An unbiased estimator is a statistical formula whose average result equals the true value it’s trying to measure. If you could repeat your sampling process infinitely many times and average all the estimates together, an unbiased estimator would land exactly on the true population value. A biased estimator, by contrast, would consistently overshoot or undershoot it.
The Core Idea
In statistics, you rarely have access to an entire population. Instead, you take a sample and use it to estimate some characteristic of the whole group, like its average or its spread. The formula you use to calculate that estimate from your sample data is called an estimator.
An estimator is unbiased when its expected value equals the true parameter. “Expected value” here means the long-run average: if you drew thousands of different random samples and applied the same formula each time, the average of all those estimates would equal the real population value. Any single estimate might be too high or too low, but the errors balance out over many repetitions. There’s no systematic lean in one direction.
When that balance doesn’t hold, the estimator is biased. A biased estimator consistently misses in one direction. It might, on average, underestimate the true value by a predictable amount.
The Sample Mean: A Classic Example
The simplest unbiased estimator is the sample mean. If you want to know the average height of adults in a country, you can measure a random sample and compute their average. That sample average is an unbiased estimator of the population average. No matter how large or small your sample, the expected value of the sample mean equals the true population mean. With a tiny sample, any individual estimate could be far off, but there’s no systematic tendency to be too high or too low.
The sample proportion works the same way. If 40% of voters support a policy and you randomly survey people, the proportion in your sample is an unbiased estimator of that 40% figure.
Why You Divide by n Minus 1 for Variance
This is where unbiasedness gets interesting and counterintuitive. When you calculate how spread out a population is, you sum up the squared differences from the mean and divide by the population size. But when you estimate that spread from a sample, dividing by the sample size gives you a biased result. It systematically underestimates the true population variance.
The math shows exactly how much it misses. If your sample has n data points, dividing by n produces an estimate whose expected value is only (n-1)/n times the true variance. For a sample of 10, that means you’d underestimate by about 10% on average. For a sample of 100, you’d underestimate by about 1%.
The fix is simple: divide by n-1 instead of n. This adjustment, called Bessel’s correction, inflates the estimate just enough to cancel out the systematic underestimate. The corrected formula produces an unbiased estimator of population variance. This is why your calculator or statistics software defaults to dividing by n-1 when computing sample variance.
Here’s a related quirk: even though dividing by n-1 gives you an unbiased estimate of variance, taking the square root of that result does not give you an unbiased estimate of standard deviation. Unbiasedness doesn’t automatically transfer through mathematical operations.
Maximum Likelihood Estimators Are Often Biased
Maximum likelihood estimation is one of the most popular methods for fitting statistical models. It finds the parameter values that make your observed data most probable. These estimators have many appealing properties, but being unbiased isn’t always one of them.
The variance example illustrates this directly. The maximum likelihood estimator for variance in a normal distribution divides by n rather than n-1. Its expected value is (n-1)/n times the true variance, not the true variance itself. It’s biased. The bias shrinks as your sample grows, but it’s always present in finite samples.
This is a useful reminder that “maximum likelihood” and “unbiased” are separate properties. An estimator can be one without being the other.
Unbiased vs. Consistent
These two terms sound similar but describe different things. An unbiased estimator has no systematic error at any sample size. A consistent estimator gets closer and closer to the true value as your sample grows, eventually converging on it.
An estimator can be unbiased but not consistent. Consider estimating the upper bound of a uniform distribution. If you just take one observation and double it (ignoring the rest of your data), the expected value equals the true parameter, so it’s unbiased. But it never improves no matter how many additional observations you collect, because it ignores them. That makes it inconsistent.
An estimator can also be biased but consistent. The maximum likelihood estimator for the uniform distribution’s upper bound has a bias of roughly 1/(n+1) times the true value. It’s biased at every finite sample size, but as n grows, that bias shrinks toward zero and the estimate converges on the truth. Statisticians call this “asymptotically unbiased.”
The ideal is both: unbiased at every sample size and consistent as the sample grows. The sample mean for a uniform distribution has both properties.
The Bias-Variance Tradeoff
Unbiasedness sounds like an obviously good property, but it’s not always the most important one. The total error of an estimator, measured by mean squared error, breaks down into two components: the squared bias plus the variance. An unbiased estimator has zero bias, but it might have high variance, meaning individual estimates scatter widely around the true value.
Sometimes a slightly biased estimator with much lower variance produces estimates that are, on average, closer to the truth than an unbiased one that bounces around wildly. This tradeoff matters especially in prediction and machine learning, where a small amount of deliberate bias can dramatically reduce overall error.
In traditional statistics, unbiasedness is treated as a strong desirable property. In applied settings, it’s one consideration among several. Whether it matters most depends on your sample size, your goals, and how much variance you’re willing to tolerate.
The Minimum Variance Unbiased Estimator
If unbiasedness is your requirement, the next question is: among all unbiased estimators, which one has the smallest variance? That “best” option is called the minimum variance unbiased estimator, or MVUE. It gives you the tightest possible estimates without any systematic error.
Finding the MVUE involves identifying a sufficient statistic, one that captures all the information in the data about the parameter you’re estimating, and then confirming that it meets a mathematical completeness condition. When these conditions hold, there’s exactly one unbiased function of that statistic, and it’s the MVUE. In practice, for common distributions like the normal distribution, these estimators have already been worked out. The sample mean is the MVUE for the population mean of a normal distribution, and the sample variance (with the n-1 denominator) is the MVUE for the population variance.