A clearer way to read research.

Start with five questions

A research result becomes easier to read when you ask:

1. Who? Which patients and clinical setting?
2. Compared with what? A treatment, a policy, or an exposure—and which alternative?
3. What happened? Which outcome, measured over how long?
4. How much, and how uncertain? The effect size and its interval.
5. Does it fit my question? Similar patients, resources, and decisions?

Keep those answers together. “The intervention worked” is less useful than “it reduced this outcome, in these patients, compared with this alternative, over this period.”

Study design: what kind of comparison is this?

Randomized trial

Investigators assign the intervention by chance. Randomization helps balance measured and unmeasured influences on average; it does not guarantee identical groups or remove every source of bias.

Cohort study

Researchers compare outcomes after exposures or treatment choices that were not randomly assigned. A difference can reflect the exposure, other differences between groups, or both.

Cross-sectional study

Exposure and outcome are assessed around the same point in time. It can describe prevalence or associations, but establishing which came first may be difficult.

Systematic review and meta-analysis

A systematic review uses explicit methods to find and assess studies. A meta-analysis statistically combines results. Pooling does not repair biased studies, and a pooled result is not independent replication of its largest trial.

Cluster trials assign groups such as hospitals, rather than individual patients. A cluster crossover trial switches each group between policies over time. Patients within a hospital or period can resemble one another, so the analysis must account for that structure. A policy effect also differs from the effect of actually receiving a drug.

Methods context: CONSORT, nonrandomized-study guidance, and Cochrane, including cluster designs.

Outcomes: a measurement is not always the complication

A primary outcome is the main prespecified question. Secondary outcomes add information and can support formal conclusions under an appropriate prespecified testing plan; exploratory analyses generate or examine additional questions. Check the protocol rather than assuming that the most impressive result was the planned main result.

A surrogate endpoint stands in for a clinical benefit. Its usefulness depends on evidence that changes in it predict the benefit in the relevant setting. Not every physiologic measurement is a validated surrogate. Ask what mattered to patients and whether the study actually measured it.

A composite outcome combines events, such as death or hospital admission. A difference in the composite does not mean every component improved; a more frequent, less severe component can drive the result.

Definitions: FDA patient outcomes and surrogate endpoints.

Absolute and relative risk: both numbers belong in the story

Imagine a randomized trial with 1,000 patients per group, measuring one complication over 30 days. These numbers are invented for teaching.

• Absolute risk reduction: 10% − 5% = 5 percentage points, or five fewer complications per 100 patients.
• Risk ratio (RR): 5% ÷ 10% = 0.50.
• Relative risk reduction: (10% − 5%) ÷ 10% = 50%.

“50% lower risk” and “five percentage points lower risk” describe the same invented result. A percentage point is an absolute subtraction; a relative percentage compares that change with the starting risk.

Effect measures: Cochrane Handbook, Chapter 6.

Number needed to treat: add the outcome and the clock

For a beneficial causal treatment effect, NNT = 1 ÷ absolute risk reduction expressed as a proportion. In example A: 1 ÷ 0.05 = 20. That means about 20 people treated, on average, for one additional person to avoid the specified complication over 30 days, under the example's assumptions.

Example B gives 1 ÷ 0.01 = 100. The same relative reduction produces a different NNT because the starting risks differ. Round a noninteger NNT up to the next whole person.

An NNT is not a promise that every twentieth patient benefits, and it does not identify who benefits. Its uncertainty follows the risk difference; an interval that crosses no effect makes a single reassuring NNT especially misleading. Do not turn a confounded observational association into a causal NNT. The corresponding calculation for an absolute increase in harm is often called the number needed to harm.

Interpretation: Cochrane Handbook, Chapter 15.

Risk, odds, prevalence, and hazard ratios are different

First check the direction. Which group is the numerator, and is the outcome desirable or undesirable? A ratio below 1 can mean less of a complication—or less of a desirable outcome.

Risk ratio

Compares event probabilities over a defined period. In example A, 0.50 means the new-treatment risk is half the comparison risk.

Odds ratio

Compares event odds: events divided by nonevents. Example A gives (50/950) ÷ (100/900) ≈ 0.47. That is not exactly the risk ratio of 0.50. The two can differ substantially when events are common; “50% lower odds” is not automatically “50% lower risk.”

