Summary. We examine philosophical problems and sampling deficiencies that are associated with current Bayesian hypothesis testing methodology, paying particular attention to objective Bayes methodology. Because the prior densities that are used to define alternative hypotheses in many Bayesian tests assign non-negligible probability to regions of the parameter space that are consistent with null hypotheses, resulting tests provide exponential accumulation of evidence in favour of true alternative hypotheses, but only sublinear accumulation of evidence in favour of true null hypotheses. Thus, it is often impossible for such tests to provide strong evidence in favour of a true null hypothesis, even when moderately large sample sizes have been obtained. We review asymptotic convergence rates of Bayes factors in testing precise null hypotheses and propose two new classes of prior densities that ameliorate the imbalance in convergence rates that is inherited by most Bayesian tests. Using members of these classes, we obtain analytic expressions for Bayes factors in linear models and derive approximations to Bayes factors in large sample settings.

Summary. Additive models are popular in high dimensional regression problems owing to their flexibility in model building and optimality in additive function estimation. Moreover, they do not suffer from the so-called curse of dimensionality generally arising in non-parametric regression settings. Less known is the model bias that is incurred from the restriction to the additive class of models. We introduce a new class of estimators that reduces additive model bias, yet preserves some stability of the additive estimator. The new estimator is constructed by localizing the additivity assumption and thus is named the local additive estimator. It follows the spirit of local linear estimation but is shown to be able to relieve partially the dimensionality problem. Implementation can be easily made with any standard software for additive regression. For detailed analysis we explicitly use the smooth backfitting estimator of Mammen, Linton and Nielsen.

Summary. In likelihood inference we usually assume that the model is fixed and then base inference on the corresponding likelihood function. Often, however, the choice of model is rather arbitrary, and there may be other models which fit the data equally well. We study robustness of likelihood inference over such 'statistically equivalent' models and suggest a simple 'envelope likelihood' to capture this aspect of model uncertainty. Robustness depends critically on how we specify the parameter of interest. Some asymptotic theory is presented, illustrated by three examples.

Summary. Several procedures that are designed to reduce nonconformity in interlaboratory studies by shrinking data towards a consensus weighted mean are suggested. Some of them are shown to have a smaller quadratic risk than the vector sample means. Shrinkage towards a weighted mean in a random-effects model and a statistic appearing in models which allow for systematic errors are also considered. The results are illustrated by two examples of collaborative studies.

Summary. Competing risks problems arise in many fields of science, where two or more types of event may occur on a subject, but only the event occurring first is observed together with its occurrence time, and other events are censored. The marginal and joint distributions of event times for competing risks cannot be identified from the observed data without assuming the relationship between events. The commonly adopted independent censoring assumption may be easily violated. An alternative is to assume that the joint distribution of event times follows a known copula, which is an explicit function of the marginal distributions. On the basis of the latter assumption, we consider marginal regression analysis for dependent competing risks, with the marginal regressions performed via semiparametric transformation models, including the proportional hazards and proportional odds models. We propose a non-parametric maximum likelihood analysis, which provides explicit expressions for the score functions and information matrix, and facilitates convenient computations. Large and finite sample properties are studied. We report an illustration with data from an acquired immune deficiency syndrome clinical trial where the censoring may be dependent. The proposal can be readily used as a sensitivity analysis for assessing effects of potential dependent censoring and can incorporate external information on the association of competing risks.

Summary. We address the problem of how to conduct a minimally informative, non-parametric Bayesian analysis. The central question is how to devise a model so that the posterior distribution satisfies a few basic properties. The concept of 'local mass' provides the key to the development of the limiting Dirichlet process model. This model is then used to provide an engine for inference in the compound decision problem and for multiple-comparisons inference in a one-way analysis-of-variance setting. Our analysis in this setting may be viewed as a limit of the analyses that were developed by Escobar and by Gopalan and Berry. Computations for the analysis are described, and the predictive performance of the model is compared with that of mixture of Dirichlet processes models.

Summary. The problem of component choice in regression-based prediction has a long history. The main cases where important choices must be made are functional data analysis, and problems in which the explanatory variables are relatively high dimensional vectors. Indeed, principal component analysis has become the basis for methods for functional linear regression. In this context the number of components can also be interpreted as a smoothing parameter, and so the viewpoint is a little different from that for standard linear regression. However, arguments for and against conventional component choice methods are relevant to both settings and have received significant recent attention. We give a theoretical argument, which is applicable in a wide variety of settings, justifying the conventional approach. Although our result is of minimax type, it is not asymptotic in nature; it holds for each sample size. Motivated by the insight that is gained from this analysis, we give theoretical and numerical justification for cross-validation choice of the number of components that is used for prediction. In particular we show that cross-validation leads to asymptotic minimization of mean summed squared error, in settings which include functional data analysis.

