How to Read a Clinical Study Fast: NNT, Absolute Risk and Bias for Medical Students

"Reduces risk by 25%." You'll find that line in abstracts, drug adverts and press releases, and on its own it tells you almost nothing. This guide gives you a repeatable way to get through a clinical paper in about fifteen minutes and come out with answers to four questions. What was compared? How big is the effect in absolute terms? How certain is it? How much could bias have distorted it?

It's written for medical students preparing for journal club, a research project or written finals. It's also meant for junior doctors who need to place a new trial in context before the ward round.

Step 1: Pin down the question (PICO)

Before you look at a single number, rewrite the study question in PICO form:

• Population: who was enrolled? Age, comorbidities, setting (hospital or primary care)?
• Intervention: what exactly was given, at what dose, and for how long?
• Comparator: placebo, usual care or an active drug?
• Outcome: which endpoint, and measured at what time point?

This step alone often settles whether the trial applies to your patient. A trial of 55-year-olds with normal kidney function says little about an 85-year-old with a reduced eGFR. If you need a quick estimate, there's our eGFR calculator (CKD-EPI).

Step 2: Separate relative from absolute effects

The most useful skill in critical appraisal is converting relative effects into absolute ones. You need two numbers: the event rate in the control group (CER) and the event rate in the experimental group (EER).

Worked example

Here's a hypothetical trial that compares a new drug with placebo over three years, with stroke as the outcome.

• Control: 10 of 100 people have a stroke, so CER = 10%.
• Treatment: 8 of 100, so EER = 8%.

So:

• ARR = 10% − 8% = 2 percentage points
• RR = 8 ÷ 10 = 0.8
• RRR = 2 ÷ 10 = 20%
• NNT = 1 ÷ 0.02 = 50 (over three years)

The headline says "20% fewer strokes", which is accurate. It's just as accurate to say that 50 people must take the drug for three years to prevent one stroke. The other 49 get no benefit and still carry the side-effect risk.

Baseline risk changes everything

Now apply the same 20% RRR to a lower-risk group whose three-year stroke risk is 1%. Risk falls to 0.8%, the ARR is 0.2 percentage points and the NNT rises to 500. The relative effect is the same, but the absolute benefit is ten times smaller.

That's why guidelines tie treatment thresholds to risk scores. Whether a therapy is worthwhile depends on the patient's baseline risk. In atrial fibrillation, for example, the CHA₂DS₂-VA score plays that role.

Number needed to harm

Harms work the same way. The number needed to harm (NNH) is 1 divided by the absolute risk increase. Always put NNT and NNH side by side for the same time frame. An NNT of 50 to prevent a stroke and an NNH of 100 for a major bleed only make sense together, weighed against how much each outcome matters to the patient.

Step 3: Lead with the confidence interval, not the p-value

The p-value is probably the most misread number in medicine. In 2016 the American Statistical Association (ASA) set out six principles about it. The essentials:

• A p-value indicates how incompatible the data are with a specified statistical model, usually the null hypothesis of no effect.
• It does not measure the probability that the hypothesis is true, or that the result arose by chance alone.
• It says nothing about the size or clinical importance of an effect.
• Conclusions shouldn't rest only on whether p crosses a threshold such as 0.05.

The 95% confidence interval (CI) is more informative. It shows the range of effect sizes reasonably compatible with the data. Read it in three moves:

1. Does it include "no effect"? That's 1 for RR, odds ratio and hazard ratio, and 0 for differences.
2. How wide is it? RR 0.80 (95% CI 0.78–0.82) is precise. RR 0.80 (95% CI 0.40–1.60) is compatible with large benefit and with harm.
3. Is the less favourable end still clinically worthwhile? If it is, the benefit is robust.

NNTs need confidence intervals too. When the ARR isn't statistically significant, the NNT's interval passes through "infinity" and into harm, so an NNT quoted without its CI is only half the story.

Step 4: Check the endpoint – patient-important or surrogate?

A surrogate endpoint is a measurement that stands in for a clinical event, such as a lab value, blood pressure or the number of ectopic beats on a Holter. Surrogates are attractive because they need fewer patients and less time. They can also mislead badly.

The textbook case is the Cardiac Arrhythmia Suppression Trial (CAST). It tested whether suppressing ventricular ectopy after myocardial infarction would reduce sudden death. Only patients whose ectopy could be suppressed by the drug were randomised. Even so, the encainide and flecainide arms were stopped because of excess mortality: 43 arrhythmic deaths on active drug versus 16 on placebo. The surrogate improved, but the outcome patients actually care about got worse.

