Read beyond the noise.

Median absolute deviation measures spread from the middle twice

A reporter comparing a short row of numbers and their distances from a median on paper.
AI-generated photographic illustration; not documentary photography.

A report can describe how spread out its measurements are without letting the most extreme observation dominate the result. One method is the median absolute deviation. Its name gives the recipe: find a middle value, measure distances from it, then find the middle of those distances.

The NIST handbook’s discussion of measures of scale distinguishes this statistic from average absolute deviation and standard deviation. That distinction matters: averaging distances and taking their median are different operations.

Five invented waiting times

Suppose a small example contains waiting times of 4, 5, 6, 7 and 18 minutes. These numbers are invented for illustration, not measurements of a real service.

  1. The middle waiting time is 6 minutes.
  2. The absolute distances from 6 are 2, 1, 0, 1 and 12 minutes.
  3. Put those distances in order: 0, 1, 1, 2, 12.
  4. The middle distance is 1 minute. The unscaled median absolute deviation is therefore 1 minute.

Now change the final wait from 18 to 80 minutes. The median is still 6. The ordered distances become 0, 1, 1, 2 and 74, so their median remains 1. That stability shows why this measure can help describe the central cluster. It does not mean the 80-minute wait is unimportant.

Read the definition beside the number

Some software reports a scaled version rather than the raw median distance. NIST’s Dataplot documentation discusses this distinction. When comparing reports, check the formula and scaling convention instead of assuming that every number labelled MAD was calculated identically.

For a news story about waiting times, a central spread measure is only part of the account. You would still want the number of observations, the period covered and a description of unusually long waits. Our five-number example cannot tell you whether one service is better than another.

A careful sentence names the statistic and its units. It also avoids turning “less affected by an extreme value” into “proves there are no serious delays.” The median calculation answers a particular question about the data; it does not replace looking at the rest of the distribution.