Prognostic modelling of breast cancer patients: a benchmark of predictive models with external validation

Dissertação apresentada para obtenção do Grau de Doutor em Engenharia Electrotécnica e de Computadores – Sistemas Digitais e Percepcionais pela Universidade Nova de Lisboa, Faculdade de Ciências e Tecnologia

Modeling U.S. Inflation Dynamics: A Bayesian Nonparametric Approach

This paper uses an infinite hidden Markov model (IIHMM) to analyze U.S. inflation dynamics with a particular focus on the persistence of inflation. The IHMM is a Bayesian nonparametric approach to modeling structural breaks. It allows for an unknown number of breakpoints and is a flexible and attractive alternative to existing methods. We found a clear structural break during the recent financial crisis. Prior to that, inflation persistence was high and fairly constant.

Multilevel Modelling with Spatial Effects

In multilevel modelling, interest in modeling the nested structure of hierarchical data has been accompanied by increasing attention to different forms of spatial interactions across different levels of the hierarchy. Neglecting such interactions is likely to create problems of inference, which typically assumes independence. In this paper we review approaches to multilevel modelling with spatial effects, and attempt to connect the two literatures, discussing the advantages and limitations of various approaches.

Modelling Breaks and Clusters in the Steady States of Macroeconomic Variables

Macroeconomists working with multivariate models typically face uncertainty over which (if any) of their variables have long run steady states which are subject to breaks. Furthermore, the nature of the break process is often unknown. In this paper, we draw on methods from the Bayesian clustering literature to develop an econometric methodology which: i) finds groups of variables which have the same number of breaks; and ii) determines the nature of the break process within each group. We present an application involving a five-variate steady-state VAR.

Co-movements in Real Effective Exchange Rates: Evidence from the Dynamic Hierarchical Factor Model

We analyze and quantify co-movements in real effective exchange rates while considering the regional location of countries. More specifically, using the dynamic hierarchical factor model (Moench et al. (2011)), we decompose exchange rate movements into several latent components; worldwide and two regional factors as well as country-specific elements. Then, we provide evidence that the worldwide common factor is closely related to monetary policies in large advanced countries while regional common factors tend to be captured by those in the rest of the countries in a region. However, a substantial proportion of the variation in the real exchange rates is reported to be country-specific; even in Europe country-specific movements exceed worldwide and regional common factors.

Hierarchical Shrinkage in Time-Varying Parameter Models

In this paper, we forecast EU-area inflation with many predictors using time-varying parameter models. The facts that time-varying parameter models are parameter-rich and the time span of our data is relatively short motivate a desire for shrinkage. In constant coefficient regression models, the Bayesian Lasso is gaining increasing popularity as an effective tool for achieving such shrinkage. In this paper, we develop econometric methods for using the Bayesian Lasso with time-varying parameter models. Our approach allows for the coefficient on each predictor to be: i) time varying, ii) constant over time or iii) shrunk to zero. The econometric methodology decides automatically which category each coefficient belongs in. Our empirical results indicate the benefits of such an approach.

Modelling dependence in a ratemaking procedure with multivariate Poisson regression models

When actuaries face with the problem of pricing an insurance contract that contains different types of coverage, such as a motor insurance or homeowner's insurance policy, they usually assume that types of claim are independent. However, this assumption may not be realistic: several studies have shown that there is a positive correlation between types of claim. Here we introduce different regression models in order to relax the independence assumption, including zero-inflated models to account for excess of zeros and overdispersion. These models have been largely ignored to multivariate Poisson date, mainly because of their computational di±culties. Bayesian inference based on MCMC helps to solve this problem (and also lets us derive, for several quantities of interest, posterior summaries to account for uncertainty). Finally, these models are applied to an automobile insurance claims database with three different types of claims. We analyse the consequences for pure and loaded premiums when the independence assumption is relaxed by using different multivariate Poisson regression models and their zero-inflated versions.

Decision-theoretic analysis of forensic sampling criteria using Bayesian decision networks

Sampling issues represent a topic of ongoing interest to the forensic science community essentially because of their crucial role in laboratory planning and working protocols. For this purpose, forensic literature described thorough (Bayesian) probabilistic sampling approaches. These are now widely implemented in practice. They allow, for instance, to obtain probability statements that parameters of interest (e.g., the proportion of a seizure of items that present particular features, such as an illegal substance) satisfy particular criteria (e.g., a threshold or an otherwise limiting value). Currently, there are many approaches that allow one to derive probability statements relating to a population proportion, but questions on how a forensic decision maker - typically a client of a forensic examination or a scientist acting on behalf of a client - ought actually to decide about a proportion or a sample size, remained largely unexplored to date. The research presented here intends to address methodology from decision theory that may help to cope usefully with the wide range of sampling issues typically encountered in forensic science applications. The procedures explored in this paper enable scientists to address a variety of concepts such as the (net) value of sample information, the (expected) value of sample information or the (expected) decision loss. All of these aspects directly relate to questions that are regularly encountered in casework. Besides probability theory and Bayesian inference, the proposed approach requires some additional elements from decision theory that may increase the efforts needed for practical implementation. In view of this challenge, the present paper will emphasise the merits of graphical modelling concepts, such as decision trees and Bayesian decision networks. These can support forensic scientists in applying the methodology in practice. How this may be achieved is illustrated with several examples. The graphical devices invoked here also serve the purpose of supporting the discussion of the similarities, differences and complementary aspects of existing Bayesian probabilistic sampling criteria and the decision-theoretic approach proposed throughout this paper.

