Biometrika - recent issues

Hierarchical testing of variable importance

Meinshausen, N. — 2008-05-24

A frequently encountered challenge in high-dimensional regression is the detection of relevant variables. Variable selection suffers from instability and the power to detect relevant variables is typically low if predictor variables are highly correlated. When taking the multiplicity of the testing problem into account, the power diminishes even further. To gain power and insight, it can be advantageous to look for influence not at the level of individual variables but rather at the level of clusters of highly correlated variables. We propose a hierarchical approach. Variable importance is first tested at the coarsest level, corresponding to the global null hypothesis. The method then tries to attribute any effect to smaller subclusters or even individual variables. The smallest possible clusters, which still exhibit a significant influence on the response variable, are retained. It is shown that the proposed testing procedure controls the familywise error rate at a prespecified level, simultaneously over all resolution levels. The method has power comparable to the Bonferroni–Holm procedure on the level of individual variables and dramatically larger power for coarser resolution levels. The best resolution level is selected adaptively.

On weighted Hochberg procedures

Tamhane, A. C., Liu, L. — 2008-05-24

We consider different ways of constructing weighted Hochberg-type step-up multiple test procedures including closed procedures based on weighted Simes tests and their conservative step-up short-cuts, and step-up counterparts of two weighted Holm procedures. It is shown that the step-up counterparts have some serious pitfalls such as lack of familywise error rate control and lack of monotonicity in rejection decisions in terms of p-values. Therefore an exact closed procedure appears to be the best alternative, its only drawback being lack of simple stepwise structure. A conservative step-up short-cut to the closed procedure may be used instead, but with accompanying loss of power. Simulations are used to study the familywise error rate and power properties of the competing procedures for independent and correlated p-values. Although many of the results of this paper are negative, they are useful in highlighting the need for caution when procedures with similar pitfalls may be used.

A family of Bayes multiple testing procedures

Cohen, A., Sackrowitz, H. B., Xu, M., Buyske, S. — 2008-05-24

Under the model of independent test statistics, we propose a two-parameter family of Bayes multiple testing procedures. The two parameters can be viewed as tuning parameters. Using the Benjamini–Hochberg step-up procedure for controlling false discovery rate as a baseline for conservativeness, we choose the tuning parameters to compromise between the operating characteristics of that procedure and a less conservative procedure that focuses on alternatives that a priori might be considered likely or meaningful. The Bayes procedures do not have the theoretical and practical shortcomings of the popular stepwise procedures. In terms of the number of mistakes, simulations for two examples indicate that over a large segment of the parameter space, the Bayes procedure is preferable to the step-up procedure. Another desirable feature of the procedures is that they are computationally feasible for any number of hypotheses.

Kernel stick-breaking processes

Dunson, D. B., Park, J.-H. — 2008-05-24

We propose a class of kernel stick-breaking processes for uncountable collections of dependent random probability measures. The process is constructed by first introducing an infinite sequence of random locations. Independent random probability measures and beta-distributed random weights are assigned to each location. Predictor-dependent random probability measures are then constructed by mixing over the locations, with stick-breaking probabilities expressed as a kernel multiplied by the beta weights. Some theoretical properties of the process are described, including a covariate-dependent prediction rule. A retrospective Markov chain Monte Carlo algorithm is developed for posterior computation, and the methods are illustrated using a simulated example and an epidemiological application.

Objective Bayesian analysis for the Student-t regression model

Fonseca, T. C. O., Ferreira, M. A. R., Migon, H. S. — 2008-05-24

We develop a Bayesian analysis based on two different Jeffreys priors for the Student-t regression model with unknown degrees of freedom. It is typically difficult to estimate the number of degrees of freedom: improper prior distributions may lead to improper posterior distributions, whereas proper prior distributions may dominate the analysis. We show that Bayesian analysis with either of the two considered Jeffreys priors provides a proper posterior distribution. Finally, we show that Bayesian estimators based on Jeffreys analysis compare favourably to other Bayesian estimators based on priors previously proposed in the literature.

Multi-parameter automodels and their applications

Hardouin, C., Yao, J.-F. — 2008-05-24

Motivated by the modelling of non-Gaussian data or positively correlated data on a lattice, extensions of Besag's automodels to exponential families with multi-dimensional parameters have been proposed recently. We provide a multiple-parameter analogue of Besag's one-dimensional result that gives the necessary form of the exponential families for the Markov random field's conditional distributions. We propose estimation of parameters by maximum pseudolikelihood and give a proof of the consistency of the estimators for the multi-parameter automodel. The methodology is illustrated with examples, in particular the building of a cooperative system with beta conditional distributions. We also indicate future applications of these models to the analysis of mixed-state spatial data.

Estimating functions for inhomogeneous spatial point processes with incomplete covariate data

Waagepetersen, R. — 2008-05-24

The R package spatstat provides a very flexible and useful framework for analysing spatial point patterns. A fundamental feature is a procedure for fitting spatial point process models depending on covariates. However, in practice one often faces incomplete observation of the covariates and this leads to parameter estimation error which is difficult to quantify. In this paper, we introduce a Monte Carlo version of the estimating function used in spatstat for fitting inhomogeneous Poisson processes and certain inhomogeneous cluster processes. For this modified estimating function, it is feasible to obtain the asymptotic distribution of the parameter estimators in the case of incomplete covariate information. This allows a study of the loss of efficiency due to the missing covariate data.

