Unified Conditional Frequentist and Bayesian Testing of Composite Hypotheses

Testing of a composite null hypothesis versus a composite alternative is considered when both have a related invariance structure. The goal is to develop conditional frequentist tests that allow the reporting of data-dependent error probabilities, error probabilities that have a strict frequentist interpretation and that reflect the actual amount of evidence in the data. The resulting tests are also seen to be Bayesian tests, in the strong sense that the reported frequentist error probabilities are also the posterior probabilities of the hypotheses under default choices of the prior distribution. The new procedures are illustrated in a variety of applications to model selection and multivariate hypothesis testing.     

Testing equality of regression coefficients in heteroscedastic normal regression models

This article addresses the problem of testing whether the vectors of regression coefficients are equal for two independent normal regression models when the error variances are unknown. This problem poses severe difficulties both to the frequentist and Bayesian approaches to statistical inference. In the former approach, normal hypothesis testing theory does not apply because of the unrelated variances. In the latter, the prior distributions typically used for the parameters are improper and hence the Bayes factor-based solution cannot be used.We propose a Bayesian solution to this problem in which no subjective input is considered. We first generate “objective” proper prior distributions (intrinsic priors) for which the Bayes factor and model posterior probabilities are well defined. The posterior probability of each model is used as a model selection tool. This consistent procedure of testing hypotheses is compared with some of the frequentist approximate tests proposed in the literature.     

A direct approach to false discovery rates

Summary. Multiple-hypothesis testing involves guarding against much more complicated errors than single-hypothesis testing. Whereas we typically control the type I error rate for a single-hypothesis test, a compound error rate is controlled for multiple-hypothesis tests. For example, controlling the false discovery rate FDR traditionally involves intricate sequential p -value rejection methods based on the observed data. Whereas a sequential p -value method fixes the error rate and estimates its corresponding rejection region, we propose the opposite approach—we fix the rejection region and then estimate its corresponding error rate. This new approach offers increased applicability, accuracy and power. We apply the methodology to both the positive false discovery rate pFDR and FDR, and provide evidence for its benefits. It is shown that pFDR is probably the quantity of interest over FDR. Also discussed is the calculation of the q -value, the pFDR analogue of the p -value, which eliminates the need to set the error rate beforehand as is traditionally done. Some simple numerical examples are presented that show that this new approach can yield an increase of over eight times in power compared with the Benjamini–Hochberg FDR method.     

The Bayesian and frequentist approaches to testing a one-sided hypothesis about a multivariate mean

This paper compares the Bayesian and frequentist approaches to testing a one-sided hypothesis about a multivariate mean. First, this paper proposes a simple way to assign a Bayesian posterior probability to one-sided hypotheses about a multivariate mean. The approach is to use (almost) the exact posterior probability under the assumption that the data has multivariate normal distribution, under either a conjugate prior in large samples or under a vague Jeffreys prior. This is also approximately the Bayesian posterior probability of the hypothesis based on a suitably flat Dirichlet process prior over an unknown distribution generating the data. Then, the Bayesian approach and a frequentist approach to testing the one-sided hypothesis are compared, with results that show a major difference between Bayesian reasoning and frequentist reasoning. The Bayesian posterior probability can be substantially smaller than the frequentist p-value. A class of example is given where the Bayesian posterior probability is basically 0, while the frequentist p-value is basically 1. The Bayesian posterior probability in these examples seems to be more reasonable. Other drawbacks of the frequentist p-value as a measure of whether the one-sided hypothesis is true are also discussed.     

Efficient Stratified Testing Procedure for a False Discovery Rate

The false discovery rate (FDR) has become a popular error measure in the large-scale simultaneous testing. When data are collected from heterogenous sources and form grouped hypotheses testing, it may be beneficial to use the distinct feature of groups to conduct the multiple hypotheses testing. We propose a stratified testing procedure that uses different FDR levels according to the stratification features based on p-values. Our proposed method is easy to implement in practice. Simulations studies show that the proposed method produces more efficient testing results. The stratified testing procedure minimizes the overall false negative rate (FNR) level, while controlling the overall FDR. An example from a type II diabetes mice study further illustrates the practical advantages of this new approach.     

