About

My research interests focus on the development of precision medicine, causal inference, and statistical learning methodology for the analysis of complex observational studies. I am particularly interested in addressing various forms of population heterogeneity with the aim of improving patient health outcomes. My work in this area has involved applications in health system risk modeling and in personalizing health system intervention enrollment decisions. My research also includes methodological and computational developments with the aim of flexibly modeling highly complex and/or large-scale data.

Research topics

• Medicine
• Virology
• Internal medicine
• Pathology
• Pediatrics

Selected publications

• Heterogeneous readmission prediction with hierarchical effect decomposition and regularization

  arXiv (Cornell University) · 2026-03-20

  preprintOpen accessSenior author

  Accurately predicting hospital readmission risks using electronic health records (EHRs) is critical for effective patient management and healthcare resource allocation. Patient populations in health systems are highly heterogeneous across different primary diagnoses, necessitating tailored yet interpretable prediction models. We propose a hierarchical modeling framework incorporating hierarchical nested re-parameterization and structured regularization methods, which we call hierNest. Specifically, our approach leverages the inherent hierarchical structure present in primary diagnoses and groupings of these diagnoses into major diagnostic categories. Our methodology facilitates information borrowing across related patient subgroups and preserves interpretability at different hierarchical levels. Simulation studies demonstrate superior predictive accuracy of the proposed method, particularly with small subgroup sample sizes and varying degrees of hierarchical effects. We apply our methods to a large EHR dataset comprising Medicare patients.

  PublisherDOI

• Association of glycemic control with Long COVID in patients with type 2 diabetes: findings from the National COVID Cohort Collaborative (N3C)

  UNC Libraries · 2025-02-14

  articleOpen accessSenior author

  INTRODUCTION: Elevated glycosylated hemoglobin (HbA1c) in individuals with type 2 diabetes is associated with increased risk of hospitalization and death after acute COVID-19, however the effect of HbA1c on Long COVID is unclear. OBJECTIVE: Evaluate the association of glycemic control with the development of Long COVID in patients with type 2 diabetes (T2D). RESEARCH DESIGN AND METHODS: We conducted a retrospective cohort study using electronic health record data from the National COVID Cohort Collaborative. Our cohort included individuals with T2D from eight sites with longitudinal natural language processing (NLP) data. The primary outcome was death or new-onset recurrent Long COVID symptoms within 30-180 days after COVID-19. Symptoms were identified as keywords from clinical notes using NLP in respiratory, brain fog, fatigue, loss of smell/taste, cough, cardiovascular and musculoskeletal symptom categories. Logistic regression was used to evaluate the risk of Long COVID by HbA1c range, adjusting for demographics, body mass index, comorbidities, and diabetes medication. A COVID-negative group was used as a control. RESULTS: Among 7430 COVID-positive patients, 1491 (20.1%) developed symptomatic Long COVID, and 380 (5.1%) died. The primary outcome of death or Long COVID was increased in patients with HbA1c 8% to <10% (OR 1.20, 95% CI 1.02 to 1.41) and ≥10% (OR 1.40, 95% CI 1.14 to 1.72) compared with those with HbA1c 6.5% to <8%. This association was not seen in the COVID-negative group. Higher HbA1c levels were associated with increased risk of Long COVID symptoms, especially respiratory and brain fog. There was no association between HbA1c levels and risk of death within 30-180 days following COVID-19. NLP identified more patients with Long COVID symptoms compared with diagnosis codes. CONCLUSION: Poor glycemic control (HbA1c≥8%) in people with T2D was associated with higher risk of Long COVID symptoms 30-180 days following COVID-19. Notably, this risk increased as HbA1c levels rose. However, this association was not observed in patients with T2D without a history of COVID-19. An NLP-based definition of Long COVID identified more patients than diagnosis codes and should be considered in future studies.

