Why not average the signed deviations?

They sum to zero around the arithmetic mean, regardless of how spread out the observations are.

The mean-based measure remains sensitive to extreme observations; it should not inherit robustness claims intended for median absolute deviation.

AI Glossary / Statistics and probability / Intermediate

Mean absolute deviation (m.a.d.)

Definition

Mean absolute deviation is the average absolute distance of observations from a specified center. In the supplied glossary, mean absolute deviation, abbreviated m.a.d., specifically uses the arithmetic mean as that center. For n observations it is sum |x[i]-x-bar| divided by n. It measures spread in the same units as the observations.

What it means

The statistic describes how far observations lie from their mean on average, ignoring whether they are above or below it. Absolute values prevent positive and negative deviations from canceling.
The name needs care because MAD is also widely used for median absolute deviation, a different statistic. Some sources also discuss mean absolute deviation about the median. Always state both the center used and whether the final aggregation is a mean or a median.

How it works

Define the observations and their units. Calculate the arithmetic mean x-bar from the same data whose dispersion is being summarized. If some observations are missing or excluded, document that choice rather than silently treating them as zeros.
Subtract x-bar from every observation. These signed deviations sum to zero, which is why simply averaging them cannot measure dispersion around the mean. Take their absolute values so distances in both directions contribute positively.
Average those absolute distances using denominator n. This defines the descriptive statistic. The familiar n-minus-one correction used for an unbiased sample variance is not automatically a correction for mean absolute deviation.
Interpret the result in the original units. A mean absolute deviation of two minutes means an average absolute distance of two minutes from the selected mean. It does not imply that every observation is within two minutes or that any fixed percentage must fall in that interval.
Unlike variance, absolute deviation does not square the distances. However, the mean-based version remains sensitive to extreme observations. An extreme value affects its own absolute distance and moves the center used for every other distance.
Distinguish the three common constructions. Mean absolute deviation about the mean averages distances from x-bar. Mean absolute deviation about the median averages distances from the median. Median absolute deviation takes the median of distances from the median, sometimes followed by a distribution-specific scaling factor.
The median, rather than the mean, minimizes the sum of absolute distances over a constant center. The mean minimizes squared distances. Therefore choosing the mean as the center is a defined convention here, not the solution to an absolute-distance minimization problem.
For weighted data, specify both how the center is computed and how the distances are averaged. Using a weighted center but an unweighted distance average produces a different quantity from a fully weighted measure.
Use a distribution plot or additional summaries when shape matters. Two datasets can have the same mean and mean absolute deviation but different tails or clusters. When comparing series measured in different units, the raw deviations are not automatically comparable.
If the goal is population inference, sampling design and dependence affect uncertainty. A descriptive deviation calculated from a dataset does not by itself provide a confidence interval or establish that the dataset represents the target population.

Three practical examples

A service team reports the average absolute spread of processing times around their mean.
A production analyst summarizes variation in measured dimensions using an easily interpreted original-unit quantity.
An AI evaluation team distinguishes dispersion of observed task scores from prediction error against task-specific reference values.

Think of it this way

Imagine marking the mean on a number line and measuring the distance from every observation to that mark. Ignore left versus right and average the distances. The result measures spread around that mark; it does not say how many observations fall inside a particular band.

Worked example

An original dataset is 1, 2, 3, 4, and 10. Its mean is 20/5 = 4. Absolute deviations from the mean are 3, 2, 1, 0, and 6, whose sum is 12. Mean absolute deviation about the mean is therefore 12/5 = 2.4.
The median is 3. Absolute deviations from that median are 2, 1, 0, 1, and 7. Their mean is 11/5 = 2.2, while their median is 1. Thus the same data produce three different numbers: 2.4, 2.2, and 1, depending on the definition.
Now replace 10 with 100. The new mean is 22, and distances from it are 21, 20, 19, 18, and 78. Their sum is 156, giving a mean absolute deviation of 31.2.
This large change demonstrates that using absolute distances does not make the mean-based statistic immune to extreme observations. The calculation also shows why reporting only the acronym MAD can create substantial ambiguity.

Illustrative case study

In a fictional case, Silverfern Operations compares two dashboard reports labeled MAD. One report averages deviations from the mean; the other uses the median of deviations from the median.
The team initially interprets the difference as a data-quality problem. After reviewing the formulas, analysts rename the fields and document the center, aggregation, and units. They choose the summary that matches the operational question and retain a distribution view for unusual cases. The case illustrates a definition mismatch, not evidence that one statistic is universally superior.

Common misconceptions and mistakes

Do not confuse mean absolute deviation with standard deviation, median absolute deviation, or mean absolute prediction error. They can coincide in special cases, but their definitions differ.
Do not assume absolute values remove all outlier sensitivity. Do not apply a normal-distribution scaling factor without explaining the target quantity and assumptions. A smaller spread is not automatically better if the data represent intentionally different groups.

Business use

Mean absolute deviation can communicate variability in familiar units without squaring distances. It is useful when stakeholders need an average-distance description. Its value depends on a clear center and population definition. For processes with rare large deviations, combine it with tail summaries and avoid treating it as a guarantee about exceptional cases.

Check your understanding

Why are 2.4 and 1 both possible MAD-labeled values for the example?

Answer: The first averages distances from the mean; the second takes the median of distances from the median. Their formulas answer different questions.

Related terms

Mean
Median
Standard deviation
Mean Absolute Error
Dispersion
Absolute deviation
Median absolute deviation

Learn next

Next, compare several dispersion measures on skewed data and datasets with an extreme observation. Check each software function's definition before comparing values across reports.

Dispersion

Sources

Howell, Statistical Methods for Psychology glossary, Mean absolute deviation: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html NIST/SEMATECH e-Handbook, Measures of Scale: https://www.itl.nist.gov/div898/handbook/eda/section3/eda356.htm

Previous: Mean Next: Mean Absolute Error