USMLE Step 1 biostatistics: sensitivity, NNT and the calculations to know, worked through
Understand the 2×2 table, risk measures, p-values, confidence intervals and study design bias, with full working for every calculation.
On this page
USMLE Step 1 biostatistics comes down to a short list of calculations: sensitivity, specificity, PPV and NPV from a 2×2 table, and relative risk, odds ratio, absolute risk reduction (ARR), NNT and NNH from trial or cohort data. Add p-values, confidence intervals and the main study designs and biases, and you have the core of what this section asks. Biostatistics & Epidemiology/Population Health makes up 4 to 6% of Step 1, according to the USMLE content outline. Below, each calculation is worked through with real numbers, followed by five Medly practice questions with full answers.
Which biostatistics calculations does USMLE Step 1 test?
Step 1 tests whether you can pull the right numbers from a stem and divide by the right group. These are the formulas worth knowing by heart.
| Measure | Formula | Starts from |
|---|---|---|
| Sensitivity | TP ÷ (TP + FN) | Everyone with disease |
| Specificity | TN ÷ (TN + FP) | Everyone without disease |
| PPV | TP ÷ (TP + FP) | Everyone testing positive |
| NPV | TN ÷ (TN + FN) | Everyone testing negative |
| Relative risk | Risk exposed ÷ risk unexposed | Cohorts and trials |
| Odds ratio | (a × d) ÷ (b × c), where a, b = exposed with and without the outcome; c, d = unexposed | Case-control studies |
| ARR | Control risk − treatment risk | Trials |
| NNT | 1 ÷ ARR | Trials |
| NNH | 1 ÷ ARI | Trials and cohorts |
Source: standard epidemiology definitions; content share from the USMLE content outline. Last checked: 17 November 2026.
The wrong options in a Step 1 question are often real numbers from the stem, divided by the wrong group. You'll see that in every practice question further down. If you'd like these on one page, download the formula sheet to print.
The 2×2 table: sensitivity, specificity, PPV and NPV
All four test measures come from the same table. Medly's lesson on sensitivity and specificity uses this hypothetical evaluation of 1,000 people, with disease status confirmed by a reference standard.
| Test result | Disease present | Disease absent | Total |
|---|---|---|---|
| Positive | 180 (TP) | 40 (FP) | 220 |
| Negative | 20 (FN) | 760 (TN) | 780 |
| Total | 200 | 800 | 1,000 |
- Sensitivity = 180 ÷ (180 + 20) = 180 ÷ 200 = 0.90, so 90%.
- Specificity = 760 ÷ (760 + 40) = 760 ÷ 800 = 0.95, so 95%.
- PPV = 180 ÷ (180 + 40) = 180 ÷ 220 = 0.818, so 81.8%.
- NPV = 760 ÷ (760 + 20) = 760 ÷ 780 = 0.974, so 97.4%.
Sensitivity and specificity read down the columns: they start from people whose disease status is known. PPV and NPV read across the rows: they start from people with a particular result. That direction is the whole difference. A sensitivity of 90% tells you the test catches 90% of people with disease. It says nothing directly about how likely a patient with a positive result is to have it. That is PPV.
Two shortcuts follow. A highly sensitive test gives few false negatives, so a negative result helps rule disease out (SnNout). A highly specific test gives few false positives, so a positive result helps rule disease in (SpPin).
How prevalence changes PPV and NPV
Sensitivity and specificity belong to the test. Predictive values also depend on who you test. Hold a test at 90% sensitivity and 90% specificity, and change only the prevalence in a group of 1,000:
| Prevalence | PPV | NPV |
|---|---|---|
| 1% | 8.3% | 99.9% |
| 10% | 50.0% | 98.8% |
| 50% | 90.0% | 90.0% |
Source: illustrative calculation from Medly's lesson on predictive values. Last checked: 17 November 2026.
At 1% prevalence, 10 people have the disease and 990 don't. The test finds 9 of the 10, but 10% of the 990 healthy people also test positive: 99 false positives. So PPV = 9 ÷ (9 + 99) = 8.3%. Higher prevalence raises PPV and lowers NPV. Lower prevalence does the opposite.