  PublisherDOI

• The association between depressive symptoms and lower urinary tract symptoms and impact among women: Investigating behavioural, cognitive and physiological pathways

  British Journal of Health Psychology · 2025-01-21

  articleOpen access

  OBJECTIVE: Mechanistic studies are needed to understand why depressive symptoms are associated with poorer physical health. The objective of this study was to examine whether behavioural, cognitive and physiological factors mediated an association between depressive symptoms, measured in early adulthood, and lower urinary tract symptoms (LUTS) and their impact, a composite variable measured in mid-life adulthood, among women in the Coronary Artery Risk Development in Young Adults study, conducted in four regions of the United States. DESIGN: Prospective cohort study. METHODS: Data were examined for 871 women. Depressive symptoms were measured and averaged across Years 5, 10 and 15. Year 20 health behaviour combined information about smoking, physical activity and diet. Year 25 cognitive function combined performance on different cognitive tests. Year 25 metabolic syndrome combined standard risk criteria for waist circumference, triglycerides, high-density lipoprotein, blood pressure and glucose. A cluster analysis of urinary incontinence, other LUTS and impact data-collected two years after Year 25-was used to group women into one of four categories: no or very mild symptoms with no impact (bladder health) versus mild, moderate or severe symptoms/impact. RESULTS: Structural equation modelling showed a statistically significant direct path between depressive symptoms and LUTS/impact. Tests of indirect paths showed that health behaviours, cognitive function and metabolic syndrome did not mediate the association between depressive symptoms and LUTS/impact. CONCLUSIONS: Depressive symptoms in early adulthood appear to be associated with LUTS and their impact in mid-life adulthood over and above health behaviours, cognitive function and metabolic syndrome.

  PublisherOA PDFDOI

• Modified treatment policy effect estimation with weighted energy distance

  The Annals of Applied Statistics · 2025-12-01

  articleSenior author

  The causal effects of continuous treatments are often characterized through the average dose response function, which is challenging to estimate from observational data due to confounding and positivity violations. Modified treatment policies (MTPs) are an alternative approach that aim to assess the effect of a modification to observed treatment values and work under relaxed assumptions. Estimators for MTPs generally focus on estimating the conditional density of treatment, given covariates, and using it to construct weights. However, weighting using conditional density models has well-documented challenges. Further, MTPs with larger treatment modifications have stronger confounding, and no tools exist to help choose an appropriate modification magnitude. This paper investigates the role of weights for MTPs, showing that, to control confounding, weights should balance the weighted data to an unobserved hypothetical target population that can be characterized with observed data. Leveraging this insight, we present a versatile set of tools to enhance estimation for MTPs. We introduce a distance that measures imbalance of covariate distributions under the MTP and use it to develop new weighting methods and tools to aid in the estimation of MTPs. Using our methods, we study the effect of mechanical power of ventilation on in-hospital mortality.

  PublisherDOI

• 595 Is PHASTR faster? A target trial emulation case study in the N3C

  Journal of Clinical and Translational Science · 2025-03-25

  articleOpen access

  Objectives/Goals: Our study team won a Public Health Answers to Speed Tractable Results (PHASTR) contract to conduct a target trial emulation to answer “Does metformin show a reduction of severe outcomes of COVID-19 or of Long COVID in the N3C Data Enclave?” We quickly delivered an answer due to productive technical and collaboration support in the N3C. Methods/Study Population: Our analytic plan was updated based on helpful feedback from the PHASTR program. We performed a trial emulation analysis using the N3C data, comparing adult new users of metformin to controls prescribed fluvoxamine, fluticasone, ivermectin, or montelukast. The composite outcome was Long COVID or Death (LC/D) within 180 days of COVID infection. We used entropy balancing to estimate the average treatment effect with a weighted log-linear model. Productivity was enhanced by reusing code workbooks and validated codesets from related N3C projects. The team of 4 (physician, informaticist, data programmer, and statistician) and key unpaid advisors spent 10 weeks developing and analyzing the data. Results/Anticipated Results: Totally, 9,660 patients were identified for analysis. After weighting, there were 248 in the metformin and control groups. In the metformin group, 4.0% developed LC/D vs. 8.5% in the control group, with an adjusted risk ratio (aRR) of 0.47 (95% CI 0.25 to 0.89). Results were consistent across subgroups and sensitivity analyses. The PHASTR contract structure helped produce high-quality results quickly by not only providing funding but also requiring a compressed timeline for a small team to focus on the study. The most time was spent on contract execution, enclave provisioning, and too many last-minute download requests. A project final report was submitted in March and a full manuscript was submitted in September. Discussion/Significance of Impact: The analysis was productive because the environment made reuse easy and supported rich collaborations among clinicians, informaticists, epidemiologists, statisticians, and data developers. Advice from PHASTR advisors (Axel) and N3C diabetes domain team members was also key to a faster completion.

