The Mean, Median, and Mode

A measure of central tendency is a single number intended to represent the "center" or "typical value" of a set of measurements. Given a sample of n observations, $x_1, x_2, \dots, x_n$, from a population, the three classical measures are defined as follows. The sample mean, denoted $\bar{x}$, is the sum of all observations divided by the number of observations:

The corresponding population parameter is denoted $\mu$ (recall the Parameters versus Statistics distinction: $\bar{x}$ estimates $\mu$ from a sample). The mean is a linear, all-data-using statistic — every single value contributes to it, weighted equally. The median is the middle value of the data once ranked from smallest to largest. If $n$ is odd, the median is the value at rank $(n+1)/2$. If $n$ is even, it is the average of the two middle values, at ranks $n/2$ and $n/2 + 1$. The median is a rank-based, order-statistic — it depends only on the position of values, not their magnitude, which is precisely why extreme outliers cannot drag it far. The mode is the value (or values) that occurs with the highest frequency in the dataset. Unlike the mean and median, the mode is defined even for nominal (categorical) data — for example, the modal blood type in a population, or the modal number of offspring in a discrete count distribution. A distribution can have one mode (unimodal), two modes (bimodal), or more (multimodal); a perfectly uniform distribution has no unique mode at all. These three statistics are related to the scale of measurement of the variable. Nominal data (blood type, sex, species) admit only a mode. Ordinal data (developmental stage rank, Likert-type behavioral scores) admit a mode and a median. Interval and ratio data (mass, length, concentration, temperature) admit all three. This hierarchy matters: computing "the mean species" is meaningless, but computing "the mean body mass" is not.

Step 1: Computing the mean. Every observation is summed, and the total is divided by the count. Because every value enters the sum with equal weight $1/n$, a single very large or very small value can shift $\bar{x}$ substantially — the mean has no "resistance" built into its formula. Mathematically, the mean is the value $c$ that minimizes the sum of squared deviations $\sum(x_i - c)^2$ — it is the least-squares center, which is also why it connects naturally to the variance (foreshadowing Variance and Standard Deviation, which is built from squared deviations around this same mean).

Step 2: Computing the median. All values are sorted. For odd $n$, one simply reads off the middle value. For even $n$, the two central values are averaged. Because the median only "counts positions," an observation can be moved to any arbitrarily large or small value without changing the median at all, so long as it stays on the same side of the middle. This is the origin of the median's robustness — formally, its breakdown point (the fraction of data that must be corrupted before the statistic becomes arbitrarily wrong) is 50%, versus 0% for the mean, whose single most extreme value can shift it without limit.

Step 3: Computing the mode. A frequency table (or histogram, for continuous data binned into classes) is built, and the class with the highest count is identified. For continuous data, the mode is somewhat bin-dependent — the modal class can shift depending on how finely the data are binned — which is one reason the mode is used less often for continuous ratio-scale traits and more often for discrete counts or categories.

Step 4: How skew separates the three. Picture a right-skewed distribution — like parasite egg counts, where most hosts carry few parasites but a handful carry very many. The bulk of the data (and hence the mode, the most frequent low value) sits on the left. The median sits somewhat to the right of the mode, because half the ranked observations must lie above it, including the long tail. The mean sits furthest to the right, because the arithmetic sum is inflated by the rare extreme values. The ordering for right-skewed data is:

For left-skewed data, the inequality reverses. For perfectly symmetric unimodal distributions (such as an idealized Gaussian), all three coincide exactly: $\text{mode} = \text{median} = \text{mean}$.

Step 5: Multimodality reveals hidden structure. If a histogram of an enzyme activity assay shows two separate humps, the mode (or rather, the two modes) correctly signals that the sample may be a mixture of two biological subgroups — for example, individuals homozygous for a fast allele and individuals homozygous for a slow allele at a metabolic locus, with heterozygotes bridging the two peaks. In such a case, computing a single mean is actively misleading: it may fall in the valley between the two peaks, a value that describes almost no individual in the sample. This is a case where the mode, or better, a description of the whole shape, is biologically superior to the mean.

Step 6: Weighted and grouped extensions. In practice biologists often compute means from grouped or weighted data — for example, a mean litter size across several sites, weighted by the number of litters observed at each site (a weighted mean, $\bar{x}_w = \sum w_i x_i / \sum w_i$). The median and mode extend analogously to grouped frequency data using standard interpolation formulas, though these refinements are secondary to understanding the basic logic above.

Unlike Mendel's peas or Meselson–Stahl's density gradients, the mean/median/mode story is not a single decisive experiment but a convergence of mathematical proof and empirical demonstration across the 18th–20th centuries, showing exactly when each measure is trustworthy.

