Loading
Understanding when a sample statistic systematically hits or misses the true population parameter.
The idea that a single number computed from sample data could serve as a reliable stand-in for an unknown population characteristic is one of the most powerful—and most subtle—ideas in all of statistics. Early practitioners of probability, working with astronomical observations and census data, recognized that different formulas applied to the same data could yield systematically different estimates of the same quantity. The formal study of point estimation grew out of a need to choose, among competing formulas, those that would on average recover the true parameter value. This pursuit drove major theoretical breakthroughs from the eighteenth century onward and remains central to the AP Statistics curriculum today.
The central question that this lesson addresses is deceptively simple: if you compute a statistic from a random sample—say, the sample mean or the sample variance—does that statistic, on average across all possible samples, equal the population parameter it is meant to estimate? When the answer is yes, the estimator is unbiased; when the answer is no, the estimator is biased, and the direction and magnitude of that bias have real consequences for inference.
Before distinguishing biased from unbiased estimators, we need a clear vocabulary. A parameter is a fixed numerical characteristic of a population—such as the population mean μ or the population proportion p. A statistic is a numerical value computed from sample data—such as the sample mean x̄ or the sample proportion p̂. When we use a statistic to approximate a parameter, we call that statistic a point estimate. The theoretical long-run behavior of that statistic across all possible random samples of the same size constitutes its sampling distribution, and the properties of that distribution determine whether the estimator is biased or unbiased.
The target analogy is the single most useful mental model for bias and variability. Notice that the upper-left target (unbiased, low variance) is the ideal: every sample estimate clusters near the bullseye. The upper-right target (unbiased, high variance) still averages to the center, but individual estimates wander far—this corresponds to a small sample size or a highly variable population. The lower-left target (biased, low variance) is deceptive: the estimates are precise but systematically wrong—much like a bathroom scale that always reads 3 pounds too high. Finally, the lower-right target (biased, high variance) is the worst case, offering neither accuracy nor precision.
The formal definition of bias is built on the concept of the expected value of a statistic. If θ is a population parameter and θ̂ is a statistic computed from a random sample, the bias of θ̂ is defined as follows.
The most important unbiased estimator in introductory statistics is the sample mean. Suppose X₁, X₂, …, Xₙ are a random sample from a population with mean μ. The sample mean is x̄ = (1/n) Σ Xᵢ. Taking the expected value yields the following.
The sample variance provides the classic illustration of bias correction. The naive variance with divisor n is biased, whereas dividing by n − 1 produces an unbiased estimator of the population variance σ².
In practice you will encounter several statistics, and knowing which are unbiased for their corresponding parameters is essential for both the AP exam and real-world inference. The table below summarizes the most important cases.
| Statistic (Estimator) | Parameter | Unbiased? | Notes |
|---|---|---|---|
| x̄ (sample mean) | μ (population mean) | Yes | Always unbiased for any distribution, any n. |
| p̂ (sample proportion) | p (population proportion) | Yes | p̂ = X/n where X ~ Binomial(n, p); E(p̂) = p. |
| s² (sample variance, n−1) | σ² (population variance) | Yes | Bessel's correction makes this unbiased. |
| s (sample std dev) | σ (population std dev) | No | E(s) < σ; the square root introduces slight downward bias. |
| Sample median | Population median | Yes* | Unbiased for symmetric distributions; may be biased otherwise. |
| Sample range | Population range | No | E(range) < population range; always underestimates. |
The diagram above conveys the central idea: an unbiased estimator's sampling distribution is centered at the true parameter value, while a biased estimator's sampling distribution is shifted. The horizontal gap between the two dashed vertical lines is exactly the bias, E(θ̂) − θ. A positive bias means the estimator tends to overestimate the parameter; a negative bias means it tends to underestimate. Importantly, any single sample's estimate may fall on either side of the parameter—unbiasedness is a property of the long-run average, not of any individual sample.
A population consists of five values: {2, 4, 6, 8, 10}. We draw all possible simple random samples of size n = 2 (without replacement) and compute x̄ for each. We will verify that x̄ is an unbiased estimator of the population mean μ.
| Property | Unbiased Estimators | Biased Estimators |
|---|---|---|
| Long-run center | Centered on the true parameter | Systematically off-center |
| Individual sample accuracy | No guarantee any single estimate equals θ | Same — no single-sample guarantee |
| Variance | Not necessarily minimal | May have lower variance (bias-variance trade-off) |
| Mean Squared Error (MSE) | MSE = Var(θ̂) | MSE = Var(θ̂) + [Bias]²; can be smaller overall |
| AP Statistics emphasis | Required knowledge: x̄, p̂, s² are unbiased | Know s is biased; range is biased |
Unbiased point estimates are the foundation on which the rest of statistical inference is built. When you construct a confidence interval, the center of that interval is your point estimate—usually x̄ or p̂—and the interval extends outward by a margin of error determined by the standard error. If the point estimate were biased, every confidence interval you constructed would be systematically shifted in the wrong direction, and your stated confidence level would no longer be valid. Similarly, in hypothesis testing, test statistics are built from point estimates; bias in those estimates would distort p-values and lead to incorrect conclusions about significance.
| Concept in This Lesson | How It Extends |
|---|---|
| E(x̄) = μ (unbiasedness) | Justifies centering a confidence interval for μ at x̄ |
| E(p̂) = p (unbiasedness) | Justifies one-proportion z-tests and confidence intervals |
| E(s²) = σ² (unbiasedness) | Justifies using s in t-procedures when σ is unknown |
| Sampling variability (SE) | Determines margin of error and power of tests |
| Bias-variance trade-off | Central to regression shrinkage methods (LASSO, Ridge) in college-level courses |
As you progress to confidence intervals and hypothesis tests later in the AP course, keep this lesson's core principle in mind: every inferential procedure you learn relies on having a point estimate whose sampling distribution is centered at the parameter. Understanding why that centering matters—and recognizing the few cases where common statistics fail to be unbiased—will strengthen your reasoning on both multiple-choice and free-response questions.
A point estimate is a single-number summary computed from sample data to approximate a population parameter. An estimator is unbiased when the mean of its sampling distribution equals the parameter, that is, E(θ̂) = θ. The sample mean x̄ is unbiased for μ, the sample proportion p̂ is unbiased for p, and the sample variance s² (with the n − 1 divisor via Bessel's correction) is unbiased for σ².
A biased estimator has E(θ̂) ≠ θ; the sample standard deviation s is a common example of a statistic that is slightly biased (downward) for σ. Bias and variability are independent properties—a good estimator minimizes both. Unbiased point estimates form the foundation for confidence intervals and hypothesis tests throughout the AP Statistics curriculum.
Keep learning with more lessons from the same subject.