Work EHR Survival Analysis
One risk factor looked like it did nothing at all. It was doing plenty — just not at a constant rate.
R · survival · Cox regression · Synthea
The finding
Age came back flat and statistically null. Testing the model’s core assumption showed why: its effect drifts across the follow-up period, and averaging it into a single number had erased it. Diabetes carried roughly 2.4× the hazard on its own.
Method
Kaplan–Meier curves with a log-rank test, Cox proportional hazards adjusted for age, sex, diabetes and obesity, and a formal check of the proportional-hazards assumption.
So what
A non-significant result is not a finding of no effect until you have checked the assumption that produced it. Most of the cost of a wrong answer here is that it looks exactly like a right one.
Interactive
Toggle to split the survival curve by diabetes status; diabetic patients develop hypertension faster.
How long until hypertension?
The curve is the share of patients still hypertension-free as years pass — it only ever steps down, once per diagnosis. Split it by diabetes.
A patient still event-free just means we hadn't seen it happen yet — censored, not deleted. That's most of this cohort.
Full write-up
What I did
My first R project — a four-lap survival analysis on the same Synthea synthetic EHR data used in Synthea EHR and EHR ML Risk Model, asking a different question than either: not “does this patient have hypertension?” but “how long does it take?” I built a one-row-per-patient time-to-event dataset (1,171 patients: 302 diagnosed, 869 censored), fit Kaplan-Meier curves split by diabetes status with a log-rank test, then a Cox proportional-hazards model adjusting for age, sex, diabetes, and obesity together, and checked whether the model’s core assumption actually held.
Key finding
Diabetic patients have roughly 2.4× the hazard of a hypertension diagnosis at any given moment, even after adjusting for age, sex, and obesity — diabetes isn’t just riding along with being older or heavier, it carries its own independent signal (obesity holds up too, at ~1.4×).
The more interesting result was a mistake I almost made: age’s adjusted hazard ratio looked completely null (HR 1.00, p = 0.4). Checking the model’s proportional-hazards assumption (cox.zph) showed why — age’s effect on hypertension risk isn’t constant over time at all, it drifts across the follow-up period, and averaging a time-varying effect into one static coefficient had quietly washed it out. A flat, non-significant number isn’t automatically “no effect.”
Honest limitations
diabetes and obesity are “ever diagnosed, at any point in the record” flags rather than baseline measurements, which risks reverse time-ordering (crediting diabetes for predicting hypertension when, for some patients, hypertension came first). The proportional-hazards assumption failed globally, not just for age, so several of these hazard ratios are time-averaged summaries of effects that actually change shape. The diabetic stratum is small (76 patients). There’s no competing-risks handling for death as an alternative reason a patient stops being observed, and no external validation against a different cohort. And — being Synthea data — any relationship this recovers is designed into the simulator, not discovered in the world, so this is a methods exercise in survival analysis, not a clinical finding.
Stack
R · survival package (base R otherwise: read.csv, aggregate, merge) · Synthea synthetic EHR