My Blog
What did half a million people actually consent to?
UK Biobank participants signed a one-page consent form in 2006, years before generative AI existed. We look at what that broad consent covered, why trained models and AI-driven analysis fall outside the safeguards built around it, and what consent and governance could look like instead.
What is a directed acyclic graph (DAG)?
How directed acyclic graphs (DAGs) help epidemiologists separate causal from non-causal paths, and decide which variables to adjust for in health data research.
Why I teach a workshop on codelists
Codelists decide which patients end up in a study, but building one is rarely taught as a method. This post explains why I run a three hour codelist workshop for universities and what it covers.
What is the difference between selection bias and sampling bias?
A reference guide to the difference between selection bias and sampling bias in health data research, with worked examples of how each one distorts your results.
How Do We Handle Rare Events in Synthetic Data?
A look into how we can use synthetic data generation to increase the number of rare events in our dataset, and what are risks and benefits of this approach
Synthetic Data in Machine Learning: Augmentation and Collapse
Exploring the use of synthetic data in machine learning, focusing on augmentation, model collapse, and the implications for health research using synthetic data.
What is a confounder, really?
A reference guide to what confounding means in health data research and how it is different from mediators and colliders.
Fairness and Bias Amplification in Synthetic Data
Explore how synthetic data can amplify existing biases and affect fairness in health research. Learn why this happens and how it differs from representativeness.