal distribution, unbiasedness and consistency are examined for the conventional estimators (derived under the assumptions of simple random sampling), maximum likelihood estimators, probability-weighted estimators and conditionally unbiased estimators of parameters. Four corresponding sampled linear discriminant functions are examined. The rates of misclassification of these four discriminant functions and the effect of sample design on these four rates of misclassification are discussed. The performances of these four discriminant functions are assessed in a simulation study.     

Bayesian sample size determination for estimating binomial parameters from data subject to misclassification

We investigate the sample size problem when a binomial parameter is to be estimated, but some degree of misclassification is possible. The problem is especially challenging when the degree to which misclassification occurs is not exactly known. Motivated by a Canadian survey of the prevalence of toxoplasmosis infection in pregnant women, we examine the situation where it is desired that a marginal posterior credible interval for the prevalence of width w has coverage 1−α, using a Bayesian sample size criterion. The degree to which the misclassification probabilities are known a priori can have a very large effect on sample size requirements, and in some cases achieving a coverage of 1−α is impossible, even with an infinite sample size. Therefore, investigators must carefully evaluate the degree to which misclassification can occur when estimating sample size requirements.     

On-Line Process Control using Attributes with Misclassification Errors: An Economical Design for Short-Run Production

On-line process control consists of inspecting a single item for every m (integer and m ≥ 2) produced items. Based on the results of the inspection, it is decided whether the process is in-control (the fraction of conforming items is p ₁; State I) or out-of-control (the fraction of conforming items is p ₂ < p ₁; State II). If the inspected item is non conforming, it is determined that the process is out-of-control, and the production process is stopped for an adjustment; otherwise, production continues. As most designs of on-line process control assume a long-run production, this study can be viewed as an extension because it is concerned with short-run production and the decision regarding the process is subject to misclassification errors. The probabilistic model of the control system employs properties of an ergodic Markov chain to obtain the expression of the average cost of the system per unit produced, which can be minimised as a function of the sampling interval, m. The procedure is illustrated by a numerical example.     

Sample size implications when biases are modelled rather than ignored

Summary.  Realistic statistical modelling of observational data often suggests a statistical model which is not fully identified, owing to potential biases that are not under the control of study investigators. Bayesian inference can be implemented with such a model, ideally with the most precise prior knowledge that can be ascertained. However, as a consequence of the non-identifiability, inference cannot be made arbitrarily accurate by choosing the sample size to be sufficiently large. In turn, this has consequences for sample size determination. The paper presents a sample size criterion that is based on a quantification of how much Bayesian learning can arise in a given non-identified model. A global perspective is adopted, whereby choosing larger sample sizes for some studies necessarily implies that some other potentially worthwhile studies cannot be undertaken. This suggests that smaller sample sizes should be selected with non-identified models, as larger sample sizes constitute a squandering of resources in making estimator variances very small compared with their biases. Particularly, consider two investigators planning the same study, one of whom admits to the potential biases at hand and consequently uses a non-identified model, whereas the other pretends that there are no biases, leading to an identified but less realistic model. It is seen that the former investigator always selects a smaller sample size than the latter, with the difference being quite marked in some illustrative cases.