  PublisherOA PDFDOI

• The Effect of COVID‐19 on Incident Diabetes in Pediatric Patients: Findings From the National COVID‐19 Cohort Collaborative (N3C)

  UNC Libraries · 2025-12-11

  articleOpen access

  Studies showing increased diabetes incidence in pediatric patients after COVID‐19 are from data early in the pandemic, and some studies found conflicting results. Our objective was to evaluate trends in pediatric diabetes incidence and whether COVID‐19 was associated with increased risk across viral variant periods. We conducted a retrospective cohort study using National COVID‐19 Cohort Collaborative data to evaluate incident diabetes risk among COVID‐19‐positive pediatric patients compared to COVID‐19‐negative patients or controls with acute respiratory illness. Cohorts were weighted on demographics, data site, and body mass index percentile. The primary outcome was the cumulative incidence ratio (CIR) of incident diabetes for each viral variant era. There was no difference in the risk of incident diabetes in pediatric patients after COVID‐19 compared to patients in COVID‐19 negative or ARI control groups during any of the viral variant periods (e.g., ancestral period CIR 1.03, 95% CI 0.65–1.41). The predominant subtype of incident diabetes was T2D. Incidence rates over time followed a U‐shaped curve, with the highest incidence in the ancestral variant period. COVID‐19 was not associated with an increased risk of diabetes in pediatric patients. Incidence rates were highest early in the pandemic, and mirrored patterns of pandemic‐era healthcare utilization. The predominance of incident T2D subtype is concerning for the adverse effects of pandemic‐related lifestyle changes among pediatric patients.

  PublisherDOI

• The Effect of COVID‐19 on Incident Diabetes in Pediatric Patients: Findings From the National COVID‐19 Cohort Collaborative (N3C)

  Pediatric Diabetes · 2025-01-01

  articleOpen access

  Objective: Studies showing increased diabetes incidence in pediatric patients after COVID-19 are from data early in the pandemic, and some studies found conflicting results. Our objective was to evaluate trends in pediatric diabetes incidence and whether COVID-19 was associated with increased risk across viral variant periods. Research Design and Methods: We conducted a retrospective cohort study using National COVID-19 Cohort Collaborative data to evaluate incident diabetes risk among COVID-19-positive pediatric patients compared to COVID-19-negative patients or controls with acute respiratory illness. Cohorts were weighted on demographics, data site, and body mass index percentile. The primary outcome was the cumulative incidence ratio (CIR) of incident diabetes for each viral variant era. Results: There was no difference in the risk of incident diabetes in pediatric patients after COVID-19 compared to patients in COVID-19 negative or ARI control groups during any of the viral variant periods (e.g., ancestral period CIR 1.03, 95% CI 0.65-1.41). The predominant subtype of incident diabetes was T2D. Incidence rates over time followed a U-shaped curve, with the highest incidence in the ancestral variant period. Conclusions: COVID-19 was not associated with an increased risk of diabetes in pediatric patients. Incidence rates were highest early in the pandemic, and mirrored patterns of pandemic-era healthcare utilization. The predominance of incident T2D subtype is concerning for the adverse effects of pandemic-related lifestyle changes among pediatric patients.

  PublisherDOI

• Exploring the effects of mechanical ventilator settings with modified vector-valued treatment policies

  ArXiv.org · 2025-07-13

  preprintOpen accessSenior author

  Mechanical ventilation is critical for managing respiratory failure, but inappropriate ventilator settings can lead to ventilator-induced lung injury (VILI), increasing patient morbidity and mortality. Evaluating the causal impact of ventilator settings is challenging due to the complex interplay of multiple treatment variables and strong confounding due to ventilator guidelines. In this paper, we propose a modified vector-valued treatment policy (MVTP) framework coupled with energy balancing weights to estimate causal effects involving multiple continuous ventilator parameters simultaneously in addition to sensitivity analysis to unmeasured confounding. Our approach mitigates common challenges in causal inference for vector-valued treatments, such as infeasible treatment combinations, stringent positivity assumptions, and interpretability concerns. Using the MIMIC-III database, our analyses suggest that equal reductions in the total power of ventilation (i.e., the mechanical power) through different ventilator parameters result in different expected patient outcomes. Specifically, lowering airway pressures may yield greater reductions in patient mortality compared to proportional adjustments of tidal volume alone. Moreover, controlling for respiratory-system compliance and minute ventilation, we found a significant benefit of reducing driving pressure in patients with acute respiratory distress syndrome (ARDS). Our analyses help shed light on the contributors to VILI.

