DPhil student in Primary Health Care, University of Oxford (Nuffield Department of Primary Care Health Sciences, St Anne's College), since October 2026.
My doctoral research asks what uneven symptom recording in primary care records does to cancer risk prediction models, and what that means for different patient groups. Analyses in CPRD Aurum are planned; data access is in progress.
I work mostly in R, care about reproducibility, and like explaining a method well enough that a non-specialist walks away understanding it.
- Clinical prediction models: calibration, subgroup performance, and whether a study was powered to find what it looked for
- Missing and incomplete data in routine health records
- Genetic epidemiology and causal inference, from my earlier work: rare-variant association analysis, Mendelian randomisation
- Reproducible research: Quarto, seeded simulation, hand-checked estimators validated against established packages
Model Evaluation from Scratch »
A worked example in R of why one AUROC tells you almost nothing, built one estimator at a time. All data are simulated, and every metric is validated against wilcox.test, glm or analytic truth.
https://lexiyao.github.io/model-evaluation-from-scratch/
Mendelian Randomisation from Scratch »
A seven-chapter R walkthrough that codes the standard MR toolkit from first principles on synthetic data, then validates each estimator against ivreg / MendelianRandomization.
https://lexiyao.github.io/mendelian-randomisation-from-scratch/
Odds-Ratio Stats Converter » A dependency-free, single-file web tool for OR, log-OR, SE, confidence intervals and two-sided p, with a parser for reported strings. Math core unit-tested against known values. https://lexiyao.github.io/or-stats-converter/
Evidence Briefs » Two-page, plain-language evidence summaries: bottom line first, then the certainty and the limits. Worked example: does lowering LDL cholesterol reduce coronary heart disease? https://lexiyao.github.io/evidence-brief-ldl-chd/
R · Quarto · tidyverse / ggplot2 · prediction-model evaluation · survival & logistic regression · Mendelian randomisation · GWAS · reproducible pipelines