Heteroscedastic ANOVA: old p values, new views

The generalization of the Behrens–Fisher problem to comparing more than two means from nonhomogeneous populations has attracted the attention of statisticians for many decades. Several approaches offer different approximations to the distribution of the test statistic. The question of statistical properties of these approximations is still alive. Here, we present a brief overview of several approaches suggested in the literature and implemented in software with a focus on investigating the accuracy of p values as well as their dependence on nuisance parameters and on the underlying assumption of normality. We illustrate by simulation the behavior of p values. In addition to the Satterthwaite–Fai–Cornelius test, the Kenward–Roger test, the simple ANOVA F test, the parametric bootstrap test, and the generalized F test will be briefly discussed.     

One-way ANOVA with Unequal Variances

We consider the one-way ANOVA problem of testing the equality of several normal means when the variances are not assumed to be equal. This is a generalization of the Behrens-Fisher problem, but even in this special case there is no exact test and the actual size of any test depends on the values of the nuisance parameters. Therefore, controlling the actual size of the test is of main concern. In this article, we first consider a test using the concept of generalized p-value. Extensive simulation studies show that the actual size of this test does not exceed the nominal level, for practically all values of the nuisance parameters, but the test is not too conservative either, in the sense that the actual size of the test can be very close to the nominal level for some values of the nuisance parameters. We then use this test to propose a simple F-test, which has similar properties but avoids the computations associated with generalized p-values. Because of its simplicity, both conceptually as well as computationally, this F-test may be more useful in practice, since one-way ANOVA is widely used by practitioners who may not be familiar with the generalized p-value and its computational aspects.     

Optimality of random allocation design for the control of accidental bias in sequential experiments

In comparing two treatments, suppose the suitable subjects arrive sequentially and must be treated at once. Known or unknown to the experimenter there may be nuisance factors systematically affecting the subjects. Accidental bias is a measure of the influence of these factors in the analysis of data. We show in this paper that the random allocation design minimizes the accidental bias among all designs that allocate n, out of 2n, subjects to each treatment and do not prefer either treatment in the assignment. When the final imbalance is allowed to be nonzero, optimal and efficient designs are given. In particular the random allocation design is shown to be very efficient in this broader setup.     

The use of conditional cutoffs in a forward selection procedure

Using a forward selection procedure for selecting the best subset of regression variables involves the calculation of critical values (cutoffs) for an F-ratio at each step of a multistep search process. On dropping the restrictive (unrealistic) assumptions used in previous works, the null distribution of the F-ratio depends on unknown regression parameters for the variables already included in the subset. For the case of known σ, by conditioning the F-ratio on the set of regressors included so far and also on the observed (estimated) values of their regression coefficients, we obtain a forward selection procedure whose stepwise type I error does not depend on the unknown (nuisance) parameters. A numerical example with an orthogonal design matrix illustrates the difference between conditional cutoffs, cutoffs for the centralF-distribution, and cutoffs suggested by Pope and Webster.     

A p-value for testing the equivalence of the variances of a bivariate normal distribution

A p-value is developed for testing the equivalence of the variances of a bivariate normal distribution. The unknown correlation coefficient is a nuisance parameter in the problem. If the correlation is known, the proposed p-value provides an exact test. For large samples, the p-value can be computed by replacing the unknown correlation by the sample correlation, and the resulting test is quite satisfactory. For small samples, it is proposed to compute the p-value by replacing the unknown correlation by a scalar multiple of the sample correlation. However, a single scalar is not satisfactory, and it is proposed to use different scalars depending on the magnitude of the sample correlation coefficient. In order to implement this approach, tables are obtained providing sub-intervals for the sample correlation coefficient, and the scalars to be used if the sample correlation coefficient belongs to a particular sub-interval. Once such tables are available, the proposed p-value is quite easy to compute since it has an explicit analytic expression. Numerical results on the type I error probability and power are reported on the performance of such a test, and the proposed p-value test is also compared to another test based on a rejection region. The results are illustrated with two examples: an example dealing with the comparability of two measuring devices, and an example dealing with the assessment of bioequivalence.     

Minimum Hellinger distance estimation in a nonparametric mixture model

In this paper, we investigate the estimation problem of the mixture proportion λ$\lambda$ in a nonparametric mixture model of the form λF(x)+(1-λ)G(x)$\lambda F(x)+(1-\lambda )G(x)$ using the minimum Hellinger distance approach, where F and G are two unknown distributions. We assume that data from the distributions F and G   as well as from the mixture distribution λF+(1-λ)G$\lambda F+(1-\lambda )G$ are available. We construct a minimum Hellinger distance estimator of λ$\lambda$ and study its asymptotic properties. The proposed estimator is chosen to minimize the Hellinger distance between a parametric mixture model and a nonparametric density estimator. We also develop a maximum likelihood estimator of λ$\lambda$. Theoretical properties such as the existence, strong consistency, asymptotic normality and asymptotic efficiency of the proposed estimators are investigated. Robustness properties of the proposed estimator are studied using a Monte Carlo study. Two real data examples are also analyzed.     