Gauss and the theory of errors (early 1800s). Carl Friedrich Gauss, working on astronomical measurement error, proved that if measurement errors are normally (symmetrically, bell-curve) distributed around a true value, the arithmetic mean is the value that minimizes the sum of squared errors — and is the maximum-likelihood estimator of the true value. This gave the mean a rigorous justification, but crucially the proof assumed symmetric, light-tailed error distributions. Gauss's derivation is why the mean became the default statistic in the physical sciences: physical measurement error genuinely tends to be symmetric.

Quetelet and the "average man" (1830s–1840s). The Belgian astronomer-turned-social-statistician Adolphe Quetelet applied Gaussian reasoning to human biology, measuring chest circumferences of Scottish soldiers and heights of French conscripts. He found these anthropometric traits followed approximately bell-shaped (binomial/normal-like) distributions, and popularized the mean as representing l'homme moyen, "the average man." This was an important empirical demonstration that some biological traits (like adult height, controlled by many small additive genetic and environmental factors) are naturally symmetric — vindicating use of the mean for such traits.

Francis Galton and the discovery of skewed biological distributions (1870s–1880s). Galton, studying inheritance of traits like sweet-pea seed size and human stature, was among the first to systematically notice that many biological variables are not symmetric. He introduced the concept of the median as the value corresponding to the 50th percentile in his ogive (cumulative frequency) plots, explicitly because it worked even when a distribution was skewed and gave a value that half of individuals exceeded and half fell short of — a property invariant to the shape of the tail. Galton's work on regression toward the mean also depended on carefully distinguishing the mean of a distribution from its typical (median) member.

Karl Pearson and formal skewness measures (1890s). Pearson, building directly on Galton, formalized the empirical relationship

as a rough rule of thumb for moderately skewed unimodal distributions, and introduced quantitative coefficients of skewness based on exactly the divergence between mean and mode (or mean and median). This turned the informal observation "mean and median disagree when data are skewed" into a precise, testable, and widely used empirical/statistical tool — establishing the mean-versus-median comparison as a formal diagnostic still taught today (and foreshadowing The Shape of Distributions and Skew).

Modern robustness theory (mid-20th century, Tukey and others). John Tukey and collaborators formalized the concept of statistical robustness and the breakdown point, proving mathematically that the median's breakdown point is 50% (up to half the data can be corrupted before the median is thrown off arbitrarily) while the mean's breakdown point is 0% (a single sufficiently extreme value can move the mean without bound). This work, grounded in simulation and proof rather than a single experiment, gave rigorous backing to the long-standing empirical intuition that the median should be preferred whenever a dataset may contain heavy-tailed noise or contamination — directly relevant to biological data riddled with recording errors, rare pathological individuals, or genuine biological extremes. Together, these threads of evidence — Gauss's optimality proof for symmetric errors, Quetelet's empirical confirmation of symmetry in some traits, Galton's empirical discovery of skew in others, Pearson's quantification of the divergence, and Tukey's robustness proofs — establish exactly the conditional rule taught today: use the mean when data are (approximately) symmetric and free of extreme outliers; use the median when data are skewed or outlier-prone; use the mode for categorical or clearly multimodal data.

Example 1: A skewed clutch-size dataset (turtle ecology). A herpetologist records the number of eggs per clutch for 11 female painted turtles at a nesting site:

(The last female, an unusually large individual, laid an exceptional clutch of 22 eggs.) Mean:

Median: With $n = 11$ (odd), the median is the 6th ranked value in the already-sorted list above: 6 eggs. Mode: The most frequent value is 6, appearing three times: mode = 6 eggs. Notice the ordering $\text{mode} = \text{median} = 6 < \text{mean} \approx 7.73$, exactly the right-skew signature predicted by the mechanism above. If the researcher reported only the mean, "a typical clutch has about 7.7 eggs," this value describes no actual female in the sample — ten of eleven females laid 9 eggs or fewer. The median (6 eggs) and the mode (6 eggs) both better represent what a "typical" female actually does; the mean is being dragged upward by one exceptional individual. This is precisely the situation in which ecologists routinely report median clutch size, or note the outlier separately, rather than lead with the mean (foreshadowing the treatment of Outliers in Biological Data). If we now ask what happens to each statistic if that one outlier female is removed from the sample entirely ($n = 10$):

The mean drops from 7.73 to 6.3 — a swing of 1.43 eggs from removing a single individual out of eleven. The median, recomputed as the average of the 5th and 6th ranked values (6 and 6), stays at 6 — completely unchanged. This numerically demonstrates the median's robustness and the mean's sensitivity in a single concrete case.

