A Unified Approach to Prior and Loss Robustness

ABSTRACT

We introduce a new statistical framework in order to study Bayesian loss robustness under classes of priors distributions, thus unifying both concepts of robustness. We propose measures that capture variation with respect to both prior selection and selection of loss function and explore general properties of these measures. We illustrate the approach for the continuous exponential family. Robustness in this context is studied first with respect to prior selection where we consider several classes of priors for the parameter of interest, including unimodal and symmetric and unimodal with positive support. After prior variation has been measured we investigate robustness to loss function, using Hellinger and Linex (Linear Exponential) classes of loss functions. The methods are applied to standard examples.     

Cross-national comparison of internal migration: issues and measures

Summary. Our objectives are to identify the issues that researchers encounter when measuring internal migration in different countries and to propose key indicators that analysts can use to compare internal migration at the 'national' level. We establish the benefits to be gained by a rigorous approach to cross-national comparisons of internal migration and discuss issues that affect such comparisons. We then distinguish four dimensions of internal migration on which countries can be compared and, for each dimension, identify a series of summary measures. We illustrate the issues and measures proposed by comparing migration in Australia and Great Britain.     

A Comparison of Different Methods for Representing Categorical Data

We first compare correspondence analysis, which uses chi-square distance, and an alternative approach using Hellinger distance, for representing categorical data in a contingency table. We propose a coefficient which globally measures the similarity between these two approaches. This coefficient can be decomposed into several components, one component for each principal dimension, indicating the contribution of the dimensions to the difference between the two representations. We also make comparisons with the logratio approach based on compositional data. These three methods of representation can produce quite similar results. Two illustrative examples are given.     

Global benefit-risk assessment in designing clinical trials and some statistical considerations of the method

When characterizing a therapy, the efficacy and the safety are two major aspects under consideration. In prescribing a therapy to a patient, a clinician puts the two aspects together and makes a decision based on a consolidated thought process. The global benefit-risk (GBR) measures proposed by Chuang-Stein et al. (Stat. Med. 1991; 10:1349-1359) are useful in facilitating the thinking, and creating the framework for making statistical comparisons based on benefit-risk point of view. This article describes how a GBR linear score was defined and used as the primary outcome measure in a clinical trial design. The robustness of the definitions of 'benefit' and 'risk' are evaluated using different criteria. The sensitivity of the pre-specified weights is also analyzed using alternative weights; one of those was determined by the relative to an identified distribution integral transformation approach (Biometrics 1958; 14:18-38). Statistical considerations are illustrated using pooled data from clinical trials studying antidepressant. The pros and cons for using GBR assessments in the setting of clinical trials are discussed.     

SB-robustness of directional mean for circular distributions

In this paper we study the robustness of the directional mean (a.k.a. circular mean) for different families of circular distributions. We show that the directional mean is robust in the sense of finite standardized gross error sensitivity (SB-robust) for the following families: (1) mixture of two circular normal distributions, (2) mixture of wrapped normal and circular normal distributions and (3) mixture of two wrapped normal distributions. We also show that the directional mean is not SB-robust for the family of all circular normal distributions with varying concentration parameter. We define the circular trimmed mean and prove that it is SB-robust for this family. In general the property of SB-robustness of an estimator at a family of probability distributions is dependent on the choice of the dispersion measure. We introduce the concept of equivalent dispersion measures and prove that if an estimator is SB-robust for one dispersion measure then it is SB-robust for all equivalent dispersion measures. Three different dispersion measures for circular distributions are considered and their equivalence studied.     

Optimal and efficient crossover designs for comparing test treatments to a control treatment under various models

We study the optimality, efficiency, and robustness of crossover designs for comparing several test treatments to a control treatment. Since A-optimality is a natural criterion in this context, we establish lower bounds for the trace of the inverse of the information matrix for the test treatments versus control comparisons under various models. These bounds are then used to obtain lower bounds for efficiencies of a design under these models. Two algorithms, both guided by these efficiencies and results from optimal design theory, are proposed for obtaining efficient designs under the various models.     

Statistics in the High School Curriculum

Recent work, spearheaded by Charles Dunnett (1980a), leads to the conclusion that the Tukey-Kramer (TK) method (popularly known as “Kramer's Method”) is the recommended multiple comparisons procedure for the simultaneous estimation of all pairwise differences of means in an imbalanced one-way ANOVA design with homogeneous variances. Nine other multiple comparisons methods are compared to each other and to the TK method using the criteria of conservativeness, narrowness of confidence intervals, robustness, and ease of use. The degree of superiority of the TK method over these methods, especially over the popular Bonferroni method, is sufficient to warrant recommending its use. Because of the lack of robustness of the TK method in heterogeneous variance cases, other methods are recommended.     

