Median vs. Mode

The reduction of complex, high-dimensional empirical distributions into singular, representative metrics lies at the very heart of exploratory data analysis, inferential statistics, and scientific inquiry. In every branch of empirical research—from microeconomic wage profiling and algorithmic latency tracking to epidemiological surveillance and machine learning feature engineering—investigators face the fundamental challenge of summarizing vast arrays of observations without completely obliterating the underlying structure of the phenomenon under study. Within this analytical paradigm, measures of central tendency operate as foundational mathematical anchors, attempting to capture the typical, central, or most probable values around which a collection of data clusters. However, the conceptualization of what is truly “central” is neither self-evident nor mathematically monolithic. Depending on the geometry of the distribution, the scale of measurement, and the specific risk or loss function under consideration, central tendency reveals itself through markedly distinct mathematical manifestations.

Among these measures, the median and the mode represent two of the most indispensable yet fundamentally divergent statistical constructs. While pedagogical traditions routinely group them alongside the arithmetic mean as a standard triad of introductory summary statistics, the median and the mode derive from completely disparate mathematical philosophies. The median is rooted in order statistics, sorting permutations, and cumulative probability distribution functions, representing the exact positional division that bisects an ordered sample space into equiprobable halves. In contrast, the mode belongs to the realm of frequency mechanics, density maximization, and discrete event repetition, isolating the coordinate of maximal likelihood or highest localized empirical concentration. Consequently, the operational behaviors, algebraic properties, computational complexities, and domain applications of these two metrics diverge radically across different operational environments.

Conflating the median and the mode or deploying them interchangeably introduces substantial analytical bias, often corrupting downstream statistical modeling, misguiding macroeconomic policy decisions, and producing distorted operational conclusions. A profound grasp of their technical distinctions is therefore non-negotiable for practitioners across data science, academic research, and engineering. This comprehensive treatise offers an exhaustive, mathematically rigorous exploration of the median and the mode. Across eight extensive sections, we examine their formal definitions, historical origins, structural behaviors, computational mechanisms, scale-level constraints, non-parametric robustness properties, and practical deployment paradigms, establishing a definitive reference for rigorous empirical research and statistical reporting.

1. Introduction to Central Tendency: Median vs. Mode

1.1 Defining Measures of Central Tendency

In quantitative methodologies, summary statistics function as dimensionality reduction mechanisms designed to synthesize extensive numerical sequences or frequency distributions into compact, interpretable descriptive parameters. Rather than requiring an analyst to examine an entire set of raw, unaggregated observations, a measure of central tendency provides a single numerical anchor intended to locate the core, center, or typical point of gravitational pull within the observed data space. The mathematical conceptualization of central tendency, however, spans several discrete analytical categories: algebraic means, which compute a weighted center of mass based on additive metric values; positional markers, which identify quantile thresholds derived from rank-ordered permutations; and frequency peaks, which isolate clusters of maximal recurrent density across discrete or discretized states.

Selecting an appropriate metric of central tendency is the bedrock of rigorous exploratory data analysis (EDA) and valid inferential modeling. Every central tendency estimator implicitly or explicitly operates upon a distinct underlying mathematical objective function. When an empirical researcher summarizes an observational vector, the chosen metric fundamentally dictates what forms of error or dispersion are penalized, what scale properties are respected or violated, and how peripheral fluctuations influence the estimated parameter. Inappropriate metric selection can obscure critical distribution topologies—such as heavy right tails, polymodal cluster splits, or nominal categorizations—generating deceptive inferences that can compromise scientific validity and degrade decision-making pipelines in commercial and governmental environments.

Furthermore, central tendency cannot be studied in complete isolation from measures of statistical dispersion. The validity and explanatory power of a central metric directly depend on how tightly the surrounding observations adhere to that anchor. An unreflective computation of central tendency across a wildly dispersed or multi-clustered continuum yields a synthetic abstraction that may represent zero actual entities in the physical world. Therefore, understanding the mechanics governing how individual observations aggregate around a central value constitutes the primary prerequisite for all advanced quantitative reasoning, predictive modeling, and empirical hypothesis testing.

1.2 The Fundamental Distinction Between Median and Mode

The structural divergence between the median and the mode stems from their foundational reliance on entirely separate computational primitives: ordinal ranking versus occurrence frequency. The median ($M$ or $Mdn$) is an inherently order-dependent, rank-based positional metric. Conceptually defined as the exact 50th percentile rank of an ordered series, the median identifies the boundary condition that divides a sorted collection of observations into two equal, equiprobable subsets, ensuring that precisely half of the data points fall at or below the median, and half fall at or above it. Its existence is contingent upon the capacity to establish a monotonic greater-than or less-than relation between observations, operating as a quantile marker that remains completely indifferent to the arithmetic magnitudes of the values residing in the peripheral extremes.

Conversely, the mode ($Mo$) is fundamentally a count-dependent, frequency-density metric. Rather than mapping the positional hierarchy or rank order of observations, the mode identifies the specific discrete value, category, or coordinate within a continuous density function that exhibits the highest absolute frequency or concentration of observations. It answers an entirely different empirical question than the median: instead of inquiring where the exact geometric or rank midpoint of the sample lies, the mode asks which individual observation or categorical state occurs with the greatest empirical probability or repetition. It reflects the peak of a distribution’s probability mass function (PMF) or probability density function (PDF), operating as a metric of commonality rather than a metric of ordinal balance.

This deep mechanical contrast means that the median and the mode can deviate sharply from one another within the exact same empirical dataset. For instance, in an asymmetric socio-economic distribution characterized by a long right-hand tail, the modal wage may correspond to the legally mandated minimum wage—since millions of workers might cluster at that specific statutory pay grade—while the median wage will locate the balance point where half the national workforce earns more and half earns less, positioning itself substantially higher than the mode. The median measures balance along an ordered continuum; the mode measures localized empirical saturation. Recognizing this distinction is vital for maintaining rigorous descriptive integrity.

1.3 Why Median and Mode Are Frequently Confused

Despite their distinct mathematical formulations, the median and the mode are persistently conflated in academic literature, business intelligence, and public discourse. A primary catalyst for this confusion is the historical pedagogical convention of teaching the “three Ms”—mean, median, and mode—as an interchangeable triad of descriptive options within introductory primary and secondary mathematics curricula. Students are routinely instructed on the procedural mechanics of finding the average, the middle, and the most frequent value across small, artificial toy datasets (such as collections of single-digit integers) without receiving a grounding in measurement scale theory, loss minimization paradigms, or distributional topology. This pedagogical oversimplification fosters the erroneous belief that these metrics are merely alternative computational routes to the same underlying construct.

