3.5.4 Friedman

Friedman Introduction Friedman refers to the set of nonparametric statistical tests developed by Milton Friedman for comparing related samples when parametric assumptions (such as normality) are not met. In the context of advanced process improvement and data-driven decision making, the most commonly used is the Friedman test for repeated measures across multiple treatments or conditions. This article focuses on the Friedman test and its essential extensions, staying within the scope necessary to apply it correctly in complex improvement projects. --- Purpose of the Friedman Test The Friedman test is a nonparametric alternative to the repeated-measures ANOVA when: - The same subjects (or blocks) are measured under several conditions. - The data are at least ordinal (ranks are meaningful). - Normality and equal variance assumptions are doubtful or clearly violated. Typical uses: - Comparing multiple process settings tested on the same units. - Evaluating different methods, tools, or operators on the same items. - Assessing multiple treatments measured on the same group over time. The test evaluates whether there is a statistically significant difference in the median rankings of three or more related conditions. --- Data Structure and Assumptions Experimental Design Requirements The Friedman test is designed for a blocked and repeated structure: - Blocks (subjects): - Each block represents a matched or repeated unit (e.g., patient, machine, part, time period). - Each block experiences all treatments. - Treatments (conditions): - These are the conditions being compared (e.g., methods, settings, or time points). - There must be at least three treatments. - Response: - Measured for each combination of block and treatment. - Can be continuous or ordinal but is analyzed through ranks. Core Assumptions For valid results, the following must hold: - Random or representative blocks: - Blocks are chosen to represent the population to which conclusions will be generalized. - Within-block comparability: - All treatments are applied within each block under comparable conditions. - Independence of blocks: - Results from one block do not influence another. - At least ordinal scale: - It must be meaningful to say that one observation is larger or smaller than another. - No missing-by-design treatment: - Each block should have a measurement for each treatment (unbalanced data require caution or alternative methods). --- Hypotheses and Statistical Logic Hypotheses The Friedman test evaluates: - Null hypothesis (H₀): - All treatments have the same distribution of responses. - Equivalently, all treatment median ranks are equal. - Alternative hypothesis (H₁): - At least one treatment differs in distribution. - Equivalently, at least one treatment median rank differs. Ranking Logic Within each block: - Responses are converted to ranks: - Smallest value → rank 1, next → rank 2, etc. - Ties share the average of their ranks. - Only relative order within each block matters. - By using ranks: - The test becomes robust to outliers and non-normality. - Differences in scale between blocks are neutralized. The test then examines the pattern of rank sums for each treatment to determine whether some treatments consistently receive higher (or lower) ranks across blocks. --- Friedman Test Statistic Notation Let: - b = number of blocks (subjects or matched sets). - k = number of treatments (conditions). - Rⱼ = sum of ranks for treatment j across all blocks. Test Statistic (Q or χ²F) A commonly used formula for the Friedman statistic is: - Q = (12 / (b·k·(k + 1))) · Σ(Rⱼ²) − 3·b·(k + 1) where the sum Σ(Rⱼ²) is taken over all treatments j = 1,…,k. Key points: - Under H₀ and for sufficiently large b and k, Q approximately follows a chi-square distribution with k − 1 degrees of freedom. - Software may label this statistic differently (e.g., Friedman’s chi-square or χ²F), but the underlying logic is the same. Handling Ties When there are ties within a block: - Ranks are assigned as average ranks. - A tie correction factor can be applied to adjust the variance of rank sums. - Many statistical packages automatically handle tie adjustments. The effect of tie correction is to maintain accurate p-values, especially when ties are numerous. --- Procedure: Step-by-Step Application Step 1 – Identify Blocks and Treatments - Define blocks as units that receive all treatments: - Example: each machine tested with each tool type. - Define treatments as conditions to compare: - Example: Tool A, Tool B, Tool C. Ensure that: - Each block has one observation per treatment. - Measurements are made under comparable conditions. Step 2 – Construct the Data Table Create a table with: - Rows = blocks. - Columns = treatments. - Cells = observed responses. Example format: - Block 1: Treatment 1, Treatment 2, Treatment 3 - Block 2: Treatment 1, Treatment 2, Treatment 3 - … Step 3 – Rank Within Each Block For each block: - Rank treatment responses from smallest (rank 1) to largest (rank k). - For ties: - Assign the average of the ranks those positions would have received. Record these ranks instead of raw data for the calculation. Step 4 – Calculate Rank Sums per Treatment Across all blocks: - Sum ranks for each treatment j: - Rⱼ = sum of ranks of treatment j over all blocks. These Rⱼ values feed into the Friedman test statistic. Step 5 – Compute the Friedman Statistic Using: - Q = (12 / (b·k·(k + 1))) · Σ(Rⱼ²) − 3·b·(k + 1) Compute Q (or use software that returns Friedman’s chi-square). Step 6 – Determine the p-Value Compare Q to a chi-square distribution: - Degrees of freedom = k − 1. - Compute the p-value: - p = P(χ² ≥ Q | df = k − 1). Interpretation: - If p is less than the chosen significance level (e.g., 0.05), reject H₀. - Conclude that at least one treatment differs in central tendency (median rank). --- Interpretation and Practical Meaning What a Significant Result Means A statistically significant Friedman test indicates: - Not all treatments are equivalent in their typical performance. - There is evidence that the central location (median) of at least one treatment’s distribution differs from the others. However: - The test does not identify which specific treatments differ. - It only states that the pattern