Bootstrap One Way Analysis of Variance (ANOVA) using SPSS

Bootstrap one-way ANOVA in SPSS is presented as a practical workaround when group data violate normality, letting researchers still test whether perceptions differ across categories like job ranks. The core workflow starts with a standard one-way ANOVA setup—Analyze → Compare Means → One-Way ANOVA—using a dependent variable such as “collaborative culture” and a grouping factor such as job rank (junior, middle, senior). Before choosing the main ANOVA test, the analysis checks homogeneity of variance via the homogeneity of variance test. If that test is insignificant, equal variances are treated as reasonable, and the analysis proceeds with the usual ANOVA main-effects table (and, if needed, post-hoc comparisons).

In the example, homogeneity of variance is insignificant, so equal variance is assumed. The main-effects ANOVA result is reported as significant but only slightly above the 0.05 threshold, leading to the conclusion that collaborative culture perceptions may differ across at least some job ranks. The effect size is described as very small, signaling that even when differences exist, they are not large.

The session then switches to bootstrapping to make the inference more robust when the data are not normal. In SPSS’s One-Way ANOVA dialog, the user selects the bootstrap option (bootstrap perform bootstrapping) and specifies a resampling count—typically 5,000 to 10,000—using 10,500 samples in this walkthrough. A bias-corrected confidence interval accelerated (BCa) approach is chosen to stabilize interval estimates. For multiple comparisons, the method is aligned with the variance assumption: when equal variance is assumed, Tukey is used; when equal variance is not assumed, Games-Howell is used.

After bootstrapping, the homogeneity-of-variance check is repeated and again remains insignificant, so the main-effects ANOVA is used with the bootstrap framework. The multiple-comparisons results are where bootstrapping changes the decision logic: the confidence intervals for pairwise mean differences are examined to see whether they cross zero. In the example, the bootstrapped multiple comparisons indicate that junior and senior differ (pair 1 vs 3), while other pairwise contrasts do not show clear separation based on whether zero lies within the confidence interval. The mean differences remain similar, but the bootstrap-adjusted confidence interval bounds determine which contrasts are treated as statistically significant.

Finally, the transcript addresses the alternative scenario: if homogeneity of variance is significant (variance inequality), equal-variance ANOVA is no longer appropriate. In that case, the analysis uses Welch’s test for the main effect and switches post-hoc comparisons to Games-Howell under the “equal variance not assumed” condition. Bootstrapped multiple comparisons then identify which group pairs differ under the more conservative variance-robust approach. Overall, the method provides a clear decision tree in SPSS: bootstrap for non-normality, then choose ANOVA vs Welch and Tukey vs Games-Howell based on homogeneity of variance results, using bootstrapped confidence intervals to judge pairwise differences.