6.12 Probability Concepts, Common Distributions, and the Central Limit Theorem

Understanding probability distributions and sampling distributions is essential for statistical modeling in Six Sigma DMAIC projects. Probability distributions provide the mathematical models required to model process variation, compute defect probabilities, and build hypothesis tests.

---

Continuous Probability Distributions

Continuous distributions model quantitative measurements that can take any real value along a continuous scale.

1. Normal (Gaussian) Distribution

The foundational distribution of parametric statistics, characterized by a symmetric bell curve defined by mean $\mu$ and variance $\sigma^2$:

$f(x) = \frac{1}{\sigma \sqrt{2\pi}} \exp \left( -\frac{(x - \mu)^2}{2\sigma^2} \right)$

• Empirical Rule (68-95-99.7 Rule):
  • $\mu \pm 1\sigma$ encompasses $68.27%$ of process data.
  • $\mu \pm 2\sigma$ encompasses $95.45%$ of process data.
  • $\mu \pm 3\sigma$ encompasses $99.73%$ of process data.

2. Standard Normal Distribution ($Z$)

A Normal distribution with $\mu = 0$ and $\sigma = 1$. Any normal variable $X$ is transformed to $Z$ via:

$Z = \frac{X - \mu}{\sigma}$

3. Exponential Distribution

Models time between independent random events (e.g., equipment mean time between failures - MTBF) with failure rate $\lambda$:

$f(t) = \lambda e^{-\lambda t} \quad (t \ge 0)$

---

Discrete Probability Distributions

Discrete distributions model countable operational events and attributes.

1. Binomial Distribution

Models the count of successes $x$ in $n$ independent Bernoulli trials with constant success probability $p$:

$P(X = x) = \binom{n}{x} p^x (1 - p)^{n - x}$

• Mean: $\mu = np$; Variance: $\sigma^2 = np(1-p)$.
• Six Sigma Application: Baseline probability model for $p$-charts and $np$-charts (defective units).

2. Poisson Distribution

Models the count of rare events $x$ occurring within a fixed interval of time or space at constant average rate $\lambda$:

$P(X = x) = \frac{e^{-\lambda} \lambda^x}{x!}$

• Mean: $\mu = \lambda$; Variance: $\sigma^2 = \lambda$.
• Six Sigma Application: Baseline probability model for $c$-charts and $u$-charts (defect counts).

---

The Central Limit Theorem (CLT)

The Central Limit Theorem is the mathematical foundation of statistical sampling.

Mathematical Statement of CLT

Regardless of the underlying population distribution shape (whether skewed, uniform, or bimodal), as sample size $n$ increases ($n \ge 30$), the sampling distribution of the sample mean $\bar{X}$ approaches a Standard Normal distribution:

$\bar{X} \sim N \left( \mu, \, \frac{\sigma^2}{n} \right)$

Standard Error of the Mean

$\text{SE}(\bar{X}) = \frac{\sigma}{\sqrt{n}}$

Operational Implications for Black Belts

1. Allows Black Belts to use normal-theory hypothesis tests ($Z$-tests, $t$-tests) on sample averages even when raw individual process data is non-normal.
2. Explains why control charts for sample means ($\bar{X}$) exhibit smooth normal control limits.

---

Worked Central Limit Theorem Numerical Example

A high-volume order processing system has an underlying lead time distribution that is heavily right-skewed with a population mean $\mu = 14.0\text{ hours}$ and standard deviation $\sigma = 6.0\text{ hours}$. A Black Belt audits a random sample of $n = 36$ orders.

1. What is the mean and standard error of the sample average $\bar{X}$? $\mu_{\bar{X}} = \mu = 14.0\text{ hours}$ $\text{SE}(\bar{X}) = \frac{\sigma}{\sqrt{n}} = \frac{6.0}{\sqrt{36}} = \frac{6.0}{6} = 1.0\text{ hour}$
2. What is the probability that the sample average lead time $\bar{X}$ exceeds $16.0\text{ hours}$?
   • Transform sample average to $Z$-score: $Z = \frac{\bar{X} - \mu_{\bar{X}}}{\text{SE}(\bar{X})} = \frac{16.0 - 14.0}{1.0} = +2.00$
   • From standard normal tables, $P(Z > 2.00) = 1 - \Phi(2.00) = 1 - 0.9772 = 0.0228$ ($2.28%$).

• Conclusion: Although individual order lead times are heavily non-normal, the sample average of 36 orders follows a Normal distribution per the Central Limit Theorem, allowing precise probability calculations.

---

Basic Probability Concepts

The Body of Knowledge requires Black Belts to describe and interpret independence, mutually exclusive events, the addition and multiplication rules, conditional probability, complementary probability, and joint occurrence.

Definitions and rules

Mutually exclusive is not the same as independent, and confusing the two is the single most common error on this topic. Two mutually exclusive events with non-zero probabilities are in fact strongly dependent: knowing that one occurred tells you with certainty that the other did not. Events cannot be both mutually exclusive and independent unless one of them has probability zero.

Worked examples

Complement. A process has a first-pass yield of 0.94. The probability a unit requires rework is $1 - 0.94 = 0.06$.

Independent multiplication. Three independent operations have first-pass yields of 0.98, 0.95, and 0.99. The probability a unit passes all three untouched is $0.98 \times 0.95 \times 0.99 = 0.9217$. This is exactly the rolled throughput yield calculation, and it depends on the independence assumption -- if a single upstream condition causes failures at more than one station, the true RTY is different.

General addition. In a month, 12% of orders are late, 7% are incomplete, and 3% are both. The probability an order has at least one problem is $0.12 + 0.07 - 0.03 = 0.16$. Failing to subtract the joint term double-counts the 3% and would give 0.19.

Conditional. Using the same figures, given that an order is late, the probability it is also incomplete is $P(\text{incomplete} \mid \text{late}) = 0.03/0.12 = 0.25$. Compare this with the unconditional 7%: lateness and incompleteness are clearly not independent, which is itself a diagnostic finding suggesting a common cause.

Dependent multiplication. A lot of 20 contains 3 defectives. Drawing two units without replacement, the probability both are defective is $\frac{3}{20} \times \frac{2}{19} = \frac{6}{380} = 0.0158$. The second factor changes because the first draw altered the population -- this is the hypergeometric situation, and treating it as $(3/20)^2 = 0.0225$ overstates it by 42%.

Why it matters in practice

• Independence underpins RTY, control chart run rules, and every t and F test. Correlated observations -- consecutive units from the same setup, repeated measures on the same part -- violate it and make significance tests overstate the evidence.
• Conditional probability is the language of detection. In FMEA terms, the escape probability is $P(\text{defect occurs}) \times P(\text{not detected} \mid \text{defect occurs})$, and treating those two as independent when the same condition drives both understates the risk.
• The complement rule is the fastest route to "at least one" problems. The probability of at least one defect in $n$ independent units is $1 - (1-p)^n$, which is far easier than summing the individual cases.