Master Biostatistics & Literature Appraisal
for USMLE Step 3

What the USMLE Step 3 Tests in Biostatistics & Literature Appraisal

USMLE Step 3 Biostatistics & Literature Appraisal tests your ability to interpret clinical studies and apply quantitative reasoning to patient management. You must calculate and interpret number needed to treat (NNT) and number needed to harm (NNH) for therapies like statins in primary prevention or anticoagulants in atrial fibrillation. You need to appraise diagnostic tests using sensitivity, specificity, and likelihood ratios for conditions such as pulmonary embolism (Wells criteria, D-dimer) or coronary artery disease (exercise stress test). You must understand study designs (RCT, cohort, case-control) and bias (selection, recall, lead-time bias in cancer screening). You will apply screening principles (e.g., USPSTF for colorectal cancer with colonoscopy every 10 years from age 45, or mammography from age 40). You must calculate absolute risk reduction (ARR) and relative risk reduction (RRR) from trial data (e.g., CARPENTER trial for antenatal corticosteroids). You need to interpret Kaplan-Meier curves and hazard ratios from oncology trials (e.g., trastuzumab in HER2+ breast cancer). Decision-making includes applying p-values (<0.05) and confidence intervals (95% CI not crossing 1.0 for odds ratio) to guide therapy changes.

High-Yield Concepts

• Sensitivity, Specificity, PPV, NPV: Sensitivity = TP/(TP+FN) — high sensitivity rules out disease (SnNOut). Specificity = TN/(TN+FP) — high specificity rules in disease (SpPIn). Positive predictive value (PPV) = TP/(TP+FP); negative predictive value (NPV) = TN/(TN+FN). For a disease with 2% prevalence (e.g., HIV in low-risk population), even a 99% specific test yields low PPV. Use likelihood ratios: LR+ = sens/(1-spec); LR- = (1-sens)/spec. Pre-test probability × LR = post-test probability (use Fagan nomogram).
• Number Needed to Treat (NNT) and Number Needed to Harm (NNH): NNT = 1/ARR, where ARR = control event rate (CER) – experimental event rate (EER). Example: in the 4S trial for simvastatin in post-MI patients, CER for death was 11.5%, EER 8.2%, ARR = 3.3%, NNT = 30. NNH = 1/attributable risk increase (ARI), where ARI = experimental adverse event rate – control adverse event rate. For rivaroxaban in atrial fibrillation (ROCKET-AF), NNH for major bleeding was ~50 over 2 years. Always use absolute numbers, not relative risk reduction, for clinical decisions.
• Study Designs and Bias: RCT: gold standard for therapy (e.g., PARADIGM-HF for sacubitril/valsartan). Cohort study: starts with exposure, follows for outcome (e.g., Framingham Heart Study for cardiovascular risk). Case-control: starts with outcome, looks back at exposure (e.g., smoking and lung cancer). Bias types: selection bias (e.g., healthy worker effect), recall bias (case-control studies), lead-time bias (screening for breast cancer), length-time bias (screening for slow-growing cancers). Confounding: e.g., hormone replacement therapy and coronary disease — women on HRT had higher socioeconomic status, which confounded the apparent protective effect.
• P-value, Confidence Interval, and Statistical Significance: P-value < 0.05 indicates <5% probability that observed difference is due to chance (type I error). 95% confidence interval (CI) for an odds ratio (OR) or hazard ratio (HR) that does not include 1.0 indicates statistical significance. Example: HR 0.80 (95% CI 0.65–0.98) for mortality with drug X means significant benefit. CI width reflects precision; narrow CI in large trials (e.g., 10,000 patients) is more reliable. Beware of multiple comparisons (Bonferroni correction: divide α by number of comparisons).
• Kaplan-Meier Curves and Hazard Ratios: Kaplan-Meier curves show survival over time with steps at events; censored patients (lost to follow-up or alive at study end) are marked with tick marks. The log-rank test compares curves. Hazard ratio (HR) from Cox regression: HR = 0.75 means 25% reduction in hazard of event per unit time. For example, in the HERA trial for trastuzumab in HER2+ breast cancer, HR for disease-free survival was 0.64 (95% CI 0.53–0.77), favouring 1 year of treatment. Always check proportional hazards assumption (parallel curves).
• Screening Test Principles: Screening reduces mortality if early detection improves outcomes (e.g., mammography for breast cancer: USPSTF recommends biennial from age 40–74). Criteria: Wilson & Jungner — disease should be important, detectable at early stage, acceptable treatment exists (e.g., colorectal cancer: FIT every year or colonoscopy every 10 years from age 45). Overdiagnosis: finding indolent cancers (e.g., prostate cancer with PSA screening). Lead-time bias: screening appears to prolong survival by advancing diagnosis date, not actual death delay. Length-time bias: screening catches slower-growing tumours more often.
• Absolute vs Relative Risk Reduction: Relative risk reduction (RRR) = (CER – EER)/CER × 100%. Absolute risk reduction (ARR) = CER – EER. Example: in the JUPITER trial for rosuvastatin in primary prevention, CER for major cardiovascular events was 1.8%, EER 0.9%, RRR = 50%, ARR = 0.9%, NNT = 111. RRR exaggerates benefit in low-risk populations. Always use ARR and NNT for patient counselling (e.g., for aspirin in primary prevention, ARR ~0.1% per year for MI, but NNH for bleeding ~0.5% per year).
• Diagnostic Test Interpretation: Likelihood Ratios and Bayes' Theorem: Pre-test probability (prevalence) × LR = post-test probability. Example: for pulmonary embolism, Wells score stratifies pre-test probability. If low probability (10%) and D-dimer negative (LR- ~0.1), post-test probability ~1% — rule out. If high probability (50%) and CT pulmonary angiogram positive (LR+ ~20), post-test probability ~95%. Use Fagan nomogram. For a test with sensitivity 90% and specificity 95%, LR+ = 18, LR- = 0.105. This is tested in scenarios like stress echocardiography for CAD or ANA for SLE.