This conflation is exacerbated by the behavior of symmetrical, unimodal bell curves—most notably the standard Gaussian (normal) distribution, the Student’s $t$-distribution, and the logistic distribution. In any perfectly symmetrical, unimodal distribution, the arithmetic mean, the positional median, and the frequency mode mathematically converge upon the exact same coordinate. Because introductory statistics heavily emphasizes Gaussian environments due to the Central Limit Theorem and classical parametric testing, practitioners frequently develop an intuitive cognitive bias that assumes central tendency metrics should always approximate one another. When applied to real-world phenomena—where distributions are routinely skewed, multi-peaked, truncated, or categorical—this intuition collapses, leaving researchers confused when the metrics diverge wildly.

Finally, broad misunderstandings surrounding measurement scale constraints, as formalized by psychologist S. S. Stevens in his seminal 1946 taxonomy, further compound this confusion. Practitioners often fail to recognize that the permissible operational domains for these metrics are strictly dictated by whether the data is nominal, ordinal, interval, or ratio. Computing the median requires an underlying order topology that nominal variables fundamentally lack; conversely, identifying the mode requires frequency counting that is conceptually trivial on categorical labels but analytically complex across continuous, non-repeating real numbers. When analysts treat datasets as generic spreadsheets of numbers without examining the scale level of their variables, the boundaries between median and mode blur, leading to inappropriate analytical applications.

2. Understanding the Median in Descriptive Statistics

2.1 Mathematical Definition and Positional Logic

In formal probability theory and descriptive statistics, the median is defined as the value that bisects a probability distribution or empirical sample space into two equal halves. Formally, for a real-valued random variable $X$ characterized by a cumulative distribution function (CDF) denoted by $F(x) = P(X le x)$, the median $m$ is any real value that simultaneously satisfies the dual inequality conditions:

$$P(X le m) ge 0.5 \quad \text{and} \quad P(X ge m) ge 0.5$$

For absolutely continuous random variables whose cumulative distribution function is strictly monotonic and continuous, the median is uniquely defined as the specific quantile solution to the inverse cumulative distribution function (the quantile function $Q(p)$) evaluated precisely at $p = 0.5$, such that $F(m) = 0.5$, or equivalently, $m = F^{-1}(0.5)$. In this continuous framework, the median represents the vertical coordinate that integrates the probability density function $f(x)$ into two equal areas of exactly 0.5 across the intervals $(-\infty, m]$ and $[m, \infty)$.

In discrete empirical samples comprising $n$ observations, the median is obtained through an algorithmic sorting procedure. Let the sample be denoted as $X = {x_1, x_2, dots, x_n}$. The data points must first be arranged in a non-decreasing monotonic sequence to yield the order statistics, denoted by $x_{(1)} le x_{(2)} le dots le x_{(n)}$, where $x_{(1)}$ represents the sample minimum and $x_{(n)}$ represents the sample maximum. The positional identification of the median depends entirely upon the parity of the sample size $n$:

• Odd Sample Sizes: When $n$ is odd, the division of the sample produces an unambiguous, unique central element. The position of the median is indexed by the rank $k = (n + 1) / 2$, and the sample median is calculated directly as:
  $$M = x_{\left(\frac{n+1}{2}\right)}$$
  In this scenario, exactly $(n – 1) / 2$ observations are smaller than or equal to $M$, and exactly $(n – 1) / 2$ observations are larger than or equal to $M$.
• Even Sample Sizes: When $n$ is even, no single central observation exists. Instead, the middle of the distribution falls between the two central order statistics indexed by ranks $k_1 = n / 2$ and $k_2 = (n / 2) + 1$. Any value within the closed real interval $[x_{(n/2)}, x_{(n/2 + 1)}]$ satisfies the formal definition of dividing the sample into two equal subsets. However, to preserve uniqueness in standard empirical applications, descriptive statistics adopts the standard linear convention of taking the arithmetic interpolation of these two middle boundaries:
  $$M = \frac{x_{\left(\frac{n}{2}\right)} + x_{\left(\frac{n}{2} + 1\right)}}{2}$$

A crucial mathematical property of the median is its direct relationship to the minimization of the $L_1$ norm. While the arithmetic mean is the parameter that minimizes the sum of squared Euclidean deviations ($L_2$ norm), the median is the optimal parameter $c$ that minimizes the sum of absolute linear deviations across all observations:

$$M = arg\min_c \sum_{i=1}^n |x_i – c|$$

This foundational property, originally highlighted by French mathematician Antoine Augustin Cournot in 1843 and subsequently expanded by Gustav Theodor Fechner in 1874, proves that the median is the optimal point predictor under a linear (or absolute) loss function. Furthermore, the median displays strict equivariance under any strictly monotonic transformation $g(\cdot)$. If an invertible function $g$ is applied to every observation such that $y_i = g(x_i)$ where $g'(x) > 0$, the median of the transformed sequence is precisely the transformation of the original median: $M(g(X)) = g(M(X))$. This property does not hold for the arithmetic mean, meaning that transforming data via logarithmic, exponential, or root functions preserves the exact rank-positional integrity of the median.

2.2 Resistance to Outliers and Distributional Skewness

The defining structural attribute of the median in applied descriptive statistics is its robust resistance to extreme outliers, heavy distributional tails, and data recording anomalies. In formal robust statistics, the resistance of an estimator is quantified through its asymptotic breakdown point—defined as the smallest proportion of arbitrary, infinitely corrupted observations that an estimator can handle before yielding an arbitrarily large or completely uninformative estimate. The arithmetic mean exhibits a catastrophic breakdown point of $1/n$ (which asymptotically approaches 0% as $n to \infty$), meaning that a single extreme outlier can pull the mean to positive or negative infinity regardless of the remaining sample size.

In contrast, the sample median possesses a breakdown point of:

$$\epsilon^* = \frac{1}{n} \left\lfloor \frac{n + 1}{2} \right\rfloor \approx 50%$$

This 50% threshold represents the highest possible breakdown point achievable by any translation-invariant location estimator. In practical terms, it means that nearly half of the observations in a dataset can be corrupted, shifted to extreme values, or replaced with gross computational errors without displacing the median beyond the range of the uncorrupted central observations. Because the median relies strictly on ordinal rank positions rather than aggregate metric sums, an observation located at the 99th percentile exerts the exact same influence on the median whether its numerical value is 100, 10,000, or 10,000,000.

