LSSBB Technical Reference • Updated November 2026
The Direct Answer: Selecting the correct statistical test in a Six Sigma Black Belt project depends on four things: the question you are asking (means, medians, variances, proportions, or relationships), Data Type (continuous vs. discrete), Sample Size, and Normality. Parametric tests like ANOVA require normal continuous data, whereas non-normal or ordinal data requires non-parametric equivalents like Kruskal-Wallis or Mann-Whitney.
Use this quick-reference matrix during the Analyze phase of DMAIC to select the right hypothesis test or modeling tool for your process inputs (X) and outputs (Y).
| Objective / Question | Data Type (Y / X) | Assumptions | Recommended Test / Tool |
|---|---|---|---|
| Compare 1 process mean to a target | Continuous | Normal distribution (non-normal: 1-Sample Wilcoxon) | 1-Sample t-Test |
| Compare 2 process groups (means) | Continuous / Discrete (2 levels) | Normal distribution (use Welch's t if variances are unequal) | 2-Sample t-Test |
| Compare before vs. after on the same units | Continuous (paired measurements) | Differences are normal (non-normal: Wilcoxon signed-rank) | Paired t-Test |
| Compare 2 groups (non-normal) | Ordinal / Non-normal Continuous / Discrete (2 levels) | Non-parametric (rank-based) | Mann-Whitney U Test |
| Compare 3+ process groups (means) | Continuous / Discrete (3+ levels) | Normal distribution, equal variance | One-Way ANOVA |
| Compare 3+ groups (non-normal) | Ordinal / Non-normal Continuous / Discrete (3+ levels) | Non-parametric (rank-based; compares medians when distribution shapes are similar) | Kruskal-Wallis Test |
| Compare variance (spread) across 2 or more groups | Continuous / Discrete (2+ levels) | Normal data for Bartlett's and the F-Test; Levene's holds up when data is non-normal | 2-Variances F-Test or Bartlett's (normal); Levene's Test (non-normal) |
| Compare a defect rate to a target, or between 2 groups | Discrete (proportion defective) | Adequate sample size (n×p and n×(1−p) ≥ 5) | 1-Proportion / 2-Proportion Test |
| Test whether 2 categorical variables are associated (defect counts by category) | Discrete (Y) / Discrete (X) | Expected cell frequency ≥ 5 | Chi-Square Test of Independence |
| Measure strength of linear association | Continuous Y / Continuous X | Linear relationship (non-normal or ordinal: Spearman) | Pearson Correlation |
| Quantify cause-and-effect relationship | Continuous Y / Continuous X | Linear relationship, normal residuals | Simple / Multiple Regression |
A Black Belt should use the Kruskal-Wallis test when comparing three or more independent process groups where the continuous output data fails normality testing (p < 0.05) or consists of ordinal rankings. While One-Way ANOVA compares population means under strict normality assumptions, Kruskal-Wallis ranks all observations and tests whether the groups differ, which is commonly interpreted as a difference in medians when the distribution shapes are similar. This keeps skewed operational data from generating false positive conclusions.
Choose Regression when your independent input variables (X) are continuous rather than discrete categories. While One-Way ANOVA evaluates the impact of distinct groups (e.g., Shift A vs. Shift B vs. Shift C), Regression builds a predictive mathematical equation to determine how changing a continuous process parameter—such as oven temperature or server memory allocation—predictably alters process output.
When process data is non-normal and cannot be normalized using Box-Cox or Johnson transformations, Black Belts must switch from parametric tests to non-parametric equivalents. Use a Mann-Whitney U test (instead of a 2-sample t-test) to compare two groups, a Kruskal-Wallis test (instead of ANOVA) for three or more groups, and Mood's Median test as an outlier-resistant alternative for comparing medians across groups. To compare process spread (variance) on non-normal data, use Levene's test rather than Bartlett's test or the F-test, which are sensitive to non-normality.
The scenarios below are hypothetical and use illustrative figures to show how the tests are applied.
An enterprise healthcare Black Belt analyzing patient door-to-doctor times across four hospital branches finds that wait time distributions are heavily right-skewed and fail Anderson-Darling normality tests. Instead of running a standard One-Way ANOVA, the Black Belt applies the Kruskal-Wallis test, which shows that at least one branch differs significantly. A post-hoc pairwise comparison (such as Dunn's test) then isolates Branch C as having a statistically significant delay compared with the other branches, without violating statistical assumptions.
A systems engineering Black Belt investigating IT ticket resolution times wants to evaluate whether server CPU throttling percentage (X) drives ticket resolution duration (Y). Because both variables are continuous and the model residuals pass normality and constant-variance checks, the Black Belt applies Multiple Linear Regression and builds a predictive model with an R-squared of 84.2%, quantifying how much each additional point of CPU throttling adds to resolution time.
Download the free LSSBB Exam Toolkit—featuring complete statistical formula reference sheets, hypothesis testing decision trees, and CSSC-aligned study guides.
Register and get the Free LSSBB Exam Toolkit →🔒 We'll email your Exam link immediately. No spam. Unsubscribe anytime.