Prevalence ratio

Compares how common a finding is at an assessment. A gastric-content finding on a scan and a future aspiration event are different endpoints, regardless of the ratio used.

Hazard ratio

Compares instantaneous event rates among people still at risk, under the analysis's assumptions. It is not simply the ratio of final event probabilities, and it does not directly tell you how many extra days someone lives.

An adjusted ratio comes from a model accounting for selected variables. It can differ from the crude ratio calculated from raw counts. Preserve the paper's effect measure and model; do not silently substitute your own calculation.

Definitions and analysis: CDC measures of association and Cochrane Handbook, Chapter 6.

Confidence intervals: read the range, not just the point

For ratios, 1 is the no-difference value. For absolute differences, it is 0. An interval spanning that value is not proof that treatments are equivalent.

A frequentist 95% confidence interval comes from a method that would cover the fixed true parameter in 95% of repeated studies under its assumptions. It is not a 95% probability that this particular interval contains that parameter, and it is not the range containing 95% of patients.

Wide intervals often signal limited precision. Narrow ones do not remove bias, poor measurement, or a mismatch with your patients. Check whether the paper reports a 90%, 95%, or adjusted interval; these are not interchangeable.

Intervals and uncertainty: Cochrane Handbook, Chapter 15.

P values: a threshold does not make the clinical decision

A P value describes how incompatible the observations are with a specified statistical model, including a null hypothesis. If the null model and other assumptions held, P = .03 means results at least as extreme as those observed would occur about 3% of the time.

It does not mean a 3% chance the null hypothesis is true, a 97% chance the treatment works, or a 3% chance the result was “caused by chance.”

A nonsignificant superiority test does not establish equivalence or noninferiority. Those questions require a suitable design and a justified prespecified margin. Conversely, a tiny difference can be statistically significant without being clinically worthwhile.

Interpretation principles: American Statistical Association statement; FDA non-inferiority guidance.

Multiple outcomes, subgroups, and early stopping

Testing many outcomes or subgroups creates more opportunities for a false-positive finding. Check the planned outcome hierarchy, multiplicity adjustment, and whether an analysis was prespecified. An exploratory result can be important without carrying the same evidential weight as the main comparison.

“Significant in one subgroup, nonsignificant in another” does not demonstrate that treatment effects differ between subgroups. Look for the interaction analysis and its uncertainty. A handful of rare events can also produce an unstable-looking large effect.

An early-stopped trial needs its stopping rule and amount of accumulated information attached to the headline. Prespecified monitoring is different from stopping when an appealing result appears; an early benefit estimate may still be exaggerated. A post hoc or Bayesian reanalysis uses existing data and additional assumptions—it is not another randomized replication.

Planning and reporting: FDA multiple-endpoint guidance and CONSORT.

Missing data, follow-up, and who was counted

Intention-to-treat preserves analysis by randomized assignment. It does not magically create missing outcomes. Imputation supplies values or estimates under stated assumptions; inspect sensitivity analyses rather than treating imputed observations as known facts.

Keep follow-up windows attached to outcomes: 72-hour delirium is not long-term cognition. Follow-up ending at discharge does not measure every later death. Procedures and patients can also have different denominators when a person contributes more than one procedure.

Participant flow and missing outcomes: CONSORT.

Applicability: can this answer my clinical question?

Compare the study's eligibility, setting, comparator, expertise, and delivered treatment with your own decision. A trial in selected ICU patients does not directly settle routine elective induction; an endoscopy trial does not automatically establish a surgical aspiration-prevention policy.

Ask whether a proposed implementation actually matches the tested intervention. Equipment, dose, timing, permitted adjuncts, and exceptions can all matter. A study's exclusions are especially important when considering a patient who was excluded for a safety concern.

Finally, separate the investigators' findings, the review author's interpretation, and a professional society's recommendation. A recommendation also weighs benefits, harms, certainty, feasibility, and patient priorities. Study funding, author relationships, corrections, and retractions deserve inspection; a disclosure does not by itself prove or disprove a result.

A useful closing sentence: “This study supports ___ for ___ compared with ___ over ___; uncertainty remains about ___.” If a review cannot fill those blanks, return to the methods before acting on the headline.