This resistance makes the median essential for modeling asymmetric, heavily skewed real-world distributions. In fields like microeconomics, wealth distributions and household incomes regularly display extreme positive (right-hand) skewness governed by Pareto-like power laws. In such environments, a tiny fraction of hyper-wealthy individuals skews the arithmetic mean upward, producing an inflated summary value that misrepresents the living standards of typical citizens. The median remains immune to these extreme right-tail dynamics, reliably reporting the exact economic threshold that bisects the population. Similarly, in site reliability engineering, network latency metrics are prone to intermittent, order-of-magnitude performance spikes caused by network timeouts or garbage-collection pauses. Tracking the 50th percentile latency (p50) ensures that infrastructure engineers can assess baseline system performance without their metrics being distorted by transient system anomalies.

2.3 Methodological Strengths and Theoretical Limitations

The primary methodological strength of the median lies in its interpretability as an empirical divider. When an analyst states that the median home price in a metropolitan area is $450,000, the empirical meaning is immediate and actionable: exactly 50% of the housing stock transacted below this figure, and 50% transacted above it. Furthermore, because the median requires only ordinal-level properties, it can be computed on ordered qualitative data (such as Likert scales, socioeconomic strata, or academic grade bands) where computing an arithmetic mean is mathematically invalid due to non-uniform intervals between categories.

Despite these strengths, the median suffers from notable theoretical and mathematical limitations that restrict its utility in advanced parametric modeling. Chief among these is its lack of algebraic manipulability. Medians cannot be aggregated across separate subgroups to calculate an overall grand median. For example, if an analyst knows the median income and sample size of five independent regional districts, it is mathematically impossible to compute the median income of the combined nation from these parameters alone; one must reconstruct the pooled order statistics from the raw data. In contrast, arithmetic means can be linearly combined via weighted averaging ($\bar{X}_{total} = \sum n_j \bar{X}_j / \sum n_j$). The non-differentiable nature of the absolute value function in the $L_1$ norm also complicates optimization, making the median much harder to handle in analytical calculus and linear regression frameworks compared to the differentiable squared errors of the $L_2$ norm.

Additionally, the median displays statistical inefficiency when applied to normally distributed data. Under ideal Gaussian conditions where the distribution is perfectly symmetrical, both the sample mean and the sample median serve as unbiased estimators of the true population center $\mu$. However, the Asymptotic Relative Efficiency (ARE) of the median relative to the mean under Gaussian noise is:

$$\text{ARE}(M, \bar{X}) = \frac{2}{\pi} \approx 0.637$$

This mathematical result indicates that the variance of the sample median is approximately 57% larger than the variance of the sample mean ($pi/2 \approx 1.571$) when sampling from a normal population. Consequently, an analyst relying on the median would need a sample size roughly 57% larger to achieve the same precision as a researcher utilizing the sample mean under ideal Gaussian conditions. While the median offers robust protection against distribution violations, it sacrifices statistical power when parametric assumptions are met.

3. Understanding the Mode in Descriptive Statistics

3.1 Mathematical Definition and Frequency Mechanics

The mode ($Mo$) is the measure of central tendency grounded purely in frequency mechanics and localized density maximization. In the context of a discrete random variable $X$ defined over a discrete support $\mathcal{X}$, the mode is formally defined as the element in the sample space that maximizes the probability mass function (PMF), denoted by $p(x) = P(X = x)$:

$$Mo = arg\max_{x in \mathcal{X}} p(x)$$

In a finite empirical sample of discrete or categorical observations $X = {x_1, x_2, dots, x_n}$, the mode is identified through absolute frequency tallies. Let $f(c) = \sum_{i=1}^n I(x_i = c)$ represent the counting function that tallies occurrences of value $c$, where $I(\cdot)$ denotes the standard indicator function. The empirical mode is the specific value $c^*$ that maximizes this tally:

$$Mo = arg\max_{c} \sum_{i=1}^n I(x_i = c)$$

From an optimization perspective, the mode is the parameter that minimizes the expected zero-one loss function ($0-1$ loss), which assigns zero penalty to an exact match and a unit penalty of one to any mismatch:

$$L_{0-1}(x, c) = I(x ne c)$$

This formulation demonstrates that if an analyst must select a single guess for an unknown observation and is penalized equally for any error regardless of magnitude, the mode represents the mathematically optimal point prediction.

Extending the mode to continuous random variables governed by a probability density function $f(x)$ introduces substantial technical nuances. In continuous distributions, the probability of observing any single specific real number is identically zero ($P(X = x) = 0$). Consequently, the continuous mode cannot be defined via discrete frequency tallies. Instead, it is defined as the coordinate $x$ at which the probability density function achieves a local or global maximum. If $f(x)$ is twice continuously differentiable, the mode satisfies the standard differential conditions:

$$f'(Mo) = 0 \quad \text{and} \quad f”(Mo) < 0$$

For finite continuous empirical samples where every observed value is unique (a common occurrence when measuring continuous real numbers to high precision), estimating the population mode requires specialized non-parametric estimation techniques. These include:

• Histogram Discretization: Partitioning the continuous domain into discrete bins of width $h$ and identifying the midpoint of the bin with the highest frequency count. This approach is highly sensitive to both bin width and bin origin placement.
• Kernel Density Estimation (KDE): Constructing a continuous smooth density estimate $\hat{f}_h(x) = \frac{1}{nh} \sum_{i=1}^n K\left(\frac{x – x_i}{h}\right)$, where $K$ is a kernel function and $h$ is the bandwidth parameter, then solving numerically for $arg\max_x \hat{f}_h(x)$.
• The Half-Sample Mode (HSM) and Venter Estimators: Non-parametric order-statistic algorithms that iteratively identify the shortest interval containing a designated fraction of observations, isolating the region of highest empirical concentration.

3.2 Modal Structures: Unimodal, Bimodal, and Multimodal Distributions

Unlike the median, which always yields a single deterministic value for any ordered finite set, the mode exhibits rich structural diversity regarding its multiplicity and existence. A distribution is described as unimodal when it possesses a single, clearly defined peak or coordinate of maximal density. Standard theoretical models—such as the Gaussian, exponential, and Student’s $t$-distributions—are classic examples of unimodal distributions, each containing a single global maximum where probability clusters.

However, empirical distributions frequently present multiple local maxima, resulting in bimodal (two peaks) or multimodal (three or more peaks) topologies. A bimodal distribution typically serves as a strong diagnostic indicator of underlying heterogeneity, signaling that the observed sample is not drawn from a single homogeneous population, but rather represents a mixture distribution composed of two or more distinct sub-populations. For example, a biometric distribution of human adult heights across a mixed-sex population often exhibits a bimodal structure, reflecting two distinct modal peaks corresponding to the female and male sub-populations. In such contexts, computing a single median or mean combines these sub-populations into an artificial central figure that may not reflect either group well; conversely, identifying both modes explicitly reveals the mixture structure of the underlying data generating process.