A Bayesian False Discovery Rate for Multiple Testing

Case-control studies of genetic polymorphisms and gene-environment interactions are reporting large numbers of statistically significant associations, many of which are likely to be spurious. This problem reflects the low prior probability that any one null hypothesis is false, and the large number of test results reported for a given study. In a Bayesian approach to the low prior probabilities, Wacholder et al. (2004) suggest supplementing the p-value for a hypothesis with its posterior probability given the study data. In a frequentist approach to the test multiplicity problem, Benjamini & Hochberg (1995) propose a hypothesis-rejection rule that provides greater statistical power by controlling the false discovery rate rather than the family-wise error rate controlled by the Bonferroni correction. This paper defines a Bayes false discovery rate and proposes a Bayes-based rejection rule for controlling it. The method, which combines the Bayesian approach of Wacholder et al. with the frequentist approach of Benjamini & Hochberg, is used to evaluate the associations reported in a case-control study of breast cancer risk and genetic polymorphisms of genes involved in the repair of double-strand DNA breaks.     

Strong control, conservative point estimation and simultaneous conservative consistency of false discovery rates: a unified approach

Summary.  The false discovery rate (FDR) is a multiple hypothesis testing quantity that describes the expected proportion of false positive results among all rejected null hypotheses. Benjamini and Hochberg introduced this quantity and proved that a particular step-up p -value method controls the FDR. Storey introduced a point estimate of the FDR for fixed significance regions. The former approach conservatively controls the FDR at a fixed predetermined level, and the latter provides a conservatively biased estimate of the FDR for a fixed predetermined significance region. In this work, we show in both finite sample and asymptotic settings that the goals of the two approaches are essentially equivalent. In particular, the FDR point estimates can be used to define valid FDR controlling procedures. In the asymptotic setting, we also show that the point estimates can be used to estimate the FDR conservatively over all significance regions simultaneously, which is equivalent to controlling the FDR at all levels simultaneously. The main tool that we use is to translate existing FDR methods into procedures involving empirical processes. This simplifies finite sample proofs, provides a framework for asymptotic results and proves that these procedures are valid even under certain forms of dependence.     

Bayesian Estimators for Small Area Models when Auxiliary Information is Measured with Error

Small area estimators in linear models are typically expressed as a convex combination of direct estimators and synthetic estimators from a suitable model. When auxiliary information used in the model is measured with error, a new estimator, accounting for the measurement error in the covariates, has been proposed in the literature. Recently, for area‐level model, Ybarra & Lohr (Biometrika, 95, 2008, 919) suggested a suitable modification to the estimates of small area means based on Fay & Herriot (J. Am. Stat. Assoc., 74, 1979, 269) model where some of the covariates are measured with error. They used a frequentist approach based on the method of moments. Adopting a Bayesian approach, we propose to rewrite the measurement error model as a hierarchical model; we use improper non‐informative priors on the model parameters and show, under a mild condition, that the joint posterior distribution is proper and the marginal posterior distributions of the model parameters have finite variances. We conduct a simulation study exploring different scenarios. The Bayesian predictors we propose show smaller empirical mean squared errors than the frequentist predictors of Ybarra & Lohr (Biometrika, 95, 2008, 919), and they seem also to be more stable in terms of variability and bias. We apply the proposed methodology to two real examples.     

A clarifying comparison of methods for controlling the false discovery rate

Traditional multiple hypothesis testing procedures fix an error rate and determine the corresponding rejection region. In 2002 Storey proposed a fixed rejection region procedure and showed numerically that it can gain more power than the fixed error rate procedure of Benjamini and Hochberg while controlling the same false discovery rate (FDR). In this paper it is proved that when the number of alternatives is small compared to the total number of hypotheses, Storey's method can be less powerful than that of Benjamini and Hochberg. Moreover, the two procedures are compared by setting them to produce the same FDR. The difference in power between Storey's procedure and that of Benjamini and Hochberg is near zero when the distance between the null and alternative distributions is large, but Benjamini and Hochberg's procedure becomes more powerful as the distance decreases. It is shown that modifying the Benjamini and Hochberg procedure to incorporate an estimate of the proportion of true null hypotheses as proposed by Black gives a procedure with superior power.     

