Central Tendency
Central tendency is the statistical idea of a single value that represents the centre or typical position of a dataset, around which the observations cluster, and the measures of central tendency are the competing ways of defining that value. The three classic measures are the mean, the median, and the mode, and the family also includes the weighted mean, geometric mean, harmonic mean, and trimmed mean. The concept itself, as distinct from any one calculation, is the question of what "typical" should mean, and the reason several answers exist is that different data, and different purposes, give that word different meanings.
Every measure of central tendency compresses a whole distribution into one number, and the choice between them is a choice about what to ignore. The mean uses every value and so responds to every change, including an extreme one. The median uses only the ordering of the values and so is nearly immune to extremes. The mode uses only the most frequent value or class and so works for categories where neither of the others can be computed. A single dataset can thus have three different centres, and none is wrong: in a small company where nine employees earn 40,000 dollars and the owner earns 1,000,000 dollars, the mean salary is 136,000 dollars, the median is 40,000 dollars, and the mode is 40,000 dollars, and which one is called "the average salary" determines whether the company sounds generous or ordinary.
The relationship between the measures is itself a diagnostic of the distribution's shape. In a symmetric, single-peaked distribution such as the normal distribution the three coincide. In a right-skewed distribution, with a long tail toward large values as in incomes, house prices, and waiting times, the mean is typically pulled above the median, and in a left-skewed distribution the order reverses, although the ordering is a tendency rather than a strict law. In a bimodal distribution no single centre describes the data well, and reporting one hides that two groups are mixed. Comparing the mean with the median is therefore a quick, cheap skewness check before any modelling begins.
The measures are tied to the scale of measurement, an idea set out by S. S. Stevens in his 1946 typology of scales: the mode is valid for nominal data, the median for ordinal data and above, and the mean for interval and ratio data. Computing a mean of category codes, such as 1 for red and 2 for blue, produces a meaningless number, and there is a long-running debate over whether the mean of ordinal ratings such as five-point survey responses is legitimate. Beyond the big three, several variants answer specific problems. The trimmed mean discards a fixed share of the smallest and largest values before averaging, giving robustness with efficiency; the weighted mean gives values unequal importance; the geometric mean suits growth rates and ratios; and the harmonic mean suits rates such as average speed over equal distances.
Practical use of central tendency runs through all of data work. Imputation of missing values commonly substitutes the mean or median of a column, and the choice affects the result on skewed variables. Baselines in machine learning predict the mean (for squared error) or the median (for absolute error) of the training targets, since each minimises its own loss, and a model has no skill unless it beats that baseline. In industry, process targets and control charts centre on the mean of a stable process, and in public policy the choice between median and mean household income shapes how inequality and growth are reported. In clinical reporting, median survival time is preferred partly because it can be estimated before everyone has experienced the event.
The leading pitfalls are treating "the average" as automatically the mean and treating the centre as if it described individuals. A mean can be a value no observation actually takes, such as 2.4 children per household, and for a heavy-tailed distribution like the Cauchy distribution no mean exists at all, so sample means never settle. An average of a group says little about any member of it, and a famous cautionary illustration is the US Air Force's 1950s study in which almost no pilot proved to be average on all of ten body dimensions measured, which is why cockpits were redesigned to be adjustable. Finally, a centre reported without a measure of dispersion is half a description: two datasets with the same mean can differ enormously in how reliably values fall near it.
Read central tendency as the problem of choosing one number to stand for many: the mean, median, and mode are three honest but different answers, the right one depends on the data's scale and shape and on the question being asked, and none of them says anything until it is paired with how widely the data scatter around it.