of ranks across treatments is unlikely under the null hypothesis of equal distributions. Practical Interpretation Guidelines When interpreting results, consider: - Direction of ranks: - Lower average ranks for a treatment typically indicate better performance if lower values are desirable (e.g., cycle time). - Higher average ranks indicate better performance if higher values are desirable (e.g., yield). - Magnitude of differences: - Compare mean or median ranks across treatments. - Examine actual metric differences in the original units for practical significance. - Context and risk: - Assess whether the observed differences are meaningful for the process or system in question, not just statistically detectable. --- Post-Hoc Comparisons for Friedman Need for Post-Hoc Tests When the Friedman test is significant: - The question remains: Which pairs of treatments differ? - Post-hoc pairwise comparisons, based on ranks, are used to answer this. Common Post-Hoc Approaches Post-hoc procedures preserve consistency with nonparametric logic: - Pairwise rank comparisons: - Compare average ranks for each pair of treatments. - Multiple-comparison adjustments: - Procedures adjust for comparing multiple pairs: - Bonferroni-type corrections based on the Friedman rank structure. - Stepwise procedures (e.g., Holm-type adjustments) adapted to this setting. Details differ among implementations, but the essentials include: - Using rank-based differences between treatments. - Maintaining control of the overall error rate when making multiple inferences. Practical Steps for Post-Hoc Analysis - Identify all treatment pairs (e.g., A vs B, A vs C, B vs C). - For each pair: - Compute the difference in average ranks. - Use the appropriate standard error formula based on b and k. - Compare standardized differences to critical values adjusted for multiple testing. - Conclude which specific treatment pairs show significant differences. Most statistical software automates these calculations. The key is understanding that: - Post-hoc comparisons remain consistent with the Friedman rank framework. - They are conducted after a significant overall Friedman test. --- Design and Validity Considerations When to Prefer Friedman Using the Friedman test is especially appropriate when: - Data are clearly non-normal or heavily skewed. - Sample sizes are modest, making parametric assumptions questionable. - There is a repeated-measures or matched-block design with multiple conditions. - The scale is ordinal or the influence of extreme values should be minimized. Common Pitfalls Avoid misusing the Friedman test in these ways: - Independent samples mistaken as blocks: - Blocks must be matched or repeated units, not unrelated groups. - Ignoring missing data structure: - Missing measurements for some treatment–block combinations can invalidate the standard Friedman procedure. - Specialized methods or imputation may be required; otherwise, conclusions weaken. - Treating it as a location-only test without context: - Relying solely on p-values without considering effect size or process relevance leads to poor decisions. - Inappropriate scale: - If the data are nominal with no order, ranks are meaningless; Friedman is not suitable. Being explicit about the design and data properties ensures valid interpretation. --- Relationship to Other Rank-Based Methods Conceptual Placement The Friedman test is part of a family of nonparametric rank-based procedures designed for different data structures: - It specifically handles multiple repeated measures across several treatments. - It generalizes the logic of rank tests for scenarios where: - The same subject or block is exposed to all conditions. - Several conditions (k ≥ 3) are compared simultaneously. Key Distinction The essential distinction is: - Within-block ranking across multiple treatments: - The focus is on consistent ranking patterns across blocks. - This differs from methods focused on independent groups or single-pair comparisons. By understanding this placement, it becomes clear when the Friedman test is the correct choice for analyzing rank-based repeated-measures data. --- Implementation Tips Data Preparation - Check completeness: - Confirm that all block–treatment combinations have valid data. - Inspect distributions: - Look for non-normality and outliers that motivate a nonparametric approach. - Confirm alignment: - Ensure that lower or higher values correspond to the desired direction of performance before interpreting ranks. Using Software When applying the Friedman test with software, ensure: - Correct specification of blocks and treatments: - Identify repeated factors correctly. - Tie handling: - Confirm the method reports tie-corrected statistics and p-values. - Post-hoc options: - Use rank-based pairwise procedures that are compatible with the Friedman framework. Reporting Results When communicating results: - State: - The Friedman statistic value. - Degrees of freedom. - p-value. - Summarize: - Average ranks for each treatment. - Key pairwise differences if post-hoc tests are performed. - Interpret: - The direction and practical relevance of differences in the process context. --- Summary The Friedman test is a nonparametric method for comparing three or more related treatments measured on the same blocks or subjects. It works by: - Ranking responses within each block. - Summing ranks for each treatment. - Evaluating whether these rank sums differ more than expected under equal-treatment assumptions. It is appropriate when: - Data are ordinal or non-normal. - Each block experiences all treatments. - Blocks are independent and comparable across treatments. A significant Friedman result indicates: - At least one treatment differs in central tendency, based on ranks. To identify specific differences, rank-based post-hoc pairwise comparisons with appropriate multiple-comparison adjustments are used. Mastering the Friedman test involves: - Recognizing when the design and data justify its use. - Correctly setting up and ranking the blocked data. - Computing and interpreting the test statistic and p-value. - Extending analysis with compatible post-hoc comparisons while maintaining a clear link between rank-based inference and practical process decisions.