Probabilistic relabelling strategies for the label switching problem in Bayesian mixture models

The label switching problem is caused by the likelihood of a Bayesian mixture model being invariant to permutations of the labels. The permutation can change multiple times between Markov Chain Monte Carlo (MCMC) iterations making it difficult to infer component-specific parameters of the model. Various so-called ‘relabelling’ strategies exist with the goal to ‘undo’ the label switches that have occurred to enable estimation of functions that depend on component-specific parameters. Existing deterministic relabelling algorithms rely upon specifying a loss function, and relabelling by minimising its posterior expected loss. In this paper we develop probabilistic approaches to relabelling that allow for estimation and incorporation of the uncertainty in the relabelling process. Variants of the probabilistic relabelling algorithm are introduced and compared to existing deterministic relabelling algorithms. We demonstrate that the idea of probabilistic relabelling can be expressed in a rigorous framework based on the EM algorithm.     

Model selection for probabilistic clustering using cross-validated likelihood

Cross-validated likelihood is investigated as a tool for automatically determining the appropriate number of components (given the data) in finite mixture modeling, particularly in the context of model-based probabilistic clustering. The conceptual framework for the cross-validation approach to model selection is straightforward in the sense that models are judged directly on their estimated out-of-sample predictive performance. The cross-validation approach, as well as penalized likelihood and McLachlan's bootstrap method, are applied to two data sets and the results from all three methods are in close agreement. The second data set involves a well-known clustering problem from the atmospheric science literature using historical records of upper atmosphere geopotential height in the Northern hemisphere. Cross-validated likelihood provides an interpretable and objective solution to the atmospheric clustering problem. The clusters found are in agreement with prior analyses of the same data based on non-probabilistic clustering techniques.     

Maximin design on non hypercube domains and kernel interpolation

In the paradigm of computer experiments, the choice of an experimental design is an important issue. When no information is available about the black-box function to be approximated, an exploratory design has to be used. In this context, two dispersion criteria are usually considered: the minimax and the maximin ones. In the case of a hypercube domain, a standard strategy consists of taking the maximin design within the class of Latin hypercube designs. However, in a non hypercube context, it does not make sense to use the Latin hypercube strategy. Moreover, whatever the design is, the black-box function is typically approximated thanks to kernel interpolation. Here, we first provide a theoretical justification to the maximin criterion with respect to kernel interpolations. Then, we propose simulated annealing algorithms to determine maximin designs in any bounded connected domain. We prove the convergence of the different schemes. Finally, the methodology is applied on a challenging real example where the black-blox function describes the behaviour of an aircraft engine.     

Bayesian Model Averaging in Proportional Hazard Models: Assessing the Risk of a Stroke

In the context of the Cardiovascular Health Study, a comprehensive investigation into the risk factors for strokes, we apply Bayesian model averaging to the selection of variables in Cox proportional hazard models. We use an extension of the leaps-and-bounds algorithm for locating the models that are to be averaged over and make available S-PLUS software to implement the methods. Bayesian model averaging provides a posterior probability that each variable belongs in the model, a more directly interpretable measure of variable importance than a P -value. P -values from models preferred by stepwise methods tend to overstate the evidence for the predictive value of a variable and do not account for model uncertainty. We introduce the partial predictive score to evaluate predictive performance. For the Cardiovascular Health Study, Bayesian model averaging predictively outperforms standard model selection and does a better job of assessing who is at high risk for a stroke.     

Probabilistic sensitivity measures applied to numerical models of flow and transport

First- and second-order reliability algorithms (FORM AND SORM) have been adapted for use in modeling uncertainty and sensitivity related to flow in porous media. They are called reliability algorithms because they were developed originally for analysis of reliability of structures. FORM and SORM utilize a general joint probability model, the Nataf model, as a basis for transforming the original problem formulation into uncorrelated standard normal space, where a first-order or second-order estimate of the probability related to some failure criterion can easily be made. Sensitivity measures that incorporate the probabilistic nature of the uncertain variables in the problem are also evaluated, and are quite useful in indicating which uncertain variables contribute the most to the probabilistic outcome. In this paper the reliability approach is reviewed and the advantages and disadvantages compared to other typical probabilistic techniques used for modeling flow and transport. Some example applications of FORM and SORM from recent research by the authors and others are reviewed. FORM and SORM have been shown to provide an attractive alternative to other probabilistic modeling techniques in some situations.     