Example 2: A bimodal enzyme-activity dataset (pharmacogenetics). A clinical researcher measures the activity of a drug-metabolizing liver enzyme (units per mg protein) in 12 patients, suspecting the gene has a common fast-metabolizer allele and a common slow-metabolizer allele:

Mean:

Median: $n = 12$ (even), so the median is the average of the 6th and 7th ranked values, 44 and 45:

Mode: Two values tie for highest frequency (14, appearing twice, and 46, appearing twice) — but visually the data clearly form two clusters, around ~14 and ~46, so the dataset is better described as bimodal (two modal regions, roughly 14 and 46) than by any single number. Here the mean (32.75) is a disaster as a summary: it falls in the empty gap between the two clusters and corresponds to zero actual patients — nobody in the sample has enzyme activity anywhere near 32.75. Neither the mean nor even the median (44.5) is fully honest about the data structure; the correct biological conclusion is that this population contains (at least) two genetically distinct metabolizer phenotypes, and the right "summary" is not one number but two group means (e.g., mean ≈ 13.6 for the slow group, mean ≈ 46.8 for the fast group) plus the proportion of patients in each. This is the clearest illustration of why a single measure of central tendency, chosen without first looking at the shape of the distribution, can misrepresent biology entirely — and why in pharmacogenetics, identifying such subpopulations (e.g., CYP2D6 poor vs. extensive metabolizers) is itself a central diagnostic goal, not a nuisance to be averaged away.

Francis Galton (1822–1911) is the pivotal figure who forced biologists to take seriously that "the average" is not a neutral, universally correct summary. A half-cousin of Charles Darwin, Galton was a polymath — explorer, meteorologist, and founder of biometrics and (controversially) eugenics — who became obsessed with quantifying heredity after reading On the Origin of Species. Working with sweet-pea seed diameters (chosen because peas self-fertilize, simplifying the genetics) and later human stature and other traits, Galton plotted cumulative frequency curves ("ogives") and needed a robust way to define the "typical" value of populations that were not always symmetric. He formalized the median as the 50th-percentile point on these curves and, in his work on regression, was careful to distinguish the median offspring value from the mean parental value — a distinction essential to his discovery of regression toward the mean. Galton's insistence on measuring real biological variation, rather than assuming it away as Quetelet's idealized "average man" had done, is why the median entered biological statistics as a serious rival to the mean, not merely an arithmetic curiosity.

Clinical medicine and pharmacology. Drug half-lives, hospital length-of-stay, and time-to-relapse data are almost always right-skewed, so clinical trial reports typically present the median survival time (with confidence intervals) rather than the mean, because a handful of long-surviving patients would otherwise inflate the mean into a number that overstates typical prognosis. Pharmacogenetic screening for drug-metabolizing enzyme variants (as in Example 2) explicitly searches for bimodal or multimodal distributions of enzyme activity to classify patients into "poor," "intermediate," and "extensive" metabolizer categories — a direct clinical application of recognizing when the mode, not the mean, carries the diagnostic signal, guiding dose adjustment for drugs like codeine and warfarin.

Ecology and conservation biology. Population size distributions, home-range areas, and dispersal-distance data are usually strongly right-skewed (most individuals disperse a short distance; a few disperse very far). Conservation biologists reporting "typical dispersal distance" for a reintroduction plan will substantially underestimate the area needed for a reserve if they mistakenly use the mean dispersal distance pulled up by rare long-distance dispersers, or conversely will over-plan if they ignore that the mean (not median) may be the ecologically relevant quantity for predicting rare colonization events — the choice of statistic must match the biological question being asked.

Genomics and gene expression. RNA-sequencing read counts and gene expression levels are typically right-skewed and are often reported using median-based normalization (e.g., median-of-ratios normalization in differential expression pipelines) precisely because a few extremely highly expressed genes would distort mean-based normalization factors across samples. Detecting bimodal expression of a gene across single cells is now a standard signal used to identify distinct cell subtypes or bistable regulatory states within an ostensibly uniform tissue sample.

Agriculture and animal breeding. Breeding programs selecting for traits like milk yield or litter size must decide whether to report and select on mean, median, or modal performance across a herd; skewed distributions (a few outstanding, high-yield animals) mean that mean-based selection differentials can be biased by a small number of exceptional individuals, and breeders often supplement mean-based indices with robust (median-based) estimates to avoid overweighting statistical flukes when choosing sires for artificial insemination programs.

Public health surveillance. Case counts, hospital admission rates, and epidemic curve data are frequently reported using both mean and median case counts over rolling windows; because outbreak reporting is prone to reporting-day artifacts and occasional data-entry outliers, public health dashboards increasingly favor 7-day medians over means specifically for their resistance to single anomalous data points, directly applying the robustness property established by Tukey's mid-20th-century theory.

Ask the tutor: If my dataset is only mildly skewed, is it ever acceptable to report the mean anyway — and how would I justify that choice in a lab report?