Statistic

A statistic is any number computed from a sample of data — a sample mean, a sample proportion, a sample standard deviation, a median, a correlation coefficient — as opposed to a parameter, which is the corresponding fixed but usually unknown quantity describing the entire population. If 1,200 voters are polled and 52% say they will vote for a candidate, the 52% is a statistic; the share of all voters who actually intend to vote that way is the parameter, and the entire enterprise of inferential statistics exists because the first is observable while the second is what anyone actually wants to know.

The defining property of a statistic is that it is a function of the data alone, and therefore a random variable before the sample is drawn. A parameter such as the population mean μ is a single fixed number (unknown, but not random); a statistic such as the sample mean x-bar changes from one random sample to the next, because each sample contains different individuals. Notation encodes the split: Greek letters (μ, σ, ρ) conventionally mark parameters, while Latin letters (x-bar, s, r, p-hat) mark their sample counterparts. The distribution of a statistic across all the samples that could have been drawn is its sampling distribution, and that distribution — not the single observed value — is what allows statements about precision, which is why the standard error exists as the standard deviation of a statistic rather than of raw data.

A statistic used to guess a parameter is called a point estimate, and the rule that produces it is an estimator, whose quality is judged by properties of its sampling distribution rather than by any one outcome. An estimator is unbiased if its average over repeated samples equals the parameter; it is consistent if it converges on the parameter as the sample grows; and it is efficient if its sampling variability is small. The sample mean is unbiased for the population mean, and the sample variance with n minus 1 in the denominator is unbiased for the population variance, which is the entire reason for Bessel's correction — dividing by n would give a statistic that systematically runs low. Ronald Fisher's work in the 1920s on estimation, sufficiency, and likelihood formalized much of this vocabulary, and his notion of a sufficient statistic — one that retains all the information the sample holds about a parameter — remains central to theory.

Statistics are not only estimators of parameters; descriptive statistics summarize a dataset on its own terms, while test statistics convert data into evidence against a hypothesis. A mean, median, or range simply describes the sample in hand and needs no population in view. A test statistic such as a z-score, a t-statistic (from William Sealy Gosset's 1908 work under the pen name Student), or Pearson's 1900 chi-square statistic is a number engineered so that its sampling distribution is known if a null hypothesis is true, allowing a p-value to be computed. In machine learning the same idea appears as summary features: batch means and variances in normalization layers, a model's validation accuracy, or an AUC are all statistics of finite data, and each carries sampling noise.

The most common pitfall is treating a statistic as if it were the parameter, forgetting that it would have come out differently under another draw. A 52% poll result is not a claim that exactly 52% of the population holds that view; a lift of 3% in an A/B test on a few thousand users may be well within what sampling variation alone produces. A second misconception is that the word means "any numerical fact" — in technical usage a figure like a national census total is a parameter-like population value, not a statistic, even though everyday speech calls both statistics. A third is that bigger samples fix every problem: a larger sample shrinks the statistic's variability, but it cannot repair bias from a flawed sampling design, as the Literary Digest poll of 1936, with millions of responses and a wrong prediction, famously showed.

Read statistic as a measurement of the sample that is only ever a noisy stand-in for something about the population: the number itself is just arithmetic on the data in hand, while its meaning lives in the sampling distribution that says how far it could plausibly be from the fixed parameter it is trying to reach.