Partitioning k multivariate normal populations according to equivalence with respect to a standard vector

We propose optimal procedures to achieve the goal of partitioning k multivariate normal populations into two disjoint subsets with respect to a given standard vector. Definition of good or bad multivariate normal populations is given according to their Mahalanobis distances to a known standard vector as being small or large. Partitioning k multivariate normal populations is reduced to partitioning k non-central Chi-square or non-central F distributions with respect to the corresponding non-centrality parameters depending on whether the covariance matrices are known or unknown. The minimum required sample size for each population is determined to ensure that the probability of correct decision attains a certain level. An example is given to illustrate our procedures.     

An Estimator of the Exponent of Regular Variation Based on K-Record Values

Let {X _n:n ≥ 1} be an i.i.d. sequence of random variables with a continuous distribution function F. Under the assumption that the upper tail of Fis regularly varying with exponent 1/α, α > 0, we study the asymptotic properties of an estimator of α based on k-record values.     

Prediction in moving average processes

For the stationary invertible moving average process of order one with unknown innovation distribution F, we construct root-n   consistent plug-in estimators of conditional expectations E(h(_Xn+1)|_X1,…,_Xn)$E(h(X_{n+1})|X_1,\dots ,X_n)$. More specifically, we give weak conditions under which such estimators admit Bahadur-type representations, assuming some smoothness of h or of F. For fixed h it suffices that h   is locally of bounded variation and locally Lipschitz in _L2(F)$L_2(F)$, and that the convolution of h and F   is continuously differentiable. A uniform representation for the plug-in estimator of the conditional distribution function P(_Xn+1?·|_X1,…,_Xn)$P(X_{n+1}?·|X_1,\dots ,X_n)$ holds if F has a uniformly continuous density. For a smoothed version of our estimator, the Bahadur representation holds uniformly over each class of functions h that have an appropriate envelope and whose shifts are F-Donsker, assuming some smoothness of F. The proofs use empirical process arguments.     

Approximate jackknife empirical likelihood method for estimating equations

It is known that the profile empirical likelihood method based on estimating equations is computationally intensive when the number of nuisance parameters is large. Recently, Li, Peng, & Qi (2011) proposed a jackknife empirical likelihood method for constructing confidence regions for the parameters of interest by estimating the nuisance parameters separately. However, when the estimators for the nuisance parameters have no explicit formula, the computation of the jackknife empirical likelihood method is still intensive. In this paper, an approximate jackknife empirical likelihood method is proposed to reduce the computation in the jackknife empirical likelihood method when the nuisance parameters cannot be estimated explicitly. A simulation study confirms the advantage of the new method. The Canadian Journal of Statistics 40: 110–123; 2012 © 2012 Statistical Society of Canada     

Rate of consistency of one sample tests of location

Let X₁,…,X_n be a sample from a population with continuous distribution function F(x?θ) such that F(x)+F(-x)=1 and 0<F(x)<1, x?R¹. It is shown that the power- function of a monotone test of H: θ=θ₀ against K: θ>θ₀ cannot tend to 1 as θ?θ₀ → ∞ more than n times faster than the tails of F tend to 0. Some standard as well as robust tests are considered with respect to this rate of convergence.     

Bahadur efficiency of the F-test

Extremal problems in large deviations of the F-statistic are considered. It is shown that the slowest rate of convergence over a specified class of distributions of the F-statistic is slower than exponential, and that the Bahadur efficiency of the F-statistic with respect to some distribution-free competitors is identically zero.     

A formula for P(bi≤Ri≤ai, 1≤i≤m) under a general class of alternatives

Let X_i≤?≤X_m and Y_i≤?≤Y_n be two sets of independent order statistics from continous distributions with distribution functions F and G respectively. Let R_i denote the rank of X_i in the combined order sample. Steck (1980) has found an expression for P(b_i≤R_i≤a_i, all i) when F = h(G), h being the incomplete beta function with parameters (α,β?α+1). An alternative expression for the same probability is obtained which is computationally a substantial improvement on Steck's result.     

