Contents

Median Absolute DeviationEdit

Median Absolute Deviation

The median absolute deviation is a robust measure of statistical dispersion. It captures how spread out a data set is by looking at the median of the absolute deviations from the data’s central location. Because it relies on medians rather than means, it is far less sensitive to outliers and skew in the data than traditional dispersion measures. In practical terms, MAD provides a dependable summary of spread when data contain unusual observations, heavy tails, or contamination, which is common in real-world datasets from finance, manufacturing, and social science alike. For a standard reference, see discussions of the median and of robust dispersion measures such as robust statistics and dispersion.

MAD is widely used because it is simple to compute and interpretable, and it scales well with linear transformations of the data. When data are shifted by a constant, or when all observations are multiplied by a positive constant, the MAD changes in a predictable way, making it a stable descriptor across different scales and units. In many statistical workflows, MAD complements the standard deviation and the interquartile range, offering a different balance of resistance to outliers and efficiency under normal-like data.

Definition

The basic, unscaled median absolute deviation of a dataset x1, x2, ..., xn is defined as the median of the absolute deviations from the dataset’s median:

MAD_raw = median(|x_i − median(x)|)

Here, median(x) denotes the central value that minimizes the sum of absolute deviations, and |x_i − median(x)| is the absolute deviation of each observation from that center. For a sample, this is typically computed as:

  • compute the sample median m = median(x1, ..., xn)
  • compute deviations d_i = |x_i − m| for i = 1, ..., n
  • MAD_raw = median(d_1, ..., d_n)

Because MAD_raw can be biased as an estimator of scale under light-tailed distributions such as the normal distribution, a scaling factor is commonly applied to make it consistent for the standard deviation under normal data. The conventional scaling constant is c ≈ 1.4826, derived from the inverse of the 75th percentile of the standard normal distribution. The scaled MAD is then:

MAD = c × MAD_raw

This scaling does not change the fundamental interpretation of MAD as a dispersion measure, but it aligns it with the standard deviation in the normal-spotlight case. See normal distribution for context on why this scaling is chosen in many applications.

Calculation and interpretation

  • Input: a sample of observations that may include outliers or skew.
  • Output: a single value representing the typical absolute deviation from the center of the data.
  • Interpretation: MAD gives a robust sense of how far observations tend to be from the central location, with the scaling c ensuring comparability to the standard deviation for Gaussian data.

MAD is translation-invariant: adding a constant to all observations leaves MAD unchanged, since both the center (median) and all deviations shift by the same constant. It is also scale-equivariant: multiplying all observations by a positive constant t scales MAD by t. See robust statistics for a broader discussion of why these properties matter for resistant measures.

Properties and robustness

  • Robustness: MAD has a breakdown point of 50%, meaning that up to half of the observations can be corrupted by outliers without forcing MAD to take arbitrarily large values. This makes MAD a preferred choice in the presence of contamination when one does not want a metric dominated by a few extreme observations.
  • Nonparametric character: MAD does not assume a particular underlying distribution; it is based on ranks and medians rather than moments. See nonparametric statistics for related ideas.
  • Efficiency: Under a normal distribution, even with the scaling, MAD is less statistically efficient than the standard deviation. In clean, well-behaved data, the standard deviation will typically offer tighter uncertainty quantification. In data with outliers or heavy tails, MAD often provides a more reliable picture of spread.
  • Alternatives in robust statistics: in some settings, researchers prefer other robust scale estimators with higher efficiency under normality, such as the Qn or Sn estimators. See Qn estimator and Sn estimator for comparisons.

Inference and practice

  • Confidence intervals and hypothesis tests: because MAD has a non-Gaussian, non-symmetric sampling distribution, standard parametric intervals based on normal theory are not appropriate in general. Resampling methods like the bootstrap are commonly used to assess uncertainty in MAD and to construct confidence intervals. See bootstrap for a practical approach to inference with robust statistics.
  • Use in regression and outlier detection: while MAD itself measures dispersion, it is often used in conjunction with robust regression and outlier-detection schemes to gauge stability and anomaly presence. In robust regression, methods that minimize absolute deviations (L1-based) are related in spirit to the robustness rationale underpinning MAD. See robust regression and least absolute deviations for related concepts.

Comparison with other measures of spread

  • Standard deviation: The standard deviation is efficient under normal data but highly sensitive to outliers. MAD trades some efficiency for robustness, making it attractive when data are messy or contaminated. See standard deviation.
  • Interquartile range (IQR): The IQR is another robust measure of spread based on quartiles. Like MAD, IQR resists outliers, but it reflects a different portion of the distribution and can be less informative about tail behavior in some contexts. See interquartile range.
  • Tail-robust measures: In some applications, combining MAD with other robust estimators or using alternative scale measures (like Qn or Sn) can yield a better balance of efficiency and resistance. See Qn estimator, Sn estimator.

Applications

  • Finance and economics: robust dispersion measures help in assessing risk without being unduly swayed by extreme market moves. MAD can be part of a toolbox for robust portfolio analysis and risk estimation when data exhibit spikes or bursts. See financial risk and volatility discussions in robust contexts.
  • Engineering and quality control: in manufacturing, MAD can serve as a simple, dependable gauge of process spread when data contain occasional faults or unusual events. It can complement process capability analyses that traditionally rely on parametric assumptions.
  • Data analysis and machine learning: MAD is used in preprocessing, as a robust scale parameter for centering and normalizing data, and as part of robust features that resist outliers. See data preprocessing and robust machine learning topics for related ideas.

Controversies and debates

  • Efficiency versus robustness: a core debate centers on whether to prioritize efficiency under ideal conditions (where the data are normal and clean) or robustness to outliers and contamination. Proponents of MAD emphasize reliability in the real world, where data rarely meet ideal assumptions; critics point to the loss of efficiency in ordinary circumstances. See statistical efficiency and robust statistics for the broader discussion.
  • Choice of estimator among robust options: because there are several robust scale estimators (MAD, Qn, Sn, and others), practitioners debate which to use in a given setting. Each has different efficiency, breakdown properties, and computational considerations. See Qn estimator and Sn estimator for comparisons.
  • Interpreting spread in skewed data: MAD, like other dispersion measures, has interpretations that may vary with skewness. In highly skewed distributions, some analysts prefer other summaries or accompany MAD with additional metrics to convey asymmetry. See skewness and robust statistics for related topics.

See also