Can two unbiased estimators have different accuracy?

Yes. Their sampling variances can differ.

Is unbiasedness the same as consistency?

No. They describe different properties.

Does one estimate above the truth prove bias?

No. Bias concerns expectation across the specified sampling process.

AI Glossary / Governance, ethics and safety / Advanced

Unbiased estimator

Definition

An unbiased estimator is a rule for estimating an unknown quantity whose expected value equals that quantity under the assumed sampling model. Across repeated samples drawn in the specified way, its estimates are centered on the target. This does not mean that every individual estimate is correct or that the method guarantees fair treatment of people.

What it means

An estimator is the rule; an estimate is the number obtained when that rule is applied to observed data. For example, “take the sample average” is an estimator, while an observed average of 12.4 minutes is an estimate.
Unbiasedness is a property of the rule's sampling distribution. Write the unknown parameter as theta and the estimator as T. Mean-unbiasedness means E[T] = theta, so the statistical bias E[T] - theta is zero. The expectation must exist, and a general claim of unbiasedness applies throughout the stated parameter model.
A single dataset normally cannot demonstrate this property by itself. The claim depends on the estimator, the target, and assumptions about how data are generated or sampled.

How it works

For observations with common finite mean mu, the sample average has expectation mu by linearity of expectation. This establishes unbiasedness for that mean. Independence is not required for this expectation calculation, although it matters for familiar variance formulas and other results.
For independent, identically distributed observations with finite variance sigma squared, the sample average's variance is sigma squared divided by the sample size. Thus, sample size affects precision even when unbiasedness already holds.
Unbiasedness differs from consistency, which concerns approaching the target as sample size increases. It also differs from low mean squared error. When the necessary moments exist, mean squared error equals variance plus squared bias. A biased estimator can sometimes have lower total squared error by reducing variance enough.

Three practical examples

An analyst averages response times from a properly designed random sample of requests. The sample mean can be unbiased for the target population mean under the sampling design. Averaging only conveniently available requests may not preserve that claim.
A quality team estimates a defect probability using the proportion of defective items in independent Bernoulli observations with a common probability. The sample proportion is unbiased for that probability, although any one sample can overestimate or underestimate it.
A researcher estimates population variance from independent, identically distributed observations. The familiar sample variance divides squared deviations from the sample mean by n - 1. Under the usual finite-variance assumptions and n greater than one, it is unbiased for population variance; taking its square root does not generally preserve unbiasedness.

Think of it this way

Imagine repeatedly measuring a target with a method whose results scatter around the correct value. Being centered correctly is different from having a tight cluster. An unbiased estimator addresses the first issue. Precision addresses the spread, and both can matter when making a decision from one sample.

Worked example

Consider a constructed population in which an observation is equally likely to be 2 or 8. Its mean is (2 + 8)/2 = 5. Draw two observations independently with replacement.
The four equally likely ordered samples are (2, 2), (2, 8), (8, 2), and (8, 8). Their sample means are 2, 5, 5, and 8. The expected sample mean is therefore (2 + 5 + 5 + 8)/4 = 5, matching the population mean.
The estimator is unbiased, but two of the four possible estimates differ from the target by three. Its mean squared error is (9 + 0 + 0 + 9)/4 = 4.5.
Now use only the first observation as an estimator of the mean. Its expectation is also 5, so it is unbiased. However, its squared error is always 9. The two-observation average is more precise in this example even though both estimators are unbiased.
This calculation enumerates every possible sample in a deliberately simple model. It illustrates the definition without suggesting that real operational data necessarily follow these assumptions.

Illustrative case study

Fictional company Cedar Support wants to estimate average handling time across all completed requests. Its analyst selects requests randomly from a complete register and uses their sample mean. The report explains the target population and the sampling method.
A manager proposes replacing the random sample with the first requests found in a convenient dashboard. The dashboard excludes some requests that required specialist support. The analyst explains that the original unbiasedness argument does not automatically apply to this new selection process.
The team retains the defined sampling procedure and reports sampling uncertainty alongside the estimate. It also checks whether recorded times measure the intended quantity.
This fictional case illustrates that a sound formula cannot rescue an unsupported sampling assumption. It does not claim that unbiasedness eliminates measurement errors, missing records, or uncertainty in an individual result.

Common misconceptions and mistakes

“Unbiased means exactly correct.” Individual estimates can be far from the target.
“Unbiased means socially fair.” Statistical centering does not establish fairness in a decision process.
“An unbiased estimator is always best.” Variance, loss, and practical requirements also matter.
“Any transformation stays unbiased.” Nonlinear transformations do not generally preserve the property.
“A larger sample removes every problem.” More observations do not automatically fix systematic selection or measurement problems.

Business use

Ask what parameter an estimate targets and which assumptions support the method. Distinguish the statistical property from claims about data quality or ethical use. Document the sampling frame, inclusion process, and treatment of missing observations.
When comparing methods, consider uncertainty and the cost of errors as well as bias. A useful report should explain how much an estimate might vary, not merely describe its estimator as unbiased. Revisit the reasoning when the data-collection process changes.

Check your understanding

Why is the worked-example sample mean unbiased?

Answer: Its expected value is 5, the population mean.

Why is using only the first observation less precise here?

Answer: Its mean squared error is 9 instead of 4.5.

What should accompany an unbiasedness claim?

Answer: A clear target, estimator, and sampling assumptions.

Related terms

Learn next

Enumerate the possible samples in a small population. Compute each estimate, its probability, and its error. Compare the expected estimate with the target and separately examine the spread of estimates.

Sources

Aalto Dictionary of Machine Learning, unbiased estimator: https://aaltodictionaryofml.github.io/ADictML_Main.pdf Marco Taboga, Estimation of the mean: https://www.statlect.com/fundamentals-of-statistics/mean-estimation Hossein Pishro-Nik, Point Estimators for Mean and Variance: https://www.probabilitycourse.com/chapter8/8_2_2_point_estimators_for_mean_and_var.php

Previous: Unawareness (to a sensitive attribute) Next: Unbounded cost search