Description
Estimate and compare the variability of outcomes across conditions, treating consistency of performance as a result in its own right rather than only an assumption check.
Quality engineering holds that a product must be consistently near target, so variability deserves inference of its own. For one group, a chi-square interval or test bounds the standard deviation. For two groups, the ratio of sample variances is referred to the F distribution. For k groups, Hartley's F-max divides the largest variance by the smallest and compares it with tabled values; Bartlett's test compares a log-variance statistic with chi-square on k - 1 df and handles unequal group sizes. Both are highly sensitive to non-normality, so normality must be checked first, and Levene or Brown-Forsythe tests (based on absolute deviations from group means or medians) are robust alternatives. When variability is itself a response across design factors, analyse ln(s) from at least four repeats per cell. Mason and colleagues advise against using a preliminary variance test to choose between pooled and Welch t-tests; use Welch directly.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Siloed
- Stage
- Evaluation
- Data
- Quantitative
- Effort
- Hours
Steps
- Compute the sample standard deviation and size of each group.
- For one group, form the chi-square interval or test for the standard deviation against a target.
- For two groups, compare the variance ratio with the F distribution.
- For k groups, compute F-max (largest over smallest variance) against tabled critical values with v = n - 1, or Bartlett's statistic against chi-square on k - 1 df.
- Check normality first; with skewed data such as task times, use Levene or Brown-Forsythe, or analyse log-transformed values.
- If variability is a design response, analyse ln(s) with at least four repeats per cell.
Sources
- Statistical Design and Analysis of Experiments (Mason et al., 2003)
- Multivariate Statistical Inference and Applications (Rencher, 1998)
Also known as
- Hartley's F-max test
- Bartlett's test
- Homogeneity of variance tests
- Variance ratio test
- Chi-square interval for a standard deviation
- Box's M test
Categories
- Analytics & Behaviour
- Statistics & Modelling