Checking the equal variance assumption (also called homoscedasticity) involves a combination of visual plots and formal statistical tests. The right approach depends on whether you’re comparing groups in an ANOVA or t-test, or checking residuals in a regression model. Getting this wrong has real consequences: in one simulation where three groups had standard deviations of 1.0, 2.0, and 3.0, the false positive rate jumped from the expected 5% to 18%, even though the group means were identical.
Why Equal Variance Matters
Parametric tests like t-tests, ANOVA, and linear regression all assume that the variability within each group or across fitted values is roughly the same. When this assumption holds, the data are called homoscedastic. When it doesn’t, the data are heteroscedastic, and your p-values become unreliable.
The damage is especially bad when sample sizes are unequal. If your smaller groups happen to come from the populations with larger variability, false positive rates inflate well beyond your chosen significance level. Conversely, if the smaller groups have less variability, the test becomes overly conservative and you lose statistical power. With balanced sample sizes the problem is less severe, but it doesn’t disappear.
Start With Visual Checks
Before running any formal test, plot your data. Visual inspection is fast, intuitive, and often reveals problems that a single test statistic can miss.
Residuals vs. Fitted Values Plot
For regression models, plot the residuals (the differences between predicted and actual values) on the y-axis against the fitted values on the x-axis. If variance is equal, the points will scatter in a roughly even horizontal band around zero. A fan or funnel shape, where the spread of residuals widens or narrows as fitted values increase, signals heteroscedasticity. This plot also helps you spot nonlinearity and outliers at the same time.
Side-by-Side Boxplots
For group comparisons (t-tests, ANOVA), boxplots let you eyeball the spread of each group simultaneously. Compare the height of the boxes, which represent the interquartile range (IQR), the middle 50% of your data. If one group’s box is noticeably taller than another’s, the groups likely differ in variability. For instance, comparing Oscar-winning ages for actors and actresses, a study found IQRs of 11 and 9.5 respectively, showing more consistency in the actresses’ ages. That kind of quick comparison is exactly what boxplots are designed for.
Formal Tests for Group Comparisons
When you’re comparing two or more groups in a t-test or ANOVA, three hypothesis tests are commonly used. Each has a null hypothesis that the group variances are equal, so a small p-value (typically below 0.05) means you should reject that assumption.
Levene’s Test
Levene’s test is the most widely used option because it doesn’t require your data to be normally distributed. It works by calculating how far each observation falls from its group mean, then running an ANOVA on those distances. This makes it a safer default when you’re unsure about the shape of your data.
Brown-Forsythe Test
The Brown-Forsythe test follows the same logic as Levene’s test but uses the group median instead of the group mean. This swap makes it more robust when your data are skewed, because medians are less influenced by extreme values. If your distributions are clearly asymmetric, this is the better choice of the two.
Bartlett’s Test
Bartlett’s test is more powerful than Levene’s, meaning it’s better at detecting real differences in variance, but only when your data are truly normal. It’s quite sensitive to departures from normality, so a significant result could reflect non-normality rather than unequal variance. Use Bartlett’s only when you have strong evidence that your data are approximately normal.
Formal Tests for Regression Models
In regression, you’re not comparing predefined groups. Instead, you’re checking whether the variance of your residuals stays constant across the range of predictor values. Two tests dominate here.
Breusch-Pagan Test
The Breusch-Pagan test checks whether the variance of residuals is related to one or more of your predictor variables. It assumes a specific form of heteroscedasticity, namely that variance changes as a linear function of the predictors. This makes it more focused and more powerful when that assumption holds, but it can miss more complex patterns.
White’s Test
White’s test makes no assumptions about the form of heteroscedasticity, which makes it more general. The trade-off is that because it’s so broad, it can flag problems that aren’t actually variance-related. A significant White’s test sometimes reflects model misspecification, like a missing interaction term, rather than unequal variance. It’s best used as a catch-all when you’re not sure what pattern the heteroscedasticity might take.
The Variance Ratio Rule of Thumb
Not every difference in variance is large enough to matter. A practical shortcut is to compute the ratio of the largest group variance to the smallest. Textbooks cite various thresholds: some say a ratio under 3 is safe, others stretch it to 4 or 5, and one common guideline allows ratios up to 10 as long as the largest sample is no more than four times the size of the smallest.
More recent research suggests these traditional thresholds may be too generous. A study examining the effect of variance ratios on ANOVA robustness found that ratios above 1.5 can already threaten the accuracy of the F-test when sample sizes are unequal. The takeaway: if your groups are balanced, moderate variance differences are tolerable. If they’re unbalanced, even small differences deserve attention.
What To Do When Variance Is Unequal
Finding heteroscedasticity doesn’t mean your analysis is doomed. You have several options, and the right one depends on your situation.
Switch to a Robust Test
For group comparisons, Welch’s t-test (for two groups) or Welch’s ANOVA (for three or more) adjusts the degrees of freedom to account for unequal variances. Welch’s method controls the false positive rate well regardless of whether sample sizes are balanced or not, and it maintains good statistical power. Research from Virginia Commonwealth University found that the Welch method outperformed both standard ANOVA and the nonparametric Kruskal-Wallis test when data were normally distributed but had unequal variances. Many statisticians now recommend using Welch’s versions by default, even before checking assumptions.
For regression, you can use heteroscedasticity-consistent standard errors (often called robust standard errors). These adjust your confidence intervals and p-values without changing the model itself.
Transform the Data
Sometimes a mathematical transformation can stabilize variance across groups or across the range of fitted values. Three transformations work well in practice. The square root transformation compresses larger values moderately and works when variance increases proportionally with the mean. The natural log transformation is more aggressive and is a good candidate when variance increases with the square of the mean, which is common in biological and financial data. The reciprocal (1/x) transformation is the strongest of the three.
If you’re not sure which transformation to use, a Box-Cox analysis can help. It searches for the optimal power transformation by finding the value that produces the most linear fit or the most consistent variance. A Box-Cox value near 0 points to a log transformation, near 0.5 suggests a square root, and near 1 means no transformation is needed.
Running These Checks in R and Python
In R, Levene’s test is available through the car package with leveneTest(), Bartlett’s test is built in with bartlett.test(), and the Breusch-Pagan test comes from the lmtest package with bptest(). Residual plots are generated with plot(model) on any fitted linear model, which automatically produces a residuals vs. fitted values plot as one of its diagnostics.
In Python, the scipy.stats module provides levene() and bartlett() for group comparisons. For regression diagnostics, the statsmodels library handles the Breusch-Pagan test through het_breuschpagan() in the statsmodels.stats.diagnostic module. You pass it the model residuals and the predictor matrix, and it returns both a Lagrange multiplier statistic and an F-statistic, each with a corresponding p-value. Residual plots can be made with matplotlib or seaborn by plotting model.resid against model.fittedvalues.
A Practical Checking Workflow
Combine visual and formal methods rather than relying on either one alone. Start with a plot: residuals vs. fitted values for regression, or side-by-side boxplots for group comparisons. Look for obvious patterns like funneling, fan shapes, or boxes of very different sizes. Then run the appropriate formal test: Levene’s or Brown-Forsythe for groups, Breusch-Pagan or White’s for regression. If both the visual and the formal test agree, you can be confident in your conclusion.
If the assumption fails, check the variance ratio and your sample sizes. With balanced groups and a ratio under 1.5, standard methods are likely fine. With unbalanced groups or larger ratios, switch to Welch’s test, apply a variance-stabilizing transformation, or use robust standard errors. These alternatives are straightforward to implement in any statistical software and carry little downside even when the assumption holds.