Therapeutic drug monitoring of imatinib: Bayesian and alternative methods to predict trough levels

Background: The imatinib trough plasma concentration (C(min)) correlates with clinical response in cancer patients. Therapeutic drug monitoring (TDM) of plasma C(min) is therefore suggested. In practice, however, blood sampling for TDM is often not performed at trough. The corresponding measurement is thus only remotely informative about C(min) exposure. Objectives: The objectives of this study were to improve the interpretation of randomly measured concentrations by using a Bayesian approach for the prediction of C(min), incorporating correlation between pharmacokinetic parameters, and to compare the predictive performance of this method with alternative approaches, by comparing predictions with actual measured trough levels, and with predictions obtained by a reference method, respectively. Methods: A Bayesian maximum a posteriori (MAP) estimation method accounting for correlation (MAP-ρ) between pharmacokinetic parameters was developed on the basis of a population pharmacokinetic model, which was validated on external data. Thirty-one paired random and trough levels, observed in gastrointestinal stromal tumour patients, were then used for the evaluation of the Bayesian MAP-ρ method: individual C(min) predictions, derived from single random observations, were compared with actual measured trough levels for assessment of predictive performance (accuracy and precision). The method was also compared with alternative approaches: classical Bayesian MAP estimation assuming uncorrelated pharmacokinetic parameters, linear extrapolation along the typical elimination constant of imatinib, and non-linear mixed-effects modelling (NONMEM) first-order conditional estimation (FOCE) with interaction. Predictions of all methods were finally compared with 'best-possible' predictions obtained by a reference method (NONMEM FOCE, using both random and trough observations for individual C(min) prediction). Results: The developed Bayesian MAP-ρ method accounting for correlation between pharmacokinetic parameters allowed non-biased prediction of imatinib C(min) with a precision of ±30.7%. This predictive performance was similar for the alternative methods that were applied. The range of relative prediction errors was, however, smallest for the Bayesian MAP-ρ method and largest for the linear extrapolation method. When compared with the reference method, predictive performance was comparable for all methods. The time interval between random and trough sampling did not influence the precision of Bayesian MAP-ρ predictions. Conclusion: Clinical interpretation of randomly measured imatinib plasma concentrations can be assisted by Bayesian TDM. Classical Bayesian MAP estimation can be applied even without consideration of the correlation between pharmacokinetic parameters. Individual C(min) predictions are expected to vary less through Bayesian TDM than linear extrapolation. Bayesian TDM could be developed in the future for other targeted anticancer drugs and for the prediction of other pharmacokinetic parameters that have been correlated with clinical outcomes.

Modelling structural zeros in compositional data

This analysis was stimulated by the real data analysis problem of householdexpenditure data. The full dataset contains expenditure data for a sample of 1224 households. The expenditure is broken down at 2 hierarchical levels: 9 major levels (e.g. housing, food, utilities etc.) and 92 minor levels. There are also 5 factors and 5 covariates at the household level. Not surprisingly, there are a small number of zeros at the major level, but many zeros at the minor level. The question is how best to model the zeros. Clearly, models that tryto add a small amount to the zero terms are not appropriate in general as at least some of the zeros are clearly structural, e.g. alcohol/tobacco for households that are teetotal. The key question then is how to build suitable conditional models. For example, is the sub-composition of spendingexcluding alcohol/tobacco similar for teetotal and non-teetotal households?In other words, we are looking for sub-compositional independence. Also, what determines whether a household is teetotal? Can we assume that it is independent of the composition? In general, whether teetotal will clearly depend on the household level variables, so we need to be able to model this dependence. The other tricky question is that with zeros on more than onecomponent, we need to be able to model dependence and independence of zeros on the different components. Lastly, while some zeros are structural, others may not be, for example, for expenditure on durables, it may be chance as to whether a particular household spends money on durableswithin the sample period. This would clearly be distinguishable if we had longitudinal data, but may still be distinguishable by looking at the distribution, on the assumption that random zeros will usually be for situations where any non-zero expenditure is not small.While this analysis is based on around economic data, the ideas carry over tomany other situations, including geological data, where minerals may be missing for structural reasons (similar to alcohol), or missing because they occur only in random regions which may be missed in a sample (similar to the durables)

Further insights on the French WISC-IV factor structure through Bayesian structural equation modeling