Modelling multiple time series via common factors

Pan, J., Yao, Q. — 2008-05-24

We propose a new method for estimating common factors of multiple time series. One distinctive feature of the new approach is that it is applicable to some nonstationary time series. The unobservable, nonstationary factors are identified by expanding the white noise space step by step, thereby solving a high-dimensional optimization problem by several low-dimensional sub-problems. Asymptotic properties of the estimation are investigated. The proposed methodology is illustrated with both simulated and real datasets.

Simultaneous confidence bands in spectral density estimation

Neumann, M. H., Paparoditis, E. — 2008-05-24

We propose a method for the construction of simultaneous confidence bands for a smoothed version of the spectral density of a Gaussian process based on nonparametric kernel estimators obtained by smoothing the periodogram. A studentized statistic is used to determine the width of the band at each frequency and a frequency-domain bootstrap approach is employed to estimate the distribution of the supremum of this statistic over all frequencies. We prove by means of strong approximations that the bootstrap estimates consistently the distribution of the supremum deviation of interest and, consequently, that the proposed confidence bands achieve asymptotically the desired simultaneous coverage probability. The behaviour of our method in finite-sample situations is investigated by simulations and a real-life data example demonstrates its applicability in time series analysis.

Least absolute deviation estimation for fractionally integrated autoregressive moving average time series models with conditional heteroscedasticity

Li, G., Li, W. K. — 2008-05-24

We consider a unified least absolute deviation estimator for stationary and nonstationary fractionally integrated autoregressive moving average models with conditional heteroscedasticity. Its asymptotic normality is established when the second moments of errors and innovations are finite. Several other alternative estimators are also discussed and are shown to be less efficient and less robust than the proposed approach. A diagnostic tool, consisting of two portmanteau tests, is designed to check whether or not the estimated models are adequate. The simulation experiments give further support to our model and the results for the absolute returns of the Dow Jones Industrial Average Index daily closing price demonstrate their usefulness in modelling time series exhibiting the features of long memory, conditional heteroscedasticity and heavy tails.

On the asymptotics of penalized splines

Li, Y., Ruppert, D. — 2008-05-24

We study the asymptotic behaviour of penalized spline estimators in the univariate case. We use B-splines and a penalty is placed on mth-order differences of the coefficients. The number of knots is assumed to converge to infinity as the sample size increases. We show that penalized splines behave similarly to Nadaraya--Watson kernel estimators with ‘equivalent’ kernels depending upon m. The equivalent kernels we obtain for penalized splines are the same as those found by Silverman for smoothing splines. The asymptotic distribution of the penalized spline estimator is Gaussian and we give simple expressions for the asymptotic mean and variance. Provided that it is fast enough, the rate at which the number of knots converges to infinity does not affect the asymptotic distribution. The optimal rate of convergence of the penalty parameter is given. Penalized splines are not design-adaptive.

Nonparametric variance estimation in the analysis of microarray data: a measurement error approach

Carroll, R. J., Wang, Y. — 2008-05-24

We investigate the effects of measurement error on the estimation of nonparametric variance functions. We show that either ignoring measurement error or direct application of the simulation extrapolation, SIMEX, method leads to inconsistent estimators. Nevertheless, the direct SIMEX method can reduce bias relative to a naive estimator. We further propose a permutation SIMEX method that leads to consistent estimators in theory. The performance of both the SIMEX methods depends on approximations to the exact extrapolants. Simulations show that both the SIMEX methods perform better than ignoring measurement error. The methodology is illustrated using microarray data from colon cancer patients.

Model diagnosis for parametric regression in high-dimensional spaces

Stute, W., Xu, W. L., Zhu, L. X. — 2008-05-24

We study tools for checking the validity of a parametric regression model. When the dimension of the regressors is large, many of the existing tests face the curse of dimensionality or require some ordering of the data. Our tests are based on the residual empirical process marked by proper functions of the regressors. They are able to detect local alternatives converging to the null at parametric rates. Parametric and nonparametric alternatives are considered. In the latter case, through a proper principal component decomposition, we are able to derive smooth directional tests which are asymptotically distribution-free under the null model. The new tests take into account precisely the ‘geometry of the model’. A simulation study is carried through and an application to a real dataset is illustrated.

Determining the dimension of the central subspace and central mean subspace

Zeng, P. — 2008-05-24

The central subspace and central mean subspace are two important targets of sufficient dimension reduction. We propose a weighted chi-squared test to determine their dimensions based on matrices whose column spaces are exactly equal to the central subspace or the central mean subspace. The asymptotic distribution of the test statistic is obtained. Simulation examples are used to demonstrate the performance of this test.

The prognostic analogue of the propensity score

Hansen, B. B. — 2008-05-24

The propensity score collapses the covariates of an observational study into a single measure summarizing their joint association with treatment conditions; prognostic scores summarize covariates' association with potential responses. As with propensity scores, stratification on prognostic scores brings to uncontrolled studies a concrete and desirable form of balance, a balance that is more familiar as an objective of experimental control. Like propensity scores, prognostic scores can reduce the dimension of the covariate, yet causal inferences conditional on them are as valid as are inferences conditional only on the unreduced covariate. As a method of adjustment unto itself, prognostic scoring has limitations not shared with propensity scoring, but it holds promise as a complement to the propensity score, particularly in certain designs for which unassisted propensity adjustment is difficult or infeasible.

