Methodology Research
My research focuses on semiparametric and causal inference methods, with emphasis on the following areas:
📌 Semiparametric Theory
- Efficient influence functions · Asymptotic variance estimation under model misspecification · Double/debiased machine learning
📌 Causal Inference
- Doubly robust estimation · Quantile treatment effects (QTE) · Probability treatment effects (PTE)
📌 Other Topics
- Semiparametric cumulative probability models · Longitudinal missingness
Below are 3 completed methodology projects from my PhD to date.
Doubly Robust Estimators of Quantile Treatment Effects with Semiparametric Cumulative Probability Models
Problem addressed:
In real-world evidence studies, outcomes are often highly skewed and subject to detection limits, making mean-based causal analyses unstable.
My contribution:
- Developed a doubly robust framework for quantile and probability treatment effects using semiparametric cumulative probability models
- Implemented efficient influence function–based estimators with valid variance estimation
Why this matters:
- Captures treatment effects across the entire outcome distribution, not just the mean
- Provides robust inference under partial model misspecification
- Directly applicable to observational studies with skewed outcomes or detection limits
Technical skills demonstrated:
- Doubly robust estimation · QTE / PTE · Semiparametric modeling
- Monte Carlo simulation and bootstrap inference
- Analysis of real biomedical data (HIV studies)
Why Double Robustness Does Not Extend to Variance Estimation Under Parametric Nuisance Models?
Problem addressed:
In practice, influence function–based variance estimators are routinely used alongside doubly robust point estimators. While double robustness guarantees consistency of point estimation under partial model misspecification, no such guarantee exists for variance estimation.
My contribution:
- Developed a formal theoretical framework explaining why double robustness does not extend to variance estimation under parametric nuisance models
- Proposed alternative variance estimation strategies, including joint inference, sample-splitting-cross-fitting, with theoretical guarantees under model misspecification
Why this matters:
- Reveals hidden risks in uncertainty quantification for causal analyses in observational studies
- Provides practical guidance for valid inference in real-world settings where working models may be misspecified
Technical skills demonstrated:
- Semiparametric theory · Influence functions · Doubly robust estimation
- Asymptotic analysis
- Monte Carlo simulation · Real-data analysis
A Simple Augmentation of Weighted Generalized Estimating Equations for Doubly Robust Estimation in Longitudinal Data with Missingness
Clinical and statistical problem addressed:
In longitudinal studies, monotone missing outcomes are common and can severely bias inference. Standard approaches, such as outcome imputation or inverse probability weighting (IPW), are often highly sensitive to model misspecification.
My contribution:
- Developed a doubly robust estimator, the Augmented Weighted Generalized Estimating Equation (AWGEE), for longitudinal responses with monotone dropout
- Unified imputation-based methods and WGEE within a single estimating equation framework
- Improved efficiency through augmentation with re-imputed outcomes
- Applied the proposed method to real longitudinal psychiatric data
Why this matters:
- Reduces sensitivity to model misspecification.
- Enables robust inference for longitudinal treatment effects in the presence of dropout.
- Reveals a common misconception in the application of doubly robust estimators to missing data problems.
Technical skills demonstrated:
- Doubly robust estimation · Longitudinal data analysis · Missing data due to dropout· Weighted Generalized estimating equations (WGEE) · Monte Carlo simulation · Real-data analysis