The superharmonic condition for simultaneous estimation of means in exponential familles

The mean vector associated with several independent variates from the exponential subclass of Hudson (1978) is estimated under weighted squared error loss. In particular, the formal Bayes and “Stein-like” estimators of the mean vector are given. Conditions are also given under which these estimators dominate any of the “natural estimators”. Our conditions for dominance are motivated by a result of Stein (1981), who treated the N_p (θ, I) case with p ≥ 3. Stein showed that formal Bayes estimators dominate the usual estimator if the marginal density of the data is superharmonic. Our present exponential class generalization entails an elliptic differential inequality in some natural variables. Actually, we assume that each component of the data vector has a probability density function which satisfies a certain differential equation. While the densities of Hudson (1978) are particular solutions of this equation, other solutions are not of the exponential class if certain parameters are unknown. Our approach allows for the possibility of extending the parametric Stein-theory to useful nonexponential cases, but the problem of nuisance parameters is not treated here.     

On the Markov property of order statistics

The order statistics from a sample of size n≥3 from a discrete distribution form a Markov chain if and only if the parent distribution is supported by one or two points. More generally, a necessary and sufficient condition for the order statistics to form a Markov chain for (n≥3) is that there does not exist any atom x₀ of the parent distribution F satisfying F(x₀-)>0 and F(x₀)<1. To derive this result a formula for the joint distribution of order statistics is proved, which is of an interest on its own. Many exponential characterizations implicitly assume the Markov property. The corresponding putative geometric characterizations cannot then be reasonably expected to obtain. Some illustrative geometric characterizations are discussed.     

Optimal Bayesian strategies for the infinite-armed Bernoulli bandit

We consider the bandit problem with an infinite number of Bernoulli arms, of which the unknown parameters are assumed to be i.i.d. random variables with a common distribution F. Our goal is to construct optimal strategies of choosing “arms” so that the expected long-run failure rate is minimized. We first review a class of strategies and establish their asymptotic properties when F is known. Based on the results, we propose a new strategy and prove that it is asymptotically optimal when F is unknown. Finally, we show that the proposed strategy performs well for a number of simulation scenarios.     

Nonparametric statistical procedures for the changepoint problem

Let X₁,…,X_r?1,X_r,X_r+1,…,X_n be independent, continuous random variables such that X_i, i = 1,…,r, has distribution function F(x), and X_i, i = r+1,…,n, has distribution function F(x?Δ), with -∞ <Δ< ∞. When the integer r is unknown, this is refered to as a change point problem with at most one change. The unknown parameter Δ represents the magnitude of the change and r is called the changepoint. In this paper we present a general review discussion of several nonparametric approaches for making inferences about r and Δ.     

Bayesian design of experiment for quantal responses: What is promised versus what is delivered

This article considers a design problem in quantal response analysis, where an experimenter must choose a set of dose levels and number of independent observations to take at these levels, subject to some total sample size, in order to minimize the expected or predicted posterior variance of some characteristics ø of the tolerance distribution F_θ, with unknown parameters θ. An exact solution to this problem is demonstrated when ø is the unknown LD50 of the one parameter logistic tolerance distribution, under the restriction that an equal number of observations are taken at each of a set of equally spaced levels. The solution is based on a combination of simulated outcomes and Monte Carlo integration to evaluate the predicted variance. The numerical results are compared to those obtained previously by asymptotic approximations in Tsutakawa (1972), (J. Amer. Statist. Assoc. 67 584–590). The wide variability in the simulated posterior variance suggests that the expected posterior variance alone is not a good criterion for design selection.     

Balanced fractional factorial designs of resolution 2l+1 for interesting effects orthogonal to some nuisance parameters: 2m1+m2 series

For 2^m₁+m₂ factorial designs, this paper investigates balanced fractional 2^m₁ factorial designs of resolution 2l+1 with some nuisance parameters concerning the second factors. They are derivable from partially balanced arrays and further permit estimation of the effects up to the l-factor interactions concerning the first factors orthogonally to the nuisance parameters.     

Simple bounds on availability in a model with unknown life and repair distributions

We take a fresh look at the classic model of a device supported by a single statistically identical spare and provision for repairs, with system failure resulting whenever the currently operating unit fails before the repair of the previously failed unit is completed to allow it to become a spare. The limiting availability A(F,G) of this system depends on the life distribution F and repair time distribution G through α=∫GdF and the expected downtime. In this paper we derive several computable and sharp bounds on A(F,G) when F,G have suitable life distribution characteristics in the sense of reliability theory but are otherwise unknown except for at most two moments. Among other results, we find a sharp bound which involves the MTBF, MTTR and the second moment of the life-distribution of the device through its coefficient of variation. This leads to a maximin result for DFR repairs and DMRL lives.