Skip to content

Chapter 15: Variance Tests

Overview

Variance tests evaluate whether observed differences in variability across groups or populations are statistically significant. These tests are fundamental tools in statistical inference, playing a critical role in validating assumptions for other methods such as ANOVA and regression analysis. This chapter covers a comprehensive suite of variance testing methods, ranging from classical parametric approaches (chi-square, F-test, Bartlett's) through robust alternatives (Levene's, Brown-Forsythe, Fligner-Killeen) to advanced computational and Bayesian techniques.

Chapter Structure

15.1 Introduction to Variance Testing

Motivation and overview of the variance testing landscape:

  • Why Test Variances -- Explains the practical importance of variance testing for checking homoscedasticity in ANOVA, validating regression assumptions, and comparing volatility across financial instruments.
  • Overview of Variance Tests -- Surveys the major variance tests (chi-square, F-test, Bartlett's, Levene's, Brown-Forsythe, Fligner-Killeen), including their test statistics, distributional assumptions, and appropriate use cases.
  • Assumptions Common to Variance Tests -- Discusses the shared requirements of independence, random sampling, and (for parametric tests) normality, and the consequences of violating these assumptions.

15.2 Chi-Square Test for Variance

A one-sample test for whether a population variance equals a specified value:

  • One-Sample Chi-Square Variance Test -- Tests \(H_0: \sigma^2 = \sigma_0^2\) using the statistic \(\chi^2 = (n-1)s^2 / \sigma_0^2\), with one-tailed and two-tailed formulations and worked examples.
  • Derivation and Distribution Theory -- Shows how the test statistic arises from the distribution of the sample variance under normality, connecting to the chi-square distribution with \(n-1\) degrees of freedom.
  • Confidence Interval for \(\sigma^2\) -- Inverts the chi-square test to construct confidence intervals for the population variance and standard deviation.

15.3 F-Test for Comparing Two Variances

A two-sample test based on the ratio of sample variances:

  • Two-Sample F-Test -- Tests \(H_0: \sigma_1^2 = \sigma_2^2\) using the statistic \(F = s_1^2 / s_2^2\), which follows an \(F\)-distribution under the null hypothesis when both populations are normal.
  • F-Distribution and Degrees of Freedom -- Details the properties of the \(F\)-distribution and how the numerator and denominator degrees of freedom are determined.
  • Sensitivity to Non-Normality -- Warns that the F-test is highly sensitive to departures from normality, often producing inflated Type I error rates with non-normal data.

15.4 Bartlett's Test

A multi-group parametric test for homogeneity of variances:

  • Bartlett's Test for Equality of Variances -- Tests \(H_0: \sigma_1^2 = \sigma_2^2 = \cdots = \sigma_k^2\) using a pooled-variance-based statistic that follows a \(\chi^2_{k-1}\) distribution under normality.
  • Derivation and Chi-Square Approximation -- Derives the Bartlett test statistic from the ratio of pooled to individual variances with the correction factor for small samples.
  • Limitations Under Non-Normality -- Emphasizes that Bartlett's test is highly sensitive to non-normality, making it unreliable when the normality assumption is questionable.

15.5 Robust Tests

Distribution-free and outlier-resistant alternatives for comparing variances:

  • Levene's Test -- Tests equality of variances by performing a one-way ANOVA on the absolute deviations of observations from their group means; robust to moderate departures from normality.
  • Brown-Forsythe Test -- A variant of Levene's test that uses deviations from group medians instead of means, providing additional robustness to skewed distributions and outliers.
  • Fligner-Killeen Test -- A non-parametric rank-based test that uses ranks of absolute deviations from group medians, offering the strongest robustness among the three methods.
  • Comparison of Robust Methods -- Side-by-side evaluation of Levene's, Brown-Forsythe, and Fligner-Killeen in terms of Type I error control, power, and robustness across different distribution shapes.

15.6 Advanced Methods

Computational and Bayesian approaches for variance comparison:

  • Bootstrap Variance Testing -- Uses resampling to construct a null distribution for variance ratios or differences, providing valid inference without distributional assumptions.
  • Bayesian Variance Testing -- Employs prior distributions on variance parameters (typically inverse-gamma) to compute posterior probabilities and Bayes factors for variance hypotheses.
  • Likelihood Ratio Test for Variances -- Compares the maximized likelihoods under the null and alternative hypotheses, providing an asymptotically chi-squared test statistic.

