Does averaging remove random error completely?

No. It usually reduces its variability under suitable assumptions.

Is a standard error the error of one observation?

No. It describes an estimator's sampling variability.

Can a result be precise but inaccurate?

Yes. Small random variation can coexist with a substantial systematic offset.

AI Glossary / Statistics and probability / Beginner

Random error

Definition

Random error is the component of an estimate's or measurement's deviation that varies unpredictably across repetitions of a specified process. It can arise because a sample contains only some population members or because repeated measurements fluctuate. Random error concerns variability, whereas systematic error concerns a persistent tendency to miss the target in a particular way. The distinction depends on the process and reference quantity being considered; describing an error as random does not mean its magnitude is unimportant.

What it means

Two properly conducted samples can give different estimates even when they target the same population. Similarly, an instrument can give slightly different readings of a stable object. These differences can reflect random variation rather than a mistake by the analyst.

The exact error of an individual estimate is often unknown because the true value is unknown. Statistical models instead describe its possible variation across repetitions. A standard error measures uncertainty associated with an estimator under a specified sampling model; it is not the actual signed error in the particular estimate observed. Measurement variability and sampling variability should also be distinguished.

How it works

First define the quantity to be measured or estimated and the repetitions that make sense. For a stable measurement target, a simple model is observed value = true value + systematic offset + random fluctuation. The random fluctuation is often modeled as having mean zero, while its variance describes repeatability under those conditions.

For an average of n independent, equally variable fluctuations with standard deviation sigma, the standard deviation of their average is sigma divided by the square root of n. This explains why repetition can improve precision. It does not imply that a systematic offset shrinks with n.

Check independence and stability before applying this rule. Readings collected close together may share disturbances, and repeated observations of one unit do not replace independent population units. Use calibration, study design, repeated measurements, and appropriate variance estimation to separate sources of uncertainty. Report which sources are represented in an uncertainty interval and which remain outside it.

Three practical examples

1. A warehouse scale gives slightly different readings for the same stable package, prompting an assessment of measurement repeatability.

2. A customer survey estimates satisfaction from a probability sample and reports sampling uncertainty because another sample would give a different result.

3. A machine-learning team evaluates performance on several appropriately independent test samples to understand variability caused by finite evaluation data.

Think of it this way

Think of arrows landing around a target. Their spread represents imprecision, while a persistent displacement of the whole group represents bias. Collecting more arrows can locate the center of the group more precisely, but it does not move that center onto the target. Real measurement systems may have both problems simultaneously.

Worked example

Consider an original measurement model for a stable package whose true mass is 100 grams. A scale has a fixed positive offset of 2 grams and independent, zero-mean measurement fluctuations with standard deviation 3 grams. Each reading therefore has expected value 102 grams.

For nine readings, the mean has random-error standard deviation 3/sqrt(9) = 1 gram. For 36 readings, that standard deviation becomes 3/sqrt(36) = 0.5 gram. Quadrupling the number of independent readings halves this component of uncertainty.

The mean still has bias of 2 grams in both cases. Relative to the true mass, its mean squared error is bias squared plus variance. With nine readings this is 2 squared + 1 squared = 5 square grams. With 36 readings it is 2 squared + 0.5 squared = 4.25 square grams. The corresponding root mean squared errors are approximately 2.236 and 2.062 grams.

These calculations assume the offset remains fixed, the target is stable, and the fluctuations are independent with the stated variance. They show why repetition helps precision but calibration is needed to address the offset. They do not describe an actual scale test.

Illustrative case study

In a fictional packaging company, operators repeatedly weigh reference objects and find that averages are highly consistent across shifts. Management initially interprets this as proof of accuracy.

A calibration comparison reveals a persistent offset. The quality team separates repeatability from calibration bias, corrects the measurement procedure, and documents uncertainty in the reference standard as well. It then chooses a reasonable number of repeated readings based on operational needs. The team avoids collecting thousands of nearly identical readings while leaving the systematic problem unresolved.

Common misconceptions and mistakes

Random error is not the same as a data-entry mistake, although mistakes can also vary unpredictably. It is not necessarily normally distributed, and an assumption of zero mean should be justified rather than treated as a fact about every error source.

Do not claim that all errors cancel in a finite sample. Cancellation is a statement about averages and their distribution under assumptions, not a guarantee for a particular dataset. More data do not automatically reduce dependence, coverage bias, nonresponse bias, or measurement drift. A narrow interval can describe a biased estimate very precisely. Finally, a model residual is not automatically a pure random error; it may contain omitted structure or measurement problems.

Business use

Understanding random error helps businesses choose sample sizes, measurement schedules, and decision thresholds. It distinguishes improvements achievable through repetition from those requiring redesigned instruments or processes. Decisions should consider the cost of uncertainty and the consequences of error, rather than pursuing precision without a clear purpose. Record assumptions about independence, calibration, stability, and the population being represented.

Check your understanding

Independent readings have random-error standard deviation 8 units. What is the random-error standard deviation of their average when n = 16?

Answer: It is 8/sqrt(16) = 2 units. This calculation does not include an uncorrected systematic offset or uncertainty from other sources.

Related terms

Learn next

Learn the distinction between standard deviation and standard error, then study how dependence changes uncertainty calculations. Compare measurement calibration with increasing sample size.

Sources

Frank Harrell, Glossary of Statistical Terms, random error: https://hbiostat.org/glossary/

The author-maintained glossary distinguishes sampling-related error from random variation in individual measurements. The measurement model, calculations, and fictional case are original teaching examples.

Previous: Random Effects Model Next: Random experiment