How to Read a Medical Paper in 10 Minutes: NNT & Absolute Risk Guide
Whether you are a clinician reviewing a newly published trial between clinic visits or a patient trying to understand a study linked in a support forum, reading a paper from start to finish is often the least effective approach. Introduction and Discussion sections are written to persuade; the Methods and baseline/outcome tables reveal what actually happened.
To appraise any medical paper in 10 minutes, ignore the abstract conclusion and check five numbers in the Results tables: (1) Study design (randomized trial vs observational cohort), (2) Primary endpoint type (hard mortality/morbidity vs surrogate biomarker), (3) Control vs experimental event rates (to compute Absolute Risk Reduction and NNT), (4) 95% confidence interval width, and (5) Attrition or cross-over rates.
What Is the Difference Between Relative Risk, Odds Ratio, and Number Needed to Treat (NNT)?
Medical journals report effects using four distinct metrics. Understanding how they diverge prevents misinterpreting modest effects as breakthroughs.
| Metric Name & Abbreviation | Mathematical Formula | Worked Value (CER 5.0% vs EER 3.5%) | How to Interpret at the Bedside |
|---|---|---|---|
| Relative Risk (RR) / Hazard Ratio (HR) | RR = EER / CER | 0.035 / 0.050 = 0.70 | The event rate in the treated group is 70% of the control rate over the trial period. |
| Relative Risk Reduction (RRR) | RRR = (CER - EER) / CER | (5.0% - 3.5%) / 5.0% = 30.0% | Headline figure ("cuts risk by 30%"). Completely blind to whether baseline risk was 50% or 0.05%. |
| Absolute Risk Reduction (ARR) | ARR = CER - EER | 5.0% - 3.5% = 1.5% (15 per 1,000) |
The true percentage-point reduction in probability for an individual patient entering the cohort. |
| Number Needed to Treat (NNT) | NNT = 1 / ARR = 100 / ARR(%) | 100 / 1.5 = 66.7 -> 67 | 67 patients must receive the therapy for the trial duration to prevent 1 adverse clinical event. |
| Number Needed to Harm (NNH) | NNH = 1 / ARI (Adverse Event Increase) | If serious harm rises from 1.0% to 3.0%: NNH = 100 / 2.0 = 50 |
When NNH (50) < NNT (67), treatment causes more serious adverse events than primary events prevented. |
Reference standard: BMJ (Trisha Greenhalgh, "How to Read a Paper" Series) · BMJ 2025 (CONSORT 2025 Statement).
Five Methodological Red Flags That Invalidate Headline Claims
Even in high-impact journals, structural design flaws can create artificial treatment benefits. Before acting on a paper, verify that none of these five biases explain the result:
1. Healthy-User & Socioeconomic Confounding in Observational Studies
When a nutritional or lifestyle cohort reports a Hazard Ratio of 0.78 (a 22% relative reduction) for coffee, resveratrol, or a specific vitamin, remember that individuals who purchase supplements also exercise more, smoke less, and have higher household incomes. In the GRADE framework, observational cohorts start at Low Certainty unless a large effect (RR > 2.0 or < 0.5) or strict dose-response gradient is proven.
2. Composite Endpoint Dilution
Trials frequently combine a hard outcome (cardiovascular death) with softer, more frequent clinician-driven events (hospitalization for unstable angina or repeat revascularization). Always look at the component breakdown table: if 80% of the composite events came from asymptomatic lab triggers or elective procedures while mortality remained flat, the headline composite HR overstates severity.
3. Per-Protocol vs. Intention-to-Treat (ITT) Switching
If 25% of patients in the experimental arm drop out due to drug toxicity and the authors only analyze "adherent completers" (Per-Protocol), they exclude the sickest and most intolerant patients, biasing the comparison. Always demand the Intention-to-Treat (ITT) primary analysis.
4. Post-Hoc Subgroup P-Hacking ("Spin")
When the overall trial primary endpoint fails (p = 0.24), authors sometimes slice the cohort into 15 subgroups until one subgroup (for example, "non-smoking women aged 55 to 64") shows p = 0.03. Without pre-specification and multiplicity adjustment, testing 20 subgroups guarantees a 64% probability of at least one false-positive p < 0.05 by chance alone.
5. Unrepresentative Control Arms
In oncology and cardiology, a new drug may look superior only because the control group received an outdated or suboptimal dose of standard-of-care therapy, or was denied cross-over upon progression.
Frequently Asked Methodological & Clinical Questions
Plain-language answers grounded in CONSORT, GRADE, WHO guidelines, and per-1,000 absolute outcome arithmetic.
What does a relative risk of 0.75 mean in a clinical study?
A relative risk (RR) of 0.75 means the experimental group experienced 75% as many events as the control group, which corresponds to a 25% Relative Risk Reduction (RRR). However, to know the actual benefit, you must check the baseline risk: if control risk was 4%, an RR of 0.75 reduces risk to 3% (an Absolute Risk Reduction of 1%, or NNT = 100).
Source: Institute for Work & Health (Absolute vs Relative Risk)
Why is an Odds Ratio (OR) larger than a Relative Risk (RR) when outcomes are common?
An Odds Ratio divides the odds of an event (p / [1 - p]) in two groups. When the baseline outcome is rare (under 5%), the Odds Ratio closely approximates Relative Risk. When the outcome is common (above 20%), the Odds Ratio exaggerates the true Relative Risk and should be converted to RR before calculating NNT.
Source: JAMA (Zhang & Yu, Correcting the Odds Ratio in Cohort Studies)
What is a narrow vs. wide 95% confidence interval?
A 95% confidence interval shows the range of plausible true effect sizes compatible with the data. A narrow interval (such as HR 0.74, 95% CI 0.68 to 0.81) indicates high statistical precision from a large sample size. A wide interval that approaches or crosses 1.0 (such as HR 0.74, 95% CI 0.49 to 1.08) indicates low precision and statistical non-significance.