a. Including Tests of Equal Variance, Normality Testing and Sample Size calculation, performing tests and interpreting results.

Including Tests of Equal Variance, Normality Testing and Sample Size calculation, performing tests and interpreting results.. Introduction This article explains how to: - Assess normality - Test equality of variance - Calculate sample size - Perform tests and interpret results All content is tightly focused on the statistical foundations required to correctly select, run, and interpret hypothesis tests related to normality, variance, and sample size for continuous and discrete data in process improvement work. --- Foundations: Data Types and Distributions Understanding Data Types Correct tests depend on correctly identifying data types. - Continuous data - Measured on a scale with meaningful intervals - Examples: time, weight, length, temperature - Commonly modeled with normal distribution (or transformed to normality) - Discrete data - Count data or categorical data - Examples: defects, number of calls, pass/fail, yes/no - Often modeled with binomial or Poisson distributions - Different sample size formulas and tests than continuous data Misclassifying data leads to incorrect test selection and wrong conclusions. Why Distribution Matters Many classical tests (t-test, ANOVA, regression with normal errors) assume: - The continuous response is approximately normally distributed - Variances are roughly equal across groups - Observations are independent Before using those tests, the data (or residuals) must be checked for: - Approximate normality - Reasonable equality (homogeneity) of variance If assumptions are severely violated, you may: - Transform data (for example, log, square root) - Use nonparametric tests (for example, Mann–Whitney, Kruskal–Wallis) - Use variance-stabilizing approaches --- Normality Testing Role of Normality in Statistical Tests Normality is most critical when: - Sample sizes are small (for example, n < 30 per group) - Using parametric tests on raw data or residuals For larger samples, parametric tests often remain robust to modest departures from normality. Severe skewness, heavy tails, or outliers can still invalidate conclusions, especially for capability analysis and confidence intervals. Visual Tools for Normality Assessment Visual checks are essential before formal tests. - Histogram - Shape should be unimodal and roughly symmetric - Look for skewness, multimodality, and outliers - Boxplot - Helps identify outliers and asymmetry - Long whiskers on one side indicate skewness - Normal probability plot (Q–Q plot) - Plots ordered data versus expected normal quantiles - If points follow a roughly straight line, normality is plausible - Systematic curvature suggests deviation: - S-shaped: heavy tails - Convex/concave: skewness Visual tools provide context and help interpret formal test results. Formal Normality Tests Common normality tests used in practice include: - Anderson–Darling test - Sensitive to deviations in tails - Frequently used in statistical software - Null hypothesis H₀: data follow a normal distribution - Alternative H₁: data do not follow a normal distribution - Shapiro–Wilk test - Powerful for small samples - Same hypotheses as Anderson–Darling - Kolmogorov–Smirnov and Lilliefors tests - Less powerful than Anderson–Darling/Shapiro–Wilk in many cases - Sometimes used historically but less common in modern quality software Most software reports: - Test statistic (e.g., A² for Anderson–Darling) - p-value Interpreting Normality Test Results Interpretation is driven by the p-value: - H₀: Data are normal - H₁: Data are not normal Using a typical significance level α = 0.05: - If p-value ≥ 0.05 - Fail to reject H₀ - Conclusion: no evidence against normality - It is acceptable to proceed with parametric methods (if other assumptions hold) - If p-value < 0.05 - Reject H₀ - Conclusion: evidence that data are not normally distributed However, context matters: - Small sample sizes - Power is low; tests may fail to detect non-normality - Rely more heavily on visual checks and subject-matter knowledge - Large sample sizes - Even trivial departures from normality can produce very small p-values - Focus on: - Magnitude of deviation on Q–Q plot - Impact on analysis (for example, residuals in regression) - Robustness of planned methods What to Do When Normality Fails Options include: - Data transformation - Log transform for right-skewed data (for example, cycle time) - Square root transform for count data with increasing variance - Box–Cox transform (software often suggests optimal λ) - Re-check normality after transformation - Nonparametric tests - For two-sample location comparisons: Mann–Whitney (Wilcoxon rank-sum) - For paired data: Wilcoxon signed-rank test - For more than two independent groups: Kruskal–Wallis test - Use distribution-appropriate methods - For counts: Poisson or negative binomial modeling - For proportions: binomial-based methods Always consider whether the lack of normality materially affects decisions; do not overreact to very small deviations in large samples. --- Tests of Equal