Statistical Distributions

Statistical Distributions

Quick Answer: A statistical distribution shows which values a process variable is likely to take and how often. For CSSGB, master the normal distribution for continuous SPC and capability, the binomial for defectives, the Poisson for defect counts, and the chi-square, Student’s t, and F distributions as the sampling distributions behind confidence intervals and many hypothesis tests.

Why Distributions Matter in Measure and Analyze

Six Sigma is quantitative. When you estimate capability, set control limits, size a sample, or test whether two means differ, you assume a probability model for the data. Choosing the wrong model produces wrong p-values, wrong control limits, and wrong decisions.

ASQ’s CSSGB BoK (III.C) expects you to understand common distributions as they apply to probability and SPC—not to derive density formulas from first principles, but to know what each models, when it is used, and how it connects to tools you already use (control charts, DPMO thinking, hypothesis tests).

A random variable is a numeric outcome of a process or experiment (thickness, defect count, cycle time). Its distribution may be:

• Discrete — countable outcomes (0, 1, 2, … defects; pass/fail)
• Continuous — measurable on a continuum (temperature, weight, time)

Normal Distribution

The normal (Gaussian) distribution is the continuous, symmetric, bell-shaped model defined by mean μ and standard deviation σ. About 68% of values lie within ±1σ of the mean, 95% within ±2σ, and 99.7% within ±3σ (the empirical rule).

Standard normal: Convert any normal value x to a z-score:

$z = \frac{x - \mu}{\sigma}$

Then use standard normal tables or software for probabilities.

Where Green Belts use it:

Worked example — fill weight: A stable filler produces bottles with mean 500.0 g and σ = 1.2 g. Spec is 497–503 g. Probability of a bottle below 497 g:

$z = \frac{497 - 500}{1.2} = -2.50$

From standard normal tables, P(Z < −2.50) ≈ 0.0062 (about 0.62%). If volume is high, that still means many short-fills over a year—capability work may be justified even if “most” bottles look fine.

Caveats: Skewed cycle times, multi-modal mixtures, and truncation at zero violate normality. Use histograms, normal probability plots, and process knowledge before trusting z-based % defective estimates. Subgroup averages are more robust via the CLT than individuals.

Binomial Distribution

The binomial distribution models the number of successes (often “defectives”) in n independent trials, each with the same probability p of success.

Assumptions: fixed n, two outcomes per trial, constant p, independence.

Mean and variance:

$\mu = np, \quad \sigma^2 = np(1-p)$

Where Green Belts use it:

• p-charts and np-charts — monitoring proportion or count of defectives
• Pass/fail sampling — lot acceptance thinking, attribute MSA “go/no-go” contexts
• Probability of k defectives in a sample of n when baseline defective rate is p

Worked example: Historical first-pass yield implies p = 0.04 defective. In a daily sample of n = 50 units, expected defectives = np = 2.0. An np-chart centerline is 2; control limits use $\sqrt{np(1-p)}$ for the standard deviation of the count. A day with 8 defectives is far above the mean and may signal a special cause—if sampling and inspection rules stayed constant.

Use binomial (or charts based on it) when you count units that are good or bad. If one unit can have multiple defects, prefer Poisson/u-chart logic.

Poisson Distribution

The Poisson distribution models the number of events (defects, calls, scratches) occurring in a fixed interval of time, area, or opportunity, when events are rare relative to opportunity and occur independently at a constant average rate λ.

$P(X = k) = \frac{e^{-\lambda} \lambda^k}{k!}, \quad \mu = \lambda, \quad \sigma^2 = \lambda$

Where Green Belts use it:

• c-charts — count of defects per constant-size unit (defects per 100 invoices of fixed length)
• u-charts — defects per unit when opportunity area varies
• Modeling arrivals, failures, or nonconformities per length/area/time

Binomial vs Poisson (exam trap):

Worked example: A packaging line averages λ = 3 surface defects per pallet. Probability of zero defects on the next pallet ≈ e^(−3) ≈ 0.05. Most pallets have some defects; reducing λ (not only “inspect harder”) is the improvement path.

Chi-Square (χ²) Distribution

The chi-square distribution is a continuous, right-skewed sampling distribution used mainly for inference about variance and for goodness-of-fit / independence tests—not for plotting daily process measurements.

Green Belt touchpoints:

• Confidence intervals and tests for a single population variance/standard deviation
• Chi-square tests of association in contingency tables (categorical Analyze work)
• Building blocks inside many software ANOVA and normality-related procedures

Degrees of freedom (df) shape the distribution; as df grows, χ² becomes more symmetric.

Student’s t Distribution

The t distribution is symmetric and bell-shaped like the normal but with heavier tails. It is used when estimating a mean and σ is unknown, replaced by sample standard deviation s, especially with small samples.

$t = \frac{\bar{x} - \mu_0}{s/\sqrt{n}}$

Where Green Belts use it:

• One-sample and two-sample t-tests (and paired t) in Analyze
• Confidence intervals for a mean when σ is estimated from data
• As n increases, t approaches the normal (z) distribution

Rule of thumb: If you know σ from a long stable history, z methods may apply; for typical project samples, use t.

F Distribution

The F distribution is right-skewed and arises as a ratio of two scaled chi-square (variance) estimates. It is the reference distribution for:

• Testing equality of two variances
• ANOVA (comparing several means via mean-square ratios)
• Some regression and DOE significance tests (Black Belt depth; Green Belts should recognize F as “variance ratio” logic)

Intuition: If two processes truly share the same variance, the ratio of sample variances should hover near 1; extreme F values suggest unequal spread or, in ANOVA, that treatment mean squares exceed error mean squares.

Choosing the Right Distribution (Decision Map)

SPC Probability Links

Control charts are practical applications of distributional thinking:

• Variable charts (I-MR, X-bar/R, X-bar/S) lean on normal/CLT ideas for limit placement.
• Attribute charts encode binomial (p, np) or Poisson (c, u) variance formulas into control limits.
• Points beyond limits are rare under the assumed model if the process is stable—that is why they signal special causes.

If the distribution assumption is badly wrong (for example, highly skewed individuals on an I-chart without transformation or alternative methods), false alarms and missed signals increase. Measure-phase distribution awareness prevents Analyze-phase misuse of tests and Control-phase blind trust in limits.

Green Belt Study Focus

For CSSGB “Understand” level on III.C:

1. Name each distribution and state its typical data type (continuous vs count).
2. Tie normal → capability/SPC variables; binomial → defectives; Poisson → defects.
3. Tie t / χ² / F to inference (means, variances, ANOVA), not to daily run charts of raw counts.
4. Never confuse defective unit (binomial) with defect count (Poisson).
5. Remember CLT: averages behave more normally than messy individuals—supporting X-bar charts.

Master these six families and you can read software output, defend chart choice, and answer exam items that describe a scenario and ask which distribution or chart family applies.