Summary. Partial least squares regression has been an alternative to ordinary least squares for handling multicollinearity in several areas of scientific research since the 1960s. It has recently gained much attention in the analysis of high dimensional genomic data. We show that known asymptotic consistency of the partial least squares estimator for a univariate response does not hold with the very large p and small n paradigm. We derive a similar result for a multivariate response regression with partial least squares. We then propose a sparse partial least squares formulation which aims simultaneously to achieve good predictive performance and variable selection by producing sparse linear combinations of the original predictors. We provide an efficient implementation of sparse partial least squares regression and compare it with well-known variable selection and dimension reduction approaches via simulation experiments. We illustrate the practical utility of sparse partial least squares regression in a joint analysis of gene expression and genomewide binding data.

Summary. In sample surveys of finite populations, subpopulations for which the sample size is too small for estimation of adequate precision are referred to as small domains. Demand for small domain estimates has been growing in recent years among users of survey data. We explore the possibility of enhancing the precision of domain estimators by combining comparable information collected in multiple surveys of the same population. For this, we propose a regression method of estimation that is essentially an extended calibration procedure whereby comparable domain estimates from the various surveys are calibrated to each other. We show through analytic results and an empirical study that this method may greatly improve the precision of domain estimators for the variables that are common to these surveys, as these estimators make effective use of increased sample size for the common survey items. The design-based direct estimators proposed involve only domain-specific data on the variables of interest. This is in contrast with small domain (mostly small area) indirect estimators, based on a single survey, which incorporate through modelling data that are external to the targeted small domains. The approach proposed is also highly effective in handling the closely related problem of estimation for rare population characteristics.

Summary. Local polynomial regression is a useful non-parametric regression tool to explore fine data structures and has been widely used in practice. We propose a new non-parametric regression technique called local composite quantile regression smoothing to improve local polynomial regression further. Sampling properties of the estimation procedure proposed are studied. We derive the asymptotic bias, variance and normality of the estimate proposed. The asymptotic relative efficiency of the estimate with respect to local polynomial regression is investigated. It is shown that the estimate can be much more efficient than the local polynomial regression estimate for various non-normal errors, while being almost as efficient as the local polynomial regression estimate for normal errors. Simulation is conducted to examine the performance of the estimates proposed. The simulation results are consistent with our theoretical findings. A real data example is used to illustrate the method proposed.

Summary. Linear pooling is by far the most popular method for combining probability forecasts. However, any non-trivial weighted average of two or more distinct, calibrated probability forecasts is necessarily uncalibrated and lacks sharpness. In view of this, linear pooling requires recalibration, even in the ideal case in which the individual forecasts are calibrated. Towards this end, we propose a beta-transformed linear opinion pool for the aggregation of probability forecasts from distinct, calibrated or uncalibrated sources. The method fits an optimal non-linearly recalibrated forecast combination, by compositing a beta transform and the traditional linear opinion pool. The technique is illustrated in a simulation example and in a case-study on statistical and National Weather Service probability of precipitation forecasts.

Summary. Formal rules governing signed edges on causal directed acyclic graphs are described and it is shown how these rules can be useful in reasoning about causality. Specifically, the notions of a monotonic effect, a weak monotonic effect and a signed edge are introduced. Results are developed relating these monotonic effects and signed edges to the sign of the causal effect of an intervention in the presence of intermediate variables. The incorporation of signed edges in the directed acyclic graph causal framework furthermore allows for the development of rules governing the relationship between monotonic effects and the sign of the covariance between two variables. It is shown that when certain assumptions about monotonic effects can be made then these results can be used to draw conclusions about the presence of causal effects even when data are missing on confounding variables.

Summary. When a treatment has a positive average causal effect (ACE) on an intermediate variable or surrogate end point which in turn has a positive ACE on a true end point, the treatment may have a negative ACE on the true end point due to the presence of unobserved confounders, which is called the surrogate paradox. A criterion for surrogate end points based on ACEs has recently been proposed to avoid the surrogate paradox. For a continuous or ordinal discrete end point, the distributional causal effect (DCE) may be a more appropriate measure for a causal effect than the ACE. We discuss criteria for surrogate end points based on DCEs. We show that commonly used models, such as generalized linear models and Cox's proportional hazard models, can make the sign of the DCE of the treatment on the true end point determinable by the sign of the DCE of the treatment on the surrogate even if the models include unobserved confounders. Furthermore, for a general distribution without any assumption of parametric models, we give a sufficient condition for a distributionally consistent surrogate and prove that it is almost necessary.