Bonett test (1 variance)

Tests the variance of one sample against a target value – also for non-normal data.

Statusvalidated
Version1.0.0
Minitab equivalentStat > Basic Statistics > 1 Variance (Bonett)

Description

The Bonett test is the robust alternative to the χ² variance test. It estimates the kurtosis of the data from a trimmed mean and uses it to correct the standard error of the log variance. The confidence interval for σ² (and σ) therefore stays approximately valid for skewed or heavy-tailed distributions, where the χ² test is badly distorted. The test has no test statistic; the p-value is obtained by inverting the confidence interval. Qprovement uses it when the Anderson-Darling test rejects normality – as Minitab does.

Formulas

c
Small-sample correction factor (z = normal quantile z_{α/2}, one-sided z_α)
γ̂
Estimated kurtosis with trimmed mean m
trim
Trim proportion per tail for m (for n ≤ 5: arithmetic mean)
se
Standard error of ln(c·s²)
KI
Confidence interval for σ²; the square roots of the limits give the CI for σ
p
p-value by CI inversion: α_L (α_U) is the smallest α at which the lower (upper) limit reaches σ₀²; one-sided α_L or α_U

Assumptions

  • Continuous data
  • Any distribution – normality is not required
  • The observations are independent

Limitations

  • Approximate procedure – for small n the coverage is only approximate
  • No confidence interval exists for n ≤ z
  • No test statistic; the p-value follows from interval inversion

References

  • Bonett, D. G. (2006), Approximate confidence interval for standard deviation of nonnormal distributions, Computational Statistics & Data Analysis, 50, 775–782
  • Banga, S. J. & Fox, G. D. (2013), Bonett's Method, Minitab white paper
  • Minitab Methods and Formulas, 1 Variance