Diagnostic measures for empirical likelihood of general estimating equations

Zhu, H., Ibrahim, J. G., Tang, N., Zhang, H. — 2008-05-24

We develop diagnostic measures for assessing the influence of individual observations when using empirical likelihood with general estimating equations, and we use these measures to construct goodness-of-fit statistics for testing possible misspecification in the estimating equations. Our diagnostics include case-deletion measures, local influence measures and pseudo-residuals. Our goodness-of-fit statistics include the sum of local influence measures and the processes of pseudo-residuals. Simulation studies are conducted to evaluate our methods, and real datasets are analyzed to illustrate the use of our diagnostic measures and goodness-of-fit statistics.

A note on deletion diagnostics for estimating equations

Preisser, J. S., Qaqish, B. F., Perin, J. — 2008-05-24

We describe an algorithm based upon the Sherman–Morrison–Woodbury formula for the inversion of matrices with special structure that occur in formulae for deletion diagnostics. Substantial computational savings relative to a method based upon Cholesky's decomposition are illustrated. The result has broad application to regression diagnostics for clustered data.

A new class of average moment matching priors

Ganesh, N., Lahiri, P. — 2008-05-24

We derive a new class of priors for the variance component in the Fay–Herriot model, a mixed regression model widely used in small area estimation. This class includes the well-known uniform or superharmonic prior. Through simulation we illustrate the use of our class of priors.

Studentization and deriving accurate p-values

Fraser, D.A.S., Rousseau, J. — 2008-02-28

We have a statistic for assessing an observed data point relative to a statistical model but find that its distribution function depends on the parameter. To obtain the corresponding p-value, we require the minimally modified statistic that is ancillary; this process is called Studentization. We use recent likelihood theory to develop a maximal third-order ancillary; this gives immediately a candidate Studentized statistic. We show that the corresponding p-value is higher-order Un(0, 1), is equivalent to a repeated bootstrap version of the initial statistic and agrees with a special Bayesian modification of the original statistic. More importantly, the modified statistic and p-value are available by Markov chain Monte Carlo simulations and, in some cases, by higher-order approximation methods. Examples, including the Behrens–Fisher problem, are given to indicate the ease and flexibility of the approach.

Distortion of effects caused by indirect confounding

Wermuth, N., Cox, D. R. — 2008-02-28

Undetected confounding may severely distort the effect of an explanatory variable on a response variable, as defined by a stepwise data-generating process. The best known type of distortion, which we call direct confounding, arises from an unobserved explanatory variable common to a response and its main explanatory variable of interest. It is relevant mainly for observational studies, since it is avoided by successful randomization. By contrast, indirect confounding, which we identify in this paper, is an issue also for intervention studies. For general stepwise-generating processes, we provide matrix and graphical criteria to decide which types of distortion may be present, when they are absent and how they are avoided. We then turn to linear systems without other types of distortion, but with indirect confounding. For such systems, the magnitude of distortion in a least-squares regression coefficient is derived and shown to be estimable, so that it becomes possible to recover the effect of the generating process from the distorted coefficient.

Population intervention models in causal inference

Hubbard, A. E., van der Laan, M. J. — 2008-02-28

We propose a new causal parameter, which is a natural extension of existing approaches to causal inference such as marginal structural models. Modelling approaches are proposed for the difference between a treatment-specific counterfactual population distribution and the actual population distribution of an outcome in the target population of interest. Relevant parameters describe the effect of a hypothetical intervention on such a population and therefore we refer to these models as population intervention models. We focus on intervention models estimating the effect of an intervention in terms of a difference and ratio of means, called risk difference and relative risk if the outcome is binary. We provide a class of inverse-probability-of-treatment-weighted and doubly-robust estimators of the causal parameters in these models. The finite-sample performance of these new estimators is explored in a simulation study.

Empirical and counterfactual conditions for sufficient cause interactions

Vanderweele, T. J., Robins, J. M. — 2008-02-28

Sufficient-component causes are discussed within the deterministic potential outcomes framework so as to formalize notions of sufficient causes, synergism and sufficient cause interactions. Doing so allows for the derivation of counterfactual and empirical conditions for detecting the presence of sufficient cause interactions. The conditions are novel in that, unlike other conditions in the literature, they make no assumption about monotonicity. The conditions can also be generalized and the conditions for three-way sufficient cause interactions are given explicitly. The statistical tests derived for sufficient cause interactions are compared with and contrasted to interaction terms in standard statistical models.

Shared parameter models under random effects misspecification

Rizopoulos, D., Verbeke, G., Molenberghs, G. — 2008-02-28

A common objective in longitudinal studies is the investigation of the association structure between a longitudinal response process and the time to an event of interest. An attractive paradigm for the joint modelling of longitudinal and survival processes is the shared parameter framework, where a set of random effects is assumed to induce their interdependence. In this work, we propose an alternative parameterization for shared parameter models and investigate the effect of misspecifying the random effects distribution in the parameter estimates and their standard errors.

