Lesson 31 of 36 · 15 min

Bias, randomization & controls

Study design determines how many alternative explanations survive. Randomization, controls and blinding do not make a study perfect; they are tools for making particular biases less likely.

Threats to inference

Ask what else could have produced the observed difference.

Selection bias

Were the people entering each group systematically different before treatment began?

Confounding

Could another variable be associated with both the intervention and the outcome?

Measurement bias

Could knowledge of treatment or inconsistent measurement change the recorded outcome?

Attrition

Did enough participants leave the study that the remaining group may no longer represent the original comparison?

Why randomize?

Randomization tries to distribute known and unknown prognostic factors without choosing who gets what.

In a sufficiently large well-conducted trial, random assignment helps make groups comparable at baseline and supports causal interpretation of between-group differences.

Randomization does not fix poor measurement, missing data, protocol deviations or selective reporting. It solves one class of problems, not every class.

Controls & blinding

The comparator tells you what the study can isolate.

A placebo control can help separate treatment effects from expectation, natural history and background care. An active comparator asks a different question: how the intervention performs against another treatment.

Blinding can reduce differences in behavior, reporting and assessment that occur when participants or investigators know the assignment.

Design rule

Every design removes some explanations and leaves others alive.

Good critical reading means identifying which explanations remain plausible after the study design is considered.

Key takeaways

✓Bias is a systematic distortion, not merely random noise.
✓Randomization improves causal inference by reducing systematic baseline differences between groups.
✓The choice of comparator determines the question a trial can answer.
✓Blinding can reduce expectation and assessment-related bias.
✓Randomized does not automatically mean low risk of bias in every domain.

Ready for the checkpoint?

The quiz opens on its own page. Get every answer correct to mark this lesson complete and continue.

Take lesson quiz →

Endpoints, effect size & uncertainty ←

Next lesson

Surrogate vs clinical outcomes →

Continue when you are ready, or come back to this lesson whenever you need a refresher.