The Shape of Distributions and Skew

A distribution's shape describes how data are spread across the range of possible values, independent of its center or spread in the ordinary sense of mean and variance. Three shape properties matter most in biology.

Symmetry means the left and right halves of the distribution (about its center) are mirror images. A symmetric distribution — the classic bell curve is one example, but not the only one — has mean, median, and mode coincide at the same value. Skewness is the departure from symmetry: one tail is longer or fatter than the other. By convention, skew is named for the direction of the long tail, not the direction of the bulk of the data. A right-skewed (positive-skew) distribution has a long right tail — most values are low, but some rare values are very high. A left-skewed (negative-skew) distribution has a long left tail — most values are high, with rare low values dragging the tail leftward.

Modality describes the number of peaks (local maxima of frequency) in a distribution. A unimodal distribution has one peak; a bimodal distribution has two; more generally, multimodal. Modality often carries direct biological meaning: a bimodal histogram of body mass in a sample of "adult birds" may signal that males and females were pooled together — two overlapping unimodal distributions summing to a two-humped shape. Recognizing modality is a diagnostic tool, prompting the biologist to ask what mixture of processes or populations produced this pattern — never to assume, without checking, that a single homogeneous population underlies the data. These shape properties connect directly to Histograms, the primary tool for visualizing them, and to The Mean, Median, and Mode, whose relative positions shift systematically with skew.

Step 1: Start from symmetry. In a perfectly symmetric unimodal distribution, the balancing point of the data (the mean, calculated as the sum of values divided by sample size) sits exactly at the middle value (the median) and at the most frequent value (the mode). All three coincide.

Step 2: Introduce a long tail. Now suppose a lower bound constrains the data — many biological variables cannot go below zero (mass, concentration, time, count) — but no comparable upper bound exists. The bulk of observations cluster near the lower bound, and a tail of increasingly rare, increasingly large values stretches rightward. This is a right-skewed distribution.

Step 3: See why the mean moves more than the median. The median is defined by rank — it only cares how many observations lie above or below, not by how much. Adding one extreme high value shifts the median by, at most, one rank position (negligible in a large sample). The mean is defined by magnitude — it is a sum divided by n, so one extreme value pulls the whole average toward it in direct proportion to its size. A single billionaire in a room of 99 average earners drags the mean income far above the median. The mean is pulled into the tail; the median resists.

Step 4: Read off the ordering. For a right-skewed unimodal distribution, the typical (though not universal) ordering is: ``` mode < median < mean ``` For a left-skewed unimodal distribution, the mirror image typically holds: ``` mean < median < mode ``` This ordering is a tendency, not a mathematical law — carefully constructed counterexamples exist — but it holds for the great majority of real biological distributions and is an extremely useful diagnostic rule of thumb (sometimes called the empirical relationship between mean, median, and mode, historically associated with Karl Pearson).

Step 5: Connect shape to biological mechanism. Skew is not a random nuisance; it is the fingerprint of the process generating the data.

• Bounded processes: reaction times, gestation lengths, and enzyme kinetics have a hard lower limit (cannot be negative or below a minimum biophysical threshold) but no firm upper limit — producing right skew.
• Multiplicative growth: many biological quantities (body mass, population size, income, gene expression levels) arise from the product of many small multiplicative factors rather than their sum. Multiplying many positive random factors together produces a log-normal distribution, which is inherently and strongly right-skewed on the raw scale (though symmetric after log-transformation).
• Ceiling effects: test scores near a maximum possible value, survival proportions near 100%, or traits under strong upper physiological constraint (e.g., maximal running speed in a well-trained population) pile up near an upper bound, producing left skew.
• Rare catastrophic or rare beneficial events: a small number of individuals surviving an epidemic, or a small number of seeds germinating under stress, generates a long tail of rare successes or failures.

Step 6: Handling skew statistically. Because the mean is sensitive to skew, biologists often (a) report the median and interquartile range instead of, or alongside, the mean and standard deviation for skewed data; (b) apply a transformation (commonly a log transform) to make a right-skewed distribution more symmetric before applying methods, like the t-test, that assume normality; or (c) use nonparametric statistical tests that make no assumption of symmetry. These strategies foreshadow later tools such as the Coefficient of Variation, which itself depends on how spread scales with the mean, and formal methods for identifying and handling Outliers in Biological Data, which are far more common in skewed distributions than in symmetric ones.

