Choose Between a Mean and a Median by the Question

Name the quantity you want

A mean and a median summarize different features of a distribution. The mean uses the magnitude of every value; the median identifies the middle of the ordered distribution. Choosing between them is not simply a contest in which one is correct whenever data are skewed.

Cochrane’s effect-measure guidance distinguishes continuous-outcome summaries and their analysis requirements. Start with the scientific purpose: average resource use, a typical patient experience and a population quantile are different targets that can require different summaries.

Inspect the distribution and data quality

Plot the observations or a suitable distribution summary. Check units, impossible values, mixture populations and measurement ceilings. A very large value may be an error, but it may also be a real contributor to total cost or burden. Deleting it because it changes the mean can remove exactly the feature the question concerns.

Describe spread on a compatible scale. An SD describes variability around the mean, while quartiles describe positions in the ordered distribution. Neither is a standard error of the summary. Preserve sample size and missingness alongside both summaries when both help interpretation.

Compare a hypothetical dataset

Consider values 1, 2, 3, 4 and 40. The mean is 50/5 = 10, while the median is 3. If these are valid resource-use amounts, five times the mean recovers the total of 50; five times the median gives 15 and does not describe total resource use.

If the largest value changes from 40 to 80, the mean becomes 90/5 = 18 while the median remains 3. That stability can be useful for describing the middle, but it also demonstrates that the median intentionally ignores much of the upper-tail magnitude. The choice should follow the research question, not whichever number appears less dramatic.

Align inference and reporting

If comparing groups, specify whether the target is a difference in means, a difference in medians or another distributional contrast. A rank-based test is not automatically a test of equal medians under all distribution shapes. Do not attach a mean-based confidence interval to a median because the software output is nearby.

Report the chosen estimand, analysis method and relevant distribution features. If the protocol selected the mean, a skewed sample should trigger assessment of the inferential method and a transparent sensitivity analysis where appropriate, rather than an undocumented switch to a different scientific question.

Source verification: 6 November 2026. Examples are hypothetical.