Variance Why Equal Variance Matters Many hypothesis tests assume equal variances between groups, such as: - One-way ANOVA - Classical two-sample t-test (pooled variance version) - Certain regression models when comparing subgroups Unequal variances can: - Inflate type I error (incorrectly finding significance) - Reduce power (missing real differences) - Distort confidence intervals Therefore, checking variance equality across groups is essential. Visual Tools for Variance Assessment Before formal tests, use plots: - Side-by-side boxplots - Compare spread and outliers among groups - Larger IQR and whiskers imply larger variance - Residual versus fitted value plots (post-model) - Look for funnel-shaped patterns - Increasing spread with fitted value suggests heteroscedasticity Visual inspection can signal obvious variance differences and guide test selection. Formal Tests for Equal Variance Common tests: - Levene’s test - Robust to non-normality - Hypotheses: - H₀: group variances are equal - H₁: at least one group variance differs - Based on absolute deviations from group medians or means - Brown–Forsythe test - Variation of Levene’s, typically using medians - More robust when distributions are skewed - F-test for equal variances (two groups only) - Compares ratio of two variances - Highly sensitive to non-normality - Use cautiously; better for well-behaved normal data Software often labels Levene-type tests as “Test for equal variances” in ANOVA or two-sample comparison dialogs. Interpreting Equal Variance Tests Given significance level α (commonly 0.05): - H₀: All group variances are equal - H₁: At least one group variance differs - If p-value ≥ α - Fail to reject H₀ - Conclusion: no evidence that variances differ - It is acceptable to use: - Pooled t-test (for two groups) - Standard ANOVA (assuming normality is reasonable) - If p-value < α - Reject H₀ - Conclusion: evidence of unequal variances What to Do When Variances Are Unequal Responses depend on the scenario: - Two-sample mean comparison - Use Welch’s t-test (separate variance t-test) - Does not assume equal variances - Software often labels it as “Do not assume equal variances” - ANOVA with multiple groups - Consider: - Welch ANOVA (unequal variances) - Transformation to stabilize variance (for example, log, square root) - Nonparametric alternative (for example, Kruskal–Wallis) - Regression modeling - Use residual plots to detect heteroscedasticity - Options: - Transform response - Weighted least squares (assign smaller weight to higher-variance observations) - Use generalized least squares or generalized linear models where appropriate In all cases, re-check assumptions after corrective actions. --- Sample Size and Power Calculations Core Concepts Sample size planning ensures that data can answer the question with acceptable risk. Key terms: - Significance level (α) - Probability of type I error: rejecting a true H₀ - Common choice: 0.05 - Power (1 − β) - Probability of correctly rejecting H₀ when the specified alternative is true - Common targets: 0.8 (80%) or 0.9 (90%) - Effect size - The smallest difference or improvement that matters practically - Examples: - Difference in means (for example, 1 second reduction in cycle time) - Ratio of standard deviations - Change in proportion defectives - Standard deviation (σ) - Variability of the data - Critical input to sample size formulas for continuous data - Often estimated from: - Historical data - Pilot study - Subject-matter knowledge Sample size decisions balance: - Risk of wrong decisions (α, β) - Required precision - Available time, cost, and capacity Sample Size for Continuous Data (Means) Common cases in process improvement: - One-sample mean - Compare process mean to a target - Inputs: - α - Desired power - Estimated σ - Practical difference from target (Δ) you want to detect - Two-sample mean (independent groups) - Compare two processes or treatments - Inputs: - α - Desired power - Common or group-specific σ - Practically important difference in means (Δ) - Allocation ratio between groups (often 1:1) - Paired mean (before–after on same units) - Compare measurements from the same units before and after change - Inputs: - α - Desired power - Standard deviation of pair differences - Target difference in mean pair difference Software typically provides a power and sample size dialog for each type of t-test. The underlying principles: - Larger σ or smaller Δ requires larger n - Higher power or lower α requires larger n - Unequal sample allocation between groups often increases total sample size required Sample Size for Proportions (Discrete Data) When working with pass/fail or defectives: - One proportion - Compare defect rate to a specification or historical rate - Inputs: - Baseline proportion (p₀) - Target proportion (p₁) that represents meaningful improvement - α and desired power - Two proportions - Compare defect rates between two processes or time periods - Inputs: - Baseline proportion (p₁) - Improved or alternative proportion (p₂) - α and desired power - Allocation ratio Approximate rules: - Rare events (small p) may require