Biased vs unbiased estimator

By Jude Wallis · Published

An estimator is unbiased when the mean of its sampling distribution equals the parameter it estimates, so it is right on average over all samples. It is biased when that mean sits above or below the parameter. The sample mean and sample proportion are unbiased; the sample maximum is biased.

AP Statistics: Unit 3 (topics 3.1 Estimators). Biased and unbiased estimators belong to Unit 3 topic 3.1, Estimators, in the Fall 2026 AP Statistics course, where an estimator is judged by the center and the spread of its sampling distribution. The sample proportion is the unbiased estimator used throughout Unit 3, and the sample mean plays the same role in Unit 4.

Biased vs unbiased estimator: the short answer

An estimator is a rule for turning a sample into a guess at a population parameter. Because samples vary, the guess varies, and the collection of guesses over all possible samples is the estimator's sampling distribution.

An estimator is unbiased when that sampling distribution is centered exactly on the parameter. Individual estimates miss high and low, and the misses balance out.

An estimator is biased when the sampling distribution is centered somewhere else. The estimates tend to land on one side of the parameter, so there is a systematic error on top of the ordinary sample-to-sample variation.

The labels describe the rule over many samples. They do not describe a single estimate, and they say nothing about how spread out the estimates are.

What makes an estimator unbiased

Write θ\theta ("theta") for the parameter and θ^\hat{\theta} ("theta-hat") for the estimator. The estimator is unbiased when

E(θ^)=θE(\hat{\theta}) = \theta

that is, when the mean of its sampling distribution equals the parameter, whatever the parameter's value happens to be.

Three estimators in the course have this property when the data come from a random sample:

  • The sample mean xˉ\bar{x} ("x-bar") for the population mean μ\mu.
  • The sample proportion p^\hat{p} ("p-hat") for the population proportion pp.
  • The sample variance s2s^2, computed with the divisor n−1n - 1, for the population variance σ2\sigma^2, when the observations are independent.

Unbiased does not mean accurate in any one sample. In the second worked example below, p^\hat{p} is unbiased for p=0.40p = 0.40 and cannot equal 0.40 in any sample. The property is about the long-run average, which is the idea topic 3.1 Estimators introduces.

What makes an estimator biased

The bias of an estimator is the gap between the center of its sampling distribution and the parameter:

bias=E(θ^)−θ\text{bias} = E(\hat{\theta}) - \theta

A negative bias means the estimator tends to come out too low, and a positive bias means too high. Three examples:

  • The sample maximum as an estimate of the population maximum. A sample cannot contain a value larger than the largest value in the population, so the sample maximum is either exactly right or too small. Its average over all samples has to fall below the true maximum.
  • The variance with divisor nn in place of n−1n - 1. For independent observations its average is n−1nσ2\frac{n-1}{n}\sigma^2, which is below σ2\sigma^2. This is the reason for the n−1n - 1, as why divide by n minus 1 shows.
  • The sample standard deviation ss as an estimate of σ\sigma. Even though s2s^2 is unbiased for σ2\sigma^2, taking the square root leaves ss slightly low on average.

In each case the cause is the formula. No amount of care in drawing the sample removes it.

The differences side by side

FeatureUnbiased estimatorBiased estimator
Center of the sampling distributionEqual to the parameterAbove or below the parameter
Bias, E(θ^)−θE(\hat{\theta}) - \theta0Not 0
A single estimateCan still miss the parameterCan still land close to it
Direction of the missesBalanced high and lowTilted to one side
Examplesxˉ\bar{x} for μ\mu, p^\hat{p} for pp, s2s^2 for σ2\sigma^2Sample maximum, variance with divisor nn, ss for σ\sigma
What it says about spreadNothingNothing

The last row matters as much as the first. Bias locates the center of the sampling distribution, and variability measures its width. An unbiased estimator with a wide sampling distribution can produce estimates far from the parameter, and a slightly biased estimator with a narrow one can produce estimates that are reliably close. The two properties are compared in bias vs variability.

When to use which

Given a choice, prefer an unbiased estimator with low variability. That is why inference in the course is built on xˉ\bar{x} and p^\hat{p}: each is centered on its parameter, and each gets less variable as the sample size grows.

A biased estimator is acceptable when the bias is small and well understood. The clearest case is ss. It runs slightly low as an estimate of σ\sigma, and it is still the standard deviation used in every tt procedure, because at ordinary sample sizes the shortfall is small.

A biased estimator is a poor choice when the bias is large relative to the precision you need. Estimating a population maximum by the sample maximum from a small sample is an example: the estimate can never be too high and is often too low, and nothing in the sample tells you by how much.

For the three biased examples on this page, the bias shrinks as the sample size increases, though it does not reach zero. That behavior belongs to these formulas and is not a general rule about bias.

The classic mix-up and how to avoid it

The first error is calling a single estimate biased. "My sample mean was 5.5 and the true mean is 5, so my estimate is biased" confuses an ordinary miss with a property of the method. One estimate has an error. Only an estimator, judged over all possible samples, has a bias.

The second error is reading unbiased as correct. An unbiased estimator is centered on the parameter and can still miss it in every individual sample.

The third error is mixing up two sources of bias. This page is about bias that comes from the formula. A separate kind comes from how the data were collected: a convenience sample or a voluntary response sample can push even xˉ\bar{x} and p^\hat{p} off center, because the sample no longer represents the population. Taking a larger sample of the same flawed kind does not fix that, as does a bigger sample fix bias explains. The statement that xˉ\bar{x} and p^\hat{p} are unbiased assumes random sampling.