Predicting future responses based on possibly mis-specified working models

Cai, T., Tian, L., Solomon, S. D., Wei, L.J. — 2008-02-28

Under a general regression setting, we propose an optimal unconditional prediction procedure for future responses. The resulting prediction intervals or regions have a desirable average coverage level over a set of covariate vectors of interest. When the working model is not correctly specified, the traditional conditional prediction method is generally invalid. On the other hand, one can empirically calibrate the above unconditional procedure and also obtain its crossvalidated counterpart. Various large and small sample properties of these unconditional methods are examined analytically and numerically. We find that the K-fold crossvalidated procedure performs exceptionally well even for cases with rather small sample sizes. The new proposals are illustrated with two real examples, one with a continuous response and the other with a binary outcome.

Flexible generalized t-link models for binary response data

Kim, S., Chen, M.-H., Dey, D. K. — 2008-02-28

A critical issue in modelling binary response data is the choice of the links. We introduce a new link based on the generalized t-distribution. There are two parameters in the generalized t-link: one parameter purely controls the heaviness of the tails of the link and the second parameter controls the scale of the link. Two major advantages are offered by the generalized t-links. First, a symmetric generalized t-link with an unknown shape parameter is much more identifiable than a Student t-link with unknown degrees of freedom and a known scale parameter. Secondly, skewed generalized t-links with both unknown shape and scale parameters provide much more flexible and improved skewed link regression models than the existing skewed links. Various theoretical properties and attractive features of the proposed links are examined and explored in detail. An efficient Markov chain Monte Carlo algorithm is developed for sampling from the posterior distribution. The deviance information criterion measure is used for guiding the choice of links. The proposed methodology is motivated and illustrated by prostate cancer data.

Analysis of least absolute deviation

Chen, K., Ying, Z., Zhang, H., Zhao, L. — 2008-02-28

We develop a unified L₁-based analysis-of-variance-type method for testing linear hypotheses. Like the classical L₂-based analysis of variance, the method is coordinate-free in the sense that it is invariant under any linear transformation of the covariates or regression parameters. Moreover, it allows singular design matrices and heterogeneous error terms. A simple approximation using stochastic perturbation is proposed to obtain cut-off values for the resulting test statistics. Both test statistics and distributional approximations can be computed using standard linear programming. An asymptotic theory is derived for the method. Special cases of one- and multi-way analysis of variance and analysis of covariance models are worked out in detail. The main results of this paper can be extended to general quantile regression. Extensive simulations show that the method works well in practical settings. The method is also applied to a dataset from General Social Surveys.

Nonparametric regression using local kernel estimating equations for correlated failure time data

Yu, Z., Lin, X. — 2008-02-28

We study nonparametric regression for correlated failure time data. Kernel estimating equations are used to estimate nonparametric covariate effects. Independent and weighted-kernel estimating equations are studied. The derivative of the nonparametric function is first estimated and the nonparametric function is then estimated by integrating the derivative estimator. We show that the nonparametric kernel estimator is consistent for any arbitrary working correlation matrix and that its asymptotic variance is minimized by assuming working independence. We evaluate the performance of the proposed kernel estimator using simulation studies, and apply the proposed method to the western Kenya parasitaemia data.

Bayesian and frequentist confidence intervals arising from empirical-type likelihoods

Chang, I. H., Mukerjee, R. — 2008-02-28

For a general class of empirical-type likelihoods for the population mean, higher-order asymptotics are developed with a view to characterizing its members which allow, for any given prior, the existence of a confidence interval that has approximately correct posterior as well as frequentist coverage. In particular, it is seen that the usual empirical likelihood always allows such a confidence interval, while many of its variants proposed in the literature do not enjoy this property. An explicit form of the confidence interval is also given.

Probability estimation for large-margin classifiers

Wang, J., Shen, X., Liu, Y. — 2008-02-28

Large margin classifiers have proven to be effective in delivering high predictive accuracy, particularly those focusing on the decision boundaries and bypassing the requirement of estimating the class probability given input for discrimination. As a result, these classifiers may not directly yield an estimated class probability, which is of interest itself. To overcome this difficulty, this article proposes a novel method for estimating the class probability through sequential classifications, by using features of interval estimation of large-margin classifiers. The method uses sequential classifications to bracket the class probability to yield an estimate up to the desired level of accuracy. The method is implemented for support vector machines and -learning, in addition to an estimated Kullback–Leibler loss for tuning. A solution path of the method is derived for support vector machines to reduce further its computational cost. Theoretical and numerical analyses indicate that the method is highly competitive against alternatives, especially when the dimension of the input greatly exceeds the sample size. Finally, an application to leukaemia data is described.

Retrospective Markov chain Monte Carlo methods for Dirichlet process hierarchical models

Papaspiliopoulos, O., Roberts, G. O. — 2008-02-28

Inference for Dirichlet process hierarchical models is typically performed using Markov chain Monte Carlo methods, which can be roughly categorized into marginal and conditional methods. The former integrate out analytically the infinite-dimensional component of the hierarchical model and sample from the marginal distribution of the remaining variables using the Gibbs sampler. Conditional methods impute the Dirichlet process and update it as a component of the Gibbs sampler. Since this requires imputation of an infinite-dimensional process, implementation of the conditional method has relied on finite approximations. In this paper, we show how to avoid such approximations by designing two novel Markov chain Monte Carlo algorithms which sample from the exact posterior distribution of quantities of interest. The approximations are avoided by the new technique of retrospective sampling. We also show how the algorithms can obtain samples from functionals of the Dirichlet process. The marginal and the conditional methods are compared and a careful simulation study is included, which involves a non-conjugate model, different datasets and prior specifications.

