Can a biased estimator be useful?

Yes. It can have lower mean squared error than an unbiased alternative in a specified setting.

Is bias the same as sampling error?

No. A realized estimate's error includes random variation; bias is the estimator's expected displacement.

What does asymptotically unbiased mean?

The bias approaches zero as sample size increases, though it can remain nonzero at each finite size.

AI Glossary / Governance, ethics and safety / Intermediate

Biased estimator

Definition

A biased estimator is a rule for estimating a population quantity whose expected result differs from that quantity under the specified sampling model. Bias concerns the average across hypothetical repeated samples, not whether one particular estimate happens to be above or below the truth.

What it means

An estimator is the calculation procedure; an estimate is the value obtained from one dataset. If the true quantity is theta and the estimator is T, its bias is E[T] − theta. The expectation describes the estimator's sampling distribution under stated assumptions.

An unbiased estimator can still give a poor result on an individual sample. A biased estimator can sometimes be closer, on average under a chosen loss, because it varies less. Bias is therefore one property to assess alongside variance, mean squared error and suitability for the task.

This statistical meaning is different from discrimination or unfair treatment. Describing a formula as unbiased does not establish that its data, target or downstream use is fair.

How it works

To determine bias analytically, specify how samples are generated and calculate the estimator's expected value. Compare that expectation with the population parameter being estimated. Simulations can illustrate the property when the data-generating process and true parameter are known, although finite simulation results are approximate.

A familiar example is estimating population variance from independent, identically distributed observations with finite variance. Averaging squared deviations from the sample mean with denominator n systematically underestimates the population variance in expectation. Using n − 1 instead gives an unbiased variance estimator for n greater than one under those assumptions.

This correction applies to variance, not automatically to every transformed quantity. Taking the square root of an unbiased variance estimate does not generally give an unbiased standard-deviation estimate. Also, an estimator's bias may shrink as sample size grows without being exactly zero at a finite sample size.

Three practical examples

An analyst checks whether a variance calculation uses a population-description formula or an estimator intended to infer population variance from a sample.
A modeling team uses shrinkage to accept some estimator bias in exchange for lower variability and better evaluated prediction.
A simulation study compares estimated averages with a known generating parameter to examine finite-sample bias under a specified model.

Think of it this way

Think of repeatedly throwing darts and looking at the center of all the landing positions. If that center is displaced from the target, the procedure has a systematic offset in this analogy.

A single dart does not reveal that center, and a widely scattered unbiased set can still contain many poor throws. Real estimation also depends on a sampling model; the analogy should not suggest that the true population quantity is always visible to the analyst.

Worked example

Enumerate all possible samples

Consider an invented population that produces 0 or 2 with equal probability. Its mean is 1 and its variance is 1. Draw two independent observations. The four ordered samples, each with probability one quarter, are (0,0), (0,2), (2,0) and (2,2).

The variance calculation using denominator n = 2 gives 0, 1, 1 and 0 respectively. Its expected value is (0 + 1 + 1 + 0)/4 = 0.5, so its bias as an estimator of variance 1 is −0.5.

Using denominator n − 1 = 1 gives 0, 2, 2 and 0, whose expectation is 1. That estimator is unbiased here, even though none of its four results equals the true variance. The example makes the repeated-sampling definition concrete and shows why unbiased does not mean exactly correct on each dataset.

Illustrative case study

Fictional Fern Packaging compares variation in package weights across small samples. Two spreadsheets disagree because one divides squared deviations by the number of observations while the other divides by one less.

The team first establishes the purpose: describe those recorded packages, or estimate the wider process variance from a sample. It documents the formula and checks the sampling assumptions. It also recognizes that repeated readings from a single batch may not represent independent observations from the whole process.

The team does not simply label one spreadsheet wrong in every context. The case illustrates that a formula's interpretation depends on the target and sampling design, and that changing the denominator cannot repair an unrepresentative data-collection process.

Common misconceptions and mistakes

A biased estimator is wrong on every sample. Bias is an expectation, so individual estimates can lie on either side or equal the truth.
Unbiasedness guarantees low error. Variability may still be large.
More observations always remove all bias. Some sources persist; asymptotic behavior depends on the estimator and assumptions.
An unbiased transformation remains unbiased. Nonlinear transformations generally require separate analysis.
Statistical unbiasedness proves fairness. These are different questions.

Business use

Businesses should ask which quantity a reported number estimates and under what sampling assumptions. This matters when comparing software summaries, interpreting small samples or selecting models that use regularization. Document enough detail to distinguish a descriptive statistic from an inferential estimate.

When choosing an estimator, consider the consequences of errors and the relevant loss rather than treating unbiasedness as the only criterion. A slight systematic offset may be acceptable in exchange for much lower instability, but that trade-off should be justified and evaluated. Where sampling or measurement is defective, a theoretically unbiased formula may still answer the wrong practical question.

Check your understanding

An estimator equals the true parameter in one observed dataset. Does that prove it is unbiased?

Answer: No. Unbiasedness concerns its expected value across the stated sampling distribution, not one successful estimate.

Related terms

Learn next

Read sampling distribution and mean squared error. Create a tiny population with known values, enumerate its possible small samples and compare the average output of two estimation rules with the true target.

Sources

Glossary of Forecasting Terms, biased estimator: https://robjhyndman.com/mwh3/FG4.pdf. NIST, Measures of Scale: https://www.itl.nist.gov/div898/handbook/eda/section3/eda356.htm. Scikit-learn, bias-variance decomposition: https://scikit-learn.org/stable/auto_examples/ensemble/plot_bias_variance.html. The enumerated population and packaging case are original teaching examples.

Previous: Bias-variance trade-off Next: Biclustering