unspurious.calculators

Core inference · Estimation

Point Estimate Calculator

Eight successes out of ten — so the true rate is 80%? Not so fast. The raw proportion x ∕ n is only one of several point estimates of a population proportion, and at small samples or near 0% and 100% it is the worst-behaved of them. This gives the maximum-likelihood estimate beside the Laplace, Jeffreys and Wilson adjustments — and is blunt that a single best guess always hides a range, never more so than when you have seen very few trials.

x is the number of times the thing happened; n is how many chances it had. The maximum-likelihood, Laplace and Jeffreys estimates don’t depend on the confidence level — only the Wilson estimate and the interval do.

Result

In plain English

A point estimate is a single best-guess number for something you can’t observe directly — here, the true proportion in a whole population, guessed from a sample. The obvious guess is the proportion you actually saw, x ∕ n. It’s a perfectly good estimate when the sample is large and the rate is nowhere near 0 or 1, but it stumbles at the edges: from 0 successes it returns 0, blandly asserting the event can never happen. The other estimates fix this by nudging the guess a little toward the middle, by an amount that fades away as the sample grows.

point estimate
A single value offered as the best guess of an unknown parameter, as opposed to an interval. Useful, but on its own it conceals how uncertain it is.
maximum likelihood (MLE)
The estimate that makes the observed data most probable. For a proportion that is simply x ∕ n — the sample proportion. Unbiased, but high-variance and badly behaved at the boundaries.
Laplace (rule of succession)
(x + 1) ∕ (n + 2): pretend you saw one extra success and one extra failure. The classic answer to “the sun has risen n days running — what’s the chance it rises tomorrow?” — never quite 1.
Jeffreys
(x + 0.5) ∕ (n + 1): a gentler nudge, from a Bayesian Beta(½, ½) prior. Often the best all-round compromise.
Wilson
(x + z² ∕ 2) ∕ (n + z²): the centre of the Wilson score interval. The size of the nudge grows with the confidence level you ask for.
shrinkage
All three adjustments pull the estimate toward ½ by adding “pseudo-counts.” The pull is large when n is tiny and vanishes as n grows — which is exactly when the raw proportion needs help, and when it doesn’t.

Frequently asked

What is a point estimate?

A point estimate is a single number put forward as the best guess of an unknown population value — a proportion, a mean, a rate — calculated from sample data. It contrasts with an interval estimate, which gives a range. For a population proportion the natural point estimate is the sample proportion x ∕ n, but it is not the only one, and not always the best. The honest practice is to quote a point estimate and a confidence interval, because the point alone says nothing about how much it might be off.

Is the sample proportion x ∕ n the best point estimate?

It is the maximum-likelihood estimate — the value that makes your data most likely — and it is unbiased, so on average it is right. But it has high variance with small samples and it behaves badly near 0 and 1, where it can land exactly on the boundary and claim impossibility or certainty. When n is small or the proportion is extreme, a shrunk estimate (Wilson or Jeffreys) is usually a better single guess. When n is large and the proportion is middling, all of them agree and x ∕ n is fine.

What do I report when there are zero successes (or all successes)?

Not 0 (or 1). Observing 0 successes in n trials does not prove the event is impossible — it just means it didn’t happen this time. The useful figure is the upper confidence bound: the rule of three says that with 95% confidence the true rate is below roughly 3 ∕ n. So 0 in 10 is “under about 30%,” 0 in 100 is “under about 3%.” Report that bound, or a shrunk point estimate such as Jeffreys, rather than a flat zero. All-successes is the mirror image.

How do the Laplace, Jeffreys and Wilson estimates differ?

All three add “pseudo-counts” that pull the estimate toward ½, but by different amounts. Laplace adds a whole success and a whole failure, (x + 1) ∕ (n + 2) — the biggest nudge. Jeffreys adds half of each, (x + 0.5) ∕ (n + 1), corresponding to a Bayesian Beta(½, ½) prior — usually the best compromise. Wilson, (x + z² ∕ 2) ∕ (n + z²), ties the nudge to the confidence level through z, so asking for 99% confidence shrinks more than 90%. As n grows, all three converge on x ∕ n.