Is noncentrality the same as effect size?

No. In common tests it combines effect size with sample size and design information.

Does zero recover the central distribution?

For the usual noncentral t, chi-squared, and F families, zero noncentrality gives the corresponding central case.

A noncentral t parameter can be negative. Conventional noncentral chi-squared and F parameters are nonnegative.

AI Glossary / Statistics and probability / Advanced

Noncentrality parameter

Definition

A noncentrality parameter is a parameter that distinguishes a noncentral probability distribution from its central counterpart. In hypothesis testing, it often describes how the sampling distribution changes under a specified alternative to the null hypothesis. Its precise definition depends on the distribution and test. It commonly combines an effect with sample size and variability, making it useful for calculating statistical power.

What it means

Under a null hypothesis, a familiar test statistic may have a central t, chi-squared, or F distribution. Under particular alternatives, its distribution belongs to a corresponding noncentral family. The noncentrality parameter helps specify that alternative distribution.

It should not universally be interpreted as a literal difference between distribution means. That interpretation can be a useful intuition, but the exact relationship depends on the family. A noncentral t parameter can be negative, whereas the conventional noncentral chi-squared parameter is nonnegative. Checking the definition prevents errors when moving between formulas or software packages.

How it works

Specify the statistical model, test statistic, and alternative scenario. Derive or obtain the distribution of the statistic under that scenario, including its degrees of freedom and noncentrality parameter. Determine the rejection region using the null distribution and the planned significance level.

Calculate power as the probability that the alternative distribution falls inside that rejection region. A two-sided test includes both relevant tails. A one-sided test uses the direction specified in the study plan.

For independent normal observations in a one-sample t-test, the noncentrality parameter is delta = (mu − mu0) × sqrt(n) / sigma. Here mu is the assumed true mean, mu0 the null mean, sigma the population standard deviation, and n the sample size. This is a planning quantity under stated assumptions, not the observed t statistic obtained by substituting a particular sample mean and sample standard deviation.

Three practical examples

An evaluation team compares candidate sample sizes by calculating power for a prespecified improvement in average task performance.
A simulation study generates noncentral chi-squared variables to examine a procedure under alternatives of different strengths.
A statistician checks whether a software function expects a signed t noncentrality parameter or a nonnegative parameter for an F distribution.

Think of it this way

Think of two maps of where repeated test results could land: one under no effect and one under a particular effect. The noncentrality parameter helps locate or shape the second map. It does not tell you where one actual study result must land, and it is not itself the probability of a successful test.

Worked example

Consider an original planning example with independent normal measurements. The null mean is 100, the proposed true mean is 104, and the population standard deviation is 10. The standardized effect is (104 − 100) / 10 = 0.4.

With n = 25 observations, delta = 0.4 × sqrt(25) = 2 and the t distribution has 24 degrees of freedom. With n = 100, delta = 0.4 × sqrt(100) = 4 and the degrees of freedom become 99. Quadrupling the sample size doubles this noncentrality parameter; it does not double the underlying standardized effect.

If the proposed true mean were 96 instead, delta would be −2 for n = 25. The sign matters for a directional test. As a separate illustration of parameter conventions, let two independent unit-variance normal variables have means 1 and 2. Their squared sum has a noncentral chi-squared distribution with two degrees of freedom and lambda = 1² + 2² = 5. Its mean is 2 + 5 = 7. These calculations illustrate different definitions; they do not equate the t parameter delta with the chi-squared parameter lambda.

Illustrative case study

Fictional case study: Cedar Evaluation plans a controlled benchmark of an AI editing tool. Its analyst specifies a practically meaningful mean improvement and several plausible variability assumptions before collecting the main evaluation data.

The initial presentation mistakenly describes the calculated noncentrality parameter as “the expected t score.” The analyst replaces that description with the appropriate distributional interpretation and calculates rejection probabilities using the full alternative distribution. Cedar compares sample-size scenarios rather than treating one parameter value as a guarantee of statistical significance. The fictional case emphasizes careful communication: effect size, sample size, noncentrality, and power are related but remain different quantities.

Common misconceptions and mistakes

“Noncentrality is the observed test statistic.” It is a distribution parameter; a test statistic is calculated from realized data.

“It always equals a mean shift.” The exact parameterization may involve a standardized effect, squared quantities, or other model-specific expressions.

“A larger value always means greater power.” Direction, test family, rejection region, and other parameters matter. A negative shift can reduce power for an upper-tailed test.

“A planning calculation establishes the true effect.” It describes a hypothetical scenario whose assumptions still require justification.

Business use

Businesses can use noncentral distributions to plan evaluations that have a reasonable chance of detecting changes they consider worthwhile. The practical value lies in connecting the proposed study design with a meaningful effect and realistic uncertainty assumptions.

Document the assumed effect, variability, sample size, test direction, and significance level. Examine several plausible scenarios rather than reporting one precise power figure without context. In AI benchmarking, check dependence between repeated tasks or observations: treating correlated measurements as independent can make the implied information and planned power overly optimistic.

Check your understanding

In the one-sample t setting, the standardized effect stays fixed while sample size increases from 36 to 144. What happens to delta?

Answer: It doubles, because the square root of sample size increases from 6 to 12. The degrees of freedom also change, so power requires the complete distribution calculation.

Related terms

Effect size
Alternative hypothesis
Degrees of freedom
Non-central F distribution
Noncentral distribution
statistical power
sample-size planning

Learn next

Study central and noncentral t distributions, then compute power from a specified rejection region. Compare signed t parameterizations with nonnegative chi-squared parameters. Finally, examine how paired, clustered, and unequal-group designs alter the relevant information and formulas.

Rejection region

Sources

David Howell, glossary, introductory noncentrality terminology: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html

R statistical documentation, noncentral t construction and one-sample parameter: https://stat.ethz.ch/R-manual/R-devel/library/stats/html/TDist.html

R statistical documentation, noncentral chi-squared construction: https://stat.ethz.ch/R-manual/R-devel/library/stats/html/Chisquare.html

Exact family-specific definitions refine the glossary's mean-departure intuition. Examples and case are original.

Previous: Noncentered parameterization Next: Nondirectional test