Big Data & Causal Inference

Emulating the trials we cannot run.

Randomized trials are the gold standard, but many urgent questions can never be trialled, they are too rare, too slow, too costly, or concern drugs already in daily use. We turn some of the world’s largest health datasets into evidence, emulating the trial we wish existed and stress-testing every answer until it holds. This is the engine behind much of our most actionable work.

Our method

Target trial emulation.

Instead of hoping observational data behaves like a trial, we design the trial first, then emulate it. Done carefully, this recovers causal answers from real-world data without the delay and cost of running the study for real.

Specify the trial Define the randomized trial we wish existed: who is eligible, which treatments to compare, and the outcome that matters.

Emulate in big data Apply that protocol to millions of real-world records, controlling confounding with matching and weighting.

Estimate the effect Recover a confounding-adjusted causal estimate, a hazard ratio with confidence interval, free of immortal-time and selection bias.

Triangulate Stress-test it: independent replication, sensitivity analyses, negative and positive controls, and prespecified modifiers.

Every result is triangulated

External replicationSensitivity analysesNegative & positive controlsPrespecified modifiers

The data

Some of the world’s largest health datasets.

180M+ TriNetX Global federated health network

500k UK Biobank Deep phenotyping and imaging

52k NACC Longitudinal dementia cohort

20+ SeLECT registry Post-stroke seizure cohorts

What it delivers

Real-world data, actionable answers.

28%potentially preventable

Accepted for publication

The right antiseizure medication could prevent a quarter of clots in epilepsy patients on blood thinners.

In over 180 million records, we used target trial emulation to show that up to 28% of thromboembolic events in people with epilepsy taking direct oral anticoagulants could be prevented by choosing a low-interaction antiseizure medication. Even drugs long assumed safe, such as levetiracetam, carried risk. A clear, immediately usable example of big data changing practice.

Under review 4.4M

Acquired epilepsy is not a bystander.

Across 4.4 million adults and 20 different brain insults, epilepsy after brain injury emerged as a distinct driver of illness and death, not just a marker of severity, and early treatment was linked to lower mortality.

Preprint 27%

A path to preventing dementia.

The same method powers our dementia work: sodium channel blocking medications linked to a 27% lower dementia risk, and antiseizure medications to lower Alzheimer’s pathology.

Seizures and Dementia

We deliberately show only work that is accepted, under review, or already public as a preprint. Plenty more is in the pipeline, so stay tuned.