Point Estimate

A point estimate is a single number computed from sample data and offered as the best guess for an unknown population parameter, such as the sample mean as a guess for the population mean, the sample proportion for a population proportion, or the sample variance for a population variance. The rule or formula that produces it is called an estimator, and the number it returns for one particular sample is the estimate; the estimator is a random variable, varying from sample to sample, while the estimate is a fixed number. If 520 of 1,000 randomly sampled voters support a measure, the point estimate of the support proportion is 0.52, a precise-looking single value that is the most common summary reported from a study.

Estimators are judged by their long-run behavior across samples, and the classic properties were set out by Ronald Fisher in 1922. An estimator is unbiased if its average over repeated samples equals the true parameter, consistent if it converges to the truth as the sample grows, and efficient if it has the smallest variance among competing estimators of its kind. These properties can conflict, so a combined yardstick, mean squared error (variance plus squared bias), is often used. The sample mean is unbiased and, for normal data, efficient for the population mean, while the sample variance divides by n − 1 rather than n (Bessel's correction) precisely to remove the downward bias that arises because deviations are measured from the sample mean rather than the true mean.

Several general methods generate point estimates. The method of moments equates sample moments to population moments; maximum likelihood, developed and popularized by Fisher in the 1910s and 1920s, chooses the parameter value that makes the observed data most probable and underlies logistic regression and most modern model fitting; least squares chooses the line that minimizes squared errors; and Bayesian analysis offers the posterior mean or median as a point summary. Each has a different justification, but all hand back one number. In machine learning, every trained parameter, from a regression slope to a neural network weight, is a point estimate, and a model's predicted value for a new case is a point prediction.

A point estimate is intentionally a single number, which is both its appeal and its main weakness: it carries no information about how far off it might be. Two studies can report the same estimate, 0.52, with entirely different reliability, one on 100 respondents and one on 100,000. That is why a point estimate is almost always paired with a standard-error, a confidence-interval or a margin-of-error, and why decisions built on a bare estimate, such as ranking items by observed click rate when some have only a handful of impressions, are prone to chasing noise. A related trap is the winner's curse: the largest of many noisy estimates tends to be inflated, so the top-ranked item typically regresses toward the average when measured again.

Unbiasedness is not the same thing as being the best, and the best estimate is not always the obvious one. Biased estimators can have lower overall error: shrinkage methods such as ridge regression, the lasso and the James–Stein estimator (Stein 1956, James and Stein 1961) deliberately pull estimates toward a common value, which reduces variance by more than it adds in bias, and James and Stein showed this beats the sample mean when estimating three or more means at once. Likewise, even unbiasedness can fail to survive transformation: the sample variance is unbiased for the variance, but its square root is slightly biased low for the standard deviation. Matching the estimator to the loss that matters, and to the sample size, is part of the craft.

Read a point estimate as the sample's single best guess at an unknown parameter, produced by an estimator whose quality is judged by bias, consistency and variance across repeated samples, and never as a finished answer: without a standard error or interval attached, it is a number without a sense of how much to trust it.