arXiv subjects

Tyler H. McCormick

Publications and source records attributed to Tyler H. McCormick.

51 records · Page 3Linked to original sources

arXiv2019

An Expectation Conditional Maximization approach for Gaussian graphical models

Bayesian graphical models are a useful tool for understanding dependence relationships among many variables, particularly in situations with external prior information. In high-dimensional settings, the space of possible graphs becomes enormous, rendering even state-of-the-art Bayesian stochastic search computationally infeasible. We propose a deterministic alternative to estimate Gaussian and Gaussian copula graphical models using an Expectation Conditional Maximization (ECM) algorithm, extending the EM approach from Bayesian variable selection to graphical model estimation. We show that the ECM approach enables fast posterior exploration under a sequence of mixture priors, and can incorporate multiple sources of information.

stat.ML↗

arXiv2019

Beyond prediction: A framework for inference with variational approximations in mixture models

Variational inference is a popular method for estimating model parameters and conditional distributions in hierarchical and mixed models, which arise frequently in many settings in the health, social, and biological sciences. Variational inference in a frequentist context works by approximating intractable conditional distributions with a tractable family and optimizing the resulting lower bound on the log-likelihood. The variational objective function is typically less computationally intensive to optimize than the true likelihood, enabling scientists to fit rich models even with extremely large datasets. Despite widespread use, little is known about the general theoretical properties of estimators arising from variational approximations to the log-likelihood, which hinders their use in inferential statistics. In this paper we connect such estimators to profile M-estimation, which enables us to provide regularity conditions for consistency and asymptotic normality of variational estimators. Our theory also motivates three methodological improvements to variational inference: estimation of the asymptotic model-robust covariance matrix, a one-step correction that improves estimator efficiency, and an empirical assessment of consistency. We evaluate the proposed results using simulation studies and data on marijuana use from the National Longitudinal Study of Youth.

stat.ME↗

arXiv2018

Bayesian factor models for probabilistic cause of death assessment with verbal autopsies

The distribution of deaths by cause provides crucial information for public health planning, response, and evaluation. About 60% of deaths globally are not registered or given a cause, limiting our ability to understand disease epidemiology. Verbal autopsy (VA) surveys are increasingly used in such settings to collect information on the signs, symptoms, and medical history of people who have recently died. This article develops a novel Bayesian method for estimation of population distributions of deaths by cause using verbal autopsy data. The proposed approach is based on a multivariate probit model where associations among items in questionnaires are flexibly induced by latent factors. Using the Population Health Metrics Research Consortium labeled data that include both VA and medically certified causes of death, we assess performance of the proposed method. Further, we estimate important questionnaire items that are highly associated with causes of death. This framework provides insights that will simplify future data collection.

stat.AP↗

arXiv2018

Quantifying the Contributions of Training Data and Algorithm Logic to the Performance of Automated Cause-assignment Algorithms for Verbal Autopsy

A verbal autopsy (VA) consists of a survey with a relative or close contact of a person who has recently died. VA surveys are commonly used to infer likely causes of death for individuals when deaths happen outside of hospitals or healthcare facilities. Several statistical and algorithmic methods are available to assign cause of death using VA surveys. Each of these methods require as inputs some information about the joint distribution of symptoms and causes. In this note, we examine the generalizability of this symptom-cause information by comparing different automated coding methods using various combinations of inputs and evaluation data. VA algorithm performance is affected by both the specific SCI themselves and the logic of a given algorithm. Using a variety of performance metrics for all existing VA algorithms, we demonstrate that in general the adequacy of the information about the joint distribution between symptoms and cause affects performance at least as much or more than algorithm logic.

stat.AP↗

arXiv2018

Using Aggregated Relational Data to feasibly identify network structure without network data

Social network data is often prohibitively expensive to collect, limiting empirical network research. Typical economic network mapping requires (1) enumerating a census, (2) eliciting the names of all network links for each individual, (3) matching the list of social connections to the census, and (4) repeating (1)-(3) across many networks. In settings requiring field surveys, steps (2)-(3) can be very expensive. In other network populations such as financial intermediaries or high-risk groups, proprietary data and privacy concerns may render (2)-(3) impossible. Both restrict the accessibility of high-quality networks research to investigators with considerable resources. We propose an inexpensive and feasible strategy for network elicitation using Aggregated Relational Data (ARD) -- responses to questions of the form "How many of your social connections have trait k?" Our method uses ARD to recover the parameters of a general network formation model, which in turn, permits the estimation of any arbitrary node- or graph-level statistic. The method works well in simulations and in matching a range of network characteristics in real-world graphs from 75 Indian villages. Moreover, we replicate the results of two field experiments that involved collecting network data. We show that the researchers would have drawn similar conclusions using ARD alone. Finally, using calculations from J-PAL fieldwork, we show that in rural India, for example, ARD surveys are 80% cheaper than full network surveys.

