Seminar Series: Dr. Osvaldo Espin-Garcia

Two-phase study design with generalized linear models: from power/sample size calculations to metaheuristic subsample selection under differential unit costs

Osvaldo Espin-Garcia

Assistant Professor
Department of Epidemiology and Biostatistics
Schulich School of Medicine & Dentistry
Western University

Assistant Professor
Department of Statistical Sciences
Dalla Lana School of Public Health
University of Toronto

Principal Biostatistician
Department of Biostatistics
University Health Network

Short Biography:
Dr. Espin-Garcia holds a PhD in Biostatistics from the University of Toronto, an MMath in Statistics-Biostatistics from the University of Waterloo and a BSc in Actuarial Sciences from the National Autonomous University of Mexico (UNAM). Prior to joining Western, he worked as Principal Biostatistician at the University Health Network.

His research program focuses on developing and implementing statistical, machine learning and computational methods for statistical omics, molecular epidemiology, and deep phenotyping of complex traits such as cancer, Crohn’s disease and osteoarthritis.

Dr. Espin-Garcia is currently a member of the Bone and Joint Institute and supports multiple training initiatives such as STAGE, the Collaborative Specialization in Machine Learning in Health and Biomedical Sciences, and the Health Data Working Group.

Abstract:
The two-phase study is cost-effective to collect and analyze expensive predictor data by accruing data in two phases. In the phase 1, observed outcomes and/or inexpensive (auxiliary) covariates for all subjects. In the phase 2, data from phase 1 are used to identify a subset of informative subjects for expensive predictor measurement. Analysis is then performed using all available data by leveraging missing data methods. Despite the appeal of the two-phase study and large body of statistical literature supporting its analysis, implementation of this strategy remains limited. This deficiency may be in part due to limited tools enabling appropriate study design and planning.

Although recent advances in optimal and quasi optimal designs have ameliorated these challenges, there has been limited investigations of another critical aspect of study planning: power and sample size calculation. In the first part of the talk, I will present ongoing work on this topic. Here, the proposed framework extends classical statistical methodologies within generalized linear models (GLMs) to semiparametric maximum likelihood estimation widely used in two-phase studies. We derive noncentrality parameters for the Wald and likelihood-ratio tests and show that, under fixed alternatives, the score statistic follows a weighted sum of noncentral chi-square variables. In simulations with four GLM families and outcome-dependent designs, predicted power closely matched empirical power and notable improvements arose in rare outcome settings when compared to alternative approaches.

In the second part of the talk, I will present recent work that investigates differential unit-cost two-phase designs, which challenges the status quo assumption of equal cost of collecting expensive data across samples. Here, the proposed approach penalizes phase 2 subsamples based on budget constraint satisfaction. We compare three metaheuristic algorithms under heterogeneous per-sample costs: Genetic Algorithm (GA), Simulated Annealing (SA), and Comprehensive Evolutionary Algorithm (CEA). The resulting designs aim to minimize the variance of the parameter of interest under the defined constraints. Simulations based on Poisson regression and inference via semiparametric maximum likelihood evaluate statistical performance across a range of design parameters. Briefly, GA and CEA consistently achieve best performance while SA exhibits greater variability and occasional convergence failures.

Date: Friday, November 13
Time: 1:30 pm - 2:30 pm
Location: PHFM 3015 (Western Centre for Public Health and Family Medicine) or Zoom (request link by email   epibio@uwo.ca )