Risk: relative risk, odds ratio, ARR, NNT and NNH
These come from trial or cohort data. Medly's lesson on treatment impact uses a one-year trial of 2,000 people, all free of a complication at the start.
| Group | Complication | No complication | Total |
|---|---|---|---|
| Treatment | 80 | 920 | 1,000 |
| Control | 120 | 880 | 1,000 |
- Treatment risk = 80 ÷ 1,000 = 0.08 (8%). Control risk = 120 ÷ 1,000 = 0.12 (12%).
- Relative risk (RR) = 0.08 ÷ 0.12 = 0.67.
- Relative risk reduction (RRR) = 1 − 0.67 = 0.33, so treatment cuts the risk by a third.
- ARR = 0.12 − 0.08 = 0.04, which is 4 percentage points, or 40 fewer complications per 1,000 treated.
- NNT = 1 ÷ 0.04 = 25. Treat 25 people for one year to prevent one extra complication.
- Odds ratio (OR) = (80 × 880) ÷ (120 × 920) = 70,400 ÷ 110,400 = 0.64.
Now suppose the same drug causes an adverse effect in 25 of 1,000 treated people and 10 of 1,000 controls. The absolute risk increase (ARI) = 0.025 − 0.010 = 0.015. NNH = 1 ÷ 0.015 = 66.7, which you round up to 67.
A few traps catch people here:
- Use proportions in the reciprocal. 1 ÷ 0.04 gives 25; 1 ÷ 4 gives nonsense. With percentage points, use 100 ÷ 4 instead.
- Round NNT and NNH up to a whole person.
- Baseline risk matters. A drop from 3% to 2% is the same one-third RRR, but ARR is only 0.01, so NNT = 100.
- The OR is close to the RR only when the outcome is rare. Here the outcome affects 8 to 12% of people, so 0.64 and 0.67 already differ.
- Case-control studies report an OR because the investigator chose how many cases to include, so risk can't be calculated.
P-values and confidence intervals: what they do and don't mean
For the trial above, the standard large-sample method gives an RR of 0.67 with a 95% confidence interval of 0.51 to 0.87, and p ≈ 0.003. The ARR of 4 percentage points has a 95% CI of 1.4 to 6.6 points.
Here's how to read those numbers:
- The p-value assumes there is no real difference (the null hypothesis). It is the probability of seeing a result at least this extreme if that were true. A p of 0.003 means that, if there were no real difference, a result this extreme would turn up about 3 times in 1,000 studies.
- A p-value is not the probability that the null hypothesis is true, and it doesn't measure how large or important an effect is. The American Statistical Association set this out in its 2016 statement on p-values.
- A 95% CI means that if the study were repeated many times, about 95% of intervals calculated this way would contain the true value.
- For a ratio (RR, OR, hazard ratio), the null value is 1. A 95% CI that excludes 1 matches p < 0.05. The interval 0.51 to 0.87 excludes 1, so the result is significant.
- For a difference (ARR, difference in means), the null value is 0. The interval 1.4 to 6.6 excludes 0, which agrees.
- A wider CI means less precision, usually from a smaller sample. A larger study narrows it.
- Statistical significance and clinical importance are separate questions. A huge trial can make a tiny ARR significant.
Errors complete the picture. A type I error (α) is finding a difference that isn't real, usually set at 0.05. A type II error (β) is missing a difference that is real. Power = 1 − β, and it rises with a larger sample, a larger true effect or a higher α.
Study designs and bias in one table
Step 1 often describes a study in a few lines and asks what it can show, or what went wrong.
| Design | Main measure | Watch for |
|---|---|---|
| Cross-sectional | Prevalence | Can't show which came first |
| Case-control | Odds ratio | Recall and selection bias |
| Cohort | Relative risk | Loss to follow-up |
| Randomised trial | RR, ARR, NNT | Poor blinding, crossover |
| Screening data | Survival time | Lead-time, length-time bias |
Source: standard study design definitions, as taught in Medly's USMLE Step 1 course. Last checked: 17 November 2026.
The biases that come up most are worth one line each:
- Confounding: a third factor linked to both exposure and outcome. Randomisation, matching and stratification reduce it.
- Selection bias: the people studied differ from the population you want to describe.
- Recall bias: cases remember past exposures more readily than controls.
- Lead-time bias: screening finds disease earlier, so survival looks longer even if death comes at the same time.
- Length-time bias: screening picks up slow-growing disease more often, which flatters outcomes.