stat.ME↗

arXiv2018

Multiresolution network models

Many existing statistical and machine learning tools for social network analysis focus on a single level of analysis. Methods designed for clustering optimize a global partition of the graph, whereas projection based approaches (e.g. the latent space model in the statistics literature) represent in rich detail the roles of individuals. Many pertinent questions in sociology and economics, however, span multiple scales of analysis. Further, many questions involve comparisons across disconnected graphs that will, inevitably be of different sizes, either due to missing data or the inherent heterogeneity in real-world networks. We propose a class of network models that represent network structure on multiple scales and facilitate comparison across graphs with different numbers of individuals. These models differentially invest modeling effort within subgraphs of high density, often termed communities, while maintaining a parsimonious structure between said subgraphs. We show that our model class is projective, highlighting an ongoing discussion in the social network modeling literature on the dependence of inference paradigms on the size of the observed graph. We illustrate the utility of our method using data on household relations from Karnataka, India.

stat.ME↗

arXiv2018

Modeling Recovery Curves With Application to Prostatectomy

We propose a Bayesian model that predicts recovery curves based on information available before the disruptive event. A recovery curve of interest is the quantified sexual function of prostate cancer patients after prostatectomy surgery. We illustrate the utility of our model as a pre-treatment medical decision aid, producing personalized predictions that are both interpretable and accurate. We uncover covariate relationships that agree with and supplement that in existing medical literature.

stat.ME↗

arXiv2018

Inferring social structure from continuous-time interaction data

Relational event data, which consist of events involving pairs of actors over time, are now commonly available at the finest of temporal resolutions. Existing continuous-time methods for modeling such data are based on point processes and directly model interaction "contagion," whereby one interaction increases the propensity of future interactions among actors, often as dictated by some latent variable structure. In this article, we present an alternative approach to using temporal-relational point process models for continuous-time event data. We characterize interactions between a pair of actors as either spurious or that resulting from an underlying, persistent connection in a latent social network. We argue that consistent deviations from expected behavior, rather than solely high frequency counts, are crucial for identifying well-established underlying social relationships. This study aims to explore these latent network structures in two contexts: one comprising of college students and another involving barn swallows.

stat.ME↗

arXiv2015

Estimating population size using the network scale up method

We develop methods for estimating the size of hard-to-reach populations from data collected using network-based questions on standard surveys. Such data arise by asking respondents how many people they know in a specific group (e.g., people named Michael, intravenous drug users). The Network Scale up Method (NSUM) is a tool for producing population size estimates using these indirect measures of respondents' networks. Killworth et al. [Soc. Netw. 20 (1998a) 23-50, Evaluation Review 22 (1998b) 289-308] proposed maximum likelihood estimators of population size for a fixed effects model in which respondents' degrees or personal network sizes are treated as fixed. We extend this by treating personal network sizes as random effects, yielding principled statements of uncertainty. This allows us to generalize the model to account for variation in people's propensity to know people in particular subgroups (barrier effects), such as their tendency to know people like themselves, as well as their lack of awareness of or reluctance to acknowledge their contacts' group memberships (transmission bias). NSUM estimates also suffer from recall bias, in which respondents tend to underestimate the number of members of larger groups that they know, and conversely for smaller groups. We propose a data-driven adjustment method to deal with this. Our methods perform well in simulation studies, generating improved estimates and calibrated uncertainty intervals, as well as in back estimates of real sample data. We apply them to data from a study of HIV/AIDS prevalence in Curitiba, Brazil. Our results show that when transmission bias is present, external information about its likely extent can greatly improve the estimates. The methods are implemented in the NSUM R package.

stat.ME↗

arXiv2015

Interpretable classifiers using rules and Bayesian analysis: Building a better stroke prediction model

We aim to produce predictive models that are not only accurate, but are also interpretable to human experts. Our models are decision lists, which consist of a series of if...then... statements (e.g., if high blood pressure, then stroke) that discretize a high-dimensional, multivariate feature space into a series of simple, readily interpretable decision statements. We introduce a generative model called Bayesian Rule Lists that yields a posterior distribution over possible decision lists. It employs a novel prior structure to encourage sparsity. Our experiments show that Bayesian Rule Lists has predictive accuracy on par with the current top algorithms for prediction in machine learning. Our method is motivated by recent developments in personalized medicine, and can be used to produce highly accurate and interpretable medical scoring systems. We demonstrate this by producing an alternative to the CHADS$_2$ score, actively used in clinical practice for estimating the risk of stroke in patients that have atrial fibrillation. Our model is as interpretable as CHADS$_2$, but more accurate.