At the opposite extreme lies the amodal distribution, in which no distinct mode exists. In a discrete uniform distribution—such as rolling a fair, six-sided die where each outcome ${1, 2, 3, 4, 5, 6}$ has an identical probability of $1/6$—every element exhibits an equal frequency tally. Under such conditions, declaring every value a mode renders the concept meaningless; hence, the distribution is characterized as having no mode. Similarly, a continuous uniform distribution across an interval $[a, b]$ exhibits a constant horizontal density function $f(x) = 1/(b – a)$, completely lacking a local maximum where $f'(x) = 0$ and $f”(x) < 0$. This susceptibility to non-existence and non-uniqueness is a defining characteristic of the mode.

3.3 Methodological Strengths and Structural Limitations

The most prominent methodological advantage of the mode is its universal applicability across all measurement scales. As S. S. Stevens demonstrated in his 1946 taxonomy, the mode is the only measure of central tendency that is mathematically valid for strictly nominal (categorical) data. Categorical variables—such as biological species, eye colors, manufacturing failure types, or political party affiliations—contain no intrinsic order or numerical distance. Computing a median on nominal data is impossible because the categories cannot be sorted into a meaningful hierarchy. The mode, however, requires only the ability to count frequencies within mutually exclusive categories, allowing analysts to extract the most common classification reliably.

In commercial, manufacturing, and operational settings, the mode provides pragmatic utility that positional and algebraic metrics cannot match. In retail inventory management, an operations manager cannot order a “median shoe size” of 8.37 or stock an abstract fractional SKU; the procurement pipeline must supply the modal shoe size (e.g., size 9) to satisfy the largest segment of customer demand. In democratic electoral systems, the winner of a plurality vote is the modal candidate—the individual capturing the highest absolute tally of ballots. Similarly, in software reliability engineering, debugging protocols focus on identifying the modal error code to resolve the defect impacting the largest volume of user interactions.

Conversely, the mode suffers from severe structural vulnerabilities, particularly when applied to continuous variables or small sample sizes. In small samples drawn from continuous distributions, slight random sampling variations can cause two identical observations to occur purely by chance, creating an unstable, spurious mode that fails to reflect the underlying population density. Furthermore, the mode is entirely insensitive to the distribution of values away from its peak. Two distributions can have identical modes while displaying vastly different variance, skewness, and tail dynamics. Because the mode minimizes the 0–1 loss function, it treats all errors identically regardless of whether an unobserved value misses the mode by one millimeter or one thousand kilometers, making it unsuitable for applications that penalize large, catastrophic deviations.

4. Key Differences Between Median and Mode

To systematically contrast the median and the mode, we must analyze their algorithmic structures, scale applicability, mathematical guarantees, and sensitivities to sample perturbations. The following table provides a comprehensive head-to-head comparison across their core statistical dimensions, establishing the foundational differences detailed in the subsections that follow.

Analytical Dimension	The Median ($M$)	The Mode ($Mo$)
Foundational Principle	Positional rank-order; identifies the exact 50th percentile rank bisecting sorted data.	Frequency density; identifies the value or category of maximal empirical occurrence.
Stevens’ Measurement Scale	Permissible on Ordinal, Interval, and Ratio scales; completely invalid on Nominal data.	Universally permissible on Nominal, Ordinal, Interval, and Ratio scales.
Optimization Loss Function	Minimizes the $L_1$ norm (sum of absolute deviations): $arg\min_c \sum |x_i – c|$.	Minimizes the $0-1$ indicator loss (misclassification risk): $arg\min_c \sum I(x_i ne c)$.
Mathematical Uniqueness	Guaranteed to exist and be uniquely defined for any finite ordered dataset.	Not guaranteed to be unique; can be unimodal, bimodal, multimodal, or entirely amodal.
Asymptotic Breakdown Point	Achieves the maximum possible breakdown point of $\approx 50%$ against arbitrary corruption.	Conceptually $0%$ for continuous kernel estimators; highly sensitive to local peak shifts.
Monotonic Equivariance	Strictly equivariant under monotonic transformations: $M(g(X)) = g(M(X))$.	Equivariant under one-to-one transformations; preserves peak location across transformations.
Continuous Estimation Mechanics	Directly derived via inverse CDF interpolation ($Q(0.5) = F^{-1}(0.5)$).	Requires differential optimization ($f'(x) = 0$), KDE, or specialized binning methods.
Computational Complexity	Selection problem solvable in $O(n)$ expected time using Quickselect, or $O(n log n)$ via sorting.	Frequency tallying solvable in $O(n)$ time using hash tables or counting arrays.
Sensitivity to Distant Outliers	Completely invariant to peripheral changes that do not shift the central rank ordering.	Completely invariant to outliers, unless an outlier forms a new peak that exceeds all others.
Primary Operational Domain	Skewed continuous data: household incomes, real estate valuations, network latency metrics.	Categorical outcomes, inventory stocking units, mixture diagnostics, voting plurality.

4.1 Algorithmic Computation and Mathematical Foundations

The mathematical and algorithmic mechanics behind computing the median and the mode reflect their underlying theoretical differences. Calculating the sample median is inherently an order-statistic selection problem. For a dataset of size $n$, the traditional naive computation involves sorting the array into monotonic order using algorithms such as Heapsort or Quicksort, which carry a computational time complexity of $O(n log n)$, followed by directly indexing the element at position $(n+1)/2$. However, computer science literature has established that sorting the entire array is computationally redundant if one only needs the 50th percentile.

Using selection algorithms such as the Quickselect algorithm (developed by Tony Hoare) or the deterministic median-of-medians algorithm (formulated by Blum, Floyd, Pratt, Rivest, and Tarjan in 1973), the median of an unsorted array can be determined in $O(n)$ expected or worst-case linear time without fully sorting the data. These selection algorithms operate by iteratively partitioning the array around a pivot element, recursively narrowing the search space to the partition containing the target rank index. In distributed computing paradigms like Apache Spark or large relational databases, computing exact medians across terabyte-scale tables remains computationally demanding because finding order statistics requires cross-node data shuffling. Consequently, distributed architectures frequently rely on approximate streaming quantile algorithms, such as the Greenwald-Khanna or t-digest algorithms, which estimate the 50th percentile within bounded error tolerances.