Robust interval forecasting algorithm based on a probabilistic cluster model

For substantiation of managerial decisions the forecasting results of dynamic indicators are used. Therefore, forecasting accuracy of these indicators must be acceptable. Consequently, forecasting algorithms are constantly improved to get the acceptable accuracy. This paper considers a variant of the method of forecasting binary outcomes. This method allows prediction of whether or not a future value of the indicator exceeds a predetermined value. This method ‘interval forecasting’ was named. In this paper a robust interval forecasting algorithm based on a probabilistic cluster model is proposed. The algorithm’s accuracy was compared with an algorithm based on logistic regression. The indicators with different statistical properties were chosen. The obtained results have shown the accuracy of both the algorithms is approximately similar in most cases. However, the cases when the algorithm based on logistic regression demonstrated unacceptable accuracy, unlike the presented algorithm have been identified. Thus, this new algorithm is more accurate.     

Generalized accelerated failure time spatial frailty model for arbitrarily censored data

Flexible incorporation of both geographical patterning and risk effects in cancer survival models is becoming increasingly important, due in part to the recent availability of large cancer registries. Most spatial survival models stochastically order survival curves from different subpopulations. However, it is common for survival curves from two subpopulations to cross in epidemiological cancer studies and thus interpretable standard survival models can not be used without some modification. Common fixes are the inclusion of time-varying regression effects in the proportional hazards model or fully nonparametric modeling, either of which destroys any easy interpretability from the fitted model. To address this issue, we develop a generalized accelerated failure time model which allows stratification on continuous or categorical covariates, as well as providing per-variable tests for whether stratification is necessary via novel approximate Bayes factors. The model is interpretable in terms of how median survival changes and is able to capture crossing survival curves in the presence of spatial correlation. A detailed Markov chain Monte Carlo algorithm is presented for posterior inference and a freely available function frailtyGAFT is provided to fit the model in the R package spBayesSurv. We apply our approach to a subset of the prostate cancer data gathered for Louisiana by the surveillance, epidemiology, and end results program of the National Cancer Institute.     

Assessing repeatability and reproducibility using hierarchical modeling: a case-study of distortion product otoacoustic emissions

We present how the repeatability and reproducibility of a measurement device can be estimated from a suitably defined hierarchical linear model. The methodology is illustrated using a collection of eight data sets which consist of the distortion product otoacoustic emission recordings collected from both ears of ten young Sprague-Dawley rats at different frequencies under eight different recording conditions. We formulate a model which extends the commonly used one-way random effects model (5) to account for an experimental setup that is more elaborated than the ones traditionally used in interlaboratory experiments. The fitted model is easily interpretable and furnishes as a by-product the frequencies at which the highest response level is achieved under the eight recording conditions. These values together with the repeatability and reproducibility limits of the protocols are crucial in contributing to the enhancement of the research capabilities on the possible causes of hearing impairment.     

Flexible mixture modelling using the multivariate skew-t-normal distribution

This paper presents a robust probabilistic mixture model based on the multivariate skew-t-normal distribution, a skew extension of the multivariate Student’s t distribution with more powerful abilities in modelling data whose distribution seriously deviates from normality. The proposed model includes mixtures of normal, t and skew-normal distributions as special cases and provides a flexible alternative to recently proposed skew t mixtures. We develop two analytically tractable EM-type algorithms for computing maximum likelihood estimates of model parameters in which the skewness parameters and degrees of freedom are asymptotically uncorrelated. Standard errors for the parameter estimates can be obtained via a general information-based method. We also present a procedure of merging mixture components to automatically identify the number of clusters by fitting piecewise linear regression to the rescaled entropy plot. The effectiveness and performance of the proposed methodology are illustrated by two real-life examples.     