The interpretation of the Wechsler Intelligence Scale for Children-Fourth Edition (WISC-IV) is based on a 4-factor model, which is only partially compatible with the mainstream Cattell-Horn-Carroll (CHC) model of intelligence measurement. The structure of cognitive batteries is frequently analyzed via exploratory factor analysis and/or confirmatory factor analysis. With classical confirmatory factor analysis, almost all crossloadings between latent variables and measures are fixed to zero in order to allow the model to be identified. However, inappropriate zero cross-loadings can contribute to poor model fit, distorted factors, and biased factor correlations; most important, they do not necessarily faithfully reflect theory. To deal with these methodological and theoretical limitations, we used a new statistical approach, Bayesian structural equation modeling (BSEM), among a sample of 249 French-speaking Swiss children (8-12 years). With BSEM, zero-fixed cross-loadings between latent variables and measures are replaced by approximate zeros, based on informative, small-variance priors. Results indicated that a direct hierarchical CHC-based model with 5 factors plus a general intelligence factor better represented the structure of the WISC-IV than did the 4-factor structure and the higher order models. Because a direct hierarchical CHC model was more adequate, it was concluded that the general factor should be considered as a breadth rather than a superordinate factor. Because it was possible for us to estimate the influence of each of the latent variables on the 15 subtest scores, BSEM allowed improvement of the understanding of the structure of intelligence tests and the clinical interpretation of the subtest scores.

Forecasting and turning point predictions in a Bayesian panel VAR model

We provide methods for forecasting variables and predicting turning points in panel Bayesian VARs. We specify a flexible model which accounts for both interdependencies in the cross section and time variations in the parameters. Posterior distributions for the parameters are obtained for a particular type of diffuse, for Minnesota-type and for hierarchical priors. Formulas for multistep, multiunit point and average forecasts are provided. An application to the problem of forecasting the growth rate of output and of predicting turning points in the G-7 illustrates the approach. A comparison with alternative forecasting methods is also provided.

Fixed and random effects in Classical and Bayesian regression

This paper proposes a common and tractable framework for analyzingdifferent definitions of fixed and random effects in a contant-slopevariable-intercept model. It is shown that, regardless of whethereffects (i) are treated as parameters or as an error term, (ii) areestimated in different stages of a hierarchical model, or whether (iii)correlation between effects and regressors is allowed, when the sameinformation on effects is introduced into all estimation methods, theresulting slope estimator is also the same across methods. If differentmethods produce different results, it is ultimately because differentinformation is being used for each methods.

Analysis, modelling and classification of geospatial data using machine learning

The research considers the problem of spatial data classification using machine learning algorithms: probabilistic neural networks (PNN) and support vector machines (SVM). As a benchmark model simple k-nearest neighbor algorithm is considered. PNN is a neural network reformulation of well known nonparametric principles of probability density modeling using kernel density estimator and Bayesian optimal or maximum a posteriori decision rules. PNN is well suited to problems where not only predictions but also quantification of accuracy and integration of prior information are necessary. An important property of PNN is that they can be easily used in decision support systems dealing with problems of automatic classification. Support vector machine is an implementation of the principles of statistical learning theory for the classification tasks. Recently they were successfully applied for different environmental topics: classification of soil types and hydro-geological units, optimization of monitoring networks, susceptibility mapping of natural hazards. In the present paper both simulated and real data case studies (low and high dimensional) are considered. The main attention is paid to the detection and learning of spatial patterns by the algorithms applied.

Wheat alleles introgress into selfing wild relatives: empirical estimates from approximate Bayesian computation in Aegilops triuncialis.

Extensive gene flow between wheat (Triticum sp.) and several wild relatives of the genus Aegilops has recently been detected despite notoriously high levels of selfing in these species. Here, we assess and model the spread of wheat alleles into natural populations of the barbed goatgrass (Aegilops triuncialis), a wild wheat relative prevailing in the Mediterranean flora. Our sampling, based on an extensive survey of 31 Ae. triuncialis populations collected along a 60 km × 20 km area in southern Spain (Grazalema Mountain chain, Andalousia, totalling 458 specimens), is completed with 33 wheat cultivars representative of the European domesticated pool. All specimens were genotyped with amplified fragment length polymorphism with the aim of estimating wheat admixture levels in Ae. triuncialis populations. This survey first confirmed extensive hybridization and backcrossing of wheat into the wild species. We then used explicit modelling of populations and approximate Bayesian computation to estimate the selfing rate of Ae. triuncialis along with the magnitude, the tempo and the geographical distance over which wheat alleles introgress into Ae. triuncialis populations. These simulations confirmed that extensive introgression of wheat alleles (2.7 × 10(-4) wheat immigrants for each Ae. triuncialis resident, at each generation) into Ae. triuncialis occurs despite a high selfing rate (Fis ≈ 1 and selfing rate = 97%). These results are discussed in the light of risks associated with the release of genetically modified wheat cultivars in Mediterranean agrosystems.