Mean Absolute Deviation: Formula, Worked Example, and How It Differs from Standard Deviation

Mean absolute deviation (MAD) — also called the average deviation from the mean, or average absolute deviation — is a measure of statistical dispersion: how far, on average, the values in a dataset sit from the dataset’s own mean. It is calculated by taking the absolute value of each observation’s deviation from the mean, then averaging those absolute deviations. Unlike variance and standard deviation, which square the deviations before averaging, MAD treats every deviation linearly, which is why it is often described as the most intuitive of the common spread statistics: it answers, in the original units of the data, “how far off from the mean is a typical observation?”

The formula for a sample or population of size n, with mean x̄ (or μ for a population), is:

MAD = (1/n) × Σ |x_i − x̄|

Why the absolute value is necessary

Deviations from the mean, left unsigned, always sum to exactly zero — that is a mathematical property of the mean itself, not a coincidence of any particular dataset. Values above the mean and values below it cancel out perfectly when added together, so a simple average of the raw (signed) deviations is always zero and carries no information about spread. Two ways around this exist: take the absolute value of each deviation before averaging (mean absolute deviation), or square each deviation before averaging (variance, and its square root, standard deviation). Both fixes remove the sign problem; they differ in what happens next, which is the comparison the rest of this guide works through.

Worked example: step-by-step calculation

The following is an illustrative dataset, not real research data, used purely to demonstrate the calculation. Take five hypothetical measurements: 4, 7, 9, 12, 13.

1. Find the mean. (4 + 7 + 9 + 12 + 13) ÷ 5 = 45 ÷ 5 = 9.
2. Find each deviation from the mean. 4−9 = −5; 7−9 = −2; 9−9 = 0; 12−9 = 3; 13−9 = 4. (Check: −5 −2 + 0 + 3 + 4 = 0, confirming the signed deviations cancel out, as they always do.)
3. Take the absolute value of each deviation. 5, 2, 0, 3, 4.
4. Average the absolute deviations. (5 + 2 + 0 + 3 + 4) ÷ 5 = 14 ÷ 5 = MAD = 2.8.

In plain terms: a typical observation in this illustrative dataset sits 2.8 units away from the mean of 9.

Mean absolute deviation vs. standard deviation

Using the same illustrative dataset, compare MAD against variance and standard deviation. Squaring the same five deviations (−5, −2, 0, 3, 4) gives 25, 4, 0, 9, 16, which sum to 54. Dividing by 5 gives a variance of 10.8, and the square root of that, the standard deviation, is approximately 3.29 — noticeably larger than the MAD of 2.8, because squaring amplifies the influence of the larger deviations (the −5 and the 4) relative to the smaller ones. That is the essential contrast between the two statistics:

• Standard deviation squares each deviation before averaging, which penalizes large deviations disproportionately — a single far-outlying value can move the standard deviation substantially. In exchange, SD has mathematical properties (it is differentiable, and variance is additive across independent variables) that make it the building block for statistical tests such as the t-test, ANOVA, and regression. This is the main reason SD dominates in practice despite MAD being easier to explain: most of inferential statistics is built on variance, not on mean absolute deviation.
• Mean absolute deviation treats every deviation linearly, so a single extreme value moves MAD by less than it moves the standard deviation. MAD is also more directly interpretable — it is literally the average distance from the mean, with no squaring-and-unsquaring step to explain — which is part of why it appears in introductory statistics teaching and in some forecasting contexts (see below), even though it is rarely the dispersion statistic reported in a research manuscript’s results section.

For data that follows a normal distribution specifically, the two statistics have a fixed relationship: MAD ≈ 0.7979 × SD (equivalently, SD ≈ 1.2533 × MAD). That ratio only holds under normality — it is one more reason to check the shape of a distribution, for example with a histogram or a formal test of the normality assumption, before assuming one dispersion statistic tells you what the other would have shown.

Mean absolute deviation vs. median absolute deviation — a genuinely different statistic

“Mean absolute deviation” and “median absolute deviation” are frequently confused, and both are commonly abbreviated MAD in different literatures — but they are calculated differently and used for different purposes.

• Mean absolute deviation (this guide’s main subject) takes deviations from the mean, then averages the absolute values of those deviations with another mean.
• Median absolute deviation (sometimes written MAD or MedAD to disambiguate) takes deviations from the median, then finds the median of those absolute deviations — a median of absolute deviations, not a mean of them. Because both the center point and the averaging step use the median rather than the mean, median absolute deviation is highly resistant to outliers, and it is the version statisticians reach for when the specific goal is robust outlier detection or a robust spread estimate for skewed or contaminated data.

Median absolute deviation is usually reported after multiplying by a scale factor of approximately 1.4826. That constant makes the scaled median absolute deviation directly comparable to a standard deviation when the underlying data is normally distributed — without it, median absolute deviation systematically understates spread relative to SD, because the median absolute deviation of a standard normal distribution is only about 0.6745, not 1.

A second illustrative dataset shows why the distinction matters in practice. Take five hypothetical measurements with one extreme value: 4, 5, 6, 7, 50.

• Mean absolute deviation: the mean is 14.4. Deviations from the mean are −10.4, −9.4, −8.4, −7.4, and 35.6; their absolute values average to 14.24. The single value of 50 has pulled both the mean and the MAD sharply upward.
• Median absolute deviation: the median is 6. Deviations from the median are −2, −1, 0, 1, and 44; their absolute values are 2, 1, 0, 1, 44, and the median of that set of absolute deviations is 1. Scaled by 1.4826, that is approximately 1.48 — a spread estimate barely affected by the outlier, because neither the median center point nor the median-of-absolute-deviations step gives the value of 50 any special weight.

