4.3 Normal Distribution & Z-Score Calculations

4.3 Normal Distribution & Z-Score Calculations

Foundations of the Normal (Gaussian) Distribution

The normal distribution (often called the Gaussian distribution or the classic bell curve) is the single most important continuous probability distribution in quality engineering. First described mathematically by Abraham de Moivre in 1733 and later expanded by Carl Friedrich Gauss, the normal distribution models natural variation resulting from the combination of many independent, random common causes (such as minor ambient temperature fluctuations, machine micro-vibrations, material alloy inconsistencies, and slight line voltage variations).

Mathematical Probability Density Function (PDF)

The continuous probability density function of a normal random variable $X$ with mean $\mu$ and standard deviation $\sigma$ is defined by:

$$f(x) = \frac{1}{\sigma \sqrt{2\pi}} e^{-\frac{1}{2}\left(\frac{x - \mu}{\sigma}\right)^2}, \quad -\infty < x < \infty$$

where $\pi \approx 3.14159$ and $e \approx 2.71828$.

Core Geometric and Metrological Properties

1. Perfect Symmetry: The distribution is unimodal and perfectly symmetrical around its central axis. The left half is a precise mirror image of the right half.
2. Coincidence of Central Measures: At the peak center of the distribution, the arithmetic mean, median, and mode are identical: $$\mu = \text{Median} = \text{Mode}$$
3. Points of Inflection: The curve changes concavity (from concave downward to concave upward) at exactly one standard deviation on either side of the mean: $\mu - \sigma$ and $\mu + \sigma$.
4. Asymptotic Tails: The curve approaches the horizontal axis asymptotically as $x$ moves toward $\pm \infty$, never quite touching zero. Theoretically, any value between $-\infty$ and $+\infty$ is possible, though values beyond $\pm 4\sigma$ are exceedingly rare.
5. Unit Area: The total area beneath the probability density curve is exactly equal to 1.0 (representing 100% total probability).

---

The Empirical Rule (The 68-95-99.7 Rule)

For any data set that is approximately normally distributed, the Empirical Rule dictates the fixed proportion of observations falling within integer standard deviation bands around the mean:

                         Empirical Rule Distribution
                                     μ
                                    /|\
                                   / | \
                                  /  |  \
                              ---/   |   \---
             ----------------/       |       \----------------
    Tail |    2.14%   |   13.59%  |34.13%|34.13%|  13.59%   |    2.14%   | Tail
   0.135%|            |           |      |      |           |            |0.135%
 --------+------------+-----------+------+------+-----------+------------+--------
        μ-3σ         μ-2σ        μ-1σ    μ     μ+1σ        μ+2σ         μ+3σ
         |                        |              |                        |
         |                        +--- 68.26% ---+                        |
         |                                                                |
         +---------------------------- 95.44% ----------------------------+
         |                                                                |
         +--------------------------- 99.73% -----------------------------+

Exact Theoretical Percentages

• $\mu \pm 1\sigma$: Contains 68.26% of all parts (approx. 68%). Exactly $34.13\%$ lies between $\mu$ and $\mu + 1\sigma$, and $34.13\%$ lies between $\mu - 1\sigma$ and $\mu$.
• $\mu \pm 2\sigma$: Contains 95.44% of all parts (approx. 95%). Exactly $47.72\%$ lies on each side of the mean.
• $\mu \pm 3\sigma$: Contains 99.73% of all parts (approx. 99.7%). Exactly $49.865\%$ lies on each side of the mean.

Tail Area Probabilities and Scrap PPM

When a process is centered between specification limits located at $\pm 3\sigma$:

• The proportion of product falling outside the limits is $1.0000 - 0.9973 = 0.0027$ (or 0.27%).
• Divided evenly between the two tails: 0.135% falls above the Upper Specification Limit ($+3\sigma$), and 0.135% falls below the Lower Specification Limit ($-3\sigma$).
• In parts-per-million: $0.0027 \times 1,000,000 = \mathbf{2,700\text{ PPM}}$ total nonconformance (1,350 PPM per tail).