Conversely, computing the mode in discrete or categorical datasets bypasses sorting entirely, relying instead on frequency hashing and counting partitions. For discrete data, an algorithm maps each observation into a hash table or direct-address counting array, incrementing a localized counter in $O(1)$ time per entry. The entire dataset can be processed in a single pass with $O(n)$ linear time complexity and $O(k)$ auxiliary space complexity, where $k$ is the number of distinct unique categories. When memory constraints are severe and the modal frequency is guaranteed to represent a strict majority (exceeding $n/2$), the Boyer-Moore Majority Vote Algorithm can identify the mode in $O(n)$ time using only $O(1)$ auxiliary memory space, requiring no sorting, no hash maps, and no distributed shuffling.

However, when computing the mode on continuous distributions, the computational profile shifts dramatically. Because unique real values rarely repeat, simple frequency counting fails. Estimating the continuous mode requires evaluating a non-parametric density function over an evaluation grid or using iterative gradient-ascent methods (such as the Mean-Shift algorithm) to locate local density maxima. The computational complexity of these continuous approaches depends heavily on the chosen bandwidth parameter $h$ and the sample size $n$, often scaling at $O(n^2)$ for naive kernel density evaluations over continuous spaces. As a result, calculating continuous modes is computationally far more complex than finding the positional median.

4.2 Scale of Measurement Applicability

The formal hierarchy of measurement scales—introduced by Harvard psychologist S. S. Stevens in his landmark 1946 paper “On the Theory of Scales of Measurement”—sets strict mathematical boundaries on where each measure of central tendency can be legitimately applied. Stevens classified empirical measurements into four distinct levels: nominal, ordinal, interval, and ratio. Each scale level is characterized by the mathematical transformations under which its empirical relationships remain invariant, directly governing the validity of summary statistics.

The mode is the only central tendency metric that operates across all four levels of Stevens’ taxonomy:

• Nominal Scale: Nominal data involves purely qualitative labels, names, or classifications (e.g., blood types $A, B, AB, O$; religious affiliations; diagnostic codes). Nominal categories have no natural ordering, meaning that relations such as “greater than” or “less than” are mathematically undefined. Because calculating a median requires a valid ordering, applying it to nominal data is completely impossible. The mode, requiring only equivalence relations ($=$ or $ne$) and counting, provides the sole legitimate measure of central tendency for nominal data.
• Ordinal Scale: Ordinal data introduces a meaningful monotonic ranking between categories, but lacks uniform, metric distances between consecutive points (e.g., Likert survey scales: strongly disagree, disagree, neutral, agree, strongly agree; military ranks; socioeconomic status). Because an ordinal hierarchy exists, the median becomes fully permissible, identifying the central rank position. The mode remains fully valid as well, identifying the most frequently selected rank.
• Interval and Ratio Scales: Interval scales provide uniform, measurable intervals between numeric points without a true absolute zero (e.g., temperature in Celsius or Fahrenheit), while ratio scales possess a non-arbitrary, true absolute zero (e.g., mass, length, Kelvin temperature, economic currency). Both the median and the mode remain fully permissible across these quantitative scales, where they are joined by the arithmetic mean.

Applying the median to nominal data is not merely a technical oversight; it is a fundamental category error. Converting arbitrary nominal strings to integers (e.g., assigning Red=1, Blue=2, Green=3) and calculating a “median” produces an arbitrary figure that changes completely if the arbitrary numeric assignments are altered. The mode remains the only mathematically coherent indicator of typicality when numerical ordering cannot be established.

4.3 Uniqueness and Existence Properties

A crucial theoretical difference between the median and the mode lies in their mathematical existence and uniqueness guarantees. For any non-empty, finite empirical dataset $X$ consisting of real numbers or ordered values, the median is mathematically guaranteed to exist and is uniquely defined under standard descriptive conventions. Even when a dataset contains an even number of observations and lacks a single physical central value, the standard arithmetic interpolation convention ($M = (x_{(n/2)} + x_{(n/2+1)})/2$) ensures a single, deterministic summary statistic. This guarantee makes the median a dependable metric for automated data pipelines, programmatic reporting systems, and unsupervised machine learning algorithms, where missing values or structural ambiguity can crash downstream processes.

The mode, in contrast, offers no such guarantees of existence or uniqueness. An empirical dataset can be unimodal, bimodal, multimodal, or entirely amodal:

• Multimodality: If two or more distinct values share the exact same maximum frequency tally, the dataset is multimodal. For example, in the discrete sequence ${2, 4, 4, 7, 9, 9, 12}$, both 4 and 9 appear with a maximal frequency of two. The dataset possesses two distinct modes. In automated analytical systems, multimodal outputs create architectural ambiguity: an algorithm expecting a single scalar metric must either handle an array of modes, arbitrarily select one, or fail.
• Amodality (Non-Existence): When every observation in a sample occurs with identical frequency, no distinct peak exists, and the mode does not exist. This is common in continuous datasets measured with high-precision sensors, where every floating-point number is unique ($f(x_i) = 1$ for all $i$). Under such conditions, standard frequency counting reports that every observation is a mode, rendering the metric useless unless the analyst implements manual binning or continuous density estimation.

This structural instability makes the mode less reliable for automated data pipelines that require a single, deterministic scalar value. While the discovery of multimodality provides valuable diagnostic insight to a human statistician, it introduces engineering hurdles in production systems designed to process scalar summary statistics.

4.4 Sensitivity to Outliers and Data Perturbations

The median and the mode react to data perturbations and extreme values in fundamentally different ways, reflecting their underlying loss functions and mathematical structures. As established by its 50% breakdown point, the median is exceptionally stable against peripheral outliers. An analyst can manipulate the most extreme observation in an ordered dataset—increasing it by multiple orders of magnitude—without moving the median by a single decimal place, provided the rank order of the central elements remains undisturbed. The median only shifts when the central values themselves are altered, or when enough points are added or removed to shift the middle rank index.

The mode exhibits a distinct form of outlier invariance: it is completely indifferent to the magnitude of numerical values across the entire distribution, whether peripheral or central, because it depends entirely on localized frequency counts. An extreme outlier added to a dataset will have zero effect on the mode, provided the outlier’s frequency does not exceed the peak frequency of the existing modal value. However, the mode displays high sensitivity to localized perturbations around regions of high density. In discrete datasets with low overall frequencies, altering the value of just one or two observations can suddenly create a new modal peak elsewhere, causing the mode to jump discontinuously across the sample space:

Consider the sequence $A = {12, 12, 14, 15, 18, 22, 25}$. Here, the mode is unambiguously 12 ($f=2$), and the median is 15. If a single observation is perturbed by recording error such that one 12 becomes a 14, the sequence becomes $B = {12, 14, 14, 15, 18, 22, 25}$. The median shifts smoothly from 15 to 14, reflecting a minor rank change. The mode, however, jumps from 12 to 14. If a further minor perturbation occurs such that the values become $C = {12, 13, 14, 15, 18, 22, 25}$, the mode disappears entirely, as the dataset becomes completely amodal. This sensitivity to tiny localized changes makes the mode volatile in small sample sizes, even while it remains immune to extreme tail outliers.

