11.1 Statistics Fundamentals & Normal Distribution

11.1 Statistics Fundamentals & Normal Distribution

Populations vs. Samples: Parameters vs. Statistics

In quality inspection and manufacturing metrology, understanding the distinction between a population and a sample is the bedrock of all data-driven decisions. An inspector rarely has the luxury or economic justification to measure every single part produced. Instead, decisions regarding millions of dollars of inventory are made by measuring representative samples.

• Population: The complete collection of all possible items, parts, features, or observations under consideration. A population can be finite (e.g., a discrete purchase order lot of 5,000 heat-treated fasteners) or infinite (e.g., an ongoing continuous stamping operation running indefinitely). Characteristics that describe an entire population are called parameters and are traditionally denoted by Greek letters.
• Sample: A finite subset of observations selected from the population using a defined sampling protocol (such as random sampling or rational subgrouping). Characteristics computed from sample data are called statistics and are denoted by Roman letters.

+-----------------------------------------------------------------------------+
|                        POPULATION VS. SAMPLE INFERENCE                      |
+-----------------------------------------------------------------------------+
|  POPULATION (Lot Size N = 10,000)                                           |
|  - True Mean: μ (Unknown Parameter)                                         |
|  - True Spread: σ (Unknown Parameter)                                       |
|                                                                             |
|             |  [Representative Random Sampling (n = 30)]                    |
|             v                                                               |
|                                                                             |
|  SAMPLE (Sample Size n = 30)                                                |
|  - Measured Sample Mean: X̄ = 1.2504 in (Sample Statistic)                  |
|  - Measured Sample Spread: s = 0.0008 in (Sample Statistic)                 |
|                                                                             |
|             |  [Statistical Estimation & Hypothesis Testing]                |
|             v                                                               |
|                                                                             |
|  INFERENCE: Estimate with 95% Confidence that μ lies in [1.2501, 1.2507]   |
+-----------------------------------------------------------------------------+

The Quality Inspector's Dilemma

Why not simply inspect the entire population? In modern high-reliability manufacturing, 100% inspection is neither economically feasible nor metrologically foolproof:

1. Destructive Testing: For tensile strength testing, metallurgical case-depth etching, weld micro-sectioning, or burst pressure testing, 100% inspection would destroy the entire production output, leaving zero product to ship.
2. Human Fatigue and Inspection Error: Studies in industrial psychology demonstrate that 100% manual visual or dimensional inspection is typically only 80% to 85% effective. Monotony, sensory fatigue, and cognitive lapses lead inspectors to inadvertently pass defective parts or reject good parts.
3. Economic Cost: The labor and equipment wear required to measure 100% of non-critical characteristics exponentially increases cost without adding customer value.

Consequently, the quality inspector relies on sample statistics to draw valid, mathematically defensible inferences about population parameters.

---

Measures of Central Tendency

A measure of central tendency identifies the central location or balance point of a dataset. When an inspector measures a series of parts from a CNC lathe, three primary statistics describe where the dimensions cluster: the Mean, the Median, and the Mode.

1. Arithmetic Mean ($\bar{X}$)

The arithmetic mean (commonly referred to as the average) is the sum of all individual observations divided by the sample size $n$: $\bar{X} = \frac{\sum_{i=1}^{n} X_i}{n} = \frac{X_1 + X_2 + \dots + X_n}{n}$

• Metrological Characteristics: The mean is the mathematical balance point of the data. Its primary advantage is that it incorporates every single numerical observation in the sample. Its primary drawback is its high sensitivity to extreme outliers or measurement errors.
• Example: Consider five pin diameters (in inches): $0.501, 0.502, 0.501, 0.502, 0.525$ (where the last part was improperly loaded into the fixture). The mean is $(0.501 + 0.502 + 0.501 + 0.502 + 0.525) / 5 = 0.5062\text{ in}$. The single flawed reading has drastically pulled the average upward, misrepresenting the true center of the conforming production stream.

2. Median ($\tilde{X}$)

The median is the middle value of a dataset when the observations are arranged in ascending or descending numerical order. It divides the distribution into two equal halves (50% above and 50% below).

