3.1.2 Classes of Distributions

Classes of Distributions Introduction Understanding classes of distributions is essential for correctly modeling data, interpreting statistical measures, selecting valid tests, and making sound process decisions. This article focuses on the main distribution classes relevant to statistical analysis in process improvement: - Discrete vs continuous distributions - Symmetric vs skewed distributions - Common continuous distributions - Common discrete distributions - Practical implications for analysis and problem solving The goal is to recognize which class fits your data, what assumptions follow, and how that affects your conclusions. --- Core Concepts: What Is a Distribution? A distribution describes how values of a random variable are spread across possible outcomes. - Random variable: A variable whose value is determined by chance (e.g., number of defects, time to serve a customer). - Probability distribution: A rule assigning probabilities to each possible value (discrete) or to ranges of values (continuous). - Shape: How data are arranged along the measurement scale (center, spread, skew, tails, number of modes). Classes of distributions group similar shapes and behaviors, helping you match models to data quickly and appropriately. --- Discrete vs Continuous Distributions Discrete Distributions A discrete distribution applies when the variable takes countable, separate values. - Examples: - Number of defects per unit - Number of calls in 10 minutes - Number of patients who no-show in a day - Key characteristics: - Counts or categories - Gaps between possible values - Probabilities assigned to each exact value Common discrete classes: - Bernoulli - Binomial - Poisson - Geometric (occasionally relevant for “trials until first success”) Continuous Distributions A continuous distribution applies when the variable can take any value within an interval. - Examples: - Time to complete a task - Diameter of a part - Temperature, weight, pressure - Key characteristics: - Measured on a scale - Infinite possible values within any range - Probabilities assigned to intervals, not single exact points Common continuous classes: - Normal - Uniform - Exponential - Weibull - Lognormal Recognizing whether your data are discrete or continuous is the first classification step and determines which families are reasonable candidates. --- Symmetry, Skewness, and Tails Symmetric vs Skewed Distributions The shape of a distribution affects which models and statistical methods are appropriate. - Symmetric distribution: - Left and right sides around the center look similar. - Mean and median are close or equal. - Many statistical tools assume symmetry (especially normality). - Skewed distribution: - One tail is longer or heavier. - Right-skewed (positively skewed): long right tail, many small values, few large values (e.g., time to repair). - Left-skewed (negatively skewed): long left tail, many large values, few small values (e.g., time until failure when many fail early). Tails and Outliers - Tails: The thin ends of the distribution where extreme values lie. - Heavy tails: More extreme values than a normal distribution would predict. - Light tails: Fewer extreme values than a normal distribution would predict. Tails matter because: - They affect risk of rare but critical events. - They influence which distribution fits best. - They can affect validity of statistical inferences (e.g., confidence intervals). --- Classes of Continuous Distributions Normal Distribution The normal distribution is a symmetric, bell-shaped distribution that underlies many statistical methods. - Key properties: - Symmetric around its mean. - Mean = median = mode. - Defined by two parameters: mean (μ) and standard deviation (σ). - Approximately 68% of values fall within ±1σ, 95% within ±2σ, and 99.7% within ±3σ. - When it appears: - Measurement data with many small, independent sources of variation. - Aggregated performance metrics (cycle times, dimensions, weights) when the process is stable. - Implications: - Many parametric tests assume normal (or near-normal) data. - Capability indices for continuous data (e.g., based on Z) are often interpreted in a normal-distribution framework. - Deviations from normality (heavy skew, strong outliers, multimodality) may require transformations or non-normal models. Uniform Distribution (Continuous) A continuous uniform distribution assigns equal probability to all values in an interval. - Key properties: - Every value between a minimum (a) and maximum (b) is equally likely. - Rectangular shape. - Mean = (a + b) / 2. - When it appears: - Randomization within fixed limits. - Simplified modeling when detailed structure is unknown but bounded. - Implications: - Useful as a baseline or worst-case assumption for variation. - Not common as a natural long-term process distribution, but helpful in simulations and approximations. Exponential Distribution The exponential distribution models the time between independent events occurring at a constant average rate. - Key properties: - Right-skewed, with a long right tail. - Defined by rate parameter λ (or mean = 1/λ). - Memoryless: future probability does not depend on past elapsed time. - When it appears: - Time between arrivals in a queue when arrivals are random and independent. - Time until an event when failure is equally likely at any moment and rate is constant. - Implications: - Useful for modeling waiting times, interarrival times, and simple reliability scenarios. - Forms part of queuing models and some reliability analyses. Weibull Distribution The Weibull distribution is a flexible right-skewed distribution widely used in reliability and life data analysis. - Key properties: - Defined by shape parameter (k) and scale parameter (λ). - Shape can represent increasing, constant, or decreasing failure rates. - Right-skewed, with tail behavior depending on k. - When it appears: - Time to failure of components, equipment, or systems. - Lifetime data where failure behavior changes over time (e.g., early failures vs wear-out). - Implications: - Can model a wide range of life behaviors (infant mortality, random failures, wear-out). - More flexible than exponential; often preferred when data show non-constant failure rates. Lognormal Distribution The lognormal distribution models positive data where the logarithm of the variable is normally distributed. - Key properties: - Right-skewed, only positive values. - If Y = ln(X) is normal, then X is lognormal. - Mean, median, and mode are all different, with mean > median > mode. - When it appears: - Time or cost when multiplicative factors dominate (e.g., series of percentage increases). - Cycle time with occasional very long delays. - Size-related measures where growth is multiplicative. - Implications: - A natural candidate for skewed, strictly positive, continuous data. - Log transformation often normalizes lognormal data, enabling normal-based methods. --- Classes of Discrete Distributions Bernoulli and Binomial Distributions These model data with two possible outcomes per trial, such as success/failure or pass/fail. - Bernoulli distribution: - Single trial with: - Probability p of success (e.g., defect present). - Probability 1 − p of failure (e.g., no defect). - Binomial distribution: - Number of successes in n independent trials, each with probability p. - Examples: - Number of defective units in a sample of 50. - Number of customers who accept an offer out of 100. - Conditions for binomial: - Fixed number of trials (n). - Only two outcomes (success/failure). - Constant probability p across trials. - Independent trials. - Implications: - Underlies p-charts (proportion defective) and np-charts (number defective). - Central limit theorem often allows normal approximation for large n and p not too close to 0 or 1. Poisson Distribution The Poisson distribution models the count of events in a fixed space or time when events occur independently at a constant average rate. - Key properties: - Parameter λ is the mean number of events per interval. - Mean = variance = λ. - Often right-skewed, especially for small λ. - When it appears: - Number of defects on a unit. - Number of errors per page. - Number of customer arrivals per minute, when arrivals are rare and independent. - Conditions for Poisson: - Events occur independently. - Average rate λ is constant. - Events are rare relative to the possible opportunities. - Implications: - Underlies c-charts (count of defects per unit) and u-charts (defects per unit, varying opportunities). - For larger λ, Poisson can be approximated by a normal distribution. --- Comparing Classes: How to Recognize and Choose Key Questions to Classify a Distribution To determine the appropriate class, start with these questions: - Nature of data: - Are values counts (0, 1, 2, …) or measurements (can be any real number within a range)? - Possible values: - Are negative values possible? - Do values have a natural upper bound or only a lower bound (e.g., time ≥ 0)? - Shape: - Is the histogram symmetric or skewed? - Do you see one peak or multiple modes? - Are there long tails? - Context: - Are you counting occurrences in a fixed interval (Poisson-like)? - Are you counting successes out of fixed trials (binomial-like)? - Are you measuring time to event (exponential, Weibull, lognormal)? Approximation Relationships Some distribution classes approximate others under specific conditions: - Binomial → Normal: - When n is large and p not extremely small or large, the binomial distribution is well approximated by a normal distribution with: - Mean = np - Variance = np(1 − p) - Poisson → Normal: - When λ is moderately large, Poisson counts can be approximated by a normal distribution with: - Mean = λ - Variance = λ - Lognormal ↔ Normal (via transformation): - If log(data) are normal, then data are lognormal. - Applying a log transformation can make skewed data approximately normal, allowing normal-based methods. These relationships explain why normality appears so frequently in practice, even when the underlying mechanism is discrete or skewed. --- Practical Effects of Distribution Classes Impact on Choice of Statistical Methods The class of distribution drives method selection and interpretation. - For continuous, near-normal data: - Many parametric tests (means, regression, ANOVA) assume a normal error distribution. - Capability analysis often relies on normal or transformed-normal assumptions. - For skewed or non-normal continuous data: - Consider transformations (e.g., logarithm) if appropriate. - Consider non-normal models (Weibull, lognormal, exponential) especially for life data and cycle times. - For discrete data: - Use binomial and Poisson-based models for counts and proportions. - Control charts and inferential methods for proportions or counts are derived from these distributions. Impact on Risk and Decision-Making Different classes imply different risk patterns: - Heavy right tails: - Higher probability of extreme long delays, large costs, or major failures. - May require more conservative limits or contingency planning. - Symmetric distributions: - Deviations above and below the mean are similar. - Average performance is more representative of typical behavior. - Bounded distributions (e.g., uniform): - Clear minimum and maximum define the worst and best cases. - Less uncertainty about extremes compared to heavy-tailed distributions. --- Visual and Diagnostic Cues Graphical Assessment by Class You can often infer the class by visual inspection: - Histograms: - Normal: bell-shaped, symmetric. - Lognormal/Weibull: right-skewed, long right tail. - Uniform: flat across the range. - Poisson: right-skewed integer counts, especially when mean is small. - Binomial: finite range of integer counts from 0 to n, with discrete peaks. - Probability plots: - Straight-line pattern on a normal probability plot suggests normality. - Straight line on a Weibull or lognormal probability plot suggests that specific model fits. Recognizing patterns speeds up model selection and helps verify assumptions behind statistical analyses. --- Summary Classes of distributions provide a structured way to understand how data behave and which statistical tools are appropriate. - Data can be discrete (counts) or continuous (measurements). - Distributions can be symmetric (often normal) or skewed (lognormal, Weibull, exponential, Poisson). - Key continuous classes: - Normal for symmetric measurement data. - Uniform for equal-likelihood ranges. - Exponential for time between independent events with constant rate. - Weibull for flexible life and reliability modeling. - Lognormal for positive, multiplicative, right-skewed data. - Key discrete classes: - Bernoulli/Binomial for success/failure counts in fixed trials. - Poisson for event counts in a fixed interval with constant rate. - Approximation relationships (binomial → normal, Poisson → normal, lognormal ↔ normal via log) explain why normal-based methods are widely used. - Correctly identifying the class of distribution ensures that assumptions match reality, leading to valid conclusions and better decisions.