large n to detect changes - Detecting small relative improvements requires larger n than detecting large improvements Sample Size for Variance and Capability Sometimes the focus is on estimating variability or capability: - Estimating σ accurately - Required n increases rapidly as desired relative precision tightens - Useful when planning gauge R&R or pilot studies - Detecting change in variance - Two-sample tests for variance ratios (for example, F-test or Levene-type designs) - Sample size depends on: - α - Power - Ratio of variances to detect (for example, 1.5× reduction) Statistical software can compute sample sizes for variance tests, but the underlying theme remains: specifying the smallest variance change that matters practically. Interpreting Power and Sample Size Outputs Software usually provides: - Required sample size for given: - α, power, effect size, and variability - Power achieved for given: - n, α, effect size, and variability Interpretation guidelines: - If required n is too large: - Reassess: - Is the effect size too small to be practically meaningful? - Can α be slightly increased or power slightly reduced without unacceptable risk? - Consider: - Sequential or staged sampling - Focusing on larger, more meaningful improvements - Always align: - The planned effect size with what is practical and important - The risk levels with stakeholder expectations --- Combining Normality, Equal Variance, and Sample Size in Practice Typical Workflow Before a Parametric Test For example, comparing means of two processes: - Step 1: Plan sample size - Define: - Desired detectable difference in means - Variability estimate - α and power - Use sample size tool for two-sample t-test to determine n per group - Step 2: Collect data - Ensure random sampling where possible - Maintain independence (avoid repeated measures unless using paired designs) - Step 3: Check normality - Use: - Histogram and boxplots - Normal probability plots - Anderson–Darling or Shapiro–Wilk test - Decide if normality assumption is reasonable - Step 4: Check equal variances - Use: - Side-by-side boxplots - Levene’s test or Brown–Forsythe test - Decide if pooled-variance methods are appropriate - Step 5: Choose and run the test - If normality and equal variances hold: - Use pooled two-sample t-test or ANOVA - If normality fails but variances okay: - Transform data or use nonparametric test - If variances unequal: - Use Welch’s t-test or robust alternatives - Step 6: Interpret results - Focus on: - p-values relative to α - Confidence intervals for effect size - Practical significance, not only statistical significance Consistency Between Planning and Analysis When planning: - Use assumptions about normality, variance, and effect size When analyzing: - Verify if those assumptions are roughly met - If they are not, document: - The nature of deviation (for example, strong skewness, variance doubling) - Any adjustments made (transformations, alternative tests) - How these affect interpretation of results This closes the loop between planned and executed study conditions. --- Interpreting Results and Making Decisions From p-Values to Decisions For all these tests (normality, equal variance, and main hypothesis tests): - Compare p-value to α - Use the result to: - Decide whether assumptions hold - Decide whether there is evidence of a difference (or change) Avoid common pitfalls: - Not checking assumptions before interpreting the main test - Using p-values alone without: - Confidence intervals - Practical significance - Context of variation and process understanding Confidence Intervals and Practical Significance Alongside p-values: - Confidence intervals show the plausible range of: - Mean differences - Proportion differences - Standard deviation or variance Decision-making considerations: - If interval for mean difference excludes 0: - Evidence of difference is present - Check whether the entire interval is practically important - If interval is wide: - The study may be underpowered - Consider whether additional data collection is feasible or necessary Alignment of statistical and practical significance is essential for sound conclusions. --- Summary Including tests of equal variance, normality testing, and sample size calculation in an analysis ensures that: - The selected statistical tests are appropriate for the data - The risk of incorrect conclusions is controlled - The study is designed with sufficient power to detect meaningful changes Key points: - Use visual and formal tests to assess normality and equal variance before applying parametric methods. - Interpret p-values in the context of sample size, plots, and practical impact. - Plan sample size using effect size, variability, α, and desired power for both continuous and discrete data. - When assumptions fail, use transformations or robust/nonparametric methods and re-interpret results accordingly. - Anchor all decisions in both statistical evidence and practical significance. Mastering these elements enables rigorous design, execution, and interpretation of hypothesis tests, ensuring trustworthy conclusions from process data.