Welch ANOVA | Analysis of Variance | Biostatistics Study Notes

Welch ANOVA compares the population means of multiple independent groups, but does not require that the variances of each group be equal. It is better suited than traditional one-way ANOVA when the standard deviations are different and the sample sizes are also unbalanced.

When to consider Welch ANOVA#

Compare population means of two or more independent groups
The dependent variable is quantitative and the average is meaningful.
Group variances are different or there is insufficient evidence for equal variances
Unbalanced sample size, especially small sample size for high variance groups

Welch's method is not a universal replacement for non-normality. Independence is still necessary, and for strong distortions, extreme outliers, and small samples, we also consider graphical diagnostics, transformations, robust methods, and nonparametric methods that fit the research question.

Hypotheses and Assumptions#

null hypothesis

H_0:\mu_1=\mu_2=\cdots=\mu_k

alternative hypothesis

H_1:\text{Not all }\mu_j\text{ are equal}

conditions	Welch ANOVA requirements
independence	Each observation is independent and belongs to only one group
dependent variable	Quantitative variables whose averages have meaning
normality	The errors in each group are approximately normal. Relatively robust in large samples
homoscedasticity	Not required

Give More Weight to More Precise Groups#

The estimated variance of the j-th group mean is approximately sⱼ²/nⱼ. Groups with larger sample sizes or smaller sample variances have more accurate mean estimates, so they are given larger weights in proportion to their inverse variances.

weight of the jth group

w_j=\frac{n_j}{s_j^2}

total weight

W=\sum_{j=1}^{k}w_j

weighted grand average

\bar X_w=\frac{\sum_{j=1}^{k}w_j\bar X_j}{W}

Welch F statistic#

The difference between each group mean and the weighted grand mean is weighted, and the denominator and degrees of freedom are adjusted by a correction amount A that reflects the uncertainty in the variance estimation.

Correction amount

A=\sum_{j=1}^{k}\frac{(1-w_j/W)^2}{n_j-1}

Welch F

F_W=\frac{\dfrac1{k-1}\sum_{j=1}^{k}w_j(\bar X_j-\bar X_w)^2}{1+\dfrac{2(k-2)}{k^2-1}A}

Molecular degrees of freedom

\nu_1=k-1

Approximate denominator degrees of freedom

\nu_2=\frac{k^2-1}{3A}

Find the p-value from the right tail of the F distribution. Even if it is significant, the conclusion is that at least one group's means are different, but the location of the difference is not known.

Comparison with Classical ANOVA#

item	Traditional one-way ANOVA	Welch ANOVA
Variance assumption	equal in all groups	It's okay to be different
within-group variation	Merger into common MSE	Weighting by group with nⱼ/sⱼ²
degree of freedom	k−1、n−k	k−1, Welch–Satterthwaite approximation
Main precautions	Can be distorted due to unequal variance + imbalance	Still be wary of small samples, strong distortions, and outliers.

After a Significant Overall Test#

We do not use LSD, pooled t, or regular Tukey HSD, which assumes a common MSE. For all pairwise comparisons, the Games–Howell method, which allows for unequal variances and unbalanced samples and also adjusts for multiplicity, is often used.

Report n, mean, standard deviation, and distribution for each group
Welch F, numerator/approximate denominator degrees of freedom, p-value reported
If significant, perform Games–Howell or planned unequal variance contrasts
Includes mean difference, confidence interval, and adjusted p-value