Two-stage sampling from a prediction point of view when the cluster sizes are unknown

Bjornstad, J. F., Ytterstad, E. — 2008-02-28

We consider the problem of estimating the population total in two-stage cluster sampling when cluster sizes are known only for the sampled clusters, making use of a population model arising from a variance component model. The problem can be considered as one of predicting the unobserved part Z of the total, and the concept of predictive likelihood is studied. Prediction intervals and a predictor for the population total are derived for the normal case, based on predictive likelihood. For a more general distribution-free model, by application of an analysis of variance approach instead of maximum likelihood for parameter estimation, the predictor obtained from the predictive likelihood is shown to be approximately uniformly optimal for large sample size and large number of clusters, in the sense of uniformly minimizing the mean-squared error in a partially linear class of model-unbiased predictors. Three prediction intervals for Z based on three similar predictive likelihoods are studied. For a small number n₀ of sampled clusters, they differ significantly, but for large n₀, the three intervals are practically identical. Model-based and design-based coverage properties of the prediction intervals are studied based on a comprehensive simulation study. The simulation study indicates that for large sample sizes, the coverage measures achieve approximately the nominal level 1 – and are slightly less than 1 – for moderately large sample sizes. For small sample sizes, the coverage measures are about 1 – 2, being raised to 1 – for a modified interval based on the distribution.

Predicting cumulative incidence probability by direct binomial regression

Scheike, T. H., Zhang, M.-J., Gerds, T. A. — 2008-02-28

We suggest a new simple approach for estimation and assessment of covariate effects for the cumulative incidence curve in the competing risks model. We consider a semiparametric regression model where some effects may be time-varying and some may be constant over time. Our estimator can be implemented by standard software. Our simulation study shows that the estimator works well and has finite-sample properties comparable with the subdistribution approach. We apply the method to bone marrow transplant data and estimate the cumulative incidence of death in complete remission following a bone marrow transplantation. Here death in complete remission and relapse are two competing events.

Nonparametric estimation of bivariate failure time associations in the presence of a competing risk

Bandeen-Roche, K., Ning, J. — 2008-02-28

Most research on the study of associations among paired failure times has either assumed time invariance or been based on complex measures or estimators. Little has accommodated competing risks. This paper targets the conditional cause-specific hazard ratio, henceforth called the cause-specific cross ratio, a recent modification of the conditional hazard ratio designed to accommodate competing risks data. Estimation is accomplished by an intuitive, nonparametric method that localizes Kendall's tau. Time variance is accommodated through a partitioning of space into ‘bins’ between which the strength of association may differ. Inferential procedures are developed, small-sample performance is evaluated, and the methods are applied to the investigation of familial association in dementia onset.

Nonparametric estimation of cause-specific cross hazard ratio with bivariate competing risks data

Cheng, Y., Fine, J. P. — 2008-02-28

We propose an alternative representation of the cause-specific cross hazard ratio for bivariate competing risks data. The representation leads to a simple plug-in estimator, unlike an existing ad hoc procedure. The large sample properties of the resulting inferences are established. Simulations and a real data example demonstrate that the proposed methodology may substantially reduce the computational burden of the existing procedure, while maintaining similar efficiency properties.

A note on path-based variable selection in the penalized proportional hazards model

Zou, H. — 2008-02-28

We propose an efficient and adaptive shrinkage method for variable selection in the Cox model. The method constructs a piecewise-linear regularization path connecting the maximum partial likelihood estimator and the origin. Then a model is selected along the path. We show that the constructed path is adaptive in the sense that, with a proper choice of regularization parameter, the fitted model works as well as if the true underlying submodel were given in advance. A modified algorithm of the least-angle-regression type efficiently computes the entire regularization path of the new estimator. Furthermore, we show that, with a proper choice of shrinkage parameter, the method is consistent in variable selection and efficient in estimation. Simulation shows that the new method tends to outperform the lasso and the smoothly-clipped-absolute-deviation estimators with moderate samples. We apply the methodology to data concerning nursing homes.

Testing hypotheses in order

Rosenbaum, P. R. — 2008-02-28

In certain circumstances, one wishes to test one hypothesis only if certain other hypotheses have been rejected. This ordering of hypotheses simplifies the task of controlling the probability of rejecting any true hypothesis. In an example from an observational study, a treated group is shown to be further from both of two control groups than the two control groups are from each other.

A Note on repeated p-values for group sequential designs

Posch, M., Wassmer, G., Brannath, W. — 2008-02-28

One-sided confidence intervals and overall p-values for group-sequential designs are typically based on a sample space ordering which determines both the overall p-value and the corresponding confidence bound. Accordingly, the strength of evidence against the null hypothesis is consistently measured by both quantities such that the order of the p-values of two distinct sample points is consistent with the order of the respective confidence bounds. An exception is the commonly used repeated p-values and repeated confidence intervals. We show that they are not ordering-consistent in the above sense and propose an alternative repeated p-value which is ordering-consistent and has the monitoring property of the classical repeated p-value in being valid even when deviating from the prefixed stopping rule.

