Smarter Data Integration: Statistical Methods for Multi-phase and Multi-source Data

Tuberculosis (TB) is notoriously difficult to diagnose among people with HIV, so many are started on anti-TB medications without a positive diagnosis. There is interest in assessing the association between TB diagnosis (positive, negative, indeterminant) and successful TB treatment completion (alive, without recurrence, and off TB medications 18 months after treatment start). To address this question, we have access to a large multinational cohort of 22,588 people with HIV who started TB treatment. However, these data were routinely collected (e.g., based on electronic health records) and are known to be prone to errors. I will describe our experience designing, carrying out, and analyzing data from a multi-wave, two-phase study. In short, approximately 950 records were selected for internal validation (i.e., chart review). Records were selected in three sampling waves to minimize the variance of coefficient estimates; information from prior waves was used to inform sampling in the next wave. Analyses were performed using a sieve maximum likelihood estimator, which is semiparametric efficient and makes minimal assumptions on nuisance models for the errors. 

Multisite longitudinal analyses of electronic health record (EHR) data can reveal real-world disease trajectories, but privacy constraints, high communication costs, outcome-specific missingness, and cross-site heterogeneity often prevent principled pooling and standard longitudinal modeling. In response, we propose PEAL (Privacy-preserving Efficient Aggregation for Longitudinal data), a single-round federated algorithm for multi-level linear mixed-effects models that yields estimates identical to pooled individual-level analysis. We then extend to MV-PEAL, a one-shot federated framework for multivariate mixed-effects models that reconstructs the global likelihood from pattern-specific summaries, enabling efficient estimation of fixed effects, covariance components, and cross-outcome correlations while accommodating outcome missingness in a secure manner. Extensive simulations show gold-standard-equivalent accuracy and improved efficiency over complete-case and standard imputation strategies. By applying the methods to the Johns Hopkins and University of Pittsburge systemic sclerosis cohorts, our methods recover clinically plausible single- and multi-biomarker trajectories, illustrating their utility for distributed research networks studying rare diseases and other time-evolving outcomes. 

Large observational datasets compiled from electronic health records are promising resources for medical research but are often affected by measurement error. Two-phase, multi-wave sampling with generalized raking (GR) offers a robust and efficient solution, though existing work has largely focused on estimation of a single target parameter. Motivated by two recent studies, we discuss extensions of this framework to the multiple parameters setting. We propose practical allocation strategies, including an integer-valued A-optimal method, and evaluate their performance through simulations and an application to a clinical HIV Cohort. Our results show that tailored multi-parameter designs for GR estimators yield marked efficiency gains over traditional case-control or IPW-optimal designs, with patterns that differ meaningfully from the single-parameter setting. These findings provide practical guidance for future two-phase studies using error-prone data. 

The increasing number of observational studies in which electronic health records(EHR) are the primary source of data has led to a greater focus on methods to correct for error-prone data. More recently, three-phase studies have been employed for measurement error correction in which there are error-prone EHR data for the entire cohort (phase-1), audited (i.e., chart-reviewed) EHR data for a sub-cohort (phase-2), and reference standard data not part of routine care that is collected for a sub-sample of patients (phase-3). An ongoing study is focused on factors associated with frailty among people living with HIV (PWH) in Latin America and East Africa. Researchers have previously collected phase-1 and phase-3 data, along with phase-2 data for records in phase-3. Generalized raking (GR) estimators that incorporate phase-1 and phase-2 data by calibrating inverse probability weighted (IPW) estimators with estimated influence functions will be used for our analyses. We can improve the efficiency of GR estimators by collecting additional phase-2 data, which is relatively inexpensive. We derive optimal phase-2 sampling strategies. We show the impact of our sampling designs relative to comparators with extensive simulations, and we illustrate our design applied to the study of frailty in PWH.  

In survival analysis, longitudinal data related to a time-to-event outcome are often incorporated into regression and prediction frameworks. Depending on data quality, these measures may be treated as time-varying covariates or modeled through a longitudinal sub-model within a joint modeling framework. Classic examples include longitudinal CD4 counts in studies of HIV disease progression or repeated PHQ assessments in studies of depression recovery. In this work, we extend these ideas to the semiparametric framework for interval-censored time-to-event data, where only the interval of an event is observed rather than the exact event time. Rather than using longitudinal biomarkers solely as predictors of event time, we investigate their role in improving estimation of nuisance parameters—such as the hazard function—and, in turn, enhancing the accuracy of cumulative incidence curve estimation. Our approach also enables estimation of biomarker distributions in both healthy and diseased populations. We will present results from a comprehensive simulation study and demonstrate the utility of our proposed biomarker-augmented method through an application to a tuberculosis (TB) study among people living with HIV.