Study guide · BEP
Biostatistics & Epidemiology/Population Health/Interpretation of Medical Literature
3 to 5% of Step 2 CK · 13 of the 320 questions on our mock · 9 sections
This system covers biostatistics, epidemiology, population health, and reading medical research. It is about 3% to 5% of Step 2 CK, and its items test practice-based learning and improvement or, for screening, health maintenance. Questions often give a 2 x 2 table, a trial result, or a short journal abstract, and then ask what it means for one patient in front of you. You must calculate a measure, name a design or a bias, and decide whether a result is valid and applies to your patient. Set up the table carefully, check which row or column is the denominator, and always ask what the number changes in patient care.
A study feature written by ExamDragon, not part of the real exam. The headings follow the USMLE Content Outline for this system, in our words. Step 2 CK asks what you would do next: the best test, the first-line treatment, the safe choice.
At a glance
BEPof Step 2 CK, the published range
- On our mock
- 13
- Sections
- 9
- Key terms
- 13
What this guide covers
- Diagnostic tests: sensitivity, specificity and predictive values
- Likelihood ratios and post-test probability
- Measures of risk and treatment effect
- Study designs and the strength of evidence
- Bias, confounding and effect modification
- Statistical inference: p values, confidence intervals and power
- Screening principles and screening biases
- Outbreak investigation and population health
- Reading a clinical trial abstract
1Diagnostic tests: sensitivity, specificity and predictive values
- Set up the 2 x 2 table with disease (by the reference standard) in columns and the test result in rows. Cells: true positive (TP), false positive (FP), false negative (FN), true negative (TN).
- Sensitivity = TP / (TP + FN), read down the disease column. Specificity = TN / (TN + FP), read down the no-disease column.
- Positive predictive value (PPV) = TP / (TP + FP) and negative predictive value (NPV) = TN / (TN + FN), read across the rows. They answer the patient's question: 'I tested positive; do I have it?'
- A highly sensitive test, when negative, rules a disease out (SnNout). Use it for screening and for diseases you must not miss. A highly specific test, when positive, rules a disease in (SpPin). Use it to confirm.
- Sensitivity and specificity belong to the test. Predictive values change with prevalence: lower prevalence means lower PPV and higher NPV.
- Moving a cutoff to call more people positive raises sensitivity and lowers specificity; moving it the other way does the reverse. The ROC curve plots this trade-off, and a larger area under the curve means a better test.
- Worked exampleA test for a disease with 10% prevalence has a sensitivity of 90% and a specificity of 90%. In 1,000 people: 100 have the disease, so TP = 90 and FN = 10; 900 do not, so TN = 810 and FP = 90. PPV = 90 / 180 = 50%. NPV = 810 / 820, about 99%. Half of the positive results are false, so a positive screen needs a confirmatory test.
- Where questions turnThe commonest error is reading the wrong direction. Sensitivity and specificity are read down the disease columns. Predictive values are read across the test rows. If the stem asks about a patient with a positive test, the answer is a predictive value.
2Likelihood ratios and post-test probability
- LR+ = sensitivity / (1 - specificity). LR- = (1 - sensitivity) / specificity. Likelihood ratios do not depend on prevalence.
- An LR+ above 10 or an LR- below 0.1 changes probability a lot. An LR near 1 barely changes it.
- Likelihood ratios work on odds. Odds = probability / (1 - probability). Probability = odds / (1 + odds).
- Post-test odds = pretest odds x LR. Then convert back to a probability.
- The pretest probability comes from the history, the examination and the local prevalence. Clinical decision rules (for example, Wells criteria for pulmonary embolism) give it a number.
- Order a test only if its result can move the probability across a threshold where you would act differently.
- Worked exampleA woman has a 30% pretest probability of pulmonary embolism. CT angiography is negative, with an LR- of 0.1. Pretest odds = 0.30 / 0.70, about 0.43. Post-test odds = 0.43 x 0.1 = 0.043. Probability = 0.043 / 1.043, about 4%. The result has moved her below the level at which most clinicians treat.
- Where questions turnDo not multiply a probability by a likelihood ratio. Convert to odds first, multiply, then convert back. For small probabilities the error is small, but at 20% or more it gives wrong answers.
3Measures of risk and treatment effect
- Relative risk (RR) = risk in exposed / risk in unexposed. It is used in cohort studies and trials. RR = 1 means no effect.
- Odds ratio (OR) = (a x d) / (b x c) from the 2 x 2 table. It is used in case-control studies. When the disease is rare, the OR is close to the RR.
- Absolute risk reduction (ARR) = control risk - treated risk. Relative risk reduction (RRR) = ARR / control risk = 1 - RR.
- Number needed to treat (NNT) = 1 / ARR. Number needed to harm (NNH) = 1 / absolute risk increase. Round both up to the next whole person, and state the time period.
- Attributable risk = risk in exposed - risk in unexposed. Attributable risk percent = (RR - 1) / RR. Population attributable risk tells how much disease in the whole population is due to the exposure.
- Relative measures sound larger than absolute ones. A drug that lowers risk from 2% to 1% has an RRR of 50% but an ARR of only 1% (NNT 100). Patients decide best with absolute numbers.