Asymptotic inference for a nonstationary double AR(1) model

Ling, S., Li, D. — 2008-02-28

We investigate the nonstationary double ar(1) model, where > 0, > 0, the _t are independent standard normal random variables and Elog | + _t| >= 0. We show that the maximum likelihood estimator of (, ) is consistent and asymptotically normal. Combination of this result with that in Ling ([11]) for the stationary case gives the asymptotic normality of the maximum likelihood estimator of for any in the real line, with a root-n rate of convergence. This is in contrast to the results for the classical ar(1) model, corresponding to = 0.

Bayesian Nonparametric Estimation of the Probability of Discovering New Species

Lijoi, A., Mena, R. H., Prunster, I. — 2007-12-18

We consider the problem of evaluating the probability of discovering a certain number of new species in a new sample of population units, conditional on the number of species recorded in a basic sample. We use a Bayesian nonparametric approach. The different species proportions are assumed to be random and the observations from the population exchangeable. We provide a Bayesian estimator, under quadratic loss, for the probability of discovering new species which can be compared with well-known frequentist estimators. The results we obtain are illustrated through a numerical example and an application to a genomic dataset concerning the discovery of new genes by sequencing additional single-read sequences of cdna fragments.

Population-Based Reversible Jump Markov Chain Monte Carlo

Jasra, A., Stephens, D. A., Holmes, C. C. — 2007-12-18

We present an extension of population-based Markov chain Monte Carlo to the transdimensional case. A major challenge is that of simulating from high- and transdimensional target measures. In such cases, Markov chain Monte Carlo methods may not adequately traverse the support of the target; the simulation results will be unreliable. We develop population methods to deal with such problems, and give a result proving the uniform ergodicity of these population algorithms, under mild assumptions. This result is used to demonstrate the superiority, in terms of convergence rate, of a population transition kernel over a reversible jump sampler for a Bayesian variable selection problem. We also give an example of a population algorithm for a Bayesian multivariate mixture model with an unknown number of components. This is applied to gene expression data of 1000 data points in six dimensions and it is demonstrated that our algorithm outperforms some competing Markov chain samplers. In this example, we show how to combine the methods of parallel chains (Geyer, 1991), tempering (Geyer & Thompson, 1995), snooker algorithms (Gilks et al., 1994), constrained sampling and delayed rejection (Green & Mira, 2001).

Generalized Spatial Dirichlet Process Models

Duan, J. A., Guindani, M., Gelfand, A. E. — 2007-12-18

Many models for the study of point-referenced data explicitly introduce spatial random effects to capture residual spatial association. These spatial effects are customarily modelled as a zero-mean stationary Gaussian process. The spatial Dirichlet process introduced by Gelfand et al. (2005) produces a random spatial process which is neither Gaussian nor stationary. Rather, it varies about a process that is assumed to be stationary and Gaussian. The spatial Dirichlet process arises as a probability-weighted collection of random surfaces. This can be limiting for modelling and inferential purposes since it insists that a process realization must be one of these surfaces. We introduce a random distribution for the spatial effects that allows different surface selection at different sites. Moreover, we can specify the model so that the marginal distribution of the effect at each site still comes from a Dirichlet process. The development is offered constructively, providing a multivariate extension of the stick-breaking representation of the weights. We then introduce mixing using this generalized spatial Dirichlet process. We illustrate with a simulated dataset of independent replications and note that we can embed the generalized process within a dynamic model specification to eliminate the independence assumption.

Monte Carlo Estimation for Nonlinear Non-Gaussian State Space Models

Jungbacker, B., Koopman, S. J. — 2007-12-18

We develop a proposal or importance density for state space models with a nonlinear non-Gaussian observation vector y ~ p(y¦) and an unobserved linear Gaussian signal vector ~ p(). The proposal density is obtained from the Laplace approximation of the smoothing density p(¦y). We present efficient algorithms to calculate the mode of p(¦y) and to sample from the proposal density. The samples can be used for importance sampling and Markov chain Monte Carlo methods. The new results allow the application of these methods to state space models where the observation density p(y¦) is not log-concave. Additional results are presented that lead to computationally efficient implementations. We illustrate the methods for the stochastic volatility model with leverage.

Estimation of Regression Models for the Mean of Repeated Outcomes Under Nonignorable Nonmonotone Nonresponse

Vansteelandt, S., Rotnitzky, A., Robins, J. — 2007-12-18

We propose a new class of models for making inference about the mean of a vector of repeated outcomes when the outcome vector is incompletely observed in some study units and missingness is nonmonotone. Each model in our class is indexed by a set of unidentified selection-bias functions which quantify the residual association of the outcome at each occasion t and the probability that this outcome is missing after adjusting for variables observed prior to time t and for the past nonresponse pattern. In particular, selection-bias functions equal to zero encode the investigator's a priori belief that nonresponse of the next outcome does not depend on that outcome after adjusting for the observed past. We call this assumption sequential explainability. Since each model in our class is nonparametric, it fits the data perfectly well. As such, our models are ideal for conducting sensitivity analyses aimed at evaluating the impact that different degrees of departure from sequential explainability have on inference about the marginal means of interest. Although the marginal means are identified under each of our models, their estimation is not feasible in practice because it requires the auxiliary estimation of conditional expectations and probabilities given high-dimensional variables. We henceforth discuss the estimation of the marginal means under each model in our class assuming, additionally, that at each occasion either one of the following two models holds: a parametric model for the conditional probability of nonresponse given current outcomes and past recorded data or a parametric model for the conditional mean of the outcome on the nonrespondents given the past recorded data. We call the resulting procedure 2^T-multiply robust as it protects at each of the T time points against misspecification of one of these two working models, although not against simultaneous misspecification of both. We extend our proposed class of models and estimators to incorporate data configurations which include baseline covariates and a parametric model for the conditional mean of the vector of repeated outcomes given the baseline covariates.

