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 8: Statistics Deep-Dive for the Skeptical Techie

Part 8 — every earlier lesson leaned on p-values and confidence intervals. Time to open that black box properly.

You've been using p < 0.05 and confidence intervals since Lesson 3 as a kind of pass/fail signal. That's enough to get through most appraisals, but it also invites a specific kind of overconfidence — treating "statistically significant" as a stamp of truth rather than what it actually is: a statement about how surprising the data would be under a specific, narrow assumption. This lesson takes the black box apart.

Why this matters for techies: a p-value is closer to a p-value from hypothesis testing in a controlled experiment than to a confidence score from a trained model — it doesn't tell you the probability your hypothesis is true, only how unusual your data would look if a specific null assumption held. Conflating the two is one of the most common and consequential misreadings in the applied sciences, and it's an easy trap to fall into if your intuitions come from ML metrics that actually do report something like a calibrated probability.

What a P-Value Actually Answers

Here is the precise definition, worth sitting with because almost every common misreading comes from skipping a clause in it:

P-value = the probability of seeing a result at least as extreme as the one observed, if the null hypothesis were true (i.e., if there were truly no effect).

Notice what it does not say. It is not the probability that the null hypothesis is true. It is not the probability that your finding is a fluke. It says nothing at all about the size or clinical importance of an effect — a trivial, meaningless difference can produce a tiny p-value if the sample is large enough, and a large, important-looking difference can produce a big p-value if the sample is small.

Common misreadingWhat a p-value of 0.03 actually means
"There's a 97% chance the treatment works"If the treatment had truly no effect, you'd see a result this extreme (or more extreme) 3% of the time by chance alone
"There's a 3% chance this is a false positive"Says nothing about that probability directly — that depends on the prior probability the effect is real, which the p-value doesn't know about (this is exactly the PPV-vs-prevalence trap from Lesson 5, transposed onto statistics)
"The effect is real and important"Says nothing about effect size — only about how compatible the data are with "no effect"

That middle row deserves a callback: it's structurally the same mistake as confusing sensitivity with PPV in Lesson 5. A p-value is like sensitivity — a property of the test procedure under an assumption — while "is this effect actually real" is like PPV — it depends on how plausible the effect was to begin with, information the p-value alone can't supply.

For non-medical readers: a p-value is not "probability the drug works." It only answers: if the drug truly did nothing, how surprising would data this extreme be? That's closer to a null-hypothesis significance test in an A/B testing framework than to a confidence score on the actual hypothesis you care about.

Confidence Intervals: More Information, Same Underlying Machinery

A confidence interval is built from the same statistical logic as a p-value, but it reports something more clinically useful: a range of effect sizes compatible with the data, not just a single accept/reject verdict.

A 95% confidence interval means: if you repeated this exact study many times, 95% of the intervals calculated this way would contain the true effect. It does not mean there's a 95% chance the true effect lies in this specific interval — the true effect either is or isn't in it; the 95% describes the reliability of the method across repetition, not a probability about this one result.

Reading Three Confidence Intervals at a Glance

no effect (RR = 1.0) Study 1 narrow, precise, significant Study 2 wide, underpowered Study 3 narrow, genuinely null

Study 1's interval is narrow and sits entirely away from "no effect" — a precise, statistically solid finding. Study 2's interval is wide and straddles "no effect" — this is very likely underpowered, and "not significant" here means "we couldn't tell," not "there's no effect." Study 3's interval is narrow and centered right on "no effect" — this is a study with enough precision to say, with some confidence, that any true effect is probably small. Width matters as much as position: "not significant" and "no effect" are not the same claim, and only a narrow interval straddling the line lets you say the second one with any confidence.

Statistical Significance vs. Clinical Significance

This distinction has come up implicitly since Lesson 3's NNT discussion, and it's worth making explicit: a result can be statistically significant and clinically trivial, or clinically important and statistically non-significant, and these are genuinely different failure modes.