• If the sample size $n$ is odd, the median is the single middle observation at position $(n + 1) / 2$.
• If the sample size $n$ is even, the median is the arithmetic mean of the two central observations at positions $n/2$ and $(n/2) + 1$.
• Resistance to Outliers: The median is a robust statistic. In the pin diameter dataset above ($0.501, 0.501, 0.502, 0.502, 0.525$), the ranked middle value is $0.502\text{ in}$. The gross outlier ($0.525\text{ in}$) has zero effect on the median, making the median a superior indicator of typical performance when data is skewed or corrupted by setup glitches.

3. Mode

The mode is the specific numerical value that occurs with the highest frequency in a dataset.

• A dataset can have one mode (unimodal), two modes (bimodal), multiple modes (multimodal), or no mode if all values occur with equal frequency.
• Shop Floor Significance: Bimodal distributions are an immediate diagnostic red flag for quality inspectors. If a histogram of a single machined diameter exhibits two distinct peaks, it almost invariably indicates that parts from two different streams have been mixed together—such as parts from two different machine spindles, two mold cavities, two parallel fixtures, or two different operator shifts.

---

Measures of Dispersion and Variation

Central tendency describes where a process is centered, but in precision manufacturing, dispersion (spread or variation) is equally critical. Two processes can share an identical mean of $1.000\text{ in}$, yet one process may produce parts tightly grouped between $0.999$ and $1.001\text{ in}$, while the other produces parts scattered dangerously from $0.990$ to $1.010\text{ in}$.

1. Range ($R$)

The sample range is the simplest metric of dispersion, defined as the mathematical difference between the maximum and minimum observed values in a sample: $R = X_{\text{max}} - X_{\text{min}}$

• Advantages: Extremely simple and rapid to calculate on the shop floor without digital processors; highly intuitive for machine operators.
• Limitations: The range utilizes only the two most extreme values in the sample and completely ignores all intermediate data. For large sample sizes ($n > 10$), the range becomes statistically inefficient and erratic because the probability of capturing extreme outliers increases with $n$. Therefore, range is strictly reserved for small subgroup sizes ($n = 2$ to $6$).

2. Sample Variance ($s^2$) and Degrees of Freedom ($n - 1$)

The sample variance measures the average squared deviation of each observation from the sample mean: $s^2 = \frac{\sum_{i=1}^{n} (X_i - \bar{X})^2}{n - 1}$

Why Divide by $n - 1$? (Bessel's Correction and Degrees of Freedom)

A universal question tested on the ASQ CQI examination is why the denominator of the sample variance formula contains $n - 1$ rather than $n$.

• Degrees of Freedom ($df$): Degrees of freedom represent the number of unconstrained, independent pieces of data available to estimate a parameter. When calculating sample variance, the true population mean $\mu$ is unknown, so the sample mean $\bar{X}$ must be used in its place.
• The Constraint: By mathematical definition, the sum of deviations about the arithmetic mean is always exactly zero: $\sum (X_i - \bar{X}) = 0$. Consequently, once $n - 1$ deviations are known, the final ($n$-th) deviation is mathematically fixed and cannot vary freely. One degree of freedom has been consumed to estimate $\bar{X}$.
• Bessel's Correction: If we divided by $n$, the resulting value would systematically underestimate the true population variance $\sigma^2$ because sample observations naturally cluster closer to their own sample mean $\bar{X}$ than to the population mean $\mu$. Dividing by $n - 1$ mathematically inflates the result just enough to eliminate this downward bias, making $s^2$ an unbiased estimator of $\sigma^2$.

3. Standard Deviation ($s$ vs. $\sigma$)

The variance expresses dispersion in squared units (e.g., $\text{in}^2$ or $\text{mm}^2$), which is physically unintuitive for dimensional inspection. Taking the square root of the variance returns the metric to the original linear units of measurement:

• Sample Standard Deviation ($s$): $s = \sqrt{s^2} = \sqrt{\frac{\sum_{i=1}^{n} (X_i - \bar{X})^2}{n - 1}}$
• Population Standard Deviation ($\sigma$): $\sigma = \sqrt{\frac{\sum_{i=1}^{N} (X_i - \mu)^2}{N}}$

Step-by-Step Calculation Example (n = 5 Shaft Diameters in inches):
Data: 0.500, 0.502, 0.498, 0.501, 0.499

Step 1: Calculate Mean (X̄):
  Sum = 0.500 + 0.502 + 0.498 + 0.501 + 0.499 = 2.500
  X̄ = 2.500 / 5 = 0.5000 in