- Observer bias and the Hawthorne effect: knowing who is treated, or being watched, changes the results. Blinding helps.
USMLE biostatistics practice questions
These are Medly practice questions, numbered 1 to 5 and each worth 1 mark. Try them before opening the answer.
Question 1: (1 mark)
In a randomized trial, 600 patients with prior transient ischemic attacks and no history of stroke receive either a new preventive treatment or a control treatment. During 2 years of follow-up, stroke occurs in 18 of 300 patients in the new treatment group and 45 of 300 patients in the control group. All patients complete follow-up. Which of the following is the absolute risk reduction for stroke with the new treatment over 2 years?
A. 6%
B. 9%
C. 15%
D. 21%
E. 60%
Answer
B. 9% earns the mark.
- Control risk = 45 ÷ 300 = 0.15, so 15%.
- Treatment risk = 18 ÷ 300 = 0.06, so 6%.
- ARR = 15% − 6% = 9 percentage points.
A is the treatment group's risk and C is the control group's risk. D adds the two risks. E is the relative risk reduction: 9 ÷ 15 = 60%. If you were then asked for the NNT, it would be 1 ÷ 0.09 = 11.1, rounded up to 12 over 2 years.
Question 2: (1 mark)
A cross-sectional study evaluates a novel serum matrix metalloproteinase-7 (MMP-7) assay for idiopathic pulmonary fibrosis. MMP-7 is an enzyme that breaks down proteins in the extracellular matrix. The study includes 500 consecutive adults with interstitial lung findings. Surgical lung biopsy confirms idiopathic pulmonary fibrosis in 100 participants and excludes it in 400. The serum assay is positive in 145 participants, including 85 with biopsy-confirmed disease. Which of the following is the false-positive rate of the serum assay in this study?
A. 4.2%
B. 12%
C. 15%
D. 41%
E. 85%
Answer
C. 15% earns the mark.
- TP = 85. FP = 145 − 85 = 60.
- FN = 100 − 85 = 15. TN = 400 − 60 = 340.
- False-positive rate = FP ÷ everyone without disease = 60 ÷ 400 = 0.15, so 15%.
A is 15 ÷ 355, the share of negative results that are false negatives. B divides by all 500 participants. D is 60 ÷ 145, the share of positive results that are false. E is the specificity (340 ÷ 400), which also equals the sensitivity here (85 ÷ 100).
Question 3: (1 mark)
In a hypothetical randomized, double-blind trial, investigators compare two novel antiplatelet regimens with placebo in patients at high risk of a first ischemic stroke. All participants complete 5 years of follow-up. The table shows the number of participants who experience a first nonfatal ischemic stroke in each group.
| Group | At risk | With stroke |
|---|---|---|
| Placebo | 1,000 | 120 |
| Regimen X | 1,000 | 40 |
| Regimen Y | 1,000 | 70 |
Which of the following is the absolute risk percent for regimen X, the absolute risk reduction for regimen X compared with placebo, and the regimen providing greater benefit based on the number needed to treat (NNT), respectively?
A. 0.04%; 0.08%; X (smaller NNT)
B. 12%; 4%; Y (lower NNT)
C. 4%; 66.7%; X (smaller NNT)
D. 4%; 8%; X (smaller NNT)
E. 4%; 8%; Y (larger NNT)
Answer
D earns the mark.
- Risk with regimen X = 40 ÷ 1,000 = 0.04, so 4%.
- Placebo risk = 120 ÷ 1,000 = 12%. ARR for X = 12% − 4% = 8 percentage points.
- NNT for X = 1 ÷ 0.08 = 12.5, rounded up to 13.
- ARR for Y = 12% − 7% = 5 points, so NNT for Y = 1 ÷ 0.05 = 20.
- X has the smaller NNT, so it gives greater benefit.
A forgets to multiply the decimals by 100. B uses the placebo risk and picks the wrong regimen. C gives the relative risk reduction (8 ÷ 12 = 66.7%) instead of the ARR. E has the right numbers but reads a larger NNT as more benefit, when it means less.
