Bias (statistics)

In statistical inference, bias measures how far an Estimator systematically deviates from the true value it's meant to capture. It's the persistent lean in one direction—like a scale that always reads 2 pounds heavier than reality.

A biased estimator produces results that cluster around a wrong answer rather than the right one. This differs from Variance, which describes how scattered those results are. You might hit your target consistently (low bias, low variance), wildly but centered on average (high variance, low bias), or miss systematically in one direction (high bias)—the worst outcome.

Bias arises from Sampling methods, Measurement error, model assumptions, or incomplete data. Selection bias, where certain groups are overrepresented, is particularly insidious in social research. Measurement bias creeps in through faulty instruments or encoding errors.

Understanding bias matters because unbiased doesn't always mean better—sometimes accepting a small bias reduces overall error when combined with variance control. This trade-off shapes modern machine learning and experimental design.

The concept became rigorous through 20th-century statistics, though practical awareness of systematic error runs far deeper in human history.

Related

Estimator, Variance (statistics), Sampling methods, Confounding variable, Statistical inference