arXiv subjects

Sylvia Richardson

Publications and source records attributed to Sylvia Richardson.

At least 37 records · Page 2Linked to original sources

arXiv2020

Kernel learning approaches for summarising and combining posterior similarity matrices

When using Markov chain Monte Carlo (MCMC) algorithms to perform inference for Bayesian clustering models, such as mixture models, the output is typically a sample of clusterings (partitions) drawn from the posterior distribution. In practice, a key challenge is how to summarise this output. Here we build upon the notion of the posterior similarity matrix (PSM) in order to suggest new approaches for summarising the output of MCMC algorithms for Bayesian clustering models. A key contribution of our work is the observation that PSMs are positive semi-definite, and hence can be used to define probabilistically-motivated kernel matrices that capture the clustering structure present in the data. This observation enables us to employ a range of kernel methods to obtain summary clusterings, and otherwise exploit the information summarised by PSMs. For example, if we have multiple PSMs, each corresponding to a different dataset on a common set of statistical units, we may use standard methods for combining kernels in order to perform integrative clustering. We may moreover embed PSMs within predictive kernel models in order to perform outcome-guided data integration. We demonstrate the performances of the proposed methods through a range of simulation studies as well as two real data applications. R code is available at https://github.com/acabassi/combine-psms.

stat.ME↗

arXiv2020

A global-local approach for detecting hotspots in multiple-response regression

We tackle modelling and inference for variable selection in regression problems with many predictors and many responses. We focus on detecting hotspots, i.e., predictors associated with several responses. Such a task is critical in statistical genetics, as hotspot genetic variants shape the architecture of the genome by controlling the expression of many genes and may initiate decisive functional mechanisms underlying disease endpoints. Existing hierarchical regression approaches designed to model hotspots suffer from two limitations: their discrimination of hotspots is sensitive to the choice of top-level scale parameters for the propensity of predictors to be hotspots, and they do not scale to large predictor and response vectors, e.g., of dimensions $10^3-10^5$ in genetic applications. We address these shortcomings by introducing a flexible hierarchical regression framework that is tailored to the detection of hotspots and scalable to the above dimensions. Our proposal implements a fully Bayesian model for hotspots based on the horseshoe shrinkage prior. Its global-local formulation shrinks noise globally and hence accommodates the highly sparse nature of genetic analyses, while being robust to individual signals, thus leaving the effects of hotspots unshrunk. Inference is carried out using a fast variational algorithm coupled with a novel simulated annealing procedure that allows efficient exploration of multimodal distributions.

stat.AP↗

arXiv2020

High-dimensional regression in practice: an empirical study of finite-sample prediction, variable selection and ranking

Penalized likelihood approaches are widely used for high-dimensional regression. Although many methods have been proposed and the associated theory is now well-developed, the relative efficacy of different approaches in finite-sample settings, as encountered in practice, remains incompletely understood. There is therefore a need for empirical investigations in this area that can offer practical insight and guidance to users. In this paper we present a large-scale comparison of penalized regression methods. We distinguish between three related goals: prediction, variable selection and variable ranking. Our results span more than 2,300 data-generating scenarios, including both synthetic and semi-synthetic data (real covariates and simulated responses), allowing us to systematically consider the influence of various factors (sample size, dimensionality, sparsity, signal strength and multicollinearity). We consider several widely-used approaches (Lasso, Adaptive Lasso, Elastic Net, Ridge Regression, SCAD, the Dantzig Selector and Stability Selection). We find considerable variation in performance between methods. Our results support a `no panacea' view, with no unambiguous winner across all scenarios or goals, even in this restricted setting where all data align well with the assumptions underlying the methods. The study allows us to make some recommendations as to which approaches may be most (or least) suitable given the goal and some data characteristics. Our empirical results complement existing theory and provide a resource to compare methods across a range of scenarios and metrics.