Worked Example

A blood-pressure drug trial with 50,000 participants finds a mean systolic pressure reduction of 1.2 mmHg, p < 0.001. The huge sample makes even a clinically negligible difference statistically detectable. Compare that to a trial of 40 participants finding a 15 mmHg reduction with a p-value of 0.08 — a much larger, potentially meaningful effect that the small sample simply wasn't powered to confirm with confidence.

Why this matters: "statistically significant" answers "is this probably not zero?" It says nothing about "is this big enough to matter to a patient?" That second question needs the effect size itself — ARR, NNT, mean difference — interpreted against what actually changes a patient's health or experience, exactly the kind of number Lesson 3 taught you to demand.

Why 0.05 Is a Convention, Not a Law of Nature

The threshold of p < 0.05 traces back to a somewhat arbitrary suggestion by the statistician Ronald Fisher in the 1920s, and it has since calcified into something treated as a bright line between "real" and "not real." Two consequences follow directly from taking that convention too literally:

P-Hacking and Multiple Comparisons

That second consequence is exploitable, deliberately or not, and it has a name: p-hacking — analyzing data in enough different ways, or testing enough different subgroups and outcomes, until something clears the significance threshold, then reporting only that result as if it were the pre-specified question.

Familiar territory: this is the same failure mode as tuning hyperparameters against your test set until performance looks good, then reporting that number as if it were a fair held-out evaluation. The fix is the same idea in both fields — decide your primary analysis in advance and hold it fixed, rather than letting the data pick the story after the fact.

Common forms worth watching for when appraising a paper:

The single best defense as a reader is checking whether the trial was pre-registered (on a registry like ClinicalTrials.gov) with a stated primary outcome and analysis plan, and then comparing that registration against what actually got published. A mismatch between the two is one of the clearest tells in the whole appraisal toolkit.

Surrogate Endpoints: A Different Kind of Trap

Separate from p-value games, there's a design-level issue worth flagging on its own: many trials measure a surrogate endpoint — a stand-in outcome that's faster or cheaper to measure than the outcome patients actually care about — instead of a patient-important outcome.

Surrogate endpointPatient-important outcome it's standing in for
LDL cholesterol levelHeart attack, stroke, death
Tumor shrinkage on imagingSurvival, quality of life
Blood glucose (HbA1c)Diabetic complications, death
Bone mineral densityFracture

Surrogates are attractive because trials using them can be smaller and shorter — you don't have to wait years to count strokes if you can measure cholesterol in weeks. The trap is that a drug can reliably move the surrogate without moving the outcome that matters, or can even move the surrogate in the "right" direction while causing harm overall. A surrogate is only trustworthy as a stand-in when it's been well validated as closely tracking the real outcome across many prior trials — and even then, treat a trial that reports only a surrogate as an incomplete answer to a therapy question, not a substitute for one.

Reading a Limitations Section Like It Matters

Most papers include a limitations paragraph, and most readers skim past it. It's often the most honest part of the paper — look for these specifically, and be equally suspicious of a limitations section that's suspiciously thin:

A Skeptical Reading Checklist

  1. Was the primary outcome pre-specified, and does the published outcome match it?
  2. Is the confidence interval reported, not just the p-value — and is it narrow or wide?
  3. Is the effect size clinically meaningful, independent of whether it's statistically significant?
  4. Is the outcome a surrogate, or something the patient would actually notice?
  5. Does the limitations section undercut the abstract's headline claim?

Homework for Lesson 8

  1. Return to the RCT you appraised in Lesson 3. Find its confidence interval for the primary outcome. Is it narrow or wide? Does it cross the line of no effect? Restate what the p-value does and does not tell you about that trial in your own words.
  2. Check whether that trial was pre-registered, and whether the published primary outcome matches the registration.
  3. Identify whether the trial's primary outcome is a surrogate or a patient-important outcome. If it's a surrogate, note whether the paper (or other evidence) establishes how well that surrogate tracks the real outcome.
  4. Read the paper's limitations section in full. Write two or three sentences on whether it changes how confidently you'd act on the abstract's conclusion.

Resources for This Lesson

ClinicalTrials.gov — trial pre-registration EQUATOR Network — reporting guidelines (CONSORT and others) CEBM — statistics calculators PubMed — search medical research papers