15.7 Applications

Practical use cases where variance testing is essential:

  • Pre-Test for ANOVA Homoscedasticity -- Uses Levene's or Bartlett's test to verify the equal-variance assumption before conducting ANOVA, with Welch's ANOVA as a fallback when variances differ.
  • Variance Testing in Regression -- Applies variance tests (e.g., Breusch-Pagan) to regression residuals to detect heteroscedasticity, with remedies including weighted least squares and robust standard errors.
  • Financial Volatility Comparisons -- Compares the volatility of different assets, portfolios, or time periods to assess risk differences and inform portfolio construction decisions.

15.8 Code

Complete Python implementations:

  • Chi-Squared Test for Variance -- One-sample variance test with critical values and \(p\)-value computation.
  • F-Test of Equality of Variances -- Two-sample F-test implementation with visualization of the rejection region.
  • Bartlett's Test -- Multi-group homogeneity test using scipy and manual computation.
  • Levene's Test -- Robust variance equality test with mean-based deviations.
  • Chi-Square Distribution -- Visualization of the chi-square distribution for different degrees of freedom.
  • Robust Variance Tests Comparison -- Side-by-side comparison of Levene, Brown-Forsythe, and Fligner-Killeen on the same data.
  • F-Test Tail Region Visualization -- Plots the F-distribution with shaded rejection regions.
  • F-Test Normality Sensitivity and Robust Alternatives -- Simulation showing how the F-test's Type I error inflates under non-normality, with robust alternatives performing correctly.
  • F-Test Power Simulation -- Monte Carlo study of F-test power as a function of variance ratio and sample size.
  • Bartlett Test Non-Normality Sensitivity -- Simulation demonstrating Bartlett's poor performance with skewed or heavy-tailed data.
  • Levene Test Normal vs Skewed Simulation -- Compares Levene's test performance across normal and non-normal distributions.
  • Brown-Forsythe Test (scipy) -- Implementation using scipy's median-based Levene variant.
  • Fligner-Killeen Test (scipy) -- Non-parametric rank-based variance test.
  • Bootstrap Variance Test -- Resampling-based approach for comparing variances without distributional assumptions.
  • Bayesian Variance Test -- Posterior inference on variance parameters using conjugate priors.

15.9 Exercises

Practice problems covering F-test computation and interpretation, Levene's test application, comparison of classical and robust test results, chi-square confidence intervals for variance, and practical decision-making when test results conflict.

Prerequisites

This chapter builds on:

  • Chapter 5 (Sampling Distributions) -- The chi-square, \(F\), and normal distributions and their roles in inference about variances.
  • Chapter 8 (Confidence Intervals) -- Confidence interval construction, particularly intervals for \(\sigma^2\) using the chi-square distribution.
  • Chapter 9 (Hypothesis Testing) -- The general framework of null and alternative hypotheses, test statistics, \(p\)-values, and decision rules.
  • Chapter 14 (Normality Tests) -- Methods for checking whether the normality assumption required by classical variance tests is satisfied.

Key Takeaways

  1. Classical variance tests (chi-square, F-test, Bartlett's) are powerful under normality but highly sensitive to non-normal data, often producing misleading results when the normality assumption is violated.
  2. Robust alternatives (Levene's, Brown-Forsythe, Fligner-Killeen) maintain proper Type I error rates across a wider range of distributions and should be preferred when normality is uncertain.
  3. The Brown-Forsythe test (median-based Levene) offers the best balance of robustness and power for most practical applications, while Fligner-Killeen provides the strongest non-parametric guarantees.
  4. Bootstrap and Bayesian methods provide modern alternatives that avoid distributional assumptions entirely or incorporate prior information, respectively.
  5. Variance testing is not just an end in itself -- it serves as a prerequisite check for ANOVA (homoscedasticity), regression (constant error variance), and financial analysis (volatility comparison), making it an essential step in many analysis workflows.