Robust multiple comparisons

The use of several robust estimators of location with their associated variance estimates in a modified T-method for pairwise multiple comparisons between treatment means was compared with the sample mean and variance and with the k-sample rank sum test. The methods were compared with respect to the stability of their experimentwise error rates under a variety of non-normal situations (robustness of validity) and their average confidence interval lengths (robustness of efficiency).     

Robust and scale-free effect sizes for non-Normal two-sample comparisons,with applications in e-commerce

The effect size (ES) has been mainly introduced and investigated for changes in location under an assumption of Normality for the underlying population. However, there are many circumstances where populations are non-Normal, or depend on scale and shape and not just a location parameter. Our motivating application from e-commerce requires an ES which is appropriate for long-tailed distributions. We review some common ES measures. We then introduce two novel alternative ES for two-sample comparisons, one scale-free and one on the original scale of measurement, and analyse some theoretical properties. We examine these ES for two-sample comparison studies under an assumption of Normality and investigate what happens when both location and scale parameters differ. We explore ES for phenomena for non-Normal situations, using the Weibull family for illustration. Finally, for an application, we assess differences in customer behaviour when browsing E-commerce websites.     

Robust projection indices

Loosely speaking a robust projection index is one that prefers projections involving true clusters over projections consisting of a cluster and an outlier. We introduce a mathematical definition of one-dimensional index robustness and describe a numerical experiment to measure it. We design five new indices based on measuring divergence from Student's t -distribution which are intended to be especially robust: the experiment shows that they are more robust than several established indices. The experiment also reveals more generally that the robustness of moment indices depends on the number of approximation terms, providing additional practical guidance for existing projection pursuit implementations. We investigate the theoretical properties of one new Student t -index and Hall's index and show that the new index automatically adapts its robustness to the degree of outlier contamination. We conclude by outlining the possibilities for extending our experiments to both higher dimensions and other new indices.     

Trimmed and Winsorized means based on a scaled deviation

Trimmed (and Winsorized) means based on a scaled deviation are introduced and studied. The influence functions of the estimators are derived and their limiting distributions are established via asymptotic representations. As a main focus of the paper, the performance of the estimators with respect to various robustness and efficiency criteria is evaluated and compared with leading competitors including the ordinary Tukey trimmed (and Winsorized) means. Unlike the Tukey trimming which always trims a fixed fraction of sample points at each end of data, the trimming scheme here only trims points at one or both ends that have a scaled deviation beyond some threshold. The resulting trimmed (and Winsorized) means are much more robust than their predecessors. Indeed they can share the best breakdown point robustness of the sample median for any common trimming thresholds. Furthermore, for appropriate trimming thresholds they are highly efficient at light-tailed symmetric models and more efficient than their predecessors at heavy-tailed or contaminated symmetric models. Detailed comparisons with leading competitors on various robustness and efficiency aspects reveal that the scaled deviation trimmed (Winsorized) means behave very well overall and consequently represent very favorable alternatives to the ordinary trimmed (Winsorized) means.     

Comparison of Goodness-of-Fit Measures in Probit Regression Model

This article examines several goodness-of-fit measures in the binary probit regression model. Existing pseudo-R ² measures are reviewed, two modified and one new pseudo-R ² measure are proposed. For the probit regression model, empirical comparisons are made for different goodness-of-fit measures with the squared sample correlation coefficient of the observed response and the predicted probabilities. As an illustration, the goodness-of-fit measures are applied to a “paid labor force” data set.     

Robust designs for one-point extrapolation

We consider the construction of designs for the extrapolation of a regression response to one point outside of the design space. The response function is an only approximately known function of a specified linear function. As well, we allow for variance heterogeneity. We find minimax designs and corresponding optimal regression weights in the context of the following problems: (P1) for nonlinear least squares estimation with homoscedasticity, determine a design to minimize the maximum value of the mean squared extrapolation error (MSEE), with the maximum being evaluated over the possible departures from the response function; (P2) for nonlinear least squares estimation with heteroscedasticity, determine a design to minimize the maximum value of MSEE, with the maximum being evaluated over both types of departures; (P3) for nonlinear weighted least squares estimation, determine both weights and a design to minimize the maximum MSEE; (P4) choose weights and design points to minimize the maximum MSEE, subject to a side condition of unbiasedness. Solutions to (P1)–(P4) are given in complete generality. Numerical comparisons indicate that our designs and weights perform well in combining robustness and efficiency. Applications to accelerated life testing are highlighted.     