  PublisherOA PDFDOI

• Financial Strain Across 25 years and Men’s Lower Urinary Tract Symptoms: A Life Course Perspective

  American Journal of Men s Health · 2025-03-01 · 1 citations

  articleOpen access

  This research utilizes Coronary Artery Risk Development in Young Adults (CARDIA) cohort study data to examine whether financial strain is associated with subsequent lower urinary tract symptoms among men and whether healthcare barriers, health risk behaviors, and comorbid conditions explain this association. CARDIA recruited Black and White participants aged 18 to 30 years at baseline (1985–1986) from four United States cities. The analytic sample was comprised of men with complete data for analyses involving financial strain trajectories across 7 assessments ( n = 602) and mediation tests of data collected at 4 assessments ( n = 634). The outcome variable, assessed when the mean age of men was 50 years, was the American Urologic Association Symptom Index score, recoded into four symptom categories: none (6.3%); mild (62.6%), moderate (28.5%), and severe (2.6%). Symptom category was regressed on financial strain variables, adjusting for age, race, education, and self-reported benign prostatic hyperplasia. Regression analyses and structural equation modeling tested potential mediators. Compared to not being financially strained across early and midlife adulthood, experiencing more than one shift in financial strain was associated with 84% greater odds (95% confidence interval [1.24, 2.75]) of being categorized into a worse symptom category. Structural equation modeling showed that both difficulty receiving healthcare and depressive symptoms explained an association between difficulty paying for medical care and worse symptoms. Additional research is needed to confirm findings and examine other mechanisms that may further explain associations between financial strain and symptoms, such as stress responses. Accumulated evidence may inform future prevention interventions, including integrated healthcare approaches.

  PublisherDOI

• Evaluation of potential approaches for counting person-time in instances where no active comparator is present

  American Journal of Epidemiology · 2025-12-03

  articleOpen accessSenior author

  Emulating the target trial framework in pharmacoepidemiology is challenging when there is no active comparator. We evaluate six approaches to finding surrogate index dates for untreated patients with the goal of identifying one or more solutions that indicate they would give potentially unbiased results. This numerical experiment used 73 070 patients from the MarketScan administrative databases (2013-2019) with type 2 diabetes, first-line therapy with metformin, and second-line therapy with either sodium-glucose cotransporter 2 inhibitors (SGLT2is) or sulfonylureas. Patients taking sulfonylureas were converted into an experimental "untreated" arm. Part 1 sought to find surrogate index dates for the untreated arm. Part 2 compared the experimental estimates of the effect of SGLT2is on cardiovascular disease compared to sulfonylureas, using the surrogate index dates, to the reference estimate. The reference hazard ratio was 0.69. The hazard ratios after the respective approaches for selecting surrogate index dates are as follows: rejection sampling: 0.61, 0.63; median: 1.10, 1.15; prediction model: 0.96; matching algorithm: 1.07. Only the rejection sampling approaches for selecting a surrogate index date provided results which indicate low amounts of potential bias. Extreme care should be taken when making study design decisions for observational research questions that lack an active comparator group.

  PublisherOA PDFDOI

Frequent coauthors

• John B. Buse

  University of North Carolina at Chapel Hill

  32 shared

• Carolyn T. Bramante

  University of Minnesota Medical Center

  32 shared

• Hrishikesh Belani

  Institute of Accelerating Systems and Applications

  24 shared

• Ken Cohen

  Optum (United States)

  20 shared

• Kenneth J. Wilkins

  19 shared

• Menggang Yu

  University of Wisconsin–Madison

  19 shared

• Til Stürmer‎

  18 shared

• Michael A. Puskarich

  Hennepin County Medical Center

  18 shared

Education

• PhD, Statistics

  University of Wisconsin Madison

  2017

Awards & honors

• McKnight Presidential Fellow