So always ask:

• Is the primary endpoint patient-important (death, stroke, hospitalisation, quality of life)?
• For composite endpoints, which component drives the result? If the hard component (death) is unchanged and only a softer one (repeat revascularisation) moves, the benefit is smaller than the headline suggests.
• Was the primary endpoint pre-registered, or has a favourable secondary outcome been promoted?

Step 5: Assess bias systematically (RoB 2)

For randomised trials, Cochrane's RoB 2 tool looks at five domains where bias can creep in:

The overall judgement has three levels. Low risk means low risk in every domain. Some concerns means at least one domain raises concerns but none is high risk. High risk means at least one domain is high risk, or concerns across several domains substantially lower confidence in the result.

For a quick read, a short version is enough. Is randomisation with concealed allocation described? Who was blinded? How many were lost to follow-up? Does the paper match its registry entry (ClinicalTrials.gov or the EU trials register)?

Step 6: Use reporting guidelines as a checklist

Good papers follow reporting guidelines, collected by the EQUATOR Network. As a reader, you can use them as checklists:

• CONSORT 2025 for randomised trials: a 30-item checklist plus a flow diagram of participants through the trial. The update adds, among other things, a section on open science.
• PRISMA 2020 for systematic reviews: a 27-item checklist, a separate abstract checklist and revised flow diagrams.

A trial without a CONSORT flow diagram, or a meta-analysis whose search strategy can't be reproduced, should make you more cautious.

Step 7: Rate the certainty of the body of evidence (GRADE)

One trial is rarely the final word. Guidelines and Cochrane reviews grade the overall evidence with GRADE, which has four levels of certainty: high, moderate, low and very low.

• Randomised trials start at high, observational studies at low.
• Evidence is rated down for risk of bias, inconsistency, indirectness, imprecision and publication bias.
• Evidence (mostly observational) is rated up for a large effect, a dose–response gradient, or when plausible confounding would have shrunk the observed effect.

When an ESC or NICE recommendation says the evidence is low certainty, it means the true effect may be substantially different from the estimate.

The fifteen-minute checklist

1. Abstract and question: write down PICO.
2. Methods: design, randomisation, blinding, registration.
3. Table 1: are groups comparable, and who are these patients really?
4. Primary endpoint: patient-important or surrogate? Composite?
5. Results: find CER and EER, calculate ARR and NNT yourself, read the CI.
6. Harms: NNH over the same period.
7. Bias: run through the five RoB 2 domains.
8. Context: what do the guideline and the latest systematic review say?

Where to practise

• Cochrane Library: systematic reviews with plain-language summaries.
• Cochrane Handbook: chapters 8 (risk of bias) and 14 (GRADE) are free and readable.
• EQUATOR Network: every major reporting guideline in one place.
• European and national guidelines (ESC, NICE, AWMF): see how evidence ratings become recommendations.

Bottom line

Ask for absolute numbers first and p-values last. An RRR without the baseline risk is incomplete, and so is an NNT without a time frame and confidence interval, or a surrogate result without a clinical outcome. PICO, ARR and NNT, confidence intervals, RoB 2 and GRADE work just as well in journal club, in exams and on the ward.

Medelix answers clinical questions with citations to the literature and guidelines. Open the original papers it links and run them through these same steps.

Sources

1. Echt DS, Liebson PR, Mitchell LB, et al. Mortality and morbidity in patients receiving encainide, flecainide, or placebo. The Cardiac Arrhythmia Suppression Trial. N Engl J Med. 1991;324:781–788. doi:10.1056/NEJM199103213241201
2. Wasserstein RL, Lazar NA. The ASA Statement on p-Values: Context, Process, and Purpose. Am Stat. 2016;70:129–133. doi:10.1080/00031305.2016.1154108; press release: magazine.amstat.org
3. Sterne JAC, Savović J, Page MJ, et al. RoB 2: a revised tool for assessing risk of bias in randomised trials. BMJ. 2019;366:l4898. doi:10.1136/bmj.l4898
4. Cochrane Handbook for Systematic Reviews of Interventions, chapter 8: Assessing risk of bias in a randomized trial. cochrane.org
5. Cochrane Handbook, chapter 14: Completing 'Summary of findings' tables and grading the certainty of the evidence. cochrane.org
6. Hopewell S, Chan AW, Collins GS, et al. CONSORT 2025 Statement: Updated Guideline for Reporting Randomized Trials. JAMA. 2025. doi:10.1001/jama.2025.4347
7. Page MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ. 2021;372:n71. doi:10.1136/bmj.n71
8. Guyatt GH, Oxman AD, Vist GE, et al. GRADE: an emerging consensus on rating quality of evidence and strength of recommendations. BMJ. 2008;336:924–926. doi:10.1136/bmj.39489.470347.AD
9. EQUATOR Network – reporting guidelines. equator-network.org
10. CONSORT and SPIRIT statements. consort-spirit.org