Historical Context & Motivation
Long before modern biostatistics emerged as a formal discipline, scientists and administrators grappled with the challenge of summarizing large collections of numerical observations into single representative values. The need was intensely practical: governments required efficient ways to report population data, astronomers needed to reconcile discrepant measurements of celestial positions, and physicians sought to characterize 'normal' vital signs. The intellectual journey toward the mean, median, and mode spans several centuries and reflects evolving perspectives on what it means to capture the 'center' of a data distribution.
This history underscores a fundamental question that persists in modern biostatistical practice: which single number best represents the center of a dataset, and under what conditions does each measure yield the most faithful summary? The answer depends on the shape of the distribution, the type of variable, and the inferential goals of the analysis — all considerations we explore in the sections that follow.
Core Principles & Definitions
Measures of central tendency serve a singular purpose: to distill a collection of observations into one value that captures where data points congregate. In biostatistics, where datasets may range from a handful of patient lab values to genome-wide expression arrays containing millions of entries, choosing the appropriate measure is not merely a computational decision — it is an inferential one. Each of the three classical measures — mean, median, and mode — encodes different assumptions about the data-generating process and responds differently to distributional features such as skewness, outliers, and multimodality.
Arithmetic Mean (x̄)
Median (x̃)
Mode
Sensitivity to Outliers
Relationship Under Symmetry & Skew
Visual Explanation — Distribution Shape and the Three Measures
The diagram above illustrates the fundamental geometric relationship among the three measures. Biostatistical datasets are frequently right-skewed — consider hospital costs, viral load counts, or time to disease recurrence — meaning that the mean will systematically overstate the 'typical' patient experience relative to the median. In clinical reporting, regulatory agencies such as the FDA therefore often request both the mean and median to ensure transparency, particularly when evaluating endpoints like overall survival or healthcare utilization. Recognizing the skewness signature at a glance, simply by noting how far the mean sits from the median, is one of the most useful quick-check skills in applied biostatistics.
Mathematical Framework
Each measure of central tendency can be defined formally, enabling rigorous derivation of its statistical properties. Below we present the formulas for sample computations — the versions most commonly applied in biostatistical practice — alongside key mathematical properties that motivate their use.
From an estimation-theoretic perspective, the mean is the maximum likelihood estimator (MLE) of the population center when the underlying distribution is normal. Its variance equals σ²/n, meaning it becomes more precise as the sample size grows — a property formalized by the law of large numbers and the central limit theorem. The median, by contrast, has an asymptotic variance of π/(2n) × σ² ≈ 1.57σ²/n under normality, making it roughly 57% less efficient than the mean for Gaussian data. However, for heavy-tailed distributions common in biological data (e.g., lognormal drug concentrations), the median can actually be more efficient than the mean — a fact that underscores why the choice between these measures is fundamentally distribution-dependent.
Choosing the Right Measure — A Decision Framework
Selecting the appropriate measure of central tendency is one of the first decisions in any biostatistical analysis, and the choice is governed by three interrelated factors: the measurement scale (nominal, ordinal, interval, ratio), the distributional shape (symmetric versus skewed), and the presence of outliers or censored observations. The flowchart below codifies these decision rules into a practical algorithm.
| Data Scenario | Recommended Measure | Biostatistical Example |
|---|---|---|
| Nominal variable | Mode | Most common blood type in a donor registry (e.g., O+) |
| Ordinal (e.g., pain scale 0–10) | Median | Median self-reported pain score after surgery |
| Symmetric, interval/ratio | Mean (± SD) | Systolic blood pressure in a normotensive population |
| Right-skewed, interval/ratio | Median (IQR) | Hospital length of stay, healthcare expenditures, viral load |
| Censored survival data | Median survival time | Median overall survival in an oncology trial (via Kaplan–Meier) |
Worked Example — Serum Creatinine in a Renal Study
A nephrology research team measured serum creatinine (mg/dL) in 11 patients presenting with early-stage chronic kidney disease. The observed values were: 0.9, 1.1, 1.0, 1.3, 1.2, 1.1, 4.8, 1.0, 1.2, 1.1, 1.3. Note that one patient (4.8 mg/dL) has a markedly elevated value, suggesting advanced renal impairment — an outlier that will differentially affect the three measures.