Question 4: (1 mark)
Investigators evaluate a hypothetical rapid salivary antigen assay for Coccidioides immitis infection using bronchoalveolar lavage culture as the reference standard. The assay has a sensitivity of 90% and a specificity of 90% in both populations studied. In a general walk-in clinic, the prevalence of infection is 10% and the positive predictive value is 50%. The assay is then used to screen agricultural excavation workers, among whom the prevalence of infection is 50%. No other information about an individual worker is available. Which of the following gives the pretest probability of infection and the probability of infection given a positive result, respectively, for a worker in this cohort?
A. 10%, 50%
B. 50%, 50%
C. 50%, 81%
D. 50%, 90%
E. 90%, 90%
Answer
D earns the mark.
- With no other information, pretest probability = cohort prevalence = 50%.
- Take 1,000 workers: 500 infected and 500 not.
- True positives = 90% of 500 = 450. False positives = 10% of 500 = 50.
- PPV = 450 ÷ (450 + 50) = 450 ÷ 500 = 90%.
A gives the clinic's figures. B forgets that PPV rises with prevalence. C multiplies sensitivity by specificity, which means nothing here. E confuses pretest probability with sensitivity.
Question 5: (1 mark)
Investigators evaluate a new rapid serologic assay for systemic lupus erythematosus in a primary care screening population. The prevalence of systemic lupus erythematosus is 0.5% in this population. The assay is then applied to a tertiary rheumatology referral cohort, where the disease prevalence is 25%. In both populations, the assay has a sensitivity of 92% and a specificity of 95%. Disease status in each population is established using the same reference standard. Which of the following best describes the assay's predictive values in the referral cohort compared with the screening population?
A. Both predictive values are higher.
B. Sensitivity increases and specificity decreases, with unchanged predictive values.
C. Both predictive values are unchanged.
D. Negative predictive value increases as positive predictive value decreases.
E. Positive predictive value rises and negative predictive value falls.
Answer
E earns the mark.
- Sensitivity and specificity are fixed, so only prevalence changes.
- Screening, per 100,000 people: 500 with lupus give 460 TP and 40 FN; 99,500 without give 4,975 FP and 94,525 TN. PPV = 460 ÷ 5,435 = 8.5%. NPV = 94,525 ÷ 94,565 = 99.96%.
- Referral, per 100,000 people: 25,000 with lupus give 23,000 TP and 2,000 FN; 75,000 without give 3,750 FP and 71,250 TN. PPV = 23,000 ÷ 26,750 = 86.0%. NPV = 71,250 ÷ 73,250 = 97.3%.
- PPV rises and NPV falls.
A misses that NPV falls. B and C treat predictive values as fixed properties of the test. D has both directions reversed.
To practise more questions like these, Medly's USMLE Step 1 course has questions on each of these topics, with an explanation for every option, so you can see exactly which denominator a wrong answer used. If you're fitting biostatistics into your revision, our study plan shows where it can go in a 6 or 12 week schedule.
Frequently asked questions
How much of USMLE Step 1 is biostatistics?
Biostatistics & Epidemiology/Population Health makes up 4 to 6% of Step 1 content, according to the USMLE content outline. Step 1 is pass/fail, and our guide to the passing score explains what that means for you.
Does prevalence change sensitivity and specificity?
No. Sensitivity and specificity describe the test, so they stay the same when prevalence changes. PPV rises and NPV falls as prevalence goes up.
What do SnNout and SpPin mean?
SnNout: a negative result on a highly sensitive test helps rule disease out. SpPin: a positive result on a highly specific test helps rule disease in. Neither gives certainty on its own.
When should you use an odds ratio instead of relative risk?
Use an odds ratio for case-control studies, where you start from cases and controls and can't calculate risk. Use relative risk for cohort studies and trials. The two are close only when the outcome is rare.
What are likelihood ratios?
LR+ = sensitivity ÷ (1 − specificity), and LR− = (1 − sensitivity) ÷ specificity. For the test above (90% sensitivity, 95% specificity), LR+ = 0.90 ÷ 0.05 = 18 and LR− = 0.10 ÷ 0.95 = 0.11. An LR+ above 10 or an LR− below 0.1 shifts the probability of disease a lot.
Should you round NNT up or down?
Up. An NNT of 12.5 becomes 13, and an NNH of 66.7 becomes 67, because you can't treat part of a person.
What does a 95% confidence interval that crosses 1 mean?
For a ratio such as RR or OR, a 95% CI that includes 1 means the result is not statistically significant at p < 0.05. For a difference such as ARR, the value to check is 0.