Failure probability estimation under additional subsystem information with application to semiconductor burn-in

In the classical approach to qualitative reliability demonstration, system failure probabilities are estimated based on a binomial sample drawn from the running production. In this paper, we show how to take account of additional available sampling information for some or even all subsystems of a current system under test with serial reliability structure. In that connection, we present two approaches, a frequentist and a Bayesian one, for assessing an upper bound for the failure probability of serial systems under binomial subsystem data. In the frequentist approach, we introduce (i) a new way of deriving the probability distribution for the number of system failures, which might be randomly assembled from the failed subsystems and (ii) a more accurate estimator for the Clopper–Pearson upper bound using a beta mixture distribution. In the Bayesian approach, however, we infer the posterior distribution for the system failure probability on the basis of the system/subsystem testing results and a prior distribution for the subsystem failure probabilities. We propose three different prior distributions and compare their performances in the context of high reliability testing. Finally, we apply the proposed methods to reduce the efforts of semiconductor burn-in studies by considering synergies such as comparable chip layers, among different chip technologies.     

Nonparametric Bayes Stochastically Ordered Latent Class Models

Latent class models (LCMs) are used increasingly for addressing a broad variety of problems, including sparse modeling of multivariate and longitudinal data, model-based clustering, and flexible inferences on predictor effects. Typical frequentist LCMs require estimation of a single finite number of classes, which does not increase with the sample size, and have a well-known sensitivity to parametric assumptions on the distributions within a class. Bayesian nonparametric methods have been developed to allow an infinite number of classes in the general population, with the number represented in a sample increasing with sample size. In this article, we propose a new nonparametric Bayes model that allows predictors to flexibly impact the allocation to latent classes, while limiting sensitivity to parametric assumptions by allowing class-specific distributions to be unknown subject to a stochastic ordering constraint. An efficient MCMC algorithm is developed for posterior computation. The methods are validated using simulation studies and applied to the problem of ranking medical procedures in terms of the distribution of patient morbidity.     

Spatial modeling using frequentist approach for disease mapping

In this article, a generalized linear mixed model (GLMM) based on a frequentist approach is employed to examine spatial trend of asthma data. However, the frequentist analysis of GLMM is computationally difficult. On the other hand, the Bayesian analysis of GLMM has been computationally convenient due to the advent of Markov chain Monte Carlo algorithms. Recently developed data cloning (DC) method, which yields to maximum likelihood estimate, provides frequentist approach to complex mixed models and equally computationally convenient method. We use DC to conduct frequentist analysis of spatial models. The advantages of the DC approach are that the answers are independent of the choice of the priors, non-estimable parameters are flagged automatically, and the possibility of improper posterior distributions is completely avoided. We illustrate this approach using a real dataset of asthma visits to hospital in the province of Manitoba, Canada, during 2000–2010. The performance of the DC approach in our application is also studied through a simulation study.     

A note on multiple testing for composite null hypotheses

Multiple hypothesis testing literature has recently experienced a growing development with particular attention to the control of the false discovery rate (FDR) based on p-values. While these are not the only methods to deal with multiplicity, inference with small samples and large sets of hypotheses depends on the specific choice of the p-value used to control the FDR in the presence of nuisance parameters. In this paper we propose to use the partial posterior predictive p-value [Bayarri, M.J., Berger, J.O., 2000. p-values for composite null models. J. Amer. Statist. Assoc. 95, 1127–1142] that overcomes this difficulty. This choice is motivated by theoretical considerations and examples. Finally, an application to a controlled microarray experiment is presented.     

A note on fiducial model averaging as an alternative to checking Bayesian and frequentist models

Just as frequentist hypothesis tests have been developed to check model assumptions, prior predictive p-values and other Bayesian p-values check prior distributions as well as other model assumptions. These model checks not only suffer from the usual threshold dependence of p-values, but also from the suppression of model uncertainty in subsequent inference. One solution is to transform Bayesian and frequentist p-values for model assessment into a fiducial distribution across the models. Averaging the Bayesian or frequentist posterior distributions with respect to the fiducial distribution can reproduce results from Bayesian model averaging or classical fiducial inference.     

Robust exchangeability designs for early phase clinical trials with multiple strata

Clinical trials with multiple strata are increasingly used in drug development. They may sometimes be the only option to study a new treatment, for example in small populations and rare diseases. In early phase trials, where data are often sparse, good statistical inference and subsequent decision‐making can be challenging. Inferences from simple pooling or stratification are known to be inferior to hierarchical modeling methods, which build on exchangeable strata parameters and allow borrowing information across strata. However, the standard exchangeability (EX) assumption bears the risk of too much shrinkage and excessive borrowing for extreme strata. We propose the exchangeability–nonexchangeability (EXNEX) approach as a robust mixture extension of the standard EX approach. It allows each stratum‐specific parameter to be exchangeable with other similar strata parameters or nonexchangeable with any of them. While EXNEX computations can be performed easily with standard Bayesian software, model specifications and prior distributions are more demanding and require a good understanding of the context. Two case studies from phases I and II (with three and four strata) show promising results for EXNEX. Data scenarios reveal tempered degrees of borrowing for extreme strata, and frequentist operating characteristics perform well for estimation (bias, mean‐squared error) and testing (less type‐I error inflation). Copyright © 2015 John Wiley & Sons, Ltd.     