Aalen Additive Hazards Change-Point Model

Martinussen, T., Scheike, T. H. — 2007-12-18

We study a test comparing the full Aalen additive hazards model and the change-point model, and suggest how to estimate the parameters of the change-point model. We also study a test for no change-point effect. Both tests are provided with large sample properties and a resampling method is applied to obtain p-values. The finite-sample properties of the proposed inference procedures and estimators are assessed through a simulation study. The methods are further applied to a dataset concerning myocardial infarction.

A General Approach to the Predictability Issue in Survival Analysis with Applications

Mammen, E., Nielsen, J. P. — 2007-12-18

Very often in survival analysis one has to study martingale integrals where the integrand is not predictable and where the counting process theory of martingales is not directly applicable, as for example in nonparametric and semiparametric applications where the integrand is based on a pilot estimate. We call this the predictability issue in survival analysis. The problem has been resolved by approximations of the integrand by predictable functions which have been justified by ad hoc procedures. We present a general approach to the solution of this problem. The usefulness of the approach is shown in three applications. In particular, we argue that earlier ad hoc procedures do not work in higher-dimensional smoothing problems in survival analysis.

The Role of Pseudo Data for Robust Smoothing with Application to Wavelet Regression

Oh, H.-S., Nychka, D. W., Lee, T. C. M. — 2007-12-18

We propose a robust curve and surface estimator based on M-type estimators and penalty-based smoothing. This approach also includes an application to wavelet regression. The concept of pseudo data, a transformation of the robust additive model to the one with bounded errors, is used to derive some theoretical properties and also motivate a computational algorithm. The resulting algorithm, termed the es-algorithm, is computationally fast and provides a simple way of choosing the amount of smoothing. Moreover, it is easily described, straightforwardly implemented and can be extended to other wavelet regression settings such as irregularly spaced data and image denoising. Results from a simulation study and real data examples demonstrate the promising empirical properties of the proposed approach.

Using Hierarchical Likelihood for Missing Data Problems

Yun, S.-C., Lee, Y., Kenward, M. G. — 2007-12-18

Most statistical solutions to the problem of statistical inference with missing data involve integration or expectation. This can be done in many ways: directly or indirectly, analytically or numerically, deterministically or stochastically. Missing-data problems can be formulated in terms of latent random variables, so that hierarchical likelihood methods of Lee & Nelder (1996) can be applied to missing-value problems to provide one solution to the problem of integration of the likelihood. The resulting methods effectively use a Laplace approximation to the marginal likelihood with an additional adjustment to the measures of precision to accommodate the estimation of the fixed effects parameters. We first consider missing at random cases where problems are simpler to handle because the integration does not need to involve the missing-value mechanism and then consider missing not at random cases. We also study tobit regression and refit the missing not at random selection model to the antidepressant trial data analyzed in Diggle & Kenward (1994).

Empirical Likelihood Semiparametric Regression Analysis for Longitudinal Data

Xue, L., Zhu, L. — 2007-12-18

A semiparametric regression model for longitudinal data is considered. The empirical likelihood method is used to estimate the regression coefficients and the baseline function, and to construct confidence regions and intervals. It is proved that the maximum empirical likelihood estimator of the regression coefficients achieves asymptotic efficiency and the estimator of the baseline function attains asymptotic normality when a bias correction is made. Two calibrated empirical likelihood approaches to inference for the baseline function are developed. We propose a groupwise empirical likelihood procedure to handle the inter-series dependence for the longitudinal semiparametric regression model, and employ bias correction to construct the empirical likelihood ratio functions for the parameters of interest. This leads us to prove a nonparametric version of Wilks' theorem. Compared with methods based on normal approximations, the empirical likelihood does not require consistent estimators for the asymptotic variance and bias. A simulation compares the empirical likelihood and normal-based methods in terms of coverage accuracies and average areas/lengths of confidence regions/intervals.

A Hybrid Pairwise Likelihood Method

Kuk, A. Y. C. — 2007-12-18

A modification to the pairwise likelihood method is proposed, which aims to improve the estimation of the marginal distribution parameters. This is achieved by replacing the pairwise likelihood score equations, for estimating such parameters, by the optimal linear combinations of the marginal score functions. A further advantage of the proposed estimator of marginal parameters, over pairwise likelihood, is that it is robust to misspecification of the bivariate distributions as long as the univariate marginal distributions are correctly specified. While alternating logistic regression can be seen as a special case of the proposed method, it is shown that an existing generalization of alternating logistic regression applicable to ordinal data is not the same as and is inferior to the proposed method because it replaces certain conditional densities by pseudodensities that assume working independence. The fitting of the multivariate negative binomial distribution is another scenario involving intractable likelihood that calls for the use of pairwise likelihood methods, and the superiority of the modified method is demonstrated in a simulation study. Two examples, based on the analyses of salamander mating and patient-controlled analgesia data, demonstrate the usefulness of the proposed method. The possibility of combining optimally the pairwise, rather than marginal, scores is also considered and its difficulty and potential are discussed.