0.9, 1.0, 1.0, 1.1, 1.1, 1.1, 1.2, 1.2, 1.3, 1.3, 4.8. Sorting is essential for identifying both the median position and the frequency of repeated values.Strengths, Limitations & Comparisons
No single measure of central tendency is universally superior; each carries a distinct set of strengths and weaknesses. The table below provides a systematic comparison along dimensions that matter most in biostatistical practice: sensitivity to outliers, algebraic tractability, applicability across measurement scales, and behavior under different distributional assumptions.
| Property | Mean | Median | Mode |
|---|---|---|---|
| Uses all data points | Yes — every value contributes | No — depends only on rank order | No — depends only on frequency |
| Sensitivity to outliers | High | Low | None |
| Algebraic properties | Additive; unbiased estimator of μ; used in variance, ANOVA | Not additive; no closed-form in many distributions | Not algebraically tractable; may be non-unique |
| Measurement scales | Interval and ratio only | Ordinal, interval, ratio | All scales including nominal |
| Sampling variability | Lowest (under normality) | Slightly higher under normality; lower under heavy tails | Highly variable in small samples |
| Best suited when… | Data are symmetric with no extreme outliers | Data are skewed, censored, or contain outliers | Data are categorical or you seek the most typical category |
Connections to Advanced Statistical Theory
The mean, median, and mode are not isolated descriptive tools; they serve as conceptual and mathematical foundations for much of inferential biostatistics. Understanding how each measure connects to more advanced methods prepares you for the deeper analytical work encountered in clinical trial design, survival analysis, and machine learning for health data.
| Descriptive Foundation | Advanced Extension | Biostatistical Application |
|---|---|---|
| Sample mean (x̄) | Least-squares estimation; OLS regression coefficients minimize Σ(yᵢ − ŷᵢ)², extending the mean's squared-deviation minimization property | Modeling dose–response relationships, adjusting for confounders in RCTs |
| Sample median | Quantile regression; median regression minimizes Σ|yᵢ − ŷᵢ|, providing a robust alternative when residuals are non-Gaussian | Modeling healthcare costs, length of stay, and other right-skewed outcomes |
| Median survival time | Kaplan–Meier estimator; the median is the time at which the survival function S(t) = 0.50 | Primary endpoint in Phase III oncology trials; FDA approval decisions |
| Mode of a density | Maximum a posteriori (MAP) estimation in Bayesian inference; kernel density estimation | Identifying subpopulations in flow cytometry, pharmacokinetic modeling |
| Weighted mean | Meta-analysis (inverse-variance weighting); generalized estimating equations (GEEs) | Combining treatment effects across multi-site clinical trials |
One especially important bridge concept is the Central Limit Theorem (CLT), which states that the sampling distribution of the mean approaches normality as sample size increases, regardless of the underlying population distribution (provided finite variance). This theorem justifies the widespread use of the mean in parametric inference. For the median, an analogous result holds: its sampling distribution is asymptotically normal with variance 1/(4nf(x̃)²), where f(x̃) is the population density evaluated at the true median. The practical implication is that both confidence intervals and hypothesis tests can be constructed around either measure for sufficiently large samples — though the choice of measure still affects interpretive meaning and statistical power.
Practice Problems
Summary — Mean, Median & Mode in Biostatistics
The three classical measures of central tendency — mean, median, and mode — each answer the question 'Where is the center?' through a different lens. The arithmetic mean (x̄ = Σxᵢ/n) is the balance point that minimizes squared deviations and feeds directly into parametric methods like t-tests, ANOVA, and regression, but is sensitive to outliers and skewness. The median is the positional middle value — robust to extreme observations and the preferred summary for skewed biostatistical data such as hospital costs, survival times, and viral loads. The mode identifies the most frequent value and is the only measure applicable to nominal data (blood type, genotype).
In practice, the relationship among the three measures serves as a quick skewness diagnostic: when mean > median, the distribution is right-skewed; when mean < median, it is left-skewed; when all three coincide, the distribution is symmetric. Best practice in biostatistical reporting is to choose the measure that matches your data's measurement scale and distributional shape, report it alongside an appropriate measure of spread (SD for the mean, IQR for the median), and justify the choice explicitly — as required by CONSORT, STROBE, and other reporting guidelines.