2.2.3 Normal Distributions & Normality

Normal Distributions & Normality Understanding the Normal Distribution Shape and Key Characteristics The normal distribution is a continuous, bell-shaped probability distribution that is symmetric about its mean. - Shape: Bell-shaped, unimodal, symmetric - Center: Mean (μ), median, and mode are equal - Spread: Described by standard deviation (σ) - Tails: Extend infinitely in both directions, approaching but never touching the horizontal axis The normal distribution is fully defined by two parameters: - Mean (μ): Location of the center - Standard deviation (σ): Width or spread of the distribution A small σ produces a tall, narrow curve; a large σ produces a flatter, wider curve. The Standard Normal Distribution The standard normal distribution is a special case of the normal distribution. - Mean: 0 - Standard deviation: 1 - Notation: Z ~ N(0, 1) Any normal random variable X ~ N(μ, σ²) can be converted to Z using: - Z-score: Z = (X − μ) / σ Z-scores: - Express how many standard deviations a value is from the mean - Allow use of standard normal tables or software for probability calculations The Empirical Rule (68–95–99.7) For a normal distribution: - About 68% of values lie within ±1σ of the mean - About 95% of values lie within ±2σ of the mean - About 99.7% of values lie within ±3σ of the mean This rule is useful to: - Check whether observed data roughly follow a normal pattern - Estimate probabilities quickly without detailed calculations - Understand how “extreme” values compare to the bulk of the data Probabilities and Z-Scores Computing Z-Scores For a data point x from a normal distribution with mean μ and standard deviation σ: - Z = (x − μ) / σ Interpretation: - Z > 0: value is above the mean - Z < 0: value is below the mean - |Z| larger: value more extreme relative to the mean For sample data, when μ and σ are unknown, they are typically estimated by: - Sample mean: x̄ - Sample standard deviation: s Then a value x can be standardized approximately as: - Z ≈ (x − x̄) / s (For probability calculations assuming population normality, use μ and σ when available.) Finding Probabilities from Z To find probabilities in a normal distribution, convert x to Z, then use the standard normal distribution. Key relationships: - Left-tail probability: P(Z ≤ z) - Right-tail probability: P(Z ≥ z) = 1 − P(Z ≤ z) - Between two values: P(a ≤ Z ≤ b) = P(Z ≤ b) − P(Z ≤ a) For a general normal variable X: - P(X ≤ x) = P(Z ≤ (x − μ)/σ) - P(X ≥ x) = 1 − P(Z ≤ (x − μ)/σ) These probabilities are obtained from: - Z-tables - Statistical software - Calculator functions Using Percentiles and Cutoffs Percentiles express relative standing in the distribution. - Percentile: value below which a given percentage of observations falls - For a normal distribution, percentiles map directly to Z-scores Procedure: - To find x for a given percentile p: - Find z such that P(Z ≤ z) = p - Convert back: x = μ + zσ - To find percentile of a given x: - Convert to Z: z = (x − μ)/σ - Find P(Z ≤ z) This is used for: - Determining specification limits or control thresholds - Evaluating how extreme a measurement is in the process context Normality and Process Data Why Normality Matters Many statistical tools used in process analysis assume normality, especially when: - Modeling continuous data (e.g., times, lengths, weights) with parametric methods - Performing hypothesis tests on means - Building confidence intervals for means - Using capability indices (Cp, Cpk, Pp, Ppk) based on normal assumptions - Interpreting control chart limits (for charts that rely on normality) If the normality assumption is seriously violated: - Probability estimates and p-values may be inaccurate - Confidence intervals may be misleading - Capability indices may misrepresent actual process performance The severity of impact depends on: - Degree and type of non-normality (skewness, heavy tails, etc.) - Sample size (large samples can reduce some effects) - Specific methods used Common Sources of Non-Normality Non-normal data can arise from: - Physical limits: e.g., cannot be negative, causing skewness - Mixtures of populations: combining different operating conditions or product types - Process shifts: changes over time causing multiple peaks - Measurement issues: resolution limits, rounding, or truncation - Small sample sizes: empirical shape appears irregular Typical patterns: - Right-skewed: long tail to the right (e.g., waiting times) - Left-skewed: long tail to the left - Bimodal or multimodal: more than one peak (often indicates mixed populations) - Heavy-tailed: more extreme values than expected under normality Recognizing these patterns is a key step before relying on normal-based methods. Assessing Normality Visual Methods Visual tools provide an initial check of normality. - Histogram with normal curve - Plot a histogram of the data - Overlay a theoretical normal curve using the sample mean and standard deviation - Compare shape, symmetry, and tail behavior - Boxplot - Look for symmetry around the median - Check for extreme outliers - Long whiskers on one side suggest skewness - Normal probability plot (Q-Q plot) - Plots ordered data against expected normal quantiles - If points lie roughly along a straight line: - Data are approximately normal - Systematic curvature suggests: - S-shape: skewness - Bowed pattern: heavy or light tails Visual methods help detect substantial deviations from normality that may affect subsequent analysis. Numerical Descriptions: Skewness and Kurtosis Two numerical measures often used to describe departures from normality are skewness and kurtosis. - Skewness: - Measures asymmetry of the distribution - For a perfectly normal distribution: skewness = 0 - Positive skewness: right tail longer or heavier - Negative skewness: left tail longer or heavier - Kurtosis: - Measures tail heaviness and peak sharpness - For a normal distribution: excess kurtosis = 0 (kurtosis = 3 on some scales) - Positive excess kurtosis: heavier tails, sharper peak - Negative excess kurtosis: lighter tails, flatter peak Interpretation guidelines: - Small deviations in skewness and kurtosis may be acceptable, especially with modest sample sizes - Large values suggest clear departures from normality that may influence parametric methods Statistical Tests of Normality Formal tests provide a p-value to evaluate the hypothesis that data come from