stat.ME↗

arXiv2019

Statistical properties of sketching algorithms

Sketching is a probabilistic data compression technique that has been largely developed in the computer science community. Numerical operations on big datasets can be intolerably slow; sketching algorithms address this issue by generating a smaller surrogate dataset. Typically, inference proceeds on the compressed dataset. Sketching algorithms generally use random projections to compress the original dataset and this stochastic generation process makes them amenable to statistical analysis. We argue that the sketched data can be modelled as a random sample, thus placing this family of data compression methods firmly within an inferential framework. In particular, we focus on the Gaussian, Hadamard and Clarkson-Woodruff sketches, and their use in single pass sketching algorithms for linear regression with huge $n$. We explore the statistical properties of sketched regression algorithms and derive new distributional results for a large class of sketched estimators. A key result is a conditional central limit theorem for data oblivious sketches. An important finding is that the best choice of sketching algorithm in terms of mean square error is related to the signal to noise ratio in the source dataset. Finally, we demonstrate the theory and the limits of its applicability on two real datasets.

stat.ME↗

arXiv2018

Shrinkage estimation of large covariance matrices using multiple shrinkage targets

Linear shrinkage estimators of a covariance matrix --- defined by a weighted average of the sample covariance matrix and a pre-specified shrinkage target matrix --- are popular when analysing high-throughput molecular data. However, their performance strongly relies on an appropriate choice of target matrix. This paper introduces a more flexible class of linear shrinkage estimators that can accommodate multiple shrinkage target matrices, directly accounting for the uncertainty regarding the target choice. This is done within a conjugate Bayesian framework, which is computationally efficient. Using both simulated and real data, we show that the proposed estimator is less sensitive to target misspecification and can outperform state-of-the-art (nonparametric) single-target linear shrinkage estimators. Using protein expression data from The Cancer Proteome Atlas we illustrate how multiple sources of prior information (obtained from more than 30 different cancer types) can be incorporated into the proposed multi-target linear shrinkage estimator. In particular, it is shown that the target-specific weights can provide insights into the differences and similarities between cancer types. Software for the method is freely available as an R-package at http://github.com/HGray384/TAS.

stat.ME↗

arXiv2018

Fast Bayesian inference in large Gaussian graphical models

Despite major methodological developments, Bayesian inference for Gaussian graphical models remains challenging in high dimension due to the tremendous size of the model space. This article proposes a method to infer the marginal and conditional independence structures between variables by multiple testing of hypotheses. Specifically, we introduce closed-form Bayes factors under the Gaussian conjugate model to evaluate the null hypotheses of marginal and conditional independence between variables. Their computation for all pairs of variables is shown to be extremely efficient, thereby allowing us to address large problems with thousands of nodes. Moreover, we derive exact tail probabilities from the null distributions of the Bayes factors. These allow the use of any multiplicity correction procedure to control error rates for incorrect edge inclusion. We demonstrate the proposed approach to graphical model selection on various simulated examples as well as on a large gene expression data set from The Cancer Genome Atlas.

stat.ME↗

arXiv2017

Using Ecological Propensity Score to Adjust for Missing Confounders in Small Area Studies

Small area ecological studies are commonly used in epidemiology to assess the impact of area level risk factors on health outcomes when data are only available in an aggregated form. However the resulting estimates are often biased due to unmeasured confounders, which typically are not available from the standard administrative registries used for these studies. Extra information on confounders can be provided through external datasets such as surveys or cohorts, where the data are available at the individual level rather than at the area level; however such data typically lack the geographical coverage of administrative registries. We develop a framework of analysis which combines ecological and individual level data from different sources to provide an adjusted estimate of area level risk factors which is less biased. Our method (i) summarises all available individual level confounders into an area level scalar variable, which we call ecological propensity score (EPS), (ii) implements a hierarchical structured approach to predict the values of EPS whenever they are missing, (iii) includes the estimated and predicted EPS into the ecological regression linking the risk factors to the health outcome. Through a simulation study we show that integrating individual level data into small area analyses via EPS is a promising method to reduce the bias intrinsic in ecological studies due to unmeasured confounders; we also apply the method to a real case study to evaluate the effect of air pollution on coronary heart disease hospital admissions in Greater London.

