How to read a result.

A simple example: start with the numbers

Imagine two similar groups of 100 people followed for the same time. An event occurs in 10 people in one group and eight in the other. The absolute difference is two people per 100, or two percentage points. Compared with the starting rate of 10 per 100, that is a 20% relative reduction.

These numbers are invented to show the calculation; they are not treatment results. The difference could reflect chance. The example also says nothing about harms, other outcomes or other people.

Start with the outcome

An outcome, sometimes called an endpoint, is what a study measures to answer a question. Researchers should set the main outcome and important secondary outcomes in advance. Exploratory outcomes may also be planned or identified later; they can suggest questions without providing a definitive answer. Ask which analyses were planned and whether the study accounted for making multiple comparisons.

An outcome might describe how people feel, function or survive. It might instead be a laboratory or scan result used as a substitute, called a surrogate. A better surrogate result does not automatically mean people feel better or live longer. A statistically detectable change can also be too small to matter in daily life.234

Absolute and relative effects

Imagine that, over one year, an unwanted event happens to 20 of 100 people receiving a comparison treatment and 10 of 100 receiving a new treatment. That is 10 fewer events per 100 people: an absolute reduction of 10 percentage points. The risk ratio is 0.5 and the relative reduction is 50%.

“The risk was halved” and “10 fewer people per 100 had the event” describe the same invented numbers. The second also shows the starting risk and the size of the change. If the starting risk were 2 per 100 instead, the same relative reduction would be about 1 fewer event per 100. Always look at the starting risk alongside the relative effect. An odds ratio is not interchangeable with a risk ratio when events are common.15

Keep the time window attached

Ten events per 100 people over a week is not the same as ten over several years. A result needs a time period. An early improvement also does not prove it lasts. When a study measures people repeatedly, check whether it compares change from the start, a value at one time or a pattern over time.15

Read the estimate beside its uncertainty

A confidence interval gives a range of estimates compatible with the data under the statistical model used. A wide interval indicates less precision; a narrow one indicates more precision. Precision does not fix a biased design or make an unrepresentative sample representative.

For the usual 95% interval, if the same design and analysis were repeated many times, about 95% of the intervals produced by that method would contain the fixed value being estimated. It is not correct to say there is a 95% probability that this particular interval contains that fixed value.

For a difference in risk, an interval including zero is compatible with no difference. For a risk ratio, the equivalent value is one. A p-value is not the probability that a treatment works and does not tell you how large or useful an effect is.

Consider the range of effects that would matter to patients as well. An interval can cross zero but rule out effects large enough to matter; it can also exclude zero while the estimated effect is too small to be useful. Multiple outcomes or repeated analyses can increase the chance of a striking result unless the analysis plan accounts for them. When comparing treatments, look at the interval for the difference or ratio itself. Whether the separate groups’ intervals overlap is not a reliable substitute.54

Statistical clarity is not the same as practical importance

Imagine a trial estimates that a treatment reduces a symptom score by 2 points compared with a control. Its 95% confidence interval runs from 0.5 to 3.5 points. The interval excludes zero, so under that model the data are not compatible with no average difference.

But suppose patients and clinicians had agreed in advance that 3 points was the smallest improvement worth noticing. The interval still includes smaller changes, so the practical importance remains uncertain. The reverse can happen too: a potentially worthwhile estimate may be imprecise in a small study. Ask both what effects the data are compatible with and which effects would matter, to whom, and at what cost or risk.53

Averages do not describe every person

A mean or median summarises a group; not everyone had that amount of change. Look for how widely results varied, how many people responded and stopped treatment, and whether subgroup analyses were planned in advance.

“No statistically significant difference” does not prove two treatments are equivalent. A claim of equivalence or non-inferiority needs an appropriate study design and a justified margin, not just a non-significant p-value.6

Missing data are part of the result

Check how many people were assigned to each group, how many contributed outcome data, why data were missing and whether this differed between groups. Analyses of missing outcomes rely on assumptions that observed data alone usually cannot fully check.

In statistics, “missing at random” does not mean that missing data are unrelated to everything. It means that, after the observed information in the analysis is considered, the chance of a missing outcome does not depend on its unseen value. Leaving because of a recorded side effect does not automatically break that assumption. The assumption can fail if the unseen outcome still affects who is missing.

Read the withdrawals, main analysis and sensitivity analyses together. If the result changes substantially under plausible alternative assumptions, it is less robust.109

Read benefit and harm on comparable terms

Look for clear definitions, group sizes and time periods for both benefits and harms. A common mild symptom and a rare serious event should not be collapsed into one score without a value judgement. Check how harms were collected, whether severity and treatment stopping were reported, and whether follow-up was long enough to find delayed risks. A small trial may miss uncommon harms.

An adverse event is an unwanted occurrence during treatment or research; it does not by itself prove the treatment caused it. Timing, background rates, the comparison group and other evidence all matter. Safety analyses often include people who received at least one dose, while benefit analyses can use a different group. Exposure time and the way repeated events are counted also affect comparisons.111

The five-question check

What was measured: an outcome important to patients or a substitute? Who was included and what were the groups compared? What were the actual event rates or average results in each group? How long were people followed? How uncertain is the estimate, what assumptions were made about missing data and multiple analyses, and what harms were seen?

If the answers are unavailable, the honest conclusion may be that a result is suggestive but uncertain.127