2.2 Six Sigma Statistics

Six Sigma Statistics Statistical Thinking in Six Sigma Six Sigma statistics focuses on using data to understand, predict, and improve process performance. The objective is to distinguish between variation that is inherent to the process and variation caused by specific, removable factors. - Key idea: Decisions are based on data, not anecdotes. - Focus: Quantifying variation, locating its sources, and measuring improvement. - Core tools: Descriptive statistics, probability, statistical inference, and modeling. Everything that follows supports using data rigorously within the DMAIC roadmap, but this article limits itself to the statistical concepts and techniques. --- Data Types and Basic Concepts Measurement Scales and Data Types Understanding data types is essential because they determine which graphs, statistics, and tests are appropriate. - Continuous data: Measured on a scale with infinite possible values. - Examples: time, length, weight, temperature. - Often preferred; carries more information. - Discrete data: Counted in whole units. - Examples: defects per unit, number of calls, number of errors. - Attribute (categorical) data: - Nominal: categories with no natural order (type of defect, product line). - Ordinal: categories with a logical order (rating scales, priority levels). - Binary data: special case of attribute data with only two outcomes. - Examples: pass/fail, yes/no, defect/non-defect. Correct classification supports proper selection of: - Charts (histogram, boxplot, p-chart, u-chart, etc.). - Summary statistics (mean, proportion, median). - Statistical tests (t-test, chi-square, ANOVA, etc.). Basic Descriptive Statistics Descriptive statistics summarize and describe the main features of a data set. - Center: - Mean: arithmetic average; sensitive to outliers. - Median: middle value; robust to outliers. - Mode: most frequent value; useful with categorical data. - Spread: - Range: max – min; quick but crude. - Variance: average squared distance from the mean. - Standard deviation (σ or s): square root of variance; key measure of variation. - Shape: - Skewness: symmetry or lack of it. - Kurtosis: heaviness of tails vs. normal distribution. - Position: - Percentiles and quartiles: position of values relative to the distribution. - Interquartile range (IQR): distance between Q1 and Q3; robust spread measure. In Six Sigma work, standard deviation and mean are central to capability analysis and process performance metrics. --- Graphical Presentation of Data Common Plots and Their Uses Visualizations provide quick insight into behavior and potential issues. - Histogram: - Shows distribution of continuous data. - Helps detect skew, multimodality, outliers, and approximate normality. - Boxplot: - Summarizes median, quartiles, and potential outliers. - Useful for comparing multiple groups or conditions. - Run chart: - Data plotted in time order without control limits. - Reveals trends, shifts, cycles, and clustering. - Scatter plot: - Plots pairs of values (X vs. Y). - Detects relationships, patterns, and outliers between two variables. - Pareto chart: - Bars ordered from most to least frequent category. - Focuses attention on the “vital few” categories. These visual tools are often used before formal hypothesis testing or modeling to check assumptions and guide analysis. --- Probability and Basic Distributions Key Probability Concepts Probability quantifies uncertainty and underlies all statistical inference. - Probability: likelihood of an event (0 to 1). - Complement: P(not A) = 1 – P(A). - Joint probability: P(A and B). - Conditional probability: P(A | B) = probability of A given B. For Six Sigma statistics, probability helps: - Translate defect rates into expected counts. - Model failure behavior. - Evaluate risks of decision errors. Discrete Distributions Discrete distributions describe counts of events. - Binomial distribution: - Fixed number of trials n. - Each trial has two possible outcomes. - Probability of success p is constant. - Applications: number of defective items in a sample, pass/fail counts. - Poisson distribution: - Models counts of rare events in a fixed interval (time, space, etc.). - Characterized by rate λ. - Applications: defects per unit, calls per minute, failures per day. Understanding these distributions supports attribute control charts and defect metrics. Continuous Distributions Continuous distributions are used for measurements and time-related behavior. - Normal distribution: - Symmetric bell-shaped curve. - Defined by mean μ and standard deviation σ. - Central to many Six Sigma methods: capability, control charts, confidence intervals. - t-distribution: - Similar to normal but with heavier tails. - Used when sample size is small and σ is unknown. - Chi-square distribution: - Used for variance and standard deviation inference. - Underlies tests of independence and normality. - F-distribution: - Ratio of two variances. - Used in ANOVA and comparisons of multiple group means. - Exponential and Weibull distributions: - Used to model time-to-failure and reliability. Recognizing which distribution applies helps choose the right formula, chart, or test. --- Sampling and Data Collection Sampling Concepts Sampling is used to estimate population characteristics without measuring every item. - Population: complete set of all items or events of interest. - Sample: subset selected from the population. - Parameter: true but usually unknown population value (μ, σ, p). - Statistic: sample-based estimate of a parameter (x̄, s, p̂). Key ideas: - Random sampling: each item has a