stat.AP↗

arXiv2016

Exploring dependence between categorical variables: benefits and limitations of using variable selection within Bayesian clustering in relation to log-linear modelling with interaction terms

This manuscript is concerned with relating two approaches that can be used to explore complex dependence structures between categorical variables, namely Bayesian partitioning of the covariate space incorporating a variable selection procedure that highlights the covariates that drive the clustering, and log-linear modelling with interaction terms. We derive theoretical results on this relation and discuss if they can be employed to assist log-linear model determination, demonstrating advantages and limitations with simulated and real data sets. The main advantage concerns sparse contingency tables. Inferences from clustering can potentially reduce the number of covariates considered and, subsequently, the number of competing log-linear models, making the exploration of the model space feasible. Variable selection within clustering can inform on marginal independence in general, thus allowing for a more efficient exploration of the log-linear model space. However, we show that the clustering structure is not informative on the existence of interactions in a consistent manner. This work is of interest to those who utilize log-linear models, as well as practitioners such as epidemiologists that use clustering models to reduce the dimensionality in the data and to reveal interesting patterns on how covariates combine.

stat.ME↗

arXiv2014

Bayesian nonparametric models for spatially indexed data of mixed type

We develop Bayesian nonparametric models for spatially indexed data of mixed type. Our work is motivated by challenges that occur in environmental epidemiology, where the usual presence of several confounding variables that exhibit complex interactions and high correlations makes it difficult to estimate and understand the effects of risk factors on health outcomes of interest. The modeling approach we adopt assumes that responses and confounding variables are manifestations of continuous latent variables, and uses multivariate Gaussians to jointly model these. Responses and confounding variables are not treated equally as relevant parameters of the distributions of the responses only are modeled in terms of explanatory variables or risk factors. Spatial dependence is introduced by allowing the weights of the nonparametric process priors to be location specific, obtained as probit transformations of Gaussian Markov random fields. Confounding variables and spatial configuration have a similar role in the model, in that they only influence, along with the responses, the allocation probabilities of the areas into the mixture components, thereby allowing for flexible adjustment of the effects of observed confounders, while allowing for the possibility of residual spatial structure, possibly occurring due to unmeasured or undiscovered spatially varying factors. Aspects of the model are illustrated in simulation studies and an application to a real data set.

stat.ME↗

arXiv2014

PReMiuM: An R Package for Profile Regression Mixture Models using Dirichlet Processes

PReMiuM is a recently developed R package for Bayesian clustering using a Dirichlet process mixture model. This model is an alternative to regression models, non-parametrically linking a response vector to covariate data through cluster membership. The package allows Bernoulli, Binomial, Poisson, Normal and categorical response, as well as Normal and discrete covariates. Additionally, predictions may be made for the response, and missing values for the covariates are handled. Several samplers and label switching moves are implemented along with diagnostic tools to assess convergence. A number of R functions for post-processing of the output are also provided. In addition to fitting mixtures, it may additionally be of interest to determine which covariates actively drive the mixture components. This is implemented in the package as variable selection.

stat.CO↗

arXiv2014

Bayesian regression discontinuity designs: Incorporating clinical knowledge in the causal analysis of primary care data

The regression discontinuity (RD) design is a quasi-experimental design that estimates the causal effects of a treatment by exploiting naturally occurring treatment rules. It can be applied in any context where a particular treatment or intervention is administered according to a pre-specified rule linked to a continuous variable. Such thresholds are common in primary care drug prescription where the RD design can be used to estimate the causal effect of medication in the general population. Such results can then be contrasted to those obtained from randomised controlled trials (RCTs) and inform prescription policy and guidelines based on a more realistic and less expensive context. In this paper we focus on statins, a class of cholesterol-lowering drugs, however, the methodology can be applied to many other drugs provided these are prescribed in accordance to pre-determined guidelines. NHS guidelines state that statins should be prescribed to patients with 10 year cardiovascular disease risk scores in excess of 20%. If we consider patients whose scores are close to this threshold we find that there is an element of random variation in both the risk score itself and its measurement. We can thus consider the threshold a randomising device assigning the prescription to units just above the threshold and withholds it from those just below. Thus we are effectively replicating the conditions of an RCT in the area around the threshold, removing or at least mitigating confounding. We frame the RD design in the language of conditional independence which clarifies the assumptions necessary to apply it to data, and which makes the links with instrumental variables clear. We also have context specific knowledge about the expected sizes of the effects of statin prescription and are thus able to incorporate this into Bayesian models by formulating informative priors on our causal parameters.