The gap between 14.24 and roughly 1.48, on the same five numbers, is the entire practical case for median absolute deviation: when a dataset may contain outliers or heavy-tailed error, a median-based spread estimate keeps working when a mean-based one (whether that is standard deviation or mean absolute deviation) is dominated by the extreme values.

When each measure of spread is appropriate

• Roughly symmetric data with no major outliers: standard deviation is the conventional default, chiefly because it feeds directly into confidence intervals, t-tests, ANOVA, and regression.
• Data you need to explain to a non-technical audience, or an intuitive spread check during exploratory analysis: mean absolute deviation is easier to interpret at face value and is common in introductory statistics teaching for that reason.
• Skewed distributions, small samples, or data suspected to contain outliers or measurement errors: median absolute deviation (scaled) is the more defensible choice, and is widely used as an outlier-detection rule — flagging any observation whose absolute deviation from the median exceeds some multiple (commonly 2.5 or 3) of the scaled median absolute deviation, a more robust alternative to a rule based on the mean and standard deviation.
• Forecast accuracy in time series and demand-planning contexts: mean absolute deviation and the related mean absolute percentage error (MAPE) are standard ways to score how far forecasted values fell from actual outcomes, precisely because the linear (not squared) penalty is easier to communicate to a planning audience than a squared-error metric.

Before choosing among these, it is worth establishing what shape the data actually has and what kind of variable you are summarizing — a quantitative (interval or ratio) variable is required for any of these statistics to be meaningful in the first place, and a quick look at a histogram of the data will usually make an outlier or a skew obvious before you calculate anything.

Relationship to variance and the coefficient of variation

Mean absolute deviation, variance, and standard deviation are all built from the same raw material — deviations from the mean — and differ only in how those deviations are combined. Variance and standard deviation dominate applied statistics not because they are more “correct” than MAD, but because squared deviations have properties MAD’s absolute values lack: variance is additive across independent variables, it is the quantity that ANOVA literally partitions into between-group and within-group components, and it is differentiable everywhere, which matters for the optimization methods behind regression and maximum-likelihood estimation. MAD, built on a non-differentiable absolute-value function, does not fit into that machinery as cleanly, which is the practical reason it stays a secondary, descriptive statistic rather than the basis for inferential methods.

The coefficient of variation (CV) — standard deviation expressed as a percentage of the mean — solves a different problem than any of the statistics above: comparing relative variability across variables measured in different units or on very different scales, where comparing raw MAD, variance, or SD values directly would not be meaningful.

Calculating MAD in R, Python, and Excel

• R: mean(abs(x - mean(x))) calculates mean absolute deviation directly. For (scaled) median absolute deviation, R’s built-in mad(x) function already applies the standard 1.4826 scale factor by default; pass constant = 1 to get the unscaled version.
• Python (NumPy/SciPy/pandas): NumPy has no single built-in mean-absolute-deviation function, but it is a one-line calculation: np.mean(np.abs(x - np.mean(x))). pandas Series objects expose .mad() in older versions (removed in pandas 2.0+, so check your version), or the same NumPy one-liner works on a pandas column. SciPy’s scipy.stats.median_abs_deviation(x, scale='normal') returns the scaled median absolute deviation.
• Excel: AVEDEV(range) calculates mean absolute deviation directly. Excel has no built-in median-absolute-deviation function; it has to be built manually as MEDIAN(ABS(range - MEDIAN(range))) entered as an array formula, then multiplied by 1.4826 if the scaled version is wanted.

Frequently asked questions

Is mean absolute deviation the same as standard deviation?

No. Both measure spread around the mean, but standard deviation squares each deviation before averaging (then takes a square root to return to the original units), while mean absolute deviation takes the absolute value of each deviation and averages directly. Standard deviation is more sensitive to outliers and is the statistic used throughout inferential statistics; mean absolute deviation is more intuitive to interpret but is not the basis for common statistical tests.

Is mean absolute deviation the same as median absolute deviation?

No, and this is the most common confusion around the term. Mean absolute deviation takes deviations from the mean and averages their absolute values with a mean. Median absolute deviation takes deviations from the median and summarizes those absolute values with a median, not a mean. Median absolute deviation is far more resistant to outliers and is the version typically used for robust outlier detection.

What does a small MAD value mean?

A small mean absolute deviation relative to the scale of the data means observations tend to sit close to the mean — the data is tightly clustered. A large MAD relative to the mean means observations are widely spread. There is no single universal threshold for “small” or “large”; it depends entirely on the units and context of what is being measured.

Why isn’t mean absolute deviation used more often in research reporting?

Largely because standard deviation and variance connect directly to the rest of inferential statistics — confidence intervals, t-tests, ANOVA, and regression are all built on variance — while mean absolute deviation, based on the non-differentiable absolute-value function, does not extend into that framework in the same way. Standard deviation remains the conventional default for reporting spread alongside a mean in most research manuscripts.

When should I use MAD/MAPE instead of a squared-error metric?

In forecasting and demand-planning contexts, mean absolute deviation and mean absolute percentage error are often preferred over mean squared error specifically because the linear penalty is easier to explain to a non-statistical audience and is less dominated by a single large forecast miss — useful when the goal is a practical accuracy readout rather than an input to further statistical modeling.