A Jackknife Variance Estimator for Unistage Stratified Samples with Unequal Probabilities

Berger, Y. G. — 2007-12-18

Existing jackknife variance estimators used with sample surveys can seriously overestimate the true variance under unistage stratified sampling without replacement with unequal probabilities. A novel jackknife variance estimator is proposed which is as numerically simple as existing jackknife variance estimators. Under certain regularity conditions, the proposed variance estimator is consistent under stratified sampling without replacement with unequal probabilities. The high entropy regularity condition necessary for consistency is shown to hold for the Rao–Sampford design. An empirical study of three unequal probability sampling designs supports our findings.

Hochberg's Step-Up Method: Cutting Corners Off Holm's Step-Down Method

Huang, Y., Hsu, J. C. — 2007-12-18

Holm's method and Hochberg's method for multiple testing can be viewed as step-down and step-up versions of the Bonferroni test. We show that both are special cases of partition testing. The difference is that, while Holm's method tests each partition hypothesis using the largest order statistic, setting a critical value based on the Bonferroni inequality, Hochberg's method tests each partition hypothesis using all the order statistics, setting a series of critical values based on Simes' inequality. Geometrically, Hochberg's step-up method ‘cuts corners’ off the acceptance regions of Holm's step-down method by making assumptions on the joint distribution of the test statistics. As can be expected, partition testing making use of the joint distribution of the test statistics is more powerful than partition testing using probabilistic inequalities. Thus, if the joint distribution of the test statistics is available, through modelling for example, we recommend partition step-down testing, setting exact critical values based on the joint distribution.

Miscellanea Kernel-Type Density Estimation on the Unit Interval

Jones, M.C., Henderson, D.A. — 2007-12-18

We consider kernel-type methods for the estimation of a density on 0,1 which eschew explicit boundary correction. We propose using kernels that are symmetric in their two arguments; these kernels are conditional densities of bivariate copulas. We give asymptotic theory for the version of the new estimator using Gaussian copula kernels and report on simulation comparisons of it with the beta-kernel density estimator of Chen ([1]). We also provide automatic bandwidth selection in the form of ‘rule-of-thumb’ bandwidths for both estimators. As well as its competitive integrated squared error performance, advantages of the new approach include its greater range of possible values at 0 and 1, the fact that it is a bona fide density and that the individual kernels and resulting estimator are comprehensible in terms of a single simple picture.

Importance Sampling Via the Estimated Sampler

Henmi, M., Yoshida, R., Eguchi, S. — 2007-12-18

Monte Carlo importance sampling for evaluating numerical integration is discussed. We consider a parametric family of sampling distributions and propose the use of the sampling distribution estimated by maximum likelihood. The proposed method of importance sampling using the estimated sampling distribution is shown to improve the asymptotic variance of the ordinary method using the true sampling distribution. The argument is closely related to the discussion of the paradox in Henmi & Eguchi (2004). We focus on a condition under which the estimated integration value obtained by the proposed method has asymptotic zero variance.

Use of the Gibbs Sampler to Obtain Conditional Tests, with Applications

Lockhart, R. A., O'Reilly, F. J., Stephens, M. A. — 2007-12-18

A random sample is drawn from a distribution which admits a minimal sufficient statistic for the parameters. The Gibbs sampler is proposed to generate samples, called conditionally sufficient or co-sufficient samples, from the conditional distribution of the sample given its value of the sufficient statistic. The procedure is illustrated for the gamma distribution. Co-sufficient samples may be used to give exact tests of fit; for the gamma distribution these are compared for size and power with approximate tests based on the parametric bootstrap.

Positive Association Among Three Binary Variables and Cross-Product Ratios

Fienberg, S. E., Kim, S.-H. — 2007-12-18

We show that, when the three-way association level among the three binary variables, X, U₁ and U₂ is fixed, D_P = pr(X = 1¦U₁ = 1) – pr(X = 1¦U₁ = 0) increases as the cross-product ratio of U₁ and U₂ increases under the assumption that X is positively associated with U₁ and U₂. We then discuss some implications of this property.

Cholesky Decompositions and Estimation of A Covariance Matrix: Orthogonality of Variance Correlation Parameters

Pourahmadi, M. — 2007-12-18

Chen & Dunson ([3]) have proposed a modified Cholesky decomposition of the form = D L L'D for a covariance matrix where D is a diagonal matrix with entries proportional to the square roots of the diagonal entries of and L is a unit lower-triangular matrix solely determining its correlation matrix. This total separation of variance and correlation is definitely a major advantage over the more traditional modified Cholesky decomposition of the form LD²L', (Pourahmadi, [13]). We show that, though the variance and correlation parameters of the former decomposition are separate, they are not asymptotically orthogonal and that the estimation of the new parameters could be more demanding computationally. We also provide statistical interpretation for the entries of L and D as certain moving average parameters and innovation variances and indicate how the existing likelihood procedures can be employed to estimate the new parameters.