5. Similarities and Methodological Overlap

5.1 Non-Parametric Foundations

Despite their structural differences, the median and the mode share foundational ground as non-parametric statistics. Unlike the arithmetic mean, which serves as a natural parameter for the family of normal distributions and underpins classical parametric modeling, neither the median nor the mode requires the analyst to assume that the data follows a specific theoretical probability distribution. Both metrics can be computed directly from raw empirical observations without estimating population parameters like the mean ($\mu$) or variance ($\sigma^2$), making them robust tools for exploratory data analysis across messy, real-world data environments.

This non-parametric foundation gives both metrics resilience against common distribution anomalies that undermine parametric methods. Real-world phenomena frequently violate idealized mathematical assumptions: financial returns exhibit fat-tailed leptokurtic distributions; biological recovery times follow heavily skewed log-normal or Weibull paths; and customer support interactions often follow zero-inflated Poisson profiles. In these non-Gaussian spaces, the parametric mean can shift away from regions of actual concentration toward low-probability tail regions. Because the median focuses strictly on ordinal rank positions and the mode targets frequency peaks, both metrics remain anchored within the actual support of the data, providing realistic representations of central tendency where parametric assumptions break down.

5.2 Convergence in Symmetrical Topologies

A central mathematical link between the median, the mode, and the arithmetic mean is their convergence within perfectly symmetrical, unimodal distributions. In any continuous probability distribution that is symmetric about a central point $c$—such that the probability density function satisfies $f(c – x) = f(c + x)$ for all real $x$—and has a single unique maximum at $c$, the mean, the median, and the mode converge upon the exact same coordinate:

$$\mu = M = Mo = c$$

This triple convergence occurs in key theoretical probability models, including the Gaussian distribution, the Student’s $t$-distribution, the logistic distribution, and the continuous uniform distribution (which is symmetric, though amodal). Because symmetry forces the area under the density curve to split evenly at the center of balance, the 50th percentile rank (median) aligns with the center of gravity (mean), which in turn coincides with the peak of the probability density function (mode).

This mathematical alignment provides a valuable diagnostic tool for evaluating skewness in empirical data. When an empirical distribution is skewed, the three measures separate systematically along the horizontal axis. In a moderately right-skewed (positively skewed) unimodal continuous distribution, the metrics typically separate in the directional order:

$$Mo < M < \bar{X}$$

The mode remains anchored at the peak where data density is highest; the mean is pulled out toward the elongated right tail by extreme values; and the median positions itself between the two, balancing the area under the curve. In a left-skewed (negatively skewed) distribution, this sequence reverses:

$$\bar{X} < M < Mo$$

This predictable divergence formed the basis for Karl Pearson’s historical empirical skewness rule, formulated in 1895. Pearson observed that for moderately skewed, unimodal continuous curves, the distance between the mean and the mode is approximately three times the distance between the mean and the median:

$$\bar{X} – Mo \approx 3(\bar{X} – M)$$

Rearranging this relationship yields Pearson’s second skewness coefficient ($Sk_2$):

$$Sk_2 = \frac{3(\bar{X} – M)}{\sigma}$$

While this empirical approximation does not hold universally for all mathematical distributions—particularly those with extreme tail behavior or discrete boundaries—it illustrates how the divergence between the median and the mode serves as a foundational diagnostic marker for structural asymmetry in empirical data.

5.3 Applications in Data Imputation and Preprocessing

In modern data engineering and applied machine learning, the median and the mode serve as standard baseline methods for handling missing data. Real-world feature matrices frequently arrive with missing entries caused by sensor dropouts, survey non-responses, or database extraction failures. Because most predictive algorithms—such as generalized linear models, support vector machines, and deep neural networks—cannot natively process missing tokens (e.g., NaN or null), these gaps must be imputed during the data preprocessing phase.

The choice between median and mode imputation is governed strictly by the feature’s measurement scale and underlying loss function:

• Mode Imputation for Categorical Features: When missing values occur within categorical, discrete, or nominal feature columns (such as geographic regions, device operating systems, or user classifications), the mode is the only mathematically valid non-parametric univariate imputation statistic. Imputing with the mode replaces missing entries with the most probable categorical state, minimizing the zero-one misclassification loss across the training dataset.
• Median Imputation for Continuous Numerical Features: When missing values occur within continuous numerical features that exhibit skewness or heavy outliers (such as transaction amounts, patient blood pressure readings, or server response times), practitioners routinely choose median imputation over mean imputation. Because the median minimizes the $L_1$ absolute deviation loss, it prevents extreme peripheral outliers from distorting the imputed value, ensuring the filled values remain realistic central representations of the feature space.

While both methods offer computationally fast, univariate baseline solutions, practitioners must remain mindful of the biases univariate imputation can introduce. Imputing either metric in large volumes artificially compresses the feature’s empirical variance, distorts natural covariance structures with other variables, and introduces spikes into the distribution’s shape at the imputed coordinate. Nonetheless, in production machine learning pipelines, median and mode imputation remain essential tools for robust data preparation.

6. Practical Decision Framework: When to Use Median vs. Mode

6.1 Optimal Contexts for Implementing the Median

The median is the preferred measure of central tendency whenever an empirical investigation involves quantitative variables (interval or ratio scales) that exhibit substantial skewness, heavy tails, or extreme peripheral outliers, and where the primary analytical objective is identifying the typical, rank-balanced center of the population. In these settings, the arithmetic mean is pulled away by tail values, while the mode may reflect an unrepresentative localized cluster.

Key practical environments where the median serves as the standard metric include:

• Socioeconomic and Labor Economics: National wage earnings, household wealth, and real estate pricing distributions are classic examples of positive skewness governed by power-law dynamics. Government statistical agencies—such as the United States Census Bureau and the UK Office for National Statistics—rely on median household income and median residential sales prices rather than their arithmetic means. Using the median ensures that economic assessments reflect the lived reality of middle-tier citizens without being skewed upward by the wealth of top earners.
• Biomedical and Survival Analysis: In clinical trials and oncology research, patient survival times, post-operative recovery periods, and remission durations are routinely summarized using the median survival time. Because clinical studies often conclude before every patient has reached the study endpoint, survival distributions are frequently right-censored. The median can be calculated via the Kaplan-Meier estimator the moment exactly 50% of the cohort experiences the event, whereas computing an arithmetic mean would require waiting indefinitely for every single patient to conclude the trial.
• Site Reliability and Software Engineering: Infrastructure and platform engineers prioritize the 50th percentile latency (p50) to evaluate user experience. While latency metrics often exhibit extreme positive spikes due to network timeouts or garbage collection, tracking the median latency reveals the baseline performance experienced by the typical user, separate from the high-percentile tail latency issues (p99 or p99.9) handled by specialized infrastructure protocols.

