Causality in Health and Medicine
Welcome
Health professionals constantly make causal decisions. We ask whether an exposure causes disease, whether a treatment improves recovery, whether a policy prevents harm, and whether an observed inequality could be reduced through intervention. Data can show patterns, but responsible action requires us to understand what those patterns do—and do not—tell us about causes.
These questions are not confined to academic journals. They appear in clinics, family conversations, social media, news reports, and government decisions. Students may encounter claims such as:
- “Drinking teh tarik causes stroke.”
- “A COVID-19 vaccine caused a blood clot.”
- “Electronic cigarettes are safer, so they improve health.”
- “Hormone replacement therapy prevents heart disease.”
- “People with obesity sometimes survive illness better, so obesity may be protective.”
- “Screening saves lives because people diagnosed through screening survive longer.”
- “Social media causes anxiety and depression in young people.”
Each statement sounds simple. Each hides a more difficult causal question.
Consider teh tarik. A person who frequently drinks it may consume more sugar and energy, but may also differ from an infrequent drinker in diet, physical activity, smoking, diabetes, blood pressure, occupation, income, and healthcare use. What exactly is the intervention: avoiding teh tarik, choosing a smaller serving, reducing added sugar, replacing condensed milk, or changing the entire dietary pattern? Is stroke risk affected directly, or through weight gain, diabetes, and hypertension? A comparison between drinkers and non-drinkers cannot answer these questions unless the alternatives, pathways, and competing explanations are made clear.
Vaccine-safety questions illustrate another challenge. If a blood clot occurs after vaccination, the timing is important, but “after” does not automatically mean “because of.” Some blood clots would have occurred in the same period without vaccination. Investigators must compare observed events with an appropriate counterfactual, distinguish vaccine products and biological syndromes, account for age and other risk factors, and examine whether a highly specific pattern occurs more often than expected. Careful surveillance did identify thrombosis with thrombocytopenia syndrome as a very rare but serious adverse event associated particularly with some adenovirus-vector COVID-19 vaccines. This is not evidence that every clot following every COVID-19 vaccine was caused by vaccination; it demonstrates how credible causal investigation can separate a specific safety signal from coincidence and background disease (WHO; EMA).
Other controversies reveal different problems. People who choose preventive treatment may already be healthier than those who do not, creating healthy-user bias. A disease can cause weight loss before diagnosis, making lower weight appear harmful through reverse causation. Screening can lengthen measured survival time without postponing death, producing lead-time bias. Social-media use and poor mental health may influence each other, making the direction of causation difficult to establish. An exposure may reduce one risk while increasing another, so “safe” or “dangerous” may depend on the comparator, population, outcome, dose, and follow-up period.
These examples matter because causal errors have consequences. A false causal claim may create fear, stigma, unnecessary treatment, or ineffective policy. Failure to recognize a genuine cause may delay prevention and expose more people to harm. Causal inference provides a disciplined way to move from a dramatic claim to answerable questions:
- What precisely is the exposure or intervention?
- Compared with what alternative?
- For which population?
- Over what period?
- Which health outcome?
- What other explanations could produce the observed pattern?
- What evidence would make the causal interpretation more or less credible?
Instead of asking, “Does teh tarik cause stroke?”, we might ask: Among Malaysian adults with no previous stroke, what would be the ten-year difference in stroke risk if they replaced one daily serving of standard teh tarik with unsweetened tea, compared with continuing their usual daily serving?
The longer question is less dramatic, but it is much closer to something that can guide research and health decisions.
This book introduces causal inference for students and researchers in epidemiology, public health, clinical medicine, and related health sciences. It is written for readers who use evidence but may not have advanced mathematical training.
The guiding question is simple:
What would happen to health if we changed an exposure, treatment, policy, or other condition?
How the book is organized
The book follows the natural progression of learning causal inference.
Part I: Introduction - Thinking Causally
The opening part develops the language and logic of causality. It distinguishes association, prediction, and causation; introduces potential outcomes and causal diagrams; explains the assumptions needed for causal inference; and presents a unified workflow.
Part II: Threats to Causal Inference
The second part examines why causal conclusions can fail. It begins with confounding and inappropriate adjustment, then develops selection bias and missing data. Later chapters will address measurement error, time-related biases, mediation, interaction, and failures of design, modelling, and interpretation.
Part III: Practical Causal Inference
The final part will guide readers through a complete applied analysis: defining a question, specifying a target trial, drawing a DAG, preparing epidemiological data, choosing an estimator, checking diagnostics, conducting sensitivity analyses, and reporting results responsibly. These chapters will progressively support student assignments.
How to use this book
Each chapter contains:
- clear learning objectives;
- health and medical examples;
- plain-language explanations before formal notation;
- translations of necessary equations;
- causal diagrams;
- executable R examples;
- warnings about common mistakes;
- review questions and applied exercises; and
- assignment milestones where appropriate.
Readers new to causal inference should work through the chapters in order. Readers with prior experience may use individual chapters as references, but should first review the causal framework in Chapter 1.
A note about mathematics
Equations are used when they make a causal comparison or assumption more precise. Every important equation is explained in ordinary language. The goal is not to manipulate symbols for their own sake, but to understand:
- which people or interventions are being compared;
- which outcome is being estimated;
- what assumptions connect the data to the causal question; and
- how the result should inform health decisions.
A note about software
The practical examples use R, RStudio, and Quarto. Code is included to make analyses transparent and reproducible. The conceptual reasoning remains primary: software can implement an analysis, but it cannot decide whether the causal assumptions are credible.
The running example
Smoking and lung cancer provide a consistent example throughout the foundations of the book. The example demonstrates confounding, selection, measurement, mediation, interaction, time, and evidence appraisal. Additional health examples will be introduced to show how the same principles apply across epidemiology and medicine.
A statistical result becomes causal only through a defensible chain connecting the question, target population, intervention, causal structure, study design, assumptions, data, analysis, diagnostics, and interpretation.