Analysis of Variance (ANOVA) for Actuarial Applications
Apply ANOVA methods to compare group means in actuarial data analysis for Exam MAS-I.
One-Way ANOVA Fundamentals
Analysis of variance (ANOVA) tests whether the means of multiple groups are equal. One-way ANOVA considers a single factor with k levels (groups). The null hypothesis is H0: mu_1 = mu_2 = ... = mu_k, meaning all group means are equal. The alternative is that at least one group mean differs. ANOVA assumes that observations within each group are independent and normally distributed with equal variances across groups (homoscedasticity).
The key insight of ANOVA is that total variability in the data can be decomposed into two sources: variation between groups (due to differences in group means) and variation within groups (due to random error). If the between-group variation is large relative to the within-group variation, this provides evidence that the group means differ.
The ANOVA Table
The ANOVA table organizes the decomposition of variability. The total sum of squares SST = sum over all observations of (Y_{ij} - Y-bar)^2 decomposes into the between-group sum of squares SSB = sum over groups of n_i * (Y-bar_i - Y-bar)^2 and the within-group sum of squares SSW = sum over all observations of (Y_{ij} - Y-bar_i)^2, so that SST = SSB + SSW.
The mean squares are MSB = SSB/(k-1) and MSW = SSW/(N-k), where k is the number of groups and N is the total sample size. The F-statistic is F = MSB/MSW. Under H0, this statistic follows an F-distribution with k-1 and N-k degrees of freedom. A large F (with a small p-value) leads to rejection of H0. The degrees of freedom for SST are N-1, confirming that (k-1) + (N-k) = N-1.
Multiple Comparisons
Rejecting H0 in ANOVA tells you that at least one group mean differs from the others, but it does not identify which means differ. Multiple comparison procedures address this follow-up question while controlling the overall Type I error rate. Without correction, conducting many pairwise t-tests inflates the probability of at least one false positive.
The Bonferroni method divides the significance level alpha by the number of comparisons C(k,2) = k(k-1)/2. Each pairwise comparison uses significance level alpha / C(k,2). This method is conservative (actual Type I error rate is below alpha) but simple to apply. Tukey's Honestly Significant Difference (HSD) method uses the Studentized range distribution and is less conservative than Bonferroni while still controlling the family-wise error rate. Scheffe's method is the most conservative but allows testing any contrast (not just pairwise differences).
Two-Way ANOVA
Two-way ANOVA examines two factors simultaneously. Factor A has a levels and factor B has b levels, creating a*b cells. The model is Y_{ijk} = mu + alpha_i + beta_j + (alpha*beta)_{ij} + epsilon_{ijk}, where alpha_i is the main effect of factor A at level i, beta_j is the main effect of factor B at level j, and (alpha*beta)_{ij} is the interaction effect. The total sum of squares decomposes as SST = SSA + SSB + SSAB + SSE, where SSAB captures the interaction.
The interaction term is particularly important. An interaction means that the effect of one factor depends on the level of the other factor. For example, the effect of territory on claim costs might depend on the coverage type. If the interaction is significant, the main effects must be interpreted cautiously because the effect of each factor is not constant across levels of the other factor. Always test the interaction first; if it is not significant, proceed to interpret main effects.
Actuarial Applications
ANOVA has numerous actuarial applications. In ratemaking, ANOVA can test whether average claim costs differ significantly across territories, vehicle types, or driver age groups. In mortality studies, ANOVA can compare mortality rates across different rating classes to determine whether separate classification is justified. In claims management, ANOVA can evaluate whether different settlement strategies produce different average costs.
For example, an actuary might use one-way ANOVA to test whether average bodily injury claim costs differ across four geographic territories. With territory as the factor and claim costs as the response, a significant F-test would support maintaining separate territorial rates. A non-significant result would suggest that the territories could be combined without losing pricing accuracy.
Assumptions and Diagnostics
ANOVA assumes normality of errors, equal variances across groups (homoscedasticity), and independence of observations. Normality can be assessed using QQ plots of residuals. Levene's test or Bartlett's test checks the equal variance assumption. ANOVA is robust to moderate departures from normality, especially with balanced designs and moderate to large sample sizes.
When the equal variance assumption is violated, Welch's ANOVA provides a robust alternative that does not assume equal variances. When normality is seriously in question, the Kruskal-Wallis test offers a nonparametric alternative to one-way ANOVA, testing whether the group distributions (not just means) differ.
Common Pitfalls
A frequent mistake is performing multiple pairwise t-tests instead of ANOVA followed by a proper multiple comparison procedure. With k = 5 groups, there are 10 pairwise comparisons, and performing each at the 5% level gives a family-wise error rate of about 40%, far above the intended 5%. Another common error is interpreting a non-significant interaction as evidence that the interaction does not exist. Non-significance means there is insufficient evidence to detect an interaction, not that one is absent.
Students also sometimes forget that ANOVA requires equal variances across groups. If the group variances are very different (say, one group has variance 10 times larger than another), the standard F-test is unreliable and Welch's modification should be used.
Study Tips
Exam MAS-I tests ANOVA table construction, F-test execution, multiple comparison procedures, and interpretation of main effects and interactions. Practice constructing ANOVA tables from raw data, computing SSB, SSW, mean squares, and the F-statistic. Know how to interpret two-way ANOVA results, especially the interaction term. Be familiar with the assumptions and know which diagnostic tools check each assumption. Work through problems that require deciding between one-way and two-way ANOVA based on the study design.