6.2 Optimal Contexts for Implementing the Mode

The mode is the optimal measure of central tendency whenever an analytical investigation requires identifying the single most common, probable, or popular state within an empirical distribution. Its use is mandatory across nominal data where mathematical ordering is absent, and highly advantageous in logistics, inventory optimization, and discrete operations research where actions must target specific, non-fractional categories.

Key operational environments where the mode serves as the primary metric include:

• Supply Chain Management and Retail Merchandising: In apparel retail, footwear distribution, and consumer packaged goods, inventory stocking policies must align with physical, discrete unit requirements. An apparel manufacturer cannot produce a median shirt size of “Medium-and-a-half” or rely on a fractional calculation. Production and logistics systems must identify and manufacture the modal clothing sizes (e.g., Large) and modal colorways to meet the largest volume of consumer demand and minimize unsold surplus inventory.
• Democratic Voting and Polling Analytics: In political science and public policy analysis, electoral outcomes and public referendums are governed by modal logic. In a plurality voting system, the winning candidate or policy initiative is simply the mode of the vote tally—the categorical selection chosen by the largest cohort of voters. Similarly, qualitative polling data evaluating voter sentiment (e.g., favorable, neutral, unfavorable) is best summarized using the modal response category to guide campaign strategy.
• Quality Assurance and Industrial Reliability: In manufacturing quality control and software systems maintenance, identifying failure points requires isolation of the modal error code, modal component defect, or modal failure mode. When an industrial assembly line experiences stoppages, quality engineers do not search for a “median” defect; they isolate the modal defect—such as a specific misaligned bearing or repetitive thermal breakdown—to focus remediation efforts where they will have the greatest impact on operational uptime.

6.3 Analytical Decision Matrix

To help researchers and data practitioners select the appropriate measure of central tendency, the following decision framework provides a clear path through the selection process, mapping specific data attributes to their mathematically valid metrics:

Step 1: Evaluate Stevens’ Measurement Scale

• Is the variable strictly Nominal (unordered categories)?
  
  → Decision: The Mode is the only permissible measure of central tendency. The Median is mathematically undefined.
• Is the variable Ordinal (ordered categories lacking uniform metric intervals)?
  
  → Decision: Both the Median and the Mode are valid. Use the Median to find the central rank threshold; use the Mode to identify the most commonly selected category.
• Is the variable Interval or Ratio (quantitative, continuous, or discrete metrics)?
  
  → Decision: Proceed to Step 2.

Step 2: Evaluate Distributional Topology and Multimodality

• Does the continuous or discrete distribution exhibit multiple distinct peaks (Bimodal or Multimodal)?
  
  → Decision: Report the Modes. Multimodality signals the presence of distinct underlying sub-populations or mixture distributions. A single median or mean will obscure this heterogeneous structure.
• Is the distribution Unimodal?
  
  → Decision: Proceed to Step 3.

Step 3: Evaluate Skewness, Outliers, and the Operational Objective Function

• Is the unimodal distribution heavily skewed, fat-tailed, or prone to extreme measurement outliers?
  
  → Decision: Use the Median. Its 50% breakdown point ensures a robust, stable central anchor that accurately reflects the 50th percentile rank.
• Does the operational goal require identifying the most probable individual discrete outcome, optimizing physical inventory units, or minimizing a 0–1 classification loss?
  
  → Decision: Use the Mode.
• Is the distribution roughly symmetrical, free of extreme outliers, and destined for linear modeling or calculus-based inferential frameworks?
  
  → Decision: Consider the Arithmetic Mean for maximum statistical efficiency, while reporting the Median alongside it to verify distributional symmetry.

7. Frequently Asked Questions

7.1 Can a dataset possess both a median and multiple modes?

Yes. A dataset can easily possess both a unique median and multiple modes simultaneously. In fact, this is common when sampling from mixed populations or multimodal distributions. The median and the mode operate on distinct mathematical planes: the median identifies the single positional rank that bisects an ordered sample space into equiprobable halves, whereas the mode identifies localized frequency peaks. For any non-empty, finite dataset of real numbers, the median is guaranteed to exist as a single deterministic value (derived via linear interpolation if the sample size $n$ is even). At the same time, if two or more distinct values share the highest frequency count, the dataset is multimodal.

For example, consider the discrete sample $X = {3, 3, 3, 8, 10, 14, 19, 19, 19}$. This dataset contains $n = 9$ observations. Because the data can be monotonically sorted, the median is uniquely identified by the 5th order statistic: $M = x_{(5)} = 10$. Exactly four values are smaller than 10, and four values are larger. Concurrently, examining the frequency tallies reveals that both the value 3 and the value 19 occur with a maximal frequency of three ($f = 3$). Consequently, the distribution is bimodal, with modes at 3 and 19. In this scenario, the median ($M = 10$) anchors the center of the sample space, while the two modes reveal that the data clusters at two distinct poles, with few observations falling near the center. Reporting both metrics gives a far clearer picture than relying on either one alone.

7.2 Why is the median preferred over the mode for skewed economic data?

The median is preferred over the mode for summarizing skewed economic data—such as household income, individual net worth, and real estate valuations—because the mode often reflects regulatory, institutional, or lower-boundary clustering rather than the general population’s earning power. In national wage distributions, the mode frequently corresponds to the legally mandated minimum wage or zero, simply because millions of workers cluster at that specific statutory pay grade. Summarizing a nation’s wage profile by citing its modal income would suggest that the typical citizen earns minimum wage, which obscures broader economic realities higher up the income scale.

The median resolves this distortion by operating as a rank-based quantile metric. By identifying the exact income level where 50% of the population earns less and 50% earns more, the median provides a balanced macro-level baseline that is immune to both the high-frequency spikes at the minimum wage floor and the multi-billion-dollar outliers in the top percentiles. This property makes the median the standard foundation for defining relative poverty thresholds—such as the OECD definition of poverty as income below 50% or 60% of the national median—and for shaping equitable taxation and social welfare policies.