stat.ME↗

arXiv2014

Sampling from Dirichlet process mixture models with unknown concentration parameter: Mixing issues in large data implementations

We consider the question of Markov chain Monte Carlo sampling from a general stick-breaking Dirichlet process mixture model, with concentration parameter alpha. This paper introduces a Gibbs sampling algorithm that combines the slice sampling approach of Walker (2007) and the retrospective sampling approach of Papaspiliopoulos and Roberts (2008). Our general algorithm is implemented as efficient open source C++ software, available as an R package, and is based on a blocking strategy similar to that suggested by Papaspiliopoulos (2008) and implemented by Yau et al (2011). We discuss the difficulties of achieving good mixing in MCMC samplers of this nature and investigate sensitivity to initialisation. We additionally consider the challenges when an additional layer of hierarchy is added such that joint inference is to be made on alpha. We introduce a new label switching move and compute the marginal model posterior to help to surmount these difficulties. Our work is illustrated using a profile regression (Molitor et al, 2010) application, where we demonstrate good mixing behaviour for both synthetic and real examples.

stat.CO↗

arXiv2014

MCMC algorithms for Bayesian variable selection in the logistic regression model for large-scale genomic applications

In large-scale genomic applications vast numbers of molecular features are scanned in order to find a small number of candidates which are linked to a particular disease or phenotype. This is a variable selection problem in the "large p, small n" paradigm where many more variables than samples are available. Additionally, a complex dependence structure is often observed among the markers/genes due to their joint involvement in biological processes and pathways. Bayesian variable selection methods that introduce sparseness through additional priors on the model size are well suited to the problem. However, the model space is very large and standard Markov chain Monte Carlo (MCMC) algorithms such as a Gibbs sampler sweeping over all p variables in each iteration are often computationally infeasible. We propose to employ the dependence structure in the data to decide which variables should always be updated together and which are nearly conditionally independent and hence do not need to be considered together. Here, we focus on binary classification applications. We follow the implementation of the Bayesian probit regression model by Albert and Chib (1993) and the Bayesian logistic regression model by Holmes and Held (2006) which both lead to marginal Gaussian distributions. We in- vestigate several MCMC samplers using the dependence structure in different ways. The mixing and convergence performances of the resulting Markov chains are evaluated and compared to standard samplers in two simulation studies and in an application to a real gene expression data set.

stat.CO↗

arXiv2012

Safe preselection in lasso-type problems by cross-validation freezing

We propose a new approach to safe variable preselection in high-dimensional penalized regression, such as the lasso. Preselection - to start with a manageable set of covariates - has often been implemented without clear appreciation of its potential bias. Based on sequential implementation of the lasso with increasing lists of predictors, we find a new property of the set of corresponding cross-validation curves, a pattern that we call freezing. It allows to determine a subset of covariates with which we reach the same lasso solution as would be obtained using the full set of covariates. Freezing has not been characterized before and is different from recently discussed safe rules for discarding predictors. We demonstrate by simulation that ranking predictors by their univariate correlation with the outcome, leads in a majority of cases to early freezing, giving a safe and efficient way of focusing the lasso analysis on a smaller and manageable number of predictors. We illustrate the applicability of our strategy in the context of a GWAS analysis and on microarray genomic data. Freezing offers great potential for extending the applicability of penalized regressions to ultra highdimensional data sets. Its applicability is not limited to the standard lasso but is a generic property of many penalized approaches.