To check a claim of bias, ask what the estimator averages to over all possible random samples and compare that with the parameter.

Three estimators checked against every possible sample

A population consists of the three values 2, 4, and 9. List all samples of size 2 drawn with replacement. For each, compute the sample mean, the sample variance s2s^2 with divisor n−1n - 1, and the sample maximum. Decide which of these is an unbiased estimator of the population mean, the population variance, and the population maximum, respectively.

  1. Find the parameters. Mean: μ=2+4+93=5\mu = \frac{2 + 4 + 9}{3} = 5. Variance: σ2=(2−5)2+(4−5)2+(9−5)23=9+1+163=8.667\sigma^2 = \frac{(2-5)^2 + (4-5)^2 + (9-5)^2}{3} = \frac{9 + 1 + 16}{3} = 8.667. Maximum: 9.

  2. List the 9 equally likely samples: (2,2), (2,4), (2,9), (4,2), (4,4), (4,9), (9,2), (9,4), (9,9).

  3. Sample means, in that order: 2, 3, 5.5, 3, 4, 6.5, 5.5, 6.5, 9. Their sum is 45, so they average 459=5\frac{45}{9} = 5.

  4. Sample variances. For two values aa and bb, s2=(a−b)22s^2 = \frac{(a-b)^2}{2}, giving, in order: 0, 2, 24.5, 2, 0, 12.5, 24.5, 12.5, 0. Their sum is 78, so they average 789=8.667\frac{78}{9} = 8.667.

  5. Sample maximums, in order: 2, 4, 9, 4, 4, 9, 9, 9, 9. Their sum is 59, so they average 599=6.556\frac{59}{9} = 6.556.

  6. Compare each average with its parameter. The sample mean averages 5 and μ=5\mu = 5. The sample variance averages 8.667 and σ2=8.667\sigma^2 = 8.667. The sample maximum averages 6.556 and the population maximum is 9, a bias of 6.556−9=−2.4446.556 - 9 = -2.444.

  7. Check the divisor. Using n=2n = 2 as the divisor halves each variance in the list, so those values average 399=4.333\frac{39}{9} = 4.333, half of σ2\sigma^2 and equal to n−1nσ2=12(8.667)\frac{n-1}{n}\sigma^2 = \frac{1}{2}(8.667).

The sample mean is unbiased for μ\mu: it averages exactly 5. The sample variance with divisor n−1n - 1 is unbiased for σ2\sigma^2: it averages exactly 8.667. The sample maximum is biased: it averages 6.556 against a true maximum of 9, a bias of −2.444-2.444. None of the nine sample means equals 5, so being unbiased did not make any individual estimate correct.

An unbiased estimator that never equals the parameter

A spinner lands on red with probability p=0.40p = 0.40 on every spin. It is spun 3 times and p^\hat{p} is the proportion of spins that land on red. Find the sampling distribution of p^\hat{p} and its mean, and say whether p^\hat{p} is unbiased.

  1. The number of reds XX is binomial with n=3n = 3 and p=0.40p = 0.40, so p^=X3\hat{p} = \frac{X}{3} can be 0, 13\frac{1}{3}, 23\frac{2}{3}, or 1.

  2. P(X=0)=0.603=0.216P(X = 0) = 0.60^3 = 0.216.

  3. P(X=1)=3(0.40)(0.60)2=3(0.40)(0.36)=0.432P(X = 1) = 3(0.40)(0.60)^2 = 3(0.40)(0.36) = 0.432.

  4. P(X=2)=3(0.40)2(0.60)=3(0.16)(0.60)=0.288P(X = 2) = 3(0.40)^2(0.60) = 3(0.16)(0.60) = 0.288.

  5. P(X=3)=0.403=0.064P(X = 3) = 0.40^3 = 0.064. Check: 0.216+0.432+0.288+0.064=1.0000.216 + 0.432 + 0.288 + 0.064 = 1.000.

  6. Mean of p^\hat{p}: 0(0.216)+13(0.432)+23(0.288)+1(0.064)=0+0.144+0.192+0.064=0.4000(0.216) + \frac{1}{3}(0.432) + \frac{2}{3}(0.288) + 1(0.064) = 0 + 0.144 + 0.192 + 0.064 = 0.400.

  7. Compare with the parameter: the mean of the sampling distribution is 0.400 and p=0.40p = 0.40, so the bias is 0.

The sampling distribution of p^\hat{p} puts probabilities 0.216, 0.432, 0.288, and 0.064 on the values 0, 13\frac{1}{3}, 23\frac{2}{3}, and 1. Its mean is 0.400, equal to pp, so p^\hat{p} is unbiased. None of the four possible values equals 0.40, which shows that unbiased describes the center of the sampling distribution and makes no promise about any single estimate.

Frequently asked questions

Does unbiased mean the estimate equals the parameter?

No. Unbiased means the estimates average out to the parameter over all possible samples. Any single estimate can be too high or too low. With three spins of a spinner that lands on red 40% of the time, the sample proportion is unbiased and can only be 0, 13\frac{1}{3}, 23\frac{2}{3}, or 1.

Is the sample standard deviation an unbiased estimator?

Not of σ\sigma. The sample variance s2s^2, with divisor n−1n - 1, is unbiased for σ2\sigma^2, but taking the square root leaves ss slightly too small on average. The shortfall shrinks as the sample size grows, and ss is still what inference for means uses.

Is an unbiased estimator always better than a biased one?

No. Bias is only one of two properties. An unbiased estimator with high variability can give estimates far from the parameter, while a slightly biased estimator with low variability can be close most of the time. The ideal is no bias and low variability together.