A clinical trial reports that a new antihypertensive drug lowers blood pressure with a p-value of 0.001. The result is highly significant. But the actual reduction? Two millimeters of mercury. Would any clinician change their prescribing based on that finding? Would any patient notice the difference?

This gap between statistical significance and clinical significance is one of the most consequential misunderstandings in modern medicine. It shapes which drugs reach the market, which treatments receive guideline endorsement, and which interventions patients are offered. Yet the distinction remains poorly understood even among experienced practitioners.

The tools to evaluate both dimensions of significance are well established. The challenge is not a lack of methodology—it is a culture that has historically equated a low p-value with a meaningful result. Understanding how to read beyond the p-value is not an academic exercise. It is a core clinical skill that directly affects patient care.

A p-value answers one specific question: If there were truly no difference between groups, how likely would we be to observe a result this extreme or more extreme? That is all it tells us. A p-value of 0.05 means there is a 5% probability of obtaining the observed data—or something more extreme—under the null hypothesis. It does not tell us the probability that the treatment works, the probability that the null hypothesis is true, or the size of the effect.

This narrow definition gets routinely inflated. Researchers, clinicians, and journalists frequently interpret a statistically significant result as proof that a treatment is effective in a clinically meaningful way. The conflation is understandable. The word significant carries weight in everyday language that far exceeds its technical statistical meaning. But the consequences of this misinterpretation are not trivial.

Large sample sizes make the problem worse. With enough participants, a trial can detect vanishingly small differences between groups—differences that would never register in a consulting room. A study enrolling 50,000 patients might achieve p < 0.001 for a treatment that reduces pain scores by 0.3 points on a 10-point scale. The statistics are impeccable. The clinical relevance is negligible. Yet the result gets reported as statistically significant, and the nuance disappears.

The American Statistical Association took the unusual step in 2016 of publishing a formal statement cautioning against the mechanical use of p-value thresholds. Among its core principles: a p-value does not measure the importance of a result, and scientific conclusions should not be based solely on whether a p-value crosses 0.05. Despite this, the binary framing of significant versus non-significant continues to dominate how research is reported and consumed.

Takeaway

A p-value tells you how surprising the data would be if there were no real effect. It says nothing about how large, important, or clinically relevant that effect actually is.

If p-values measure statistical surprise, effect size metrics measure magnitude. They quantify how large the observed difference actually is—independent of sample size. Common effect size measures include Cohen's d for comparing group means, relative risk and odds ratios for binary outcomes, and number needed to treat (NNT) for translating group-level findings into individual patient relevance. Each of these answers a fundamentally different question than the p-value.

Consider the NNT. If a statin trial reports that the drug reduces cardiac events with p < 0.01, that tells us the finding is unlikely due to chance. But the NNT tells us how many patients must take the drug for one additional patient to benefit. An NNT of 20 means treating 20 patients for one to avoid a cardiac event. An NNT of 200 means 199 out of 200 patients receive no benefit. Both results might carry the same p-value. The clinical implications are profoundly different.

The concept of the minimal clinically important difference (MCID) formalizes this distinction further. The MCID represents the smallest change in an outcome measure that patients perceive as meaningful. For the six-minute walk test in heart failure, for instance, the MCID is approximately 30 meters. A trial showing a statistically significant improvement of 12 meters has produced a real but clinically imperceptible change. Without referencing the MCID, that result looks like progress. With it, the finding is properly contextualized.

Confidence intervals add another essential layer. Rather than reducing a result to a binary yes-or-no, a 95% confidence interval shows the range within which the true effect plausibly lies. A trial might report a mean improvement of 4 points with a confidence interval of 1 to 7. If the MCID is 5 points, the confidence interval tells us the true effect might be clinically important—but it might also fall well below the threshold. This kind of uncertainty is exactly what clinicians need to communicate to patients.

Takeaway

Effect sizes, NNTs, MCIDs, and confidence intervals answer the question that actually matters in the clinic: is this difference large enough to change how a patient feels, functions, or survives?

Reading a trial report with both statistical and clinical significance in mind requires a structured approach. Start with the primary outcome: is it something patients care about? Surrogate endpoints like biomarker levels or imaging findings are easier to measure but may not translate into outcomes that matter—survival, symptom relief, functional capacity. A statistically significant change in a surrogate marker is two steps removed from clinical relevance.

Next, examine the effect size alongside the p-value. Ask three questions. First, is the observed difference larger than the established MCID for this outcome measure? If no MCID exists, consider whether the magnitude of change would be perceptible to a patient. Second, what is the confidence interval, and does its lower bound still represent a meaningful effect? Third, what is the NNT, and does it represent a reasonable treatment proposition given the drug's cost, side effect profile, and duration of therapy?

Then consider the population. A trial conducted in a highly selected group—younger patients, fewer comorbidities, single-site academic centers—may produce results that do not generalize to the patients sitting in your waiting room. Effect sizes often shrink when treatments move from controlled trial environments to routine clinical practice. This attenuation is predictable but rarely discussed in the abstract or press release.

Finally, weigh the totality of evidence. A single trial, however well designed, provides one data point. Systematic reviews and meta-analyses that pool multiple studies offer more stable estimates of both statistical reliability and effect magnitude. Look for consistency across trials, subgroup analyses that test whether effects hold in different populations, and sensitivity analyses that examine whether methodological choices influenced the result. No single metric—not the p-value, not the effect size, not the NNT—provides a complete picture on its own.

Takeaway

A complete appraisal of a trial asks not just whether the result is real, but whether it is large enough to matter, stable enough to trust, and applicable enough to act on for the patient in front of you.

The distinction between statistical and clinical significance is not a methodological technicality. It is a safeguard against adopting treatments that are mathematically defensible but practically inconsequential. Every prescribing decision, every guideline recommendation, and every patient conversation about treatment options benefits from this clarity.

The tools already exist. Effect sizes, confidence intervals, NNTs, and MCIDs are reported in well-designed trials. The discipline lies in looking for them—and in refusing to let a low p-value substitute for a thoughtful assessment of whether a treatment meaningfully changes a patient's life.

Statistical significance tells us something was detected. Clinical significance tells us whether it matters. The best clinical decisions require both.