Small-sample performance of Bernoulli two-armed bandit Bayesian strategies

In this paper we examine the small-sample performance of a number of strategies for Bernoulli two-armed bandit problems with independent arms. We first investigate strategies based on a one-armed bandit threshold value (an index analogous to the ‘Gittins index’) and on upper confidence bounds for θ_i. Using backward induction and the Bayesian viewpoint, we observe that these strategies improve on the myopic strategy and get much closer to optimal in terms of total expected reward, even though for very small samples, the myopic worth itself is already close to optimal. Second, we find that the myopic strategy and the strategy based on the one-armed threshold value dominate the Bayesian optimal strategy over a region in the parameter space that can have large probability under the assumed prior. Finally, through examples we show how this has an impact on robustness: small specifications of the prior can lead to the myopic strategy performing better than the optimal strategy in terms of Bayes worth.     

Bias-robust L-estimators of a scale parameter

We derive optimal bias-robust L-estimators of a scale parameter σ based on random samples from F(( ·?θ/σ), where θ and σ are unknown and F is an unknown member of a ε-contaminated neighborhood of a fixed symmetric error distribution F ₀. Within a very general class S of L-estimators which are Fisher-consistent at F, we solve for: (i) the estimator with minimax asymptotic bias over the ε-contamination neighborhood; and (ii) the estimator with minimum gross error sensitivity at F ₀ [the limiting case of (i) as ε → 0]. The solutions to problems (i) and (ii) are shown, using a generalized method of moment spaces, to be mixtures of at most two interquantile ranges. A graphical method is presented for finding the optimal bias-robust solutions, and examples are given.     

On two random search problems

The paper is concerned with static search on a finite set. An unknown subset of cardinality k of the finite set is to be found by testing its subsets. We investigate two problems: in the first, the number of common elements of the tested and the unknown subset is given; in the second, only the information whether the tested and the unknown subset are disjoint or not is given. Both problems correspond to problems on false coins. If the unknown subset is taken from the family of k-element sets with uniform distribution, we determine the minimum of the lengths of the strategies that find the unknown element with small error probability. The strategies are constructed by probabilistic means.     

An optimal search procedure

Consider the following problem. There are exactly two defective (unknown) elements in the set X={x₁, x₂,…,x_n}, all possibilities occuring with equal probabilities. We want to identify the unknown (defective) elements by testing some subsets A of X, and for each such set A determining whether A contains any of them. The test on an individual subset A informs us that either all elements of the tested set A are good, or that at least one of them is defective (but we do not know which ones or how many). A set containing at least one defective element is said to be defective. Our aim is to minimize the maximal number of tests. For the optimal strategy, let the maximal test length be denoted by l₂(n). We obtain the value of this function for an infinite sequence of values of n.     

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.     

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.     

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.     

Sequential estimation of the mean of NEF-PVF distributions

Let F = {F₀: 0 ϵ Θ} denote the class of natural exponential family of distributions having power variance function, (NEF-PVF). We consider the problem of sequentially estimating the mean μ of F₀ ϵ F, based on i.i.d. observations from F₀. We propose an appropriate sequential estimation procedure under a combined loss of estimation error and sampling cost. We provide expansion for the regret R_a and study its asymptotic properties. We show that R_a = cv²(μ) + o(1) as a → ∞, where c > 0 is a known constant and v(μ) denotes the coefficient of variation of F₀.     

Characterization of diagonal symmetry: location unknown,and a test based on allied U-processes

For a (possibly multivariate) distribution F, a characterization of (diagonal) symmetry is made with respect to a kernel of degree 2; this is incorporated in the formulation of appropriate U-processes that provide the access to a suitable test statistic for testing the hypothesis of diagonal symmetry when the location is treated as unknown. Asymptotic properties of the test are studied.     

Admissible and optimal sampling strategy for estimating finite population mean in randomized response surveys with multiple responses

ABSTRACT

We consider the problem of estimation of a finite population mean (or proportion) related to a sensitive character under a randomized response model when independent responses are obtained from each sampled individual as many times as he/she is selected in the sample and prove the admissibility of a sampling strategy in a class of comparable linear unbiased strategies. We prove that the admissible strategy is also optimal in this class under a super-population model.     

An exact upper limit for the variance bias in a carry-over model with correlated errors

The analysis of crossover designs assuming i.i.d. errors leads to biased variance estimates whenever the true covariance structure is not spherical. As a result, the OLS F-test for the equality of the direct effects of the treatments is not valid. Bellavance et al. [1996. Biometrics 52, 607–612] use simulations to show that a modified F-test based on an estimate of the within subjects covariance matrix allows for nearly unbiased tests. Kunert and Utzig [1993. JRSS B 55, 919–927] propose an alternative test that does not need an estimate of the covariance matrix. Instead, they correct the F-statistic by multiplying by a constant based on the worst-case scenario. However, for designs with more than three observations per subject, Kunert and Utzig (1993) only give a rough upper bound for the worst-case variance bias. This may lead to overly conservative tests. In this paper we derive an exact upper limit for the variance bias due to carry-over for an arbitrary number of observations per subject. The result holds for a certain class of highly efficient balanced crossover designs.     

