VibeRounds This course is built in the spirit of VibeRounds — Socratic learning (AI that questions rather than answers) and Guided Discovery, part of the wider Clinical Cognition Operating System.
Evidence-Based Medicine · Course for Techies

Lesson 3: Critically Appraising a Randomized Controlled Trial

Part 3 — you found the paper in Lesson 2. Now: can you trust it, what does it actually say, and does it apply to your patient?

In Lesson 2, you turned a clinical worry into a PICO question and found a paper that matches it — ideally a randomized controlled trial (RCT), since that's the strongest design for therapy questions. Having the paper is not the same as trusting it. This lesson walks through the classic three-question framework used to critically appraise an RCT.

Why this matters for techies: think of appraisal as a code review. Finding the pull request isn't the job — checking whether it actually does what it claims, whether the tests are sound, and whether it's safe to merge into your specific production environment is the job. Same here: find the trial, then review it.

The Three-Question Framework

Every major appraisal checklist (CASP, CEBM, JAMA Users' Guides) boils down to three questions, asked in this order:

  1. Are the results valid? (Can I trust the methodology?)
  2. What are the results? (What did they actually find, in numbers?)
  3. Will the results help my patient? (Does this apply to the person in front of me?)

Skipping straight to question 3 — "does this help my patient?" — without first checking validity is the single most common mistake. A big effect size from a badly designed trial is worth less than a modest effect from a rigorous one.

Question 1: Are the Results Valid?

This is about ruling out bias. For an RCT, check these in order:

CheckWhat you're looking for
RandomizationWas allocation to groups truly random (computer-generated sequence), not something predictable like alternating patients or odd/even birthdates?
Allocation concealmentDid the person enrolling patients know in advance which group the next patient would land in? If yes, selection bias can creep in.
BlindingWere patients, clinicians, and outcome assessors unaware of group assignment? Single, double, or triple blind?
Similar groups at baselineWas there a "Table 1" showing the intervention and control groups were comparable at the start (age, severity, comorbidities)?
Complete follow-upWere dropouts few, and were they explained? A good rule of thumb: worry if losses exceed ~20%.
Intention-to-treat analysisWere patients analyzed in the group they were originally assigned to, even if they didn't complete treatment? This preserves the value of randomization.
Equal treatment otherwiseAside from the intervention being tested, were both groups treated the same (co-interventions, monitoring)?

If a trial fails badly on several of these, question 2 and 3 barely matter — the numbers may not reflect reality.

Question 2: What Are the Results?

Once you trust the method, extract the actual numbers. For a therapy trial with a binary outcome (e.g. "had a stroke" vs. "did not"), a few core statistics do most of the work:

Absolute Risk Reduction (ARR) = Risk in control group − Risk in treatment group

Relative Risk Reduction (RRR) = ARR ÷ Risk in control group

Number Needed to Treat (NNT) = 1 ÷ ARR

Worked Example

Suppose a trial finds: 10% of patients on placebo have a stroke over 5 years, vs. 6% of patients on the new drug.

Why this matters: "40% relative risk reduction!" sounds dramatic in a headline. "You need to treat 25 people to prevent one stroke, and the other 24 get the drug's side effects with no benefit" is the same result, stated honestly. Always ask for the absolute numbers behind a relative-risk claim.

For non-medical readers: relative risk reduction is like reporting a bug-fix rate as "cut crashes by 40%" without saying the app only crashed for 1 in 10 users to begin with. NNT (25, here) is the more honest metric — it's the "cost per unit of benefit," similar to reporting how many units you need to ship before one behaves as expected.

Also check the confidence interval around the main result. A 95% CI that's narrow and doesn't cross the "no difference" line (e.g. a relative risk CI that doesn't cross 1.0) suggests a precise, statistically solid finding. A wide CI that straddles "no effect" means the study may be underpowered — too small to be sure.

Question 3: Will This Help My Patient?

This is where the patient from Lesson 1 comes back into the picture. Ask:

This is also where your PICO question from Lesson 2 pays off: if you built it around your actual patient, this final step becomes a direct comparison rather than a guess.

Putting It Together: A Quick Checklist

Fast appraisal pass for an RCT:
  1. Randomized, concealed, blinded? → validity
  2. Groups similar at start, follow-up complete, ITT analysis? → validity
  3. ARR, RRR, NNT, confidence interval? → results
  4. Does the trial population and outcome match my patient? → applicability

Pro tip: the CASP RCT checklist and the CEBM RCT worksheet (linked in Lesson 1 and below) turn this into a fill-in-the-blanks form. Use one the first several times you appraise a trial — it becomes intuitive faster than you'd expect.

Homework for Lesson 3

  1. Take the paper you found in Lesson 2's homework. Run it through the three-question framework above.
  2. Calculate the ARR, RRR, and NNT (or NNH, Number Needed to Harm, if the outcome is a side effect) from the paper's own numbers.
  3. Write two or three sentences: would you recommend this treatment to your patient from Lesson 1, and why or why not?

Resources for This Lesson

CASP UK — RCT Checklist CEBM Oxford — Critical Appraisal Worksheets CEBM — NNT / ARR / RRR calculators PubMed — search medical research papers