Who ends up in your study? The epidemiology of selection bias

A study can have a large sample, sophisticated statistics, and a carefully specified research question — and still produce a biased

.

Sometimes the problem starts before the analysis.

It starts with who gets into the study.

🔎 Selection bias occurs when inclusion in a study, or remaining in it, is related to factors that affect the relationship we are trying to

.

Consider a cohort study looking at whether an exposure is associated with a health outcome.

If people who participate differ systematically from those who do not, the study population may no longer represent the population we intended to study.

And the problem can become even more complicated when participants are lost during follow-up.

For example, imagine a long-term study of occupational exposure and respiratory disease. If workers experiencing early symptoms are more likely to leave their jobs — and therefore leave the study — the participants who remain may have systematically different exposure and health profiles.

The final dataset can look perfectly clean.

But the process that created it may already have changed the association we observe.

📊 A bigger sample does not automatically solve this.

Increasing sample size mainly improves precision when selection is not systematically distorting the

. If the mechanism that determines who is included is related to the exposure and outcome, adding more participants selected through the same mechanism can simply give us a more precise

of a biased result.

Selection can happen in many places:

• recruitment

• eligibility criteria

• participation and non-response

• loss to follow-up

• survival to study entry

• conditioning on variables affected by both exposure and outcome

The last point is particularly important.

Sometimes researchers try to “control for” a variable because it appears relevant. But conditioning on the wrong variable can actually create or amplify an association rather than remove bias.

This is one reason causal diagrams can be useful: they force us to think about how people entered the dataset and what determined that process, rather than treating the final sample as if it appeared independently of the research question.

Before asking whether your model is appropriate, ask another question:

Who is missing from my dataset — and could their absence change the result?

That question can reveal problems that no regression model will fix.