The formal study of asymmetric distributions was launched by the English statistician Karl Pearson in the 1890s. Pearson was analyzing biometric data — crab measurements from the Bay of Naples collected by his collaborator Walter Weldon — and encountered histograms of forehead-to-body-length ratios that would not fit a single normal (Gaussian) curve. Instead of one bell curve, some samples showed a subtle asymmetry, and Weldon initially wondered whether this indicated two overlapping crab populations — potentially, incipient divergence into two species, a question Weldon posed directly to test Darwinian evolution's operation in real time. Pearson approached the puzzle mathematically. He: 1. Defined a family of frequency curves (the Pearson distributions) flexible enough to describe skewed, not just symmetric, data — a major methodological advance beyond assuming everything is normal. 2. Derived, from these curves, an empirical relationship connecting the three measures of central tendency, expressed today in the approximate rule: ``` mean − mode ≈ 3 × (mean − median) ``` equivalently rearranged as the commonly cited Pearson's second skewness coefficient: ``` skewness ≈ 3 × (mean − median) / standard deviation ``` 3. Applied this apparatus to the crab data and showed that what looked like one puzzling asymmetric hump could be statistically decomposed into a mixture of two normal distributions with different means — precisely the two-population signature that shows up as bimodality (or, when the two component peaks are close together, as skew) in a combined histogram. Pearson's analysis is historically important for two reasons. First, it demonstrated quantitatively — not just by eyeballing a histogram — that a seemingly single-peaked, skewed dataset could conceal two underlying subpopulations, giving biologists a rigorous statistical tool to detect population heterogeneity from shape alone. Second, it established the general principle that shape carries information: deviations from symmetry are not noise to be ignored but a signal about the process — mixture, boundary, multiplicative growth — that generated the data. Later empirical confirmation of these mechanisms came from countless field studies. A now-classic illustration is body-mass or wing-length data in sexually dimorphic bird and insect species: pooling males and females routinely produces distributions with two clear modes, resolved instantly by separating the sexes — precisely the diagnostic Pearson's method formalized. Similarly, studies of income, city size, and gene expression levels have repeatedly confirmed that multiplicative generating processes produce log-normal, strongly right-skewed distributions on the raw scale, becoming symmetric only after log-transformation — a pattern independently verified across economics, demography, and molecular biology throughout the 20th century.

Example 1 — Right-skewed clutch sizes and the mean/median gap. A field ecologist records the number of eggs in 11 nests of a songbird: ``` 2, 2, 3, 3, 3, 4, 4, 4, 4, 5, 15 ``` (The value 15 comes from an unusual double-clutch nest where two females laid in the same box.)

• Median: the middle (6th) value of the ordered list is 4.
• Mode: the most frequent value is 4 (appearing four times).
• Mean: sum = 2+2+3+3+3+4+4+4+4+5+15 = 49; 49 / 11 = 4.45.

Ordering: mode (4) ≈ median (4) < mean (4.45). The single unusual nest (15 eggs) pulls the mean well above the typical clutch, while the median — insensitive to the magnitude of that outlier, only its rank — stays anchored at the typical value. If the biologist reported only "mean clutch size = 4.45," a reader would wrongly conclude the typical nest has nearly 4.5 eggs; in reality, 9 of 11 nests (82%) have 4 eggs or fewer. This is exactly why field ecologists reporting clutch size, litter size, or brood size routinely lead with the median and note the skew, reserving the mean for total population-level egg production (where the sum, and hence the mean, is the biologically relevant quantity — e.g., total reproductive output across the colony).

Example 2 — Log-normal enzyme kinetics and choosing a transformation. A biochemist measures the time (in seconds) for 8 replicate reactions to reach 50% substrate conversion: ``` 12, 14, 15, 16, 18, 22, 30, 68 ```

• Mean = (12+14+15+16+18+22+30+68)/8 = 195/8 = 24.4 s
• Median = average of 4th and 5th ordered values = (16+18)/2 = 17 s
• Mode: no exact repeats; the distribution is unimodal but the bulk clusters near 12–18 s with one high outlier (68 s).