---

The Standard Normal Distribution ($Z \sim N(0, 1)$) and Z-Scores

Because every manufacturing process has its own unique mean $\mu$ and standard deviation $\sigma$, comparing raw measurements across different processes is cumbersome. The standard normal distribution is a standardized normal curve that has a mean of zero ($\mu = 0$) and a standard deviation of one ($\sigma = 1$).

Z-Score Transformation Formula

A Z-score (or standard score) converts any raw measurement $X$ into units of standard deviations relative to the process mean:

• For a Population: $$Z = \frac{X - \mu}{\sigma}$$
• For a Sample: $$Z = \frac{X - \bar{x}}{s}$$

Physical Meaning of $Z$

• If $Z = 0$, the measurement falls exactly on the process mean.
• If $Z = +1.50$, the measurement lies exactly $1.50$ standard deviations above the mean.
• If $Z = -2.25$, the measurement lies exactly $2.25$ standard deviations below the mean.

---

Reading Standard Normal Cumulative Probability Tables

Standard normal tables (Z-tables) provide the cumulative area under the curve from $-\infty$ up to a given $Z$ value, denoted $\Phi(Z) = P(Z \le z)$.

Cumulative Table Reference Snippet

Crucial Symmetry Rule for Negative $Z$

Because standard normal curves are symmetrical:

$$P(Z \le -z) = 1 - \Phi(z) = P(Z \ge z)$$

For example, $P(Z \le -1.96) = 1 - \Phi(1.96) = 1 - 0.9750 = 0.0250$.

---

Step-by-Step Probability Calculations Between Specifications

In quality assurance, customer drawings specify a Lower Specification Limit (LSL) and an Upper Specification Limit (USL). To compute expected scrap or rework:

1. Calculate Upper Z-Score ($Z_{USL}$): $$Z_{USL} = \frac{USL - \mu}{\sigma}$$
2. Calculate Lower Z-Score ($Z_{LSL}$): $$Z_{LSL} = \frac{LSL - \mu}{\sigma}$$
3. Calculate Conforming Proportion ($P_{\text{good}}$): $$P(LSL \le X \le USL) = \Phi(Z_{USL}) - \Phi(Z_{LSL})$$
4. Calculate Upper Tail Fraction (Scrap above USL): $$P(X > USL) = 1 - \Phi(Z_{USL})$$
5. Calculate Lower Tail Fraction (Scrap below LSL): $$P(X < LSL) = \Phi(Z_{LSL})$$
6. Convert to Parts Per Million (PPM): $$\text{PPM} = P_{\text{defect}} \times 1,000,000$$

---

Comprehensive Worked Metrology Example: Uncentered Precision Dowel Pins

Scenario

A precision Swiss-machining cell manufactures hardened ground steel dowel pins. The engineering drawing specifies outer diameter as:

$$\text{Outer Diameter} = 12.000 \pm 0.015\text{ mm}$$

• Lower Specification Limit ($LSL$) = $12.000 - 0.015 = 11.985\text{ mm}$
• Upper Specification Limit ($USL$) = $12.000 + 0.015 = 12.015\text{ mm}$

A capability study of 100 consecutive pieces demonstrates that the process is in statistical control and normally distributed, with:

• Process Mean: $\mu = 12.003\text{ mm}$ (process is running slightly oversized by $0.003\text{ mm}$)
• Process Standard Deviation: $\sigma = 0.004\text{ mm}$

The quality technician must calculate the expected scrap fraction, conforming yield, and defect rates in PPM.