known, nonzero chance of selection. - Sampling error: natural difference between sample statistic and true parameter. - Bias: systematic error due to flawed sampling or measurement. Central Limit Theorem The central limit theorem (CLT) is crucial to Six Sigma statistics. - For sufficiently large sample sizes: - The distribution of the sample mean tends toward normality. - This happens regardless of the shape of the population distribution (under broad conditions). - Implications: - Enables use of normal-based methods (z, t, control charts) even when raw data are not perfectly normal. - Justifies confidence intervals and many hypothesis tests. Sample Size and Power Basics Choosing sample size balances precision, cost, and risk. - Larger samples: - Reduce standard error of estimates. - Narrow confidence intervals. - Increase power to detect real differences. - Core relationships: - Standard error of the mean ≈ σ / √n. - To halve margin of error, sample size must roughly quadruple. Even simple approximations are helpful for planning data collection. --- Estimation and Confidence Intervals Point Estimates A point estimate is a single-number best guess of a parameter. - Mean: x̄ approximates μ. - Proportion: p̂ approximates p. - Standard deviation: s approximates σ. - Difference: (x̄₁ – x̄₂) approximates (μ₁ – μ₂). Point estimates are always uncertain, which motivates interval estimation. Confidence Intervals A confidence interval (CI) provides a range of plausible values for a parameter at a given confidence level (often 95%). - General form: - Estimate ± (critical value) × (standard error). - Common intervals: - Mean (σ known): z-based. - Mean (σ unknown): t-based. - Proportion: normal approximation or exact methods. - Difference in means or proportions. - Standard deviation and variance: chi-square based. Interpretation: - A 95% CI does not guarantee that the true parameter is inside the interval for any single sample. - Over many repeated samples, about 95% of such intervals will contain the true parameter. Confidence intervals are used in Six Sigma to: - Assess process performance metrics. - Quantify uncertainty around improvements. - Support decisions based on precision, not only p-values. --- Hypothesis Testing Fundamentals Logic of Hypothesis Testing Hypothesis testing evaluates whether data provide enough evidence to support a claim about a population. - Null hypothesis (H₀): usually a statement of no difference or no effect. - Alternative hypothesis (H₁): statement representing a change, difference, or effect. Process: - Assume H₀ is true. - Compute a test statistic from sample data. - Compare the test statistic to a distribution (z, t, F, chi-square). - Obtain a p-value: - Probability of observing data at least as extreme, given H₀ is true. - Decision rule: - If p-value ≤ α (significance level, often 0.05), reject H₀. - If p-value > α, do not reject H₀ (insufficient evidence against H₀). Types of Errors and Power Statistical decisions can be wrong. - Type I error (α): - Rejecting a true H₀ (false alarm). - Controlled by the chosen significance level (e.g., 0.05). - Type II error (β): - Failing to reject a false H₀ (missed detection). - Power (1 – β): - Probability of correctly rejecting a false H₀. - Increases with: - Larger effect sizes. - Larger sample sizes. - Lower variability. - Higher α. Balancing α and power is central to designing meaningful analyses and experiments. One-Tailed vs. Two-Tailed Tests The choice depends on the question. - Two-tailed: - Tests for difference in either direction (≠). - More common when direction is not pre-specified. - One-tailed: - Tests for change in a specific direction (>, <). - Requires strong justification before data collection. - Concentrates all α in one tail, increasing power for that direction only. --- Statistical Tests for Means and Medians One-Sample Tests Used when comparing a sample to a known or target value. - One-sample t-test: - Question: Is the mean equal to a target? - One-sample z-test: - Used when population σ is known (rare in practice). - One-sample sign or Wilcoxon tests: - Nonparametric alternatives when normality is doubtful or outliers are extreme. - Focus on median rather than mean. Two-Sample and Paired Tests Used to compare two conditions, processes, or treatments. - Two-sample t-test (independent samples): - Compares means of two independent groups. - Variants: - Equal variances assumed. - Unequal variances (Welch correction). - Paired t-test: - Same subjects or units measured twice (before/after, two conditions). - Tests mean of differences. - Nonparametric counterparts: - Mann–Whitney test for independent samples. - Wilcoxon signed-rank for paired data. - Used when assumptions of normality are violated. Assumptions to check: - Independence of observations. - Approximate normality of data (or differences). - Homogeneity of variances for some tests. --- Analysis of Variance (ANOVA) One-Way ANOVA ANOVA compares means across three or more groups. - Purpose: Test whether at least one group mean differs from the others. - Concept: - Total variation in data is partitioned into: - Between-group variation (due to factor). - Within-group variation (random error). - F-statistic compares these two sources: - F = MSbetween / MSwithin. - Assumptions: - Independent observations. - Approximate normality within groups. - Equal variances across groups. If overall F-test is significant, follow-up (post hoc) comparisons identify which pairs of groups differ. Two-Way and Multi-Factor ANOVA When more than one factor is involved: - Two-way ANOVA: - Evaluates main effects of