Common Traps in Biostatistics & Literature Appraisal Questions

• Confusing relative risk reduction (RRR) with absolute risk reduction (ARR); RRR can be misleadingly large when baseline risk is low (e.g., 50% RRR from 2% to 1% is only 1% ARR).
• Assuming a positive predictive value (PPV) equals sensitivity; PPV depends heavily on disease prevalence, so a highly sensitive test can still have low PPV in a low-prevalence population.
• Thinking that a p-value > 0.05 proves the null hypothesis (no difference); it only indicates insufficient evidence to reject it — type II error (beta) may exist.
• Misinterpreting a confidence interval that includes 1.0 for odds ratio as 'no effect' without considering clinical significance or study power (e.g., underpowered trial).
• Forgetting that intention-to-treat (ITT) analysis is the standard for RCTs, not per-protocol; ITT preserves randomization and avoids bias from dropouts (e.g., in drug trials for chronic diseases).
• Applying screening test results from a high-prevalence population (e.g., specialist clinic) to a low-prevalence population (e.g., general practice) without adjusting for spectrum bias.

How to Revise Biostatistics & Literature Appraisal for the USMLE Step 3

For USMLE Step 3 Biostatistics, prioritise rapid calculation of NNT/NNH and interpretation of ARR/RRR from 2x2 tables given in clinical vignettes (e.g., statin trials or anticoagulation studies). Focus on study design flaws: identify selection bias in cohort studies (e.g., healthy worker effect) and recall bias in case-control studies. Practise applying likelihood ratios to change pre-test probability using Fagan nomogram for conditions like PE or CAD. Master Kaplan-Meier curve interpretation: identify when curves separate (treatment benefit) and where censoring occurs. Expect questions that require you to choose the correct statistical test (t-test for continuous, chi-square for categorical) based on data type. Revise USPSTF screening guidelines for breast, colorectal, cervical, and lung cancer. Use the CCS algorithm: calculate ARR, then NNT, then decide if benefit outweighs harm (e.g., aspirin for primary prevention). Simulate timed practice with 2x2 tables and survival curves from NEJM-style abstracts.

Practise it: MedLumen has 50 Biostatistics & Literature Appraisal questions for the USMLE Step 3, each with a full explanation and references.