stat.AP↗

arXiv2015

Probabilistic Cause-of-death Assignment using Verbal Autopsies

In regions without complete-coverage civil registration and vital statistics systems there is uncertainty about even the most basic demographic indicators. In such areas the majority of deaths occur outside hospitals and are not recorded. Worldwide, fewer than one-third of deaths are assigned a cause, with the least information available from the most impoverished nations. In populations like this, verbal autopsy (VA) is a commonly used tool to assess cause of death and estimate cause-specific mortality rates and the distribution of deaths by cause. VA uses an interview with caregivers of the decedent to elicit data describing the signs and symptoms leading up to the death. This paper develops a new statistical tool known as InSilicoVA to classify cause of death using information acquired through VA. InSilicoVA shares uncertainty between cause of death assignments for specific individuals and the distribution of deaths by cause across the population. Using side-by-side comparisons with both observed and simulated data, we demonstrate that InSilicoVA has distinct advantages compared to currently available methods.

stat.AP↗

arXiv2015

Reactive point processes: A new approach to predicting power failures in underground electrical systems

Reactive point processes (RPPs) are a new statistical model designed for predicting discrete events in time based on past history. RPPs were developed to handle an important problem within the domain of electrical grid reliability: short-term prediction of electrical grid failures ("manhole events"), including outages, fires, explosions and smoking manholes, which can cause threats to public safety and reliability of electrical service in cities. RPPs incorporate self-exciting, self-regulating and saturating components. The self-excitement occurs as a result of a past event, which causes a temporary rise in vulner ability to future events. The self-regulation occurs as a result of an external inspection which temporarily lowers vulnerability to future events. RPPs can saturate when too many events or inspections occur close together, which ensures that the probability of an event stays within a realistic range. Two of the operational challenges for power companies are (i) making continuous-time failure predictions, and (ii) cost/benefit analysis for decision making and proactive maintenance. RPPs are naturally suited for handling both of these challenges. We use the model to predict power-grid failures in Manhattan over a short-term horizon, and to provide a cost/benefit analysis of different proactive maintenance programs.

stat.AP↗

arXiv2014

Clustering South African households based on their asset status using latent variable models

The Agincourt Health and Demographic Surveillance System has since 2001 conducted a biannual household asset survey in order to quantify household socio-economic status (SES) in a rural population living in northeast South Africa. The survey contains binary, ordinal and nominal items. In the absence of income or expenditure data, the SES landscape in the study population is explored and described by clustering the households into homogeneous groups based on their asset status. A model-based approach to clustering the Agincourt households, based on latent variable models, is proposed. In the case of modeling binary or ordinal items, item response theory models are employed. For nominal survey items, a factor analysis model, similar in nature to a multinomial probit model, is used. Both model types have an underlying latent variable structure - this similarity is exploited and the models are combined to produce a hybrid model capable of handling mixed data types. Further, a mixture of the hybrid models is considered to provide clustering capabilities within the context of mixed binary, ordinal and nominal response data. The proposed model is termed a mixture of factor analyzers for mixed data (MFA-MD). The MFA-MD model is applied to the survey data to cluster the Agincourt households into homogeneous groups. The model is estimated within the Bayesian paradigm, using a Markov chain Monte Carlo algorithm. Intuitive groupings result, providing insight to the different socio-economic strata within the Agincourt region.

stat.AP↗

arXiv2013

Latent demographic profile estimation in hard-to-reach groups

The sampling frame in most social science surveys excludes members of certain groups, known as hard-to-reach groups. These groups, or subpopulations, may be difficult to access (the homeless, e.g.), camouflaged by stigma (individuals with HIV/AIDS), or both (commercial sex workers). Even basic demographic information about these groups is typically unknown, especially in many developing nations. We present statistical models which leverage social network structure to estimate demographic characteristics of these subpopulations using Aggregated relational data (ARD), or questions of the form "How many X's do you know?" Unlike other network-based techniques for reaching these groups, ARD require no special sampling strategy and are easily incorporated into standard surveys. ARD also do not require respondents to reveal their own group membership. We propose a Bayesian hierarchical model for estimating the demographic characteristics of hard-to-reach groups, or latent demographic profiles, using ARD. We propose two estimation techniques. First, we propose a Markov-chain Monte Carlo algorithm for existing data or cases where the full posterior distribution is of interest. For cases when new data can be collected, we propose guidelines and, based on these guidelines, propose a simple estimate motivated by a missing data approach. Using data from McCarty et al. [Human Organization 60 (2001) 28-39], we estimate the age and gender profiles of six hard-to-reach groups, such as individuals who have HIV, women who were raped, and homeless persons. We also evaluate our simple estimates using simulation studies.

stat.AP↗

arXiv2012

Bayesian hierarchical rule modeling for predicting medical conditions

We propose a statistical modeling technique, called the Hierarchical Association Rule Model (HARM), that predicts a patient's possible future medical conditions given the patient's current and past history of reported conditions. The core of our technique is a Bayesian hierarchical model for selecting predictive association rules (such as "condition 1 and condition 2 $\rightarrow$ condition 3") from a large set of candidate rules. Because this method "borrows strength" using the conditions of many similar patients, it is able to provide predictions specialized to any given patient, even when little information about the patient's history of conditions is available.

stat.AP↗