stat.ME↗

arXiv2012

A Bayesian Model of NMR Spectra for the Deconvolution and Quantification of Metabolites in Complex Biological Mixtures

Nuclear Magnetic Resonance (NMR) spectra are widely used in metabolomics to obtain profiles of metabolites dissolved in biofluids such as cell supernatants. Methods for estimating metabolite concentrations from these spectra are presently confined to manual peak fitting and to binning procedures for integrating resonance peaks. Extensive information on the patterns of spectral resonance generated by human metabolites is now available in online databases. By incorporating this information into a Bayesian model we can deconvolve resonance peaks from a spectrum and obtain explicit concentration estimates for the corresponding metabolites. Spectral resonances that cannot be deconvolved in this way may also be of scientific interest so we model them jointly using wavelets. We describe a Markov chain Monte Carlo algorithm which allows us to sample from the joint posterior distribution of the model parameters, using specifically designed block updates to improve mixing. The strong prior on resonance patterns allows the algorithm to identify peaks corresponding to particular metabolites automatically, eliminating the need for manual peak assignment. We assess our method for peak alignment and concentration estimation. Except in cases when the target resonance signal is very weak, alignment is unbiased and precise. We compare the Bayesian concentration estimates to those obtained from a conventional numerical integration method and find that our point estimates have sixfold lower mean squared error. Finally, we apply our method to a spectral dataset taken from an investigation of the metabolic response of yeast to recombinant protein expression. We estimate the concentrations of 26 metabolites and compare to manual quantification by five expert spectroscopists. We discuss the reason for discrepancies and the robustness of our methods concentration estimates.

stat.ME↗

arXiv2011

Monte Carlo algorithms for model assessment via conflicting summaries

The development of statistical methods and numerical algorithms for model choice is vital to many real-world applications. In practice, the ABC approach can be instrumental for sequential model design; however, the theoretical basis of its use has been questioned. We present a measure-theoretic framework for using the ABC error towards model choice and describe how easily existing rejection, Metropolis-Hastings and sequential importance sampling ABC algorithms are extended for the purpose of model checking. Considering a panel of applications from evolutionary biology to dynamic systems, we discuss the choice of summaries which differs from standard ABC approaches. The methods and algorithms presented here may provide the workhorse machinery for an exploratory approach to ABC model choice, particularly as the application of standard Bayesian tools can prove impossible.

stat.ME↗

arXiv2011

Classification Loss Function for Parameter Ensembles in Bayesian Hierarchical Models

Parameter ensembles or sets of point estimates constitute one of the cornerstones of modern statistical practice. This is especially the case in Bayesian hierarchical models, where different decision-theoretic frameworks can be deployed to summarize such parameter ensembles. The estimation of these parameter ensembles may thus substantially vary depending on which inferential goals are prioritised by the modeller. In this note, we consider the problem of classifying the elements of a parameter ensemble above or below a given threshold. Two threshold classification losses (TCLs) --weighted and unweighted-- are formulated. The weighted TCL can be used to emphasize the estimation of false positives over false negatives or the converse. We prove that the weighted and unweighted TCLs are optimized by the ensembles of unit-specific posterior quantiles and posterior medians, respectively. In addition, we relate these classification loss functions on parameter ensembles to the concepts of posterior sensitivity and specificity. Finally, we find some relationships between the unweighted TCL and the absolute value loss, which explain why both functions are minimized by posterior medians.

stat.ME↗

arXiv2010

Evolutionary Stochastic Search for Bayesian model exploration

Implementing Bayesian variable selection for linear Gaussian regression models for analysing high dimensional data sets is of current interest in many fields. In order to make such analysis operational, we propose a new sampling algorithm based upon Evolutionary Monte Carlo and designed to work under the "large p, small n" paradigm, thus making fully Bayesian multivariate analysis feasible, for example, in genetics/genomics experiments. Two real data examples in genomics are presented, demonstrating the performance of the algorithm in a space of up to 10,000 covariates. Finally the methodology is compared with a recently proposed search algorithms in an extensive simulation study.

stat.CO↗