Step 1: Compute Upper and Lower Z-Scores

$$Z_{USL} = \frac{USL - \mu}{\sigma} = \frac{12.015 - 12.003}{0.004} = \frac{+0.012}{0.004} = +3.00$$ $$Z_{LSL} = \frac{LSL - \mu}{\sigma} = \frac{11.985 - 12.003}{0.004} = \frac{-0.018}{0.004} = -4.50$$

Step 2: Determine Tail Areas from Normal Tables

• Upper Tail ($X > USL$): For $Z_{USL} = +3.00$, $\Phi(3.00) = 0.99865$.

  $$P(X > 12.015) = 1 - \Phi(3.00) = 1 - 0.99865 = 0.00135$$

  This corresponds to oversized scrap (or reworkable pins):

  $$\text{PPM}_{\text{Upper}} = 0.00135 \times 1,000,000 = \mathbf{1,350\text{ PPM}}$$

• Lower Tail ($X < LSL$): For $Z_{LSL} = -4.50$, using the symmetry rule:

  $$P(X < 11.985) = \Phi(-4.50) = 1 - \Phi(4.50)$$

  From extended standard normal tables, $\Phi(4.50) \approx 0.9999966$.

  $$P(X < 11.985) = 1 - 0.9999966 = 0.0000034$$

  This corresponds to undersized scrap (unrecoverable material loss):

  $$\text{PPM}_{\text{Lower}} = 0.0000034 \times 1,000,000 = \mathbf{3.4\text{ PPM}}$$

Step 3: Total Nonconformance Rate and Yield

$$\text{Total Nonconformance Proportion } (P_{\text{def}}) = 0.00135 + 0.0000034 \approx 0.0013534$$ $$\text{Total PPM} = 1,350 + 3.4 = \mathbf{1,353.4\text{ PPM}}$$ $$\text{Conforming Process Yield} = (1 - 0.0013534) \times 100\% = \mathbf{99.865\%}$$

Metrology Insight: Because the process mean was shifted high by only $0.003\text{ mm}$ (less than one standard deviation!), over $99.7\%$ of all defective parts occur at the high side ($1,350\text{ PPM}$ high vs. $3.4\text{ PPM}$ low). Centering the tool offset back to $12.000\text{ mm}$ would equalize both tails at $Z = \pm 3.75$, dropping total scrap from $1,353.4\text{ PPM}$ to under $0.2\text{ PPM}$!

---

The Central Limit Theorem (CLT)

The Central Limit Theorem (CLT) is the mathematical engine powering statistical process control (SPC). It explains why control charts work across thousands of diverse manufacturing operations where individual part dimensions are not normally distributed.

Formal Statement of the Central Limit Theorem

Given a population of any arbitrary distributional shape (uniform, exponential, triangular, or skewed) with a finite mean $\mu$ and finite standard deviation $\sigma$:

1. As the sample size $n$ increases, the distribution of the sample means ($\bar{x}$) approaches a normal distribution.
2. The mean of the distribution of sample means equals the population mean: $$\mu_{\bar{x}} = \mu$$
3. The standard deviation of the distribution of sample means (the standard error of the mean) equals the population standard deviation divided by the square root of the sample size: $$\sigma_{\bar{x}} = \frac{\sigma}{\sqrt{n}}$$

Practical Rules of Thumb for Quality Technicians

• Sample Size ($n \ge 30$): If the underlying parent population is heavily skewed or multimodal, a sample size of $n \ge 30$ ensures the sample mean distribution is virtually normal.
• Moderate Sample Size ($n = 4$ or $5$): If the underlying population is reasonably symmetrical and unimodal, subgroup sizes of just $n = 4$ or $n = 5$ (the standard rational subgroup size for $\bar{X}-R$ charts) produce sample means that conform closely to the normal distribution.
• Standard Error Reduction: If a technician increases subgroup size from $n = 4$ to $n = 16$, the standard error of the mean cuts in half ($\sqrt{16} = 4$ vs. $\sqrt{4} = 2$), tightening control limits and accelerating detection of small process shifts.