Persistence and existence of stationary measures for a logistic growth model with predation

We consider a stochastic logistic growth model with a predation term, and a diffusive stochastic part with a power-type coefficient. We provide criteria for the persistence of the population and for the existence and uniqueness of a stationary measure. Furthermore, we perform a detailed study of the densities of the stationary measures resorting to the forward Kolmogorov equation. We compile our results in a stochastic bifurcation diagram, drawing comparisons with the corresponding deterministic model.     

Concentration functions and Bayesian robustness

The concentration function, extending the classical notion of Lorenz curve, is well suited for comparing probability measures. Such a feature can be useful in different issues in Bayesian robustness, when a probability measure is deemed a baseline to be compared with other measures by means of their functional forms. Neighbourhood classes Γ of probability measures, including well-known ones, can be defined through the concentration function and both prior and posterior expectations of given functions of the unknown parameter are studied. The ranges of such expectations over Γ can be found, restricting the search among the extremal measures in Γ. The concentration function can be also used as a criterion to assess posterior robustness, when considering sensitivity to changes in the likelihood and the prior.     

A statistical analysis of the screening measure of multiple recursive random number generators of orders one and two

This paper presents the results of a comprehensive empirical analysis of the screening measure of multiple recursive generators (MRGs) of orders one and two. Two kinds of screening measures are distinguished: spectral test and lattice test. With regard to these screening measures, two exhaustive searches of the twenty best MRGs of orders one and two are conducted. Some empirical comparisons reveal that the screening procedure with maximum spectral value criterion is preferred in terms of efficiency and thus, is a good way of obtaining ideal MRGs of higher orders. Several extensively tested second-order MRGs are also presented and are therefore recommended.     

A predictive approach to measuring the strength of statistical evidence for single and multiple comparisons

The normalized maximum likelihood (NML) is a recent penalized likelihood that has properties that justify defining the amount of discrimination information (DI) in the data supporting an alternative hypothesis over a null hypothesis as the logarithm of an NML ratio, namely, the alternative hypothesis NML divided by the null hypothesis NML. The resulting DI, like the Bayes factor but unlike the P‐value, measures the strength of evidence for an alternative hypothesis over a null hypothesis such that the probability of misleading evidence vanishes asymptotically under weak regularity conditions and such that evidence can support a simple null hypothesis. Instead of requiring a prior distribution, the DI satisfies a worst‐case minimax prediction criterion. Replacing a (possibly pseudo‐) likelihood function with its weighted counterpart extends the scope of the DI to models for which the unweighted NML is undefined. The likelihood weights leverage side information, either in data associated with comparisons other than the comparison at hand or in the parameter value of a simple null hypothesis. Two case studies, one involving multiple populations and the other involving multiple biological features, indicate that the DI is robust to the type of side information used when that information is assigned the weight of a single observation. Such robustness suggests that very little adjustment for multiple comparisons is warranted if the sample size is at least moderate. The Canadian Journal of Statistics 39: 610–631; 2011. © 2011 Statistical Society of Canada     

Variable Sampling Interval Cumulative Count of Conforming Chart with Runs Rules

The cumulative count of conforming (CCC) chart is effective in detecting very low fraction of nonconforming items for high yield manufacturing processes. In this study, a combination of runs rules and variable sampling interval feature is proposed to a lower sided CCC chart by inspecting the items one by one. The performance measures of the control chart are derived by using the Markov chain approach. The numerical comparisons show that the performance of the CCC chart can be improved by adding the runs rules and varying the sampling interval.     

The Gaussian rank correlation estimator: robustness properties

The Gaussian rank correlation equals the usual correlation coefficient computed from the normal scores of the data. Although its influence function is unbounded, it still has attractive robustness properties. In particular, its breakdown point is above 12%. Moreover, the estimator is consistent and asymptotically efficient at the normal distribution. The correlation matrix obtained from pairwise Gaussian rank correlations is always positive semidefinite, and very easy to compute, also in high dimensions. We compare the properties of the Gaussian rank correlation with the popular Kendall and Spearman correlation measures. A simulation study confirms the good efficiency and robustness properties of the Gaussian rank correlation. In the empirical application, we show how it can be used for multivariate outlier detection based on robust principal component analysis.     

Optimal and lead-in adaptive allocation for binary outcomes: A comparison of Bayesian methods

We compare posterior and predictive estimators and probabilities in response-adaptive randomization designs for two- and three-group clinical trials with binary outcomes. Adaptation based upon posterior estimates are discussed, as are two predictive probability algorithms: one using the traditional definition, the other using a skeptical distribution. Optimal and natural lead-in designs are covered. Simulation studies show that efficacy comparisons lead to more adaptation than center comparisons, though at some power loss, skeptically predictive efficacy comparisons and natural lead-in approaches lead to less adaptation but offer reduced allocation variability. Though nuanced, these results help clarify the power-adaptation trade-off in adaptive randomization.