Evaluation of false discovery rate and power via sample size in microarray studies

Microarray studies are now common for human, agricultural plant and animal studies. False discovery rate (FDR) is widely used in the analysis of large-scale microarray data to account for problems associated with multiple testing. A well-designed microarray study should have adequate statistical power to detect the differentially expressed (DE) genes, while keeping the FDR acceptably low. In this paper, we used a mixture model of expression responses involving DE genes and non-DE genes to analyse theoretical FDR and power for simple scenarios where it is assumed that each gene has equal error variance and the gene effects are independent. A simulation study was used to evaluate the empirical FDR and power for more complex scenarios with unequal error variance and gene dependence. Based on this approach, we present a general guide for sample size requirement at the experimental design stage for prospective microarray studies. This paper presented an approach to explicitly connect the sample size with FDR and power. While the methods have been developed in the context of one-sample microarray studies, they are readily applicable to two-sample, and could be adapted to multiple-sample studies.     

Multiple Hypothesis Testing for Variable Selection

We propose two new procedures based on multiple hypothesis testing for correct support estimation in high‐dimensional sparse linear models. We conclusively prove that both procedures are powerful and do not require the sample size to be large. The first procedure tackles the atypical setting of ordered variable selection through an extension of a testing procedure previously developed in the context of a linear hypothesis. The second procedure is the main contribution of this paper. It enables data analysts to perform support estimation in the general high‐dimensional framework of non‐ordered variable selection. A thorough simulation study and applications to real datasets using the R package mht shows that our non‐ordered variable procedure produces excellent results in terms of correct support estimation as well as in terms of mean square errors and false discovery rate, when compared to common methods such as the Lasso, the SCAD penalty, forward regression or the false discovery rate procedure (FDR).     

Default Bayesian analysis of the Behrens–Fisher problem

In the Bayesian approach, the Behrens–Fisher problem has been posed as one of estimation for the difference of two means. No Bayesian solution to the Behrens–Fisher testing problem has yet been given due, perhaps, to the fact that the conventional priors used are improper. While default Bayesian analysis can be carried out for estimation purposes, it poses difficulties for testing problems. This paper generates sensible intrinsic and fractional prior distributions for the Behrens–Fisher testing problem from the improper priors commonly used for estimation. It allows us to compute the Bayes factor to compare the null and the alternative hypotheses. This default procedure of model selection is compared with a frequentist test and the Bayesian information criterion. We find discrepancy in the sense that frequentist and Bayesian information criterion reject the null hypothesis for data, that the Bayes factor for intrinsic or fractional priors do not.     

On the correspondence between frequentist and Bayesian tests

Modern theory for statistical hypothesis testing can broadly be classified as Bayesian or frequentist. Unfortunately, one can reach divergent conclusions if Bayesian and frequentist approaches are applied in parallel to analyze the same data set. This is a serious impasse since there is a lack of consensus on when to use one approach in detriment of the other. However, this conflict can be resolved. The present paper shows the existence of a perfect equivalence between Bayesian and frequentist methods for testing. Hence, Bayesian and frequentist decision rules can always be calibrated, in both directions, in order to present concordant results.     

Objective Testing Procedures in Linear Models: Calibration of the p-values

Abstract.  An optimal Bayesian decision procedure for testing hypothesis in normal linear models based on intrinsic model posterior probabilities is considered. It is proven that these posterior probabilities are simple functions of the classical F -statistic, thus the evaluation of the procedure can be carried out analytically through the frequentist analysis of the posterior probability of the null. An asymptotic analysis proves that, under mild conditions on the design matrix, the procedure is consistent. For any testing hypothesis it is also seen that there is a one-to-one mapping – which we call calibration curve – between the posterior probability of the null hypothesis and the classical bi p -value. This curve adds substantial knowledge about the possible discrepancies between the Bayesian and the p -value measures of evidence for testing hypothesis. It permits a better understanding of the serious difficulties that are encountered in linear models for interpreting the p -values. A specific illustration of the variable selection problem is given.