Median

, where $x_{(1)} \le x_{(2)} \le \dots \le x_{(n)}$ are the sorted values (order statistics)

Deep Breakdown

What it tells me (Situation Awareness): The median is the middle value of a sorted dataset: half the observations sit at or below it, half at or above it. For a random variable it is any $m$ with $P(X \le m) \ge 0.5$ and $P(X \ge m) \ge 0.5$, i.e. the 50th percentile, $F^{-1}(0.5)$. Unlike the mean, it depends only on the rank order of the data, not on the magnitudes of the extremes, so a single outlier (a 10,000 USD order, a 30-second page load) cannot drag it around. Its breakdown point is 50%: almost half the data can be corrupted before the median becomes arbitrarily wrong, versus 0% for the mean (one bad value suffices). It is also the value that minimizes total absolute deviation $\sum_i |x_i - m|$, which is why it is the optimal point forecast under MAE / L1 loss, while the mean is optimal under squared (L2) loss. In DA/DS: reporting typical order value, salary, house price or session length (all right-skewed); robust imputation of missing values; median-based outlier rules (MAD); quantile regression at $\tau = 0.5$; comparing mean vs median as a quick skewness check (mean much larger than median means a long right tail).

Flowchart

Real-world use

Where it is used

KPI dashboards for skewed metrics (median order value, median time-to-first-purchase, median support resolution time); housing and income statistics; robust missing-value imputation (SimpleImputer(strategy="median")); median absolute deviation (MAD) for outlier detection; L1-loss regression and quantile regression; median-of-means estimators in streaming systems.

Problem it solves

The arithmetic mean is hypersensitive to extreme values. In heavy-tailed business data (revenue per customer, latency, claim sizes) a handful of whales or glitches make the mean describe nobody. The median gives a "typical" value that is stable under outliers.

How it solves the problem

It throws away magnitude information and keeps only order: after sorting, only the middle position matters. Moving the largest value from 250 to 250,000 does not change which observation is in the middle, so the median does not move.

Example in a business

An e-commerce analyst reports average order value. Five orders: 12, 15, 18, 22 and one bulk B2B order of 250 USD. Mean $= 63.4$ USD, which is larger than four of the five orders and misleads the marketing team planning free-shipping thresholds. Median $= 18$ USD, the real "typical basket". The team sets the free-shipping threshold just above the median to nudge typical customers to add one more item.

The formula, piece by piece

• $x_{(k)}$: the $k$-th smallest value after sorting (an order statistic).
• $n$: number of observations.
• $(n+1)/2$: the middle position when $n$ is odd (e.g. position 3 of 5).
• $n/2$ and $n/2 + 1$: the two middle positions when $n$ is even (e.g. positions 3 and 4 of 6).
• $\tfrac{1}{2}(\cdot + \cdot)$: average of the two middle values, the usual convention for even $n$ (any value between them minimizes absolute deviation).
• Population version: $m = F^{-1}(0.5)$, the point where the CDF crosses one half.
• Optimality: $\text{median} = \arg\min_m \sum_i |x_i - m|$.

The formula in words

"Sort the numbers and take the one in the middle; if there are two in the middle, take their average." Tiny example: orders $[12, 15, 18, 22, 30, 250]$ ($n = 6$, even). Middle positions 3 and 4 hold 18 and 22, so the median is $(18 + 22)/2 = 20$, while the mean is $347/6 = 57.83$.

Critical Concepts

• Robustness (breakdown point 50%): the median only fails when about half the data is contaminated; the mean fails with one bad point.
• Mean vs median as a skew diagnostic: mean greater than median suggests right skew (revenue, latency); mean less than median suggests left skew. For symmetric distributions they coincide.
• L1 optimality: the median minimizes the sum of absolute errors, so a model trained with MAE loss predicts conditional medians, not conditional means.
• Not additive: the median of a sum is not the sum of medians, and you cannot average medians across segments to get the overall median. Store raw data or histograms/sketches instead.
• Computation: exact median needs sorting, $O(n \log n)$, or quickselect, $O(n)$ average. For streaming or distributed data use approximate sketches (t-digest, KLL), as in APPROX_PERCENTILE in SQL engines.
• Standard error: for large $n$ and a continuous density $f$, $\text{SE}(\text{median}) \approx 1/(2 f(m)\sqrt{n})$; for Normal data that is about $1.25\,\sigma/\sqrt{n}$, so the median is less efficient than the mean when data are truly Normal. Use the bootstrap for confidence intervals in practice.

Step-by-step calculation (tiny example)

Step 1: Collect order values $[22, 12, 250, 15, 18]$ USD, $n = 5$.

Step 2: Sort $[12, 15, 18, 22, 250]$.

Step 3: Find the middle position $n$ is odd, so position $(5+1)/2 = 3$.

Step 4: Read the value $x_{(3)} = 18$. Median $= 18$ USD.

Step 5: Compare with the mean Mean $= (12+15+18+22+250)/5 = 317/5 = 63.4$ USD. The gap ($63.4$ vs $18$) flags strong right skew caused by the 250 USD order.

Code examples

Beginner example (Python)

import statistics

orders_odd = [22, 12, 250, 15, 18]
orders_even = [12, 15, 18, 22, 30, 250]

print("Sorted (odd):", sorted(orders_odd))
print("Median (odd n):", statistics.median(orders_odd))     # 18
print("Mean (odd n):", statistics.mean(orders_odd))         # 63.4
print("Median (even n):", statistics.median(orders_even))   # 20.0
print("Mean (even n):", round(statistics.mean(orders_even), 2))  # 57.83

Beginner example (R)

orders_odd <- c(22, 12, 250, 15, 18)
orders_even <- c(12, 15, 18, 22, 30, 250)
cat("Median (odd n):", median(orders_odd), "\n")    # 18
cat("Mean (odd n):", mean(orders_odd), "\n")        # 63.4
cat("Median (even n):", median(orders_even), "\n")  # 20

Professional example (Python)

import numpy as np


def robust_summary(values: np.ndarray, n_boot: int = 2000, seed: int = 0) -> dict[str, float]:
    """Median, MAD-based outlier flags and a bootstrap 95% CI for the median.

    Args:
        values: 1-D array of a skewed business metric (e.g. order value).
        n_boot: number of bootstrap resamples.
        seed: RNG seed for reproducibility.

    Returns:
        Dict with mean, median, MAD, outlier count and median CI.
    """
    rng = np.random.default_rng(seed)
    med = float(np.median(values))
    mad = float(np.median(np.abs(values - med)))
    robust_z = 0.6745 * (values - med) / mad          # MAD scaled to sigma under Normality
    boot_medians = np.median(
        rng.choice(values, size=(n_boot, values.size), replace=True), axis=1
    )
    lo, hi = np.percentile(boot_medians, [2.5, 97.5])
    return {
        "mean": float(values.mean()),
        "median": med,
        "mad": mad,
        "n_outliers_robust_z_gt_3.5": int(np.sum(np.abs(robust_z) > 3.5)),
        "median_ci_low": float(lo),
        "median_ci_high": float(hi),
    }


rng = np.random.default_rng(42)
order_values = rng.lognormal(mean=3.5, sigma=0.8, size=5_000)   # right-skewed, like real AOV
order_values[:10] = 5_000.0                                      # a few B2B whales
summary = robust_summary(order_values)
for key, val in summary.items():
    print(f"{key:>28}: {val:,.2f}")
print("Mean is pulled up by whales; the median barely moves.")

Source: domain knowledge

#probability-statistics#statistical-operations-and-random-variables#basic-statistical-operations