- Worked exampleIn a trial, stroke occurs in 5% with placebo and 3% with a drug over 3 years; bleeding occurs in 1% with placebo and 2% with the drug. ARR = 2%, so NNT = 50. Absolute risk increase for bleeding = 1%, so NNH = 100. For every 100 patients treated for 3 years, about 2 strokes are prevented and 1 extra bleed occurs.
- Where questions turnCheck the denominator before you divide. RRR divides by the control risk, not the treated risk. NNT uses the difference between groups, never the event rate in one group alone.
4Study designs and the strength of evidence
- Randomized controlled trials (RCTs) give the strongest single-study evidence of cause, because randomization balances known and unknown confounders.
- Cohort studies follow exposed and unexposed people over time and give incidence and RR. They are best for rare exposures and for studying many outcomes of one exposure.
- Case-control studies start from the outcome and look back at exposures, and give an OR. They are best for rare diseases and are fast and cheap.
- Cross-sectional studies measure exposure and disease at one time and give prevalence. They cannot show which came first.
- Ecological studies compare groups using averages. A link at the group level may not hold for individuals (ecological fallacy).
- Systematic reviews and meta-analyses pool studies. A funnel plot can show publication bias. Crossover trials, noninferiority trials and pragmatic trials are variants of the RCT.
- Worked exampleInvestigators want to know if a common antibiotic in pregnancy is linked to a rare limb defect. They choose 150 infants with the defect and 450 without, and ask mothers about past drug use. This is a case-control study, the best design for a rare outcome, and its measure is the odds ratio.
- Where questions turnDecide the design by the direction of the study, not by the words 'prospective' or 'retrospective'. If the study chooses people by outcome, it is case-control. If it chooses people by exposure and looks for outcomes, it is a cohort, even when the data are old records.
5Bias, confounding and effect modification
- Selection bias: the people studied differ from the target population (for example, volunteers, hospital patients or loss to follow-up). Randomization and complete follow-up reduce it.
- Information bias: exposure or outcome is measured wrongly. Recall bias (cases remember more), observer bias (assessors know the group) and the Hawthorne effect (people change when watched) are examples. Blinding reduces observer bias.
- Confounding: a third factor is linked to both exposure and outcome and is not on the causal path. Control it by randomization, restriction or matching in the design, and by stratification or multivariable analysis.
- Effect modification: the effect of an exposure truly differs across strata (for example, a drug works in women but not in men). It is a real finding, not a bias to remove.
- Stratify to tell them apart. If the stratum results are similar to each other but differ from the crude result, it is confounding. If the stratum results differ from each other, it is effect modification.
- Intention-to-treat analysis keeps every patient in the assigned group, which keeps the benefit of randomization. Per-protocol analysis can overstate benefit.
- Worked exampleCoffee drinkers seem to have more lung cancer (RR 2.0). In smokers alone the RR is 1.0, and in nonsmokers alone it is 1.0. The stratum results agree with each other and differ from the crude result, so smoking is a confounder. Coffee does not cause lung cancer here.
- Where questions turnDo not call every stratified result 'confounding'. When the strata disagree with each other, the answer is effect modification, and you should report the result separately for each group.
6Statistical inference: p values, confidence intervals and power
- The null hypothesis says there is no difference. The p value is the chance of a result this extreme or more if the null hypothesis is true. A p value below alpha (usually 0.05) is called significant.
- A 95% confidence interval (CI) that includes the null value (0 for a difference, 1 for a ratio) means the result is not significant at 0.05. A narrow CI means a precise estimate.
- Type I error (alpha): finding a difference that is not real. Type II error (beta): missing a real difference. Power = 1 - beta, usually set at 80% or more.
- Power rises with a larger sample, a larger true effect, less variation and a higher alpha.
- Statistical significance is not clinical importance. A large trial can find a tiny, unimportant effect. Look at the size of the effect and the whole CI.
- Many comparisons raise the chance of a false positive. Treat secondary and subgroup results as ideas for new studies, not as proof.
- Worked exampleA trial reports a hazard ratio for death of 0.92 (95% CI 0.80 to 1.06). The interval includes 1, so the result is not significant. It also allows anything from a 20% benefit to a 6% harm. The correct conclusion is that the trial did not show a difference, not that the treatments are equal.
- Where questions turnA nonsignificant result does not prove 'no effect', especially in a small trial. Say the study found no significant difference and ask whether it had enough power.
7Screening principles and screening biases
- Screening is worth doing when the disease is common and serious, has a detectable early stage, has an effective early treatment, and the test is safe, acceptable and affordable.
- Judge screening by disease-specific mortality in a randomized trial, not by survival from diagnosis.
- Lead-time bias: earlier diagnosis makes survival look longer without delaying death. Length-time bias: interval screening finds more slow-growing disease. Overdiagnosis: finding disease that would never have caused harm.
- Current USPSTF examples: colorectal cancer screening at 45 to 75 years; biennial mammography at 40 to 74 years; cervical screening at 21 to 65 years (cytology every 3 years, or from age 30 high-risk HPV testing every 5 years); annual low-dose CT at 50 to 80 years with 20 pack-years who smoke now or quit within 15 years.