---

Real-World Shop Scenarios & Common Exam Traps

Common Exam Traps for CQT Candidates

• Exam Trap 1: Forgetting to Divide by $\sqrt{n}$ for Sample Means: When an exam problem asks for the probability that an individual part exceeds a specification, use $Z = \frac{X - \mu}{\sigma}$. But when the question asks for the probability that the average of a sample subgroup ($\bar{x}$) exceeds a limit, you MUST use the standard error: $Z = \frac{\bar{x} - \mu}{\sigma / \sqrt{n}}$. Using $\sigma$ instead of $\sigma/\sqrt{n}$ for sample averages is the single most common mathematical error on the CQT exam.
• Exam Trap 2: Subtracting from 1.0 for Upper Tails: Standard normal tables provide cumulative area from the left ($-\infty$ to $+Z$). If you need the scrap fraction exceeding an Upper Specification Limit, remember to subtract the table value from 1.0 ($P(X > USL) = 1 - \Phi(Z)$).
• Exam Trap 3: Assuming All Physical Features Follow Normal Distributions: Runout, hole position error (true position), surface roughness ($R_a$), and flatness cannot be negative; their distributions are bounded at zero and naturally follow Chi-square, Rayleigh, or Weibull distributions. Technicians must never blindly apply normal Z-score tables to geometric tolerances without testing for normality.

---

Confidence Levels and Confidence Limits

Every statistic a technician calculates from a sample — a mean, a standard deviation, a $C_{pk}$, a gage bias, a percent nonconforming — is an estimate. Repeat the sampling and the number changes. A confidence interval puts a defensible range around that estimate, and the Body of Knowledge expects a technician to explain confidence levels and limits in various situations.

Point Estimate Versus Interval Estimate

• A point estimate is a single value: the sample mean $\bar{x} = 1.2504"$.
• An interval estimate is a range with a stated confidence level: "with 95 percent confidence, the true process mean lies between $1.2501"$ and $1.2507"$."
• The endpoints of that range are the confidence limits; the range between them is the confidence interval.

The confidence level (commonly 90, 95, or 99 percent) is the long-run proportion of such intervals that would contain the true parameter if the sampling were repeated many times. Its complement is $\alpha$: a 95 percent confidence level corresponds to $\alpha = 0.05$, which is the same $\alpha$ used as the significance level in the normality tests of Section 7.1 and the bias $t$-test of Section 9.3.

The Interval for a Process Mean

When the process standard deviation $\sigma$ is known — the usual situation once a control chart has been running and $\sigma$ has been estimated from $\bar{R}/d_2$ — the interval uses the standard error of the mean from the Central Limit Theorem:

$$\bar{x} \pm z_{\alpha/2} \frac{\sigma}{\sqrt{n}}$$

with $z_{\alpha/2} = 1.645$ for 90 percent, $1.96$ for 95 percent, and $2.575$ for 99 percent two-sided confidence. When $\sigma$ is unknown and the sample is small, the sample standard deviation $s$ replaces $\sigma$ and the $t$ distribution replaces $z$:

$$\bar{x} \pm t_{\alpha/2,\ n-1} \frac{s}{\sqrt{n}}$$

The $t$ value is always larger than the corresponding $z$ value, and it grows as the sample shrinks — which is simply the arithmetic saying that a small sample buys less certainty.

Worked Example

A technician measures $n = 25$ shafts from a stable process and obtains $\bar{x} = 1.2504"$ with a known process $\sigma = 0.0008"$. The 95 percent confidence interval for the process mean is:

$$1.2504 \pm 1.96 \times \frac{0.0008}{\sqrt{25}} = 1.2504 \pm 1.96 \times 0.00016 = 1.2504 \pm 0.00031$$

giving confidence limits of $1.25009"$ and $1.25071"$. If the print nominal is $1.2500"$, the interval does not contain the nominal, so the evidence supports a real centering offset rather than sampling noise, and a tool-offset adjustment is justified.

Two-Sided Versus One-Sided Limits

• Use two-sided limits when the question is "where is the parameter?" — a process mean, a gage bias.
• Use a one-sided confidence limit when only one direction matters: a lower confidence limit on a weld's break strength, an upper confidence limit on a contamination level, or a lower confidence bound on $C_{pk}$. A one-sided 95 percent limit uses $z = 1.645$, not $1.96$, because the whole $5\%$ risk sits in one tail.

What Changes the Width

Interpreting the Interval Correctly

Where a Technician Actually Meets Confidence Limits