Adaptive evolutionary Monte Carlo algorithm for optimization with applications to sensor placement problems

In this paper, we present an adaptive evolutionary Monte Carlo algorithm (AEMC), which combines a tree-based predictive model with an evolutionary Monte Carlo sampling procedure for the purpose of global optimization. Our development is motivated by sensor placement applications in engineering, which requires optimizing certain complicated “black-box” objective function. The proposed method is able to enhance the optimization efficiency and effectiveness as compared to a few alternative strategies. AEMC falls into the category of adaptive Markov chain Monte Carlo (MCMC) algorithms and is the first adaptive MCMC algorithm that simulates multiple Markov chains in parallel. A theorem about the ergodicity property of the AEMC algorithm is stated and proven. We demonstrate the advantages of the proposed method by applying it to a sensor placement problem in a manufacturing process, as well as to a standard Griewank test function.     

Learning classification trees

Algorithms for learning classification trees have had successes in artificial intelligence and statistics over many years. This paper outlines how a tree learning algorithm can be derived using Bayesian statistics. This introduces Bayesian techniques for splitting, smoothing, and tree averaging. The splitting rule is similar to Quinlan's information gain, while smoothing and averaging replace pruning. Comparative experiments with reimplementations of a minimum encoding approach,c4 (Quinlanet al., 1987) andcart (Breimanet al., 1984), show that the full Bayesian algorithm can produce more accurate predictions than versions of these other approaches, though pays a computational price.     

Inference on Polychotomous Responses in Finite Populations

Abstract. A model‐based predictive estimator is proposed for the population proportions of a polychotomous response variable, based on a sample from the population and on auxiliary variables, whose values are known for the entire population. The responses for the non‐sample units are predicted using a multinomial logit model, which is a parametric function of the auxiliary variables. A bootstrap estimator is proposed for the variance of the predictive estimator, its consistency is proved and its small sample performance is compared with that of an analytical estimator. The proposed predictive estimator is compared with other available estimators, including model‐assisted ones, both in a simulation study involving different sampling designs and model mis‐specification, and using real data from an opinion survey. The results indicate that the prediction approach appears to use auxiliary information more efficiently than the model‐assisted approach.     

A sequential split-and-conquer approach for the analysis of big dependent data in computer experiments

Massive correlated data with many inputs are often generated from computer experiments to study complex systems. The Gaussian process (GP) model is a widely used tool for the analysis of computer experiments. Although GPs provide a simple and effective approximation to computer experiments, two critical issues remain unresolved. One is the computational issue in GP estimation and prediction where intensive manipulations of a large correlation matrix are required. For a large sample size and with a large number of variables, this task is often unstable or infeasible. The other issue is how to improve the naive plug-in predictive distribution which is known to underestimate the uncertainty. In this article, we introduce a unified framework that can tackle both issues simultaneously. It consists of a sequential split-and-conquer procedure, an information combining technique using confidence distributions (CD), and a frequentist predictive distribution based on the combined CD. It is shown that the proposed method maintains the same asymptotic efficiency as the conventional likelihood inference under mild conditions, but dramatically reduces the computation in both estimation and prediction. The predictive distribution contains comprehensive information for inference and provides a better quantification of predictive uncertainty as compared with the plug-in approach. Simulations are conducted to compare the estimation and prediction accuracy with some existing methods, and the computational advantage of the proposed method is also illustrated. The proposed method is demonstrated by a real data example based on tens of thousands of computer experiments generated from a computational fluid dynamic simulator.     

Sequential Monte Carlo for counting vertex covers in general graphs

In this paper we describe a sequential importance sampling (SIS) procedure for counting the number of vertex covers in general graphs. The optimal SIS proposal distribution is the uniform over a suitably restricted set, but is not implementable. We will consider two proposal distributions as approximations to the optimal. Both proposals are based on randomization techniques. The first randomization is the classic probability model of random graphs, and in fact, the resulting SIS algorithm shows polynomial complexity for random graphs. The second randomization introduces a probabilistic relaxation technique that uses Dynamic Programming. The numerical experiments show that the resulting SIS algorithm enjoys excellent practical performance in comparison with existing methods. In particular the method is compared with cachet—an exact model counter, and the state of the art SampleSearch, which is based on Belief Networks and importance sampling.     

The predictive Lasso

We propose a shrinkage procedure for simultaneous variable selection and estimation in generalized linear models (GLMs) with an explicit predictive motivation. The procedure estimates the coefficients by minimizing the Kullback-Leibler divergence of a set of predictive distributions to the corresponding predictive distributions for the full model, subject to an l ₁ constraint on the coefficient vector. This results in selection of a parsimonious model with similar predictive performance to the full model. Thanks to its similar form to the original Lasso problem for GLMs, our procedure can benefit from available l ₁-regularization path algorithms. Simulation studies and real data examples confirm the efficiency of our method in terms of predictive performance on future observations.     