7.3 Is the mode usable on continuous variables without grouping?

In pure mathematical theory, the mode of a continuous random variable is defined as the maximum of its probability density function ($f'(x) = 0, f”(x) < 0$). However, in practical descriptive statistics, the sample mode cannot be calculated on continuous data without grouping or specialized modeling. In continuous measurements recorded with high-precision instruments (such as floating-p\oint sensor logs or financial exchange rates), the probability of observing two identical values is infinitesimally small ($P(X = x) = 0$). As a result, an empirical continuous sample of size$n$ typically consists of $n$ unique values, rendering the discrete mode nonexistent or amodal.

To extract a useful mode from continuous samples, analysts must use continuous density estimation techniques:

• Discretization (Histogram Binning): The continuous range is divided into a series of contiguous, non-overlapping intervals (bins). The mode is then estimated as the midpoint of the bin with the highest frequency count. However, the result is highly sensitive to the chosen bin width and edge placement.
• Kernel Density Estimation (KDE): A smooth, continuous density function $\hat{f}_h(x)$ is constructed by placing a kernel (typically a Gaussian curve) over each observation, using a bandwidth parameter $h$ (often selected via Silverman’s rule of thumb). The mode is then found numerically as the peak of this continuous surface.
• Half-Sample Mode (HSM): A non-parametric order-statistic algorithm that avoids arbitrary binning. The algorithm recursively identifies the shortest continuous interval that contains at least half of the remaining data points until it converges on the region of highest empirical concentration.

7.4 How do missing values affect median and mode computations?

Missing values (such as NA, null, or NaN) affect median and mode computations differently, reflecting their contrasting underlying mechanics. For the median, computing the 50th percentile rank requires an unambiguous, monotonic ordering of every element in the sample space. Because missing values lack numerical magnitude, their ordinal relationship to real numbers is undefined—a missing value cannot be evaluated as greater than, less than, or equal to five. Consequently, computing a median requires either listwise deletion (removing records containing missing values) or pairwise handling before sorting the data. If missing records are ignored without adjusting the sample size denominator, the positional rank index $(n+1)/2$ shifts, introducing analytical bias if the missingness mechanism is non-random (Missing Not At Random, or MNAR).

In contrast, the mode handles missing data with greater structural flexibility, particularly in discrete and categorical datasets. Because the mode relies on frequency counting rather than rank ordering, missingness can be explicitly coded as its own independent categorical level (e.g., "Unknown" or "Missing"). In systems like customer surveys or clinical trial records, missingness often carries valuable diagnostic meaning: if the most frequent response to an optional survey question is a non-response, "Missing" becomes the legitimate empirical mode of the variable. However, if an analyst intends to perform mode imputation to fill missing continuous fields, imputing large volumes of a single modal value will artificially deflate the feature’s variance and introduce sharp, artificial spikes into the distribution’s density profile.

8. Conclusion

8.1 Summary of Core Distinctions

The median and the mode represent two fundamentally distinct mathematical paradigms within descriptive statistics and quantitative analysis. The median is an inherently positional metric rooted in order statistics and cumulative distribution functions. By identifying the exact 50th percentile rank that splits an ordered dataset into two equiprobable subsets, the median serves as a reliable measure of central balance. It achieves a 50% breakdown point, minimizes the $L_1$ absolute error loss function, and remains invariant under strictly monotonic transformations. These properties make it the premier metric for summarizing skewed, fat-tailed, and outlier-prone quantitative distributions across interval and ratio scales, while remaining fully valid on ordinal data.

In contrast, the mode is a frequency-density metric that identifies the specific value, category, or density peak where observations cluster with the highest empirical probability. Minimizing the 0–1 indicator loss function, the mode operates as an optimal guess under equal-penalty conditions, uniquely spanning the entire hierarchy of measurement scales from nominal labels to continuous real numbers. While the mode is essential for discrete operational decisions, inventory management, categorical survey analysis, and mixture diagnostics, it lacks the existence and uniqueness guarantees of the median, exhibiting volatility in small samples and requiring specialized continuous density estimation techniques when values rarely repeat.

8.2 Strategic Guidelines for Statistical Reporting

To maintain high standards of scientific integrity and analytical rigor, researchers and data practitioners should move away from the traditional practice of reporting a single, isolated measure of central tendency. Relying on an isolated metric—whether the mean, the median, or the mode—inevitably conceals critical structural properties of the underlying distribution. When preparing statistical reports, white papers, or predictive data pipelines, practitioners should implement the following strategic guidelines:

• Report Multiple Metrics Concurrently: Whenever feasible, present the median and the mode alongside the arithmetic mean. Comparing the relative positions of these three metrics provides an immediate, intuitive diagnostic of distributional skewness and asymmetry, alerting audiences to non-Gaussian behavior.
• Pair Central Tendency with Appropriate Dispersion Metrics: A measure of central tendency is only as informative as the dispersion surrounding it. The median should be paired with the Interquartile Range (IQR) or Median Absolute Deviation (MAD), both of which share its non-parametric robustness against outliers. The mode should be reported alongside frequency-entropy metrics, variation ratios, or modal percentages that quantify how much of the sample clusters at the primary peak.
• Respect Stevens’ Scale Hierarchy: Enforce strict adherence to measurement scale constraints across all analytical workflows. Never calculate medians on unordered nominal categories, and ensure that continuous modes are derived using sound density estimation techniques rather than unreflective frequency counts on floating-point values.

By aligning metric selection with the structural properties of the data and the operational objectives of the analysis, researchers ensure that their statistical summaries provide an accurate, robust, and actionable foundation for empirical discovery and decision-making.

References

• Cournot, A. A. (1843). Exposition de la théorie des chances et des probabilités. Paris: L. Hachette. https://gallica.bnf.fr/ark:/12148/bpt6k29505h
• Fechner, G. T. (1874). Ueber den Ausgangswerth der kleinsten Abweichungssumme, dessen Bestimmung, Verwerthung und Nachwirkung. Abhandlungen der Mathematisch-Physischen Classe der Königlich Sächsischen Gesellschaft der Wissenschaften, 11, 1–76.
• Galton, F. (1881). Report of the Anthropometric Committee. Report of the British Association for the Advancement of Science, 51, 245–260.
• Pearson, K. (1895). Contributions to the mathematical theory of evolution. II. Skew variation in homogeneous material. Philosophical Transactions of the Royal Society of London A, 186, 343–414. https://doi.org/10.1098/rsta.1895.0010
• Stevens, S. S. (1946). On the theory of scales of measurement. Science, 103(2684), 677–680. https://doi.org/10.1126/science.103.2684.677