- Also: one-time abdominal aortic aneurysm ultrasound for men aged 65 to 75 who ever smoked; osteoporosis screening for women 65 and older and younger postmenopausal women at risk; hepatitis C screening once for adults 18 to 79; HIV screening at 15 to 65; prediabetes and diabetes screening at 35 to 70 with overweight or obesity.
- Screen by the organs a person has: a transgender man with a cervix needs cervical screening; a transgender woman may need prostate cancer discussion. Older adults with limited life expectancy gain little from cancer screening.
- Worked exampleA 52-year-old veteran with no symptoms has never had colorectal screening. He qualifies under the USPSTF advice for ages 45 to 75. Options include colonoscopy every 10 years, an annual fecal immunochemical test, or a stool DNA test every 1 to 3 years. A positive stool test must be followed by colonoscopy.
- Where questions turnLonger survival in screened patients does not prove screening works. Look for a fall in deaths from the disease in a randomized trial; otherwise suspect lead-time bias, length-time bias or overdiagnosis.
8Outbreak investigation and population health
- Steps: confirm the diagnosis and that an outbreak exists (more cases than expected), define a case, find and count cases, describe them by time, place and person, form a hypothesis, test it, then control and report.
- Attack rate = ill / exposed. Compare attack rates in people who did and did not eat or do something; the largest relative risk points to the source.
- An epidemic curve with one sharp peak suggests a point source. Waves that rise over several incubation periods suggest person-to-person spread.
- Incubation time is a clue: 1 to 6 hours suggests a preformed toxin (Staphylococcus aureus, Bacillus cereus); 1 to 3 days suggests infection (Salmonella, norovirus).
- Incidence = new cases / population at risk over time. Prevalence is about incidence x duration. Report notifiable diseases to the local health department.
- Prevention levels: primary prevents disease (vaccines), secondary finds it early (screening), tertiary limits harm from existing disease (rehabilitation), quaternary avoids overtreatment.
- Worked exampleTwenty people at a wedding have vomiting 3 hours after the meal. Of 40 who ate fried rice, 18 are ill (45%); of 60 who did not, 2 are ill (about 3.3%). The relative risk is 45% / 3.3%, about 14. The short incubation and the food suggest Bacillus cereus toxin from rice kept warm after cooking.
- Where questions turnDo not pick the food with the most ill people. A food that almost everyone ate will have many ill people even if it is safe. Compare attack rates between eaters and non-eaters.
9Reading a clinical trial abstract
- Start with the question: who were the patients (PICO: patients, intervention, comparison, outcome)? Read the inclusion and exclusion criteria.
- Check internal validity: randomization, concealed allocation, blinding of patients and assessors, similar groups at baseline, complete follow-up and intention-to-treat analysis.
- Find the primary outcome and its size. Turn percents into ARR, RRR, NNT and NNH yourself; abstracts often report only relative numbers.
- Check the CI and p value for the primary outcome. Secondary and subgroup results are less reliable.
- Ask about external validity: would your patient have been enrolled? If your patient belongs to an excluded group (for example, dementia, pregnancy, kidney failure), the result may not apply.
- Weigh benefit against harm and cost with the patient, and look for a conflict of interest or industry sponsorship.
- Worked exampleAn abstract reports that a drug lowered hospitalization from 15% to 10% (RR 0.67; 95% CI 0.50 to 0.89) in adults under 75 without kidney disease. ARR = 5%, so NNT = 20. Your 80-year-old patient with stage 4 chronic kidney disease would have been excluded, so the result does not clearly apply to her.
- Where questions turnDo not stop at 'statistically significant'. A strong result in the wrong population is still the wrong evidence for your patient, and a relative risk reduction can hide a small absolute benefit.
Key terms
- Sensitivity
- The share of people with a disease who test positive.
- Specificity
- The share of people without a disease who test negative.
- Predictive value
- The chance that a test result (positive or negative) is correct for the person tested.
- Likelihood ratio
- How much a test result changes the odds of disease.
- Relative risk
- The risk in exposed or treated people divided by the risk in unexposed or control people.
- Odds ratio
- The odds of exposure in cases divided by the odds of exposure in controls.
- Number needed to treat
- How many patients must be treated for a set time to prevent one bad outcome.
- Confounder
- A third factor linked to both an exposure and an outcome that distorts the link between them.
- Confidence interval
- A range that likely contains the true value of an effect.
- Power
- The chance that a study finds a real difference when one exists.
- Lead-time bias
- Survival looks longer only because screening finds the disease earlier.
- Attack rate
- The share of exposed people who become ill in an outbreak.
- External validity
- How well a study's results apply to patients outside the study.
After this guide: After this guide, a learner should be able to calculate and explain test accuracy, risk and treatment-effect measures, judge a study's design, bias and statistics, apply current screening advice, and decide whether a trial abstract applies to the patient in front of them.
Practise Biostatistics & Epidemiology/Population Health/Interpretation of Medical LiteratureSources: The USMLE Content Outline and the Step 2 CK content specifications (usmle.org); the explanations are ours. Every source this path is built on.