a normal distribution. Common tests include: - Anderson–Darling test - Shapiro–Wilk test - Kolmogorov–Smirnov–Lilliefors variants General interpretation: - Null hypothesis (H₀): Data are from a normal distribution - Alternative (H₁): Data are not from a normal distribution - p-value: - p > α (e.g., 0.05): fail to reject H₀ (no evidence against normality) - p ≤ α: reject H₀ (evidence of non-normality) Important considerations: - With very large samples, small deviations produce very small p-values - With small samples, tests may have low power and fail to detect moderate non-normality - Always combine test results with visual assessments and process knowledge Dealing with Non-Normal Data When Normality Is “Good Enough” In practice, data rarely follow a perfect normal distribution, but many methods remain robust if: - Sample size is reasonably large - Departures from normality are mild - There are no extreme outliers or severe skewness In such situations: - Parametric methods assuming normality can often still be applied - Results should be interpreted with some caution - Visual checks and sensitivity considerations support the decision Transformations Toward Normality When non-normality is substantial and normal-based methods are needed, data transformations may be used. Common transformations: - Log transformation: Y = ln(X) or log₁₀(X) - Often used for right-skewed, positive data (e.g., time, concentration) - Square root transformation: Y = √X - Sometimes applied to count data or moderate skewness - Reciprocal transformation: Y = 1/X - Used for strongly right-skewed data with positive values - Box–Cox transformation: - Family of power transformations: Y(λ) = (X^λ − 1)/λ for λ ≠ 0, and ln(X) for λ = 0 - λ is chosen to make the transformed data approximately normal Key points: - Transformations change the scale of the data - Interpretation of results must be done in the original scale if practical decisions depend on original units - Back-transformation may be needed to communicate results Alternatives to Normal-Based Methods When data remain clearly non-normal or transformation is undesirable: - Consider methods that do not require normality, such as nonparametric tests for location differences - Use techniques designed for specific distributions (e.g., exponential, Weibull) if appropriate - For capability analysis, use non-normal capability methods that fit alternative distributions Even when alternatives are used, understanding normality remains crucial for: - Choosing the right method - Interpreting results compared to standard normal-based approaches Normality in Capability and Control Normality in Process Capability Analysis Many traditional capability indices are defined under the assumption of normality. - Cp, Cpk: - Based on process standard deviation and mean - Assume the process distribution is approximately normal - Relate specification limits to the natural spread of the process (often ±3σ) If the data are non-normal: - The relationship between σ and tail probabilities is altered - Cp and Cpk may misrepresent actual defect probabilities - Very skewed distributions can show acceptable indices while still generating many out-of-spec observations on one side Before using normal-based capability indices: - Check normality visually and statistically - Consider transformations or non-normal capability methods if needed Normality in Control Charts Many control charts for continuous data operate effectively under approximate normality. - X̄ and R charts: - Use averages and ranges from subgroups - For subgroup sizes > 1, the distribution of subgroup means is closer to normal (Central Limit Theorem), even if individual measurements are somewhat non-normal - X̄ and s charts: - Similar logic, using standard deviation instead of range Considerations: - If individual data are extremely non-normal: - Control limits and false alarm rates may deviate from their theoretical values - Alternative charts or transformations may be more appropriate (e.g., charts on transformed data or distribution-based charts) Normality checks: - Examine histogram and normal probability plot of subgroup means or individuals - Investigate causes of strong non-normality, such as mixture of conditions or special causes Practical Interpretation of Normality Linking Process Knowledge and Statistical Shape Statistical normality is not an abstract property; it reflects how the process behaves. - Symmetric, stable processes with numerous small independent effects often approximate normality - Strong skewness, multimodality, or heavy tails often indicate: - Different underlying conditions or modes of operation - Process instability or special causes - Physical or operational constraints Combining: - Data shape - Process understanding - Visual checks - Formal tests supports reliable decisions about: - Validity of normal-based methods - Need for transformations or alternative approaches - Interpretation of capability, stability, and risk Balancing Precision and Practicality In practice: - Do not insist on perfect normality; instead, evaluate whether deviations are meaningful for the decision at hand - For moderate sample sizes and mild departures: - Normal-based methods often give useful approximations - For strong non-normality, critical decisions, or probability estimates in the tails: - Use more appropriate distributional assumptions - Apply transformations or nonparametric approaches The goal is to ensure that conclusions about process performance and improvement are statistically sound and practically meaningful. Summary Normal distributions provide a fundamental model for continuous data, defined entirely by mean and standard deviation. Z-scores and the standard normal distribution enable probability calculations, percentile determination, and interpretation of how extreme particular values are. Many statistical tools for analysis, capability assessment, and control rely on an assumption of normality. Assessing normality involves visual methods, numerical descriptors like skewness and kurtosis, and formal tests. When data significantly deviate from normality, transformations or alternative methods may be required, especially for accurate estimation of tail probabilities and capability indices. Interpreting normality in the context of process knowledge ensures that statistical models reflect real process behavior and support sound decision-making.