two factors. - Tests for interaction: whether the effect of one factor depends on the level of another. - General multi-factor ANOVA: - Handles more factors and interactions. In Six Sigma, ANOVA is used to: - Compare process performance across shifts, machines, suppliers, or methods. - Assess influence of factors on critical outputs. --- Regression and Correlation Correlation Correlation quantifies the strength and direction of linear association between two continuous variables. - Pearson correlation coefficient (r): - Range: -1 to +1. - r > 0: positive relationship. - r < 0: negative relationship. - r ≈ 0: little or no linear relationship. - Important cautions: - Correlation does not prove causation. - Outliers can strongly influence r. - Nonlinear relationships may yield low r despite strong association. Simple Linear Regression Simple linear regression models the relationship between one predictor (X) and one response (Y). - Model form: Y = β₀ + β₁X + ε - β₀: intercept. - β₁: slope (change in Y per unit change in X). - ε: random error. - Key outputs: - Estimated slope and intercept. - p-value for slope: evidence that X affects Y. - R² (coefficient of determination): - Proportion of variation in Y explained by X. - Assumptions: - Linearity between X and Y. - Independent errors. - Constant variance of errors (homoscedasticity). - Approximately normal error distribution. Regression is used to predict process outcomes and quantify the strength of input–output relationships. Multiple Linear Regression Multiple regression extends simple regression to multiple predictors. - Model: Y = β₀ + β₁X₁ + β₂X₂ + … + βkXk + ε - Uses: - Estimate contributions of multiple factors to a response. - Identify significant drivers of process performance. - Control for confounding factors. - Considerations: - Multicollinearity (correlated predictors) can distort coefficient estimates. - Model selection and simplification help maintain interpretability. Multiple regression is often linked with designed experiments but can also be applied to observational data. --- Nonparametric Methods When and Why to Use Nonparametric Tests Nonparametric methods are used when: - Data do not meet assumptions of normality. - Distributions are seriously skewed or heavy-tailed. - Data are ordinal or ranked. - Sample sizes are small, and transformations are ineffective. Advantages: - Fewer assumptions. - More robust to outliers and non-normality. - Applicable to ranked or categorical order data. Common Nonparametric Tests Typical tests relevant to Six Sigma statistics: - Mann–Whitney: - Alternative to two-sample t-test for independent groups. - Wilcoxon signed-rank: - Alternative to paired t-test. - Kruskal–Wallis: - Alternative to one-way ANOVA for multiple groups. - Mood’s median test: - Tests for equality of medians across groups. These tests rely on rankings rather than actual values and focus on location differences in distributions. --- Statistical Process Control Basics Common Cause vs. Special Cause Variation Control charts distinguish between two types of variation: - Common cause: - Natural, inherent fluctuation in a stable process. - Random, predictable within statistical limits. - Special cause: - Due to specific, identifiable factors. - Causes signals beyond what is expected by random variation. The goal is to achieve and maintain a stable system with only common cause variation before capability analysis or major changes. Structure of Control Charts Control charts monitor processes over time using plotted statistics. - Center line (CL): typical process level (mean, proportion, etc.). - Upper and lower control limits (UCL, LCL): - Calculated statistically, usually at ±3 standard errors from the CL. - Not specification limits. - Points: - Each point is a statistic from a subgroup or time period. - Rules for signals (examples): - Point outside control limits. - Run of several points on one side of CL. - Trends or cycles. - Near-constant or highly erratic patterns. Types of Control Charts Selection depends on data type and subgrouping. - For continuous data: - X̄-R chart: for subgroup sizes typically 2–10. - X̄-S chart: for larger subgroups (n > 10). - Individuals–Moving Range (I-MR) chart: for n = 1 (no subgroups). - For attribute data: - p-chart: proportion defective with varying sample sizes. - np-chart: number defective with constant sample size. - c-chart: count of defects per unit when area of opportunity is fixed. - u-chart: defects per unit when area of opportunity varies. Control charts are used to verify stability before estimating process capability and to sustain gains after improvements. --- Measurement System Analysis (MSA) Measurement Error and Variation Any data analysis assumes that the measurement system is adequate. - Total observed variation = process variation + measurement variation. - Excessive measurement error can: - Obscure true changes. - Create false signals. - Mislead capability and performance studies. Measurement system quality is assessed for: - Repeatability (equipment variation). - Reproducibility (appraiser variation). - Stability over time. - Linearity and bias across operating range. Gage R&R for Variables Data Gage Repeatability and Reproducibility (Gage R&R) quantifies measurement variation for continuous data. - Basic design: - Multiple appraisers measure multiple parts multiple times. - Decomposition of variation: - Part-to-part variation. - Repeatability (same appraiser, same part). - Reproducibility (differences between appraisers). - Key metrics: - % Gage R&R relative to total variation. - Number of distinct categories (NDC) the system can reliably distinguish. Guidance: - Lower % Gage R&R and higher NDC indicate a better measurement system. Attribute Agreement Analysis For attribute (classification) data, attribute agreement analysis evaluates: - Within-appraiser consistency. - Between-appraiser agreement. - Agreement with a reference standard or known truth. Common outputs: - Percent agreement (with self, with others, with standard). - Kappa statistics for agreement beyond chance. Reliable attribute measurement is essential when using defect counts, pass/fail data, or categorization in improvement work. --- Process Capability and Performance Defects, DPMO, and Sigma Level Process capability in Six Sigma is often expressed as sigma level. - Defect: any failure to meet a requirement. - Defective unit: unit with one or more defects. - Defects per million opportunities (DPMO): - DPMO = (Number of defects / (Units × Opportunities per unit)) × 1,000,000 - Sigma level: - Converts defect rates or DPMO into an equivalent standard deviation distance from the target or specification. - Higher sigma level implies fewer defects. These measures allow comparison of different processes on a common scale. Capability Indices for Continuous Data Capability indices compare the spread and centering of a process to specification limits. - Cp: - Cp = (USL – LSL) / (6σ) - Measures potential capability assuming process is centered. - Cpk: - Cpk = min[(USL – μ) / (3σ), (μ – LSL) / (3σ)] - Accounts for both spread and centering. - Sensitive to process mean shift. - Pp and Ppk: - Similar to Cp and Cpk but based on overall standard deviation (including between-subgroup variation). - Reflect long-term performance rather than within-subgroup capability. Interpretation: - Indices > 1 indicate the process spread is narrower than the specification width. - Higher values represent better capability and fewer nonconforming items. Data Requirements and Assumptions Capability analysis requires: - Stable, in-control process (verified by control charts). - Appropriate distribution model (often normal). - Meaningful specifications and measurement system. For non-normal data, transformations or alternative distribution-based analyses (e.g., lognormal, Weibull) may be applied. --- Design of Experiments (Statistical Foundation) Purpose of Designed Experiments Designed experiments (DOE) use structured data collection to understand how factors affect a response. Key statistical ideas: - Control and deliberate variation of inputs (factors). - Randomization to protect against hidden biases. - Replication to estimate experimental error. - Blocking to handle known sources of variability. Factorial Designs Factorial designs study multiple factors simultaneously. - Full factorial: - All combinations of factor levels are tested. - Allows estimation of: - Main effects (impact of each factor). - Interactions (combined effects of factors). - Fractional factorial: - Uses subset of all combinations. - More efficient but may confound certain effects. Statistical analysis uses ANOVA and regression: - Identify significant factors and interactions. - Quantify their effects on the response. - Build predictive models. Response Surface Methods (Brief Foundation) When relationships are curved rather than purely linear: - Response surface methods (RSM) fit quadratic models. - Designs such as central composite and Box–Behnken allow estimation of curvature. - The statistical focus is: - Estimation of linear, interaction, and squared terms. - Optimization of responses (maximizing or minimizing performance). All DOE analyses rely heavily on regression, ANOVA, residual analysis, and the standard assumptions listed earlier. --- Reliability and Life Data Basics Life Distributions When analyzing time to failure or survival data: - Common distributions: - Exponential. - Weibull. - Lognormal. - Parameters characterize: - Shape of failure risk over time. - Scale (time dimension). These models allow estimation of: - Probability of survival to a given time. - Mean time to failure (MTTF) or mean time between failures (MTBF). Censoring and Estimation Reliability data often include censored observations: - Right-censoring: - Item has not failed by the end of the study. - Estimation methods: - Maximum likelihood estimation tailored for censored data. - Graphical methods (e.g., Weibull plots) to check model fit. Reliability statistics support failure prediction, warranty analysis, and maintenance strategies. --- Summary Six Sigma statistics provides a toolkit for understanding and improving process performance through data. Core capabilities include: - Classifying data types and summarizing them with descriptive statistics and graphical methods. - Applying probability and distributions to model variation and uncertainty. - Designing samples, estimating parameters, and constructing confidence intervals. - Conducting hypothesis tests to evaluate differences, relationships, and effects. - Using ANOVA, correlation, and regression (including multiple regression) to analyze complex data structures. - Applying nonparametric tests when parametric assumptions fail. - Monitoring and stabilizing processes using control charts. - Assessing measurement systems to ensure data reliability. - Quantifying process capability and performance via indices and sigma levels. - Using designed experiments and reliability analyses grounded in solid statistical principles. Mastery of these statistical concepts enables rigorous diagnosis of process behavior, valid evaluation of changes, and confident, data-based decision making within Six Sigma projects.