Biostatistics and epidemiology for USMLE Step 1

A study follows 20,000 people over 15 years, recording smoking status at baseline and subsequent incidence of bladder cancer. Which measure of association can be calculated directly from this design?

1. Prevalence ratio Wrong. A prevalence ratio belongs to a cross-sectional study, which measures existing cases at one point in time rather than new cases developing over time.
2. Relative risk correct Correct. This is a prospective cohort study. Because new cases (incidence) are measured directly in both the exposed and unexposed groups, the risk in each group, and therefore the relative risk, can be calculated directly.
3. Hazard ratio only, with no other measure available Wrong. A hazard ratio can be estimated from time-to-event data, but it is not the only measure this study design allows, and relative risk is what the study permits most directly.
4. Odds ratio only Wrong. Being limited to only an odds ratio is what happens in a CASE-CONTROL study, where people are chosen based on outcome, so incidence cannot be measured directly. An odds ratio can also be calculated from this cohort study, but that is not what makes this design special.
5. Attributable risk cannot be calculated from this design Wrong. Attributable risk is exactly what a cohort design supports. It is the difference in incidence between the exposed and unexposed groups.

The point: The study design determines which measure you can calculate. Cohort studies give incidence, relative risk, and attributable risk. Case-control studies give only an odds ratio. Cross-sectional studies give prevalence. Randomized trials let you draw causal conclusions. When a disease is rare, the odds ratio becomes a good estimate of the relative risk.

Source: USMLE Content Outline — biostatistics and epidemiology NBME / FSMB (USMLE programme) · tier 0, exam blueprint / regulator

A screening test for a disease with a prevalence of 1% has a sensitivity of 99% and a specificity of 90%. It is applied to 10,000 people. Approximately what is the positive predictive value?

1. About 90% Wrong. About 90% confuses PPV with specificity. PPV depends on how common the disease is (prevalence), while specificity does not.
2. About 99% Wrong. About 99% confuses PPV with sensitivity, and it would mean almost every positive test result is a true case, which is not what happens here.
3. About 9% correct Correct. Out of 100 people with the disease, 99 test positive. Out of the 9,900 without it, 10%, or 990 people, test positive by mistake (false positives). Positive predictive value equals 99 divided by (99 plus 990), which is about 9%. In a population where the disease is rare, even an excellent test still produces mostly false positives. This is one of the most tested ideas in USMLE biostatistics.
4. About 50% Wrong. A PPV of about 50% would need the number of true positives and false positives to be roughly equal, which needs either a much higher disease prevalence or a much better specificity than this test has.
5. It cannot be calculated without the incidence Wrong. Prevalence is already given in the question, and that is exactly what is needed to calculate PPV. Incidence is not needed for predictive value.

The point: Sensitivity and specificity are fixed properties of the TEST itself. PPV and NPV depend on how common the disease is (PREVALENCE) in the population being tested. As prevalence falls, PPV falls and NPV rises. This is why a positive screening test should be confirmed with a highly specific test, which is the reasoning behind testing with ELISA first and then confirming with a Western blot.

Source: USMLE Content Outline — biostatistics and epidemiology NBME / FSMB (USMLE programme) · tier 0, exam blueprint / regulator

A cohort study finds that coffee drinkers have a higher rate of lung cancer than non-drinkers. Coffee drinkers in this population are also far more likely to smoke. After adjustment for smoking status, the association with coffee disappears. Which phenomenon does this illustrate?

1. Confounding correct Correct. This is confounding. Smoking is linked to the exposure (coffee drinking) and, separately, causes the outcome (lung cancer) on its own, and it is not a step in the pathway between coffee and cancer. Adjustment, stratification, matching, or randomization can all remove this kind of distortion.
2. Recall bias Wrong. Recall bias is an error in remembering past exposure that differs between groups, and it affects case-control studies far more than a cohort study like this one, where exposure was measured at the start.
3. Berkson bias Wrong. Berkson bias is a selection bias that comes from using hospitalized patients as controls, which does not apply to a study of a community-based cohort like this one.
4. Effect modification Wrong. In effect modification, the association is DIFFERENT across groups, real in one group and absent in another, and this should be reported separately rather than adjusted away. Here, the association disappeared completely once smoking was accounted for, which points to confounding, not effect modification.
5. Lead-time bias Wrong. Lead-time bias is the apparent survival benefit that comes only from diagnosing a disease earlier, without actually changing how the disease progresses. It applies to screening studies, not to this one.

The point: The test for a confounder: it is associated with the exposure, it independently causes the outcome, and it is not a step in between the two. Fix confounding either at the design stage (randomization, restriction, matching) or at the analysis stage (stratification, multivariable adjustment). Effect modification is not a bias, it should just be reported separately for each group (stratum).

Source: USMLE Content Outline — biostatistics and epidemiology NBME / FSMB (USMLE programme) · tier 0, exam blueprint / regulator