A pre-test like estimator dominating the least-squares method

We develop a pre-test type estimator of a deterministic parameter vector β$\mathit{\beta }$ in a linear Gaussian regression model. In contrast to conventional pre-test strategies, that do not dominate the least-squares (LS) method in terms of mean-squared error (MSE), our technique is shown to dominate LS when the effective dimension is greater than or equal to 4. Our estimator is based on a simple and intuitive approach in which we first determine the linear minimum MSE (MMSE) estimate that minimizes the MSE. Since the unknown vector β$\mathit{\beta }$ is deterministic, the MSE, and consequently the MMSE solution, will depend in general on β$\mathit{\beta }$ and therefore cannot be implemented. Instead, we propose applying the linear MMSE strategy with the LS substituted for the true value of β$\mathit{\beta }$ to obtain a new estimate. We then use the current estimate in conjunction with the linear MMSE solution to generate another estimate and continue iterating until convergence. As we show, the limit is a pre-test type method which is zero when the norm of the data is small, and is otherwise a non-linear shrinkage of LS.     

Robust designs for nearly linear regression

In this paper we seek designs and estimators which are optimal in some sense for multivariate linear regression on cubes and simplexes when the true regression function is unknown. More precisely, we assume that the unknown true regression function is the sum of a linear part plus some contamination orthogonal to the set of all linear functions in the L₂ norm with respect to Lebesgue measure. The contamination is assumed bounded in absolute value and it is shown that the usual designs for multivariate linear regression on cubes and simplices and the usual least squares estimators minimize the supremum over all possible contaminations of the expected mean square error. Additional results for extrapolation and interpolation, among other things, are discussed. For suitable loss functions optimal designs are found to have support on the extreme points of our design space.     

Multistage bandit problems

Consider two or more treatments with dichotomous responses. The total number N of experimental units are to be allocated in a fixed number r of stages. The problem is to decide how many units to assign to each treatment in each stage. Responses from selections in previous stages are available and can be considered but responses in the current stage are not available until the next group of selections is made. Information is updated via the Bayes theorem after each stage. The goal is to maximize the overall expected number of successes in the N units.Two forms of prior information are considered: (i) All arms have beta priors, and (ii) prior distributions have continuous densities. Various characteristics of optimal decisions are presented. For example, in most cases of (i) and (ii), the rate of the optimal size of the first stage cannot be greater than √N when r = 2.     

On the adaptive wavelet deconvolution of a density for strong mixing sequences

This paper studies the estimation of a density in the convolution density model from strong mixing observations. The ordinary smooth case is considered. Adopting the minimax approach under the mean integrated square error over Besov balls, we explore the performances of two wavelet estimators: a linear one based on projections and a non-linear one based on a hard thresholding rule. The feature of the non-linear one is to be adaptive, i.e., it does not require any prior knowledge of the smoothness class of the unknown density in its construction. We prove that it attains a fast rate of convergence which corresponds to the optimal one obtained in the standard i.i.d. case up to a logarithmic term.     

A new characterization of Elfving's method for high dimensional computation

We give a new characterization of Elfving's (1952) method for computing c-optimal designs in k dimensions which gives explicit formulae for the k unknown optimal weights and k unknown signs in Elfving's characterization. This eliminates the need to search over these parameters to compute c-optimal designs, and thus reduces the computational burden from solving a family of optimization problems to solving a single optimization problem for the optimal finite support set. We give two illustrative examples: a high dimensional polynomial regression model and a logistic regression model, the latter showing that the method can be used for locally optimal designs in nonlinear models as well.     

On the construction of orthogonal F-squares of order n from an orthogonal array (n, k, s, 2) and an OL(s, t) set

Complete sets of orthogonal F-squares of order n = s^p, where g is a prime or prime power and p is a positive integer have been constructed by Hedayat, Raghavarao, and Seiden (1975). Federer (1977) has constructed complete sets of orthogonal F-squares of order n = 4t, where t is a positive integer. We give a general procedure for constructing orthogonal F-squares of order n from an orthogonal array (n, k, s, 2) and an OL(s, t) set, where n is not necessarily a prime or prime power. In particular, we show how to construct sets of orthogonal F-squares of order n = 2s^p, where s is a prime or prime power and p is a positive integer. These sets are shown to be near complete and approach complete sets as s and/or p become large. We have also shown how to construct orthogonal arrays by these methods. In addition, the best upper bound on the number t of orthogonal F(n, λ₁), F(n, λ₂), …, F(n, λ₁) squares is given.     

On testing more IFRA ordering-II

ABSTRACT

Suppose F and G are two life distribution functions. It is said that F is more IFRA (increasing failure rate average) than G (written by F ? _*G) if G^{? 1}F(x) is star-shaped on (0, ∞). In this paper, the problem of testing H₀: F = _*G against H₁: F ? _*G and F ≠ _*G is considered in both cases when G is known and when G is unknown. We propose a new test based on U-statistics and obtain the asymptotic distribution of the test statistics. The new test is compared with some well-known tests in the literature. In addition, we apply our test to a real data set in the context of reliability.     

Choice experiments for estimating main effects and interactions

Choice-based conjoint experiments are used when choice alternatives can be described in terms of attributes. The objective is to infer the value that respondents attach to attribute levels. This method involves the design of profiles on the basis of attributes specified at certain levels. Respondents are presented sets of profiles and asked to select the one they consider best. However if choice sets have too many profiles, they may be difficult to implement. In this paper we provide strategies for reducing the number of profiles in choice sets. We consider situations where only a subset of interactions is of interest, and we obtain connected main effect plans with smaller choice sets that are capable of estimating subsets of two-factor and three-factor interactions in 2ⁿ and 3ⁿ plans. We also provide connected main effect plans for mixed level designs.