{
  "slug": "median-absolute-deviation",
  "term": "Median absolute deviation",
  "aliases": [
    "MAD"
  ],
  "definition": "Median absolute deviation is the median of the absolute distances between observations and their sample median.",
  "note": "Raw MAD = median(|x - median(x)|), in the observations' units. It describes robust central spread, not the slowest requests. [nist-mad]\n\nNormal scaling divides raw MAD by about 0.67449 to estimate a normal population's standard deviation consistently as the sample grows. Report whether this adjustment was used; it is not a universal conversion for skewed latency data. [scipy-mad]\n\nSynthetic completed latencies [1, 1, 1, 1, 100] ms have median 1 ms and raw MAD 0 ms despite one 100 ms request. A zero MAD can therefore coexist with a long observed tail. The mean of absolute distances is a different statistic. [scipy-mad] [hesela-spread]",
  "related": [
    "coefficient-of-variation",
    "quantile",
    "tail-latency",
    "standard-deviation"
  ],
  "citations": [
    {
      "id": "nist-mad",
      "title": "NIST Dataplot: Median Absolute Deviation",
      "url": "https://www.itl.nist.gov/div898/software/dataplot/refman2/auxillar/mad.htm",
      "accessed": "2026-11-19"
    },
    {
      "id": "scipy-mad",
      "title": "SciPy: median_abs_deviation, scale and center conventions",
      "url": "https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.median_abs_deviation.html",
      "accessed": "2026-11-19"
    },
    {
      "id": "hesela-spread",
      "title": "Hesela: robust and relative latency spread, fixed arithmetic and reproducible code",
      "url": "https://hesela.com/analyses/storage-latency-coordinated-omission/#robust-relative-spread",
      "accessed": "2026-11-19"
    }
  ],
  "status": "research",
  "advantages": [],
  "limitations": [],
  "supersedes": [],
  "supersededBy": [],
  "firstSeen": null,
  "lastSeen": null,
  "updated": "2026-11-19"
}
