Point Estimate Calculator
Estimate a population proportion from sample data. Enter the number of successes and the number of trials, pick a confidence level (90%, 95%, or 99% for the Wilson method), and get the best point estimate along with all four standard estimators.
Example: with Number of successes (x) 8 · Number of trials (n) 10 · Confidence level (z for Wilson) 95% (z = 1.960) → Best point estimate: 0.8000 — maximum likelihood estimate (x / n).
- Maximum likelihood (x / n)0.8000
- Wilson estimate0.7167
- Laplace estimate0.7500
Computed by the calculator below using its default values. Change any input to see your own numbers.
Which estimator is "best" follows the common rule of thumb: Wilson when x/n ≤ 0.5, MLE for middling proportions, Jeffreys near 1, Laplace when every trial succeeded. It is a convention, not a law.
Why there is more than one point estimate
The obvious estimate of a proportion is x / n — the maximum likelihood estimate. It is unbiased and ideal for middling results, but it behaves badly at the extremes: 0 successes in 10 trials gives an estimate of exactly 0%, which claims impossible certainty from ten observations.
The other estimators fix that by adding pseudo-observations. Laplace adds one success and one failure ((x + 1) / (n + 2)); Jeffreys adds half of each ((x + 0.5) / (n + 1)); Wilson pulls the estimate toward 0.5 by an amount tied to the z value ((x + z²/2) / (n + z²)). All three shrink extreme results toward the middle, which usually predicts future data better with small samples. As n grows, all four converge on x / n.
How it’s calculated
MLE = x / n. Laplace = (x + 1) / (n + 2). Jeffreys = (x + 0.5) / (n + 1). Wilson = (x + z²/2) / (n + z²) with z = 1.645, 1.960, or 2.576 for 90%, 95%, 99%. The "best" pick follows the common heuristic: x/n ≤ 0.5 → Wilson; 0.5 < x/n < 0.9 → MLE; 0.9 ≤ x/n < 1 → Jeffreys; x/n = 1 → Laplace. Default example: 8 of 10 gives MLE 0.8000, Wilson 0.7167, Laplace 0.7500, Jeffreys 0.7727.
Assumes independent trials with a constant success probability (binomial sampling); the best-estimator rule is a widely used convention, not a theorem.
Four estimators of a proportion
| Method | Formula | Typically preferred when |
|---|---|---|
| Maximum likelihood | x / n | Middling proportions (0.5 < x/n < 0.9) |
| Wilson | (x + z²/2) / (n + z²) | Small samples or x/n ≤ 0.5 |
| Jeffreys | (x + 0.5) / (n + 1) | Proportions near 1 (0.9 ≤ x/n < 1) |
| Laplace | (x + 1) / (n + 2) | Every trial succeeded (x = n) |
Formulas from standard binomial estimation (Laplace rule of succession, Jeffreys prior, Wilson score center).
Common mistakes
- Reporting 0% or 100% from a small sample — with n = 10, a result of 10/10 is better summarized by Laplace’s 0.9167 than by a flat 1.0.
- Entering percentages instead of counts: the inputs are the raw number of successes and trials, not 80 and 100 percent.
- Confusing the point estimate with a confidence interval — the estimate is a single best guess and says nothing about its own uncertainty.
- Using x > n (more successes than trials), which makes the proportion meaningless.
Frequently asked questions
What is the point estimate formula?
The basic one is p̂ = x / n, the maximum likelihood estimate. Alternatives smooth extremes: Laplace (x + 1) / (n + 2), Jeffreys (x + 0.5) / (n + 1), and Wilson (x + z²/2) / (n + z²).
Which point estimate is best?
A common rule of thumb: use Wilson when x/n ≤ 0.5, plain x/n for middling values, Jeffreys when x/n is 0.9 or above, and Laplace when x = n. With large samples the four differ by very little.
Is a point estimate the same as a confidence interval?
No. The point estimate is a single number — the best guess. A confidence interval wraps a margin of error around it. Report both when the sample is small.
Why not just use x / n every time?
Because at the extremes it claims certainty it does not have: 0/10 gives exactly 0%. The smoothed estimators add fractional pseudo-counts so small samples produce sensible, slightly moderated estimates.
What does the z value change?
Only the Wilson estimate. A larger z (higher confidence level) adds more pseudo-observations, pulling the estimate further toward 0.5. At 95%, z = 1.960 adds about 3.84 trials worth of smoothing.