Mean (24.4) far exceeds median (17): a classic right-skew signature, driven by the single slow replicate at 68 s (perhaps a well with a pipetting error, or genuine biological variability in enzyme activity). Reaction-time data of this kind is frequently log-normal because reaction rate depends multiplicatively on temperature, enzyme concentration, and substrate concentration together. Taking the natural log of each value: ``` ln(12)=2.48, ln(14)=2.64, ln(15)=2.71, ln(16)=2.77, ln(18)=2.89, ln(22)=3.09, ln(30)=3.40, ln(68)=4.22 ``` The log-transformed mean is 3.02 and log-transformed median is 2.83 — still not identical (real data is never perfectly log-normal), but the ratio of the discrepancy shrinks enormously (from a raw gap of 7.4 s to a log-scale gap of 0.19 log-units), showing the transformation has pulled the distribution much closer to symmetric. A biostatistician would report geometric-mean reaction time (the back-transformed mean of the logs, e^3.02 ≈ 20.5 s) as a more representative "typical" value than the raw arithmetic mean of 24.4 s, and would use the log-transformed data for any t-test or ANOVA comparing enzyme conditions, since those tests assume roughly symmetric, normally distributed residuals.

Karl Pearson (1857–1936) was a British applied mathematician who became one of the founders of modern mathematical statistics. Trained originally in law and mathematics at Cambridge, Pearson turned his attention to biology and evolution after being recruited into Francis Galton's circle of biometricians, who sought to quantify Darwinian variation and heredity with rigorous statistical tools rather than descriptive natural history. Pearson founded the journal Biometrika in 1901 (with Weldon and Galton) specifically to give quantitative evolutionary biology a home, and he built much of the mathematical machinery — correlation coefficients, chi-squared goodness-of-fit tests, and the Pearson system of frequency curves — that biology and other empirical sciences still rely on. His empirical work on skewed, non-normal biological data (crab shell measurements, human anthropometric data) directly motivated his development of skewness coefficients and his classification of distributions beyond the normal curve, cementing the idea that shape, not just mean and variance, is a fundamental statistical property worth measuring precisely.

Distribution shape is a working diagnostic in nearly every quantitative branch of biology and medicine. In clinical laboratory medicine, reference ranges for markers like triglycerides, C-reactive protein, and viral load are reported using percentiles or medians rather than means, precisely because these measurements are right-skewed in healthy populations — reporting a mean would overstate the "normal" level and cause over-flagging of healthy patients. In conservation biology and ecology, skewed distributions of population sizes, dispersal distances, or reproductive success (many individuals contribute little, a few contribute enormously — a phenomenon central to effective population size calculations) directly shape extinction-risk models and harvest quotas; ignoring the skew and using a raw mean can drastically misestimate a sustainable catch or the true breeding population. In genomics and gene expression analysis, RNA sequencing read counts and expression levels are classically log-normal or negative-binomial (both right-skewed), which is why differential expression software (e.g., DESeq2, edgeR) explicitly models or log-transforms count data rather than assuming raw normality. In epidemiology, incubation periods, viral load trajectories, and hospital length-of-stay are right-skewed, and public-health reporting increasingly favors median incubation period (as famously reported for COVID-19, roughly 5–6 days median versus a longer mean pulled by rare extended cases) over means to avoid misleading the public about typical risk timelines. In agriculture, crop yield distributions across fields are often left-skewed near an achievable maximum yield ceiling, informing breeding targets and irrigation thresholds. In every case, the first analytic step is the same one this article teaches: plot the histogram, read its shape, and only then decide whether the mean, median, or a transformed statistic honestly represents the biology — a judgment call that becomes especially important once outliers or unusually variable samples are in play, the subject of Outliers in Biological Data.

Ask the tutor: If a dataset is bimodal, is it ever still appropriate to compute a single overall mean and standard deviation for it?