Propensity score prediction for electronic healthcare databases using super learner and high-dimensional propensity score methods

ABSTRACT

The optimal learner for prediction modeling varies depending on the underlying data-generating distribution. Super Learner (SL) is a generic ensemble learning algorithm that uses cross-validation to select among a ‘library’ of candidate prediction models. While SL has been widely studied in a number of settings, it has not been thoroughly evaluated in large electronic healthcare databases that are common in pharmacoepidemiology and comparative effectiveness research. In this study, we applied and evaluated the performance of SL in its ability to predict the propensity score (PS), the conditional probability of treatment assignment given baseline covariates, using three electronic healthcare databases. We considered a library of algorithms that consisted of both nonparametric and parametric models. We also proposed a novel strategy for prediction modeling that combines SL with the high-dimensional propensity score (hdPS) variable selection algorithm. Predictive performance was assessed using three metrics: the negative log-likelihood, area under the curve (AUC), and time complexity. Results showed that the best individual algorithm, in terms of predictive performance, varied across datasets. The SL was able to adapt to the given dataset and optimize predictive performance relative to any individual learner. Combining the SL with the hdPS was the most consistent prediction method and may be promising for PS estimation and prediction modeling in electronic healthcare databases.     

Probabilistic text understanding

We discuss a new framework for text understanding. Three major design decisions characterize this approach. First, we take the problem of text understanding to be a particular case of the general problem of abductive inference. Second, we use probability theory to handle the uncertainty which arises in this abductive inference process. Finally, all aspects of natural language processing are treated in the same framework, allowing us to integrate syntactic, semantic and pragmatic constraints. In order to apply probability theory to this problem, we have developed a probabilistic model of text understanding. To make it practical to use this model, we have devised a way of incrementally constructing and evaluating belief networks. We have written a program,wimp3, to experiment with this framework. To evaluate this program, we have developed a simple single-blind testing method.     

Bayesian networks of customer satisfaction survey data

A Bayesian network (BN) is a probabilistic graphical model that represents a set of variables and their probabilistic dependencies. Formally, BNs are directed acyclic graphs whose nodes represent variables, and whose arcs encode the conditional dependencies among the variables. Nodes can represent any kind of variable, be it a measured parameter, a latent variable, or a hypothesis. They are not restricted to represent random variables, which form the “Bayesian” aspect of a BN. Efficient algorithms exist that perform inference and learning in BNs. BNs that model sequences of variables are called dynamic BNs. In this context, [A. Harel, R. Kenett, and F. Ruggeri, Modeling web usability diagnostics on the basis of usage statistics, in Statistical Methods in eCommerce Research, W. Jank and G. Shmueli, eds., Wiley, 2008] provide a comparison between Markov Chains and BNs in the analysis of web usability from e-commerce data. A comparison of regression models, structural equation models, and BNs is presented in Anderson et al. [R.D. Anderson, R.D. Mackoy, V.B. Thompson, and G. Harrell, A bayesian network estimation of the service–profit Chain for transport service satisfaction, Decision Sciences 35(4), (2004), pp. 665–689]. In this article we apply BNs to the analysis of customer satisfaction surveys and demonstrate the potential of the approach. In particular, BNs offer advantages in implementing models of cause and effect over other statistical techniques designed primarily for testing hypotheses. Other advantages include the ability to conduct probabilistic inference for prediction and diagnostic purposes with an output that can be intuitively understood by managers.     

A Statistical Analysis of Probabilistic Counting Algorithms

Abstract. This article considers the problem of cardinality estimation in data stream applications. We present a statistical analysis of probabilistic counting algorithms, focusing on two techniques that use pseudo‐random variates to form low‐dimensional data sketches. We apply conventional statistical methods to compare probabilistic algorithms based on storing either selected order statistics, or random projections. We derive estimators of the cardinality in both cases, and show that the maximal‐term estimator is recursively computable and has exponentially decreasing error bounds. Furthermore, we show that the estimators have comparable asymptotic efficiency, and explain this result by demonstrating an unexpected connection between the two approaches.