Study design and bias questions are among the most learnable on the USMLE, and they appear at every Step, from Step 1 epidemiology to the research abstracts that clinical exams ask you to interpret. They look like vocabulary tests, but the exam rarely asks for a definition. It describes what researchers did and asks you to name it. One question, asked first, settles most of them: how were the participants chosen?
Start with how participants were selected
Read the vignette for the selection step and the rest usually follows.
Selected at random into groups by the researchers. A randomised controlled trial.
Selected by exposure, then followed for outcomes. A cohort study.
Selected by outcome, then asked about past exposures. A case-control study.
Selected as a population and measured once. A cross-sectional study.
Grouped by population rather than by individual. An ecological study.
The trap is the word “retrospective”. A retrospective cohort study uses old records but still groups people by exposure and looks forward through those records for outcomes. It remains a cohort study. Direction of selection decides the design, not the date of the data.
Each design supports particular conclusions
Randomised controlled trials are the strongest design for showing that an intervention causes an outcome, because randomisation balances confounders, including ones nobody thought to measure. Crossover trials, in which each participant receives both treatments in sequence, let people act as their own controls but need a washout period so the first treatment does not affect the second.
Cohort studies measure incidence and support relative risk. They suit common outcomes and rare exposures, and they establish that exposure came before outcome.
Case-control studies suit rare outcomes, because you can recruit everyone who has the disease. They cannot measure incidence, so they report an odds ratio rather than a relative risk.
Cross-sectional studies measure prevalence at one point in time. Because exposure and outcome are measured together, they cannot show which came first.
Meta-analyses pool results from several studies to increase power, but they inherit the weaknesses of the studies they include, including any bias in which studies were published.
If an answer choice offers a measure the design cannot produce, such as a relative risk from a case-control study, it is wrong.
Selection bias happens at enrolment
Several named biases describe who ended up in the study.
Selection bias in general means the study population differs systematically from the population the conclusions are meant for. A trial recruiting only from a specialist centre is a common example.
Attrition bias arises when people drop out unevenly between groups, for instance when patients with side effects leave the treatment arm.
Berkson bias comes from selecting cases and controls from hospital patients, who differ from the general population.
Healthy worker effect describes occupational cohorts appearing healthier than the general population because people must be well enough to work.
None of these can be fixed afterwards by blinding. They are prevented by careful recruitment, random sampling and complete follow-up.
Information bias happens during measurement
Recall bias occurs when people with an outcome remember past exposures differently from those without it, which makes it the characteristic weakness of case-control studies.
Observer bias occurs when the person measuring the outcome knows which group a participant is in. Blinding the assessors addresses it.
Hawthorne effect describes participants changing their behaviour because they know they are being studied.
Measurement bias covers any systematic error in how data are collected, such as a miscalibrated instrument used in one group only.
Screening has its own pair of biases
Lead-time bias makes survival appear longer after screening simply because the disease is diagnosed earlier. If death occurs at the same moment either way, nothing has been gained, and the correct comparison is mortality rather than survival from diagnosis.
Length-time bias arises because screening at intervals is more likely to catch slow-growing disease, which has a better outlook regardless of treatment.
Questions about a new screening programme that “improves survival” are very often testing one of these two.
Confounding and effect modification are not the same
A confounder is associated with both the exposure and the outcome and distorts the link between them. Coffee drinking appears to cause lung cancer until smoking is taken into account. Randomisation, restriction, matching and stratified analysis all address it.
Effect modification is different: the exposure genuinely has a different effect in different subgroups. Stratifying reveals it rather than removing it. If the stratified results differ from each other, suspect effect modification; if they agree with each other but differ from the crude result, suspect confounding.
Practising across the sequence
These concepts do not change between Steps, but the presentation does. Step 1 tends to describe a study in a few sentences. Step 2 CK and Step 3 are more likely to show an abstract or a results table and ask what it supports. Practise both forms. When you miss one, note whether you misread the selection step, chose a measure the design cannot produce, or confused two named biases, because each has a different fix.
Frequently asked questions
How do you tell a cohort study from a case-control study?
Look at how participants were chosen. A cohort study starts with people grouped by exposure and follows them for outcomes, while a case-control study starts with people who have and do not have the outcome and looks back for exposures. The direction of selection decides it, not whether the data are old or new.
What is the difference between lead-time and length-time bias?
Lead-time bias is an apparent survival gain caused only by diagnosing disease earlier. Length-time bias arises because screening tends to find slow-growing disease, which has a better prognosis anyway. Both make screening look more effective than it is.
Is confounding a type of bias?
It is usually treated separately. Confounding happens when a third factor is associated with both the exposure and the outcome, distorting the apparent link between them. It can be reduced by randomisation, matching, restriction or statistical adjustment.