Step 2: Calculate Deviations (X - X̄) and Squared Deviations (X - X̄)^2:
  (0.500 - 0.5000)^2 = ( 0.0000)^2 = 0.00000000
  (0.502 - 0.5000)^2 = (+0.0020)^2 = 0.00000400
  (0.498 - 0.5000)^2 = (-0.0020)^2 = 0.00000400
  (0.501 - 0.5000)^2 = (+0.0010)^2 = 0.00000100
  (0.499 - 0.5000)^2 = (-0.0010)^2 = 0.00000100
  Sum of Squared Deviations = 0.00001000 in^2

Step 3: Divide by Degrees of Freedom (n - 1 = 4):
  s^2 = 0.00001000 / 4 = 0.00000250 in^2

Step 4: Take the Square Root:
  s = sqrt(0.00000250) = 0.00158 in

---

The Normal (Gaussian) Distribution

The Normal Distribution (frequently termed the Gaussian distribution or bell curve) is the cornerstone of statistical process control, process capability analysis, and dimensional tolerancing.

Mathematical Properties of the Normal Curve

1. Symmetry: The curve is perfectly symmetrical about the central mean $\mu$. The left half is a precise mirror image of the right half.
2. Equivalence of Central Tendency: At the center of a true normal distribution, the Mean, Median, and Mode are mathematically identical ($\text{Mean} = \text{Median} = \text{Mode}$).
3. Asymptotic Tails: The tails of the curve extend infinitely in both directions ($-\infty$ to $+\infty$), approaching the horizontal axis asymptotically without ever touching it. In practical manufacturing, however, the area beyond $\pm 3\sigma$ or $\pm 4\sigma$ represents an infinitesimally small fraction of production.
4. Total Area: The total area under the probability density curve is strictly equal to $1.0000$ (or $100%$).
5. Inflection Points: The curve transitions from concave downward to concave upward at exactly one standard deviation from the mean ($\mu \pm 1\sigma$).

+-----------------------------------------------------------------------------+
|                        THE GAUSSIAN NORMAL DISTRIBUTION                     |
+-----------------------------------------------------------------------------+
|                                    |                                        |
|                                  .---.                                      |
|                                 /  |  \                                     |
|                                /   |   \                                    |
|                              .'    |    '.                                  |
|                            .'      |      '.                                |
|                          .'        |        '.                              |
|                       .-'          |          '-.                           |
|                  .---'             |             '---.                      |
|      .----------'                  |                  '----------.          |
| ----+--------------+---------------+---------------+--------------+------- |
|   μ - 3σ         μ - 2σ          μ - 1σ          μ + 1σ         μ + 2σ      μ + 3σ 
|                    |<--- 68.26% --->|                                       |
|             |<------------ 95.44% ------------>|                            |
|      |<------------------- 99.73% ------------------->|                     |
+-----------------------------------------------------------------------------+

The Central Limit Theorem (CLT)

Inspectors frequently ask: "What if my raw parts are not normally distributed?" The Central Limit Theorem resolves this dilemma. The CLT proves that regardless of the underlying population distribution (even if it is uniform, skewed, or bimodal), the distribution of sample means ($\bar{X}$) will approach a normal distribution as the sample size $n$ increases (typically $n \ge 4$ or $5$ for moderately symmetric distributions, and $n \ge 30$ for severely non-normal distributions). Furthermore, the standard deviation of sample averages—known as the Standard Error of the Mean—is narrower than the individual part spread: $\sigma_{\bar{X}} = \frac{\sigma}{\sqrt{n}}$ This foundational theorem explains why control charts for averages ($\bar{X}$ charts) work reliably across virtually all manufacturing processes.

---

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

For any dataset that follows a normal distribution, the Empirical Rule provides exact, predictable percentages of product that fall within integer standard deviation intervals from the mean:

Shop Floor Metrology Takeaway

The fact that $99.73%$ of all natural, common-cause variation falls within $\pm 3\sigma$ is the precise mathematical rationale why Dr. Walter Shewhart established 3-sigma control limits on statistical process control charts. If a process is in statistical control, finding a part outside $\pm 3\sigma$ by pure random chance occurs only 27 times out of 10,000 observations ($0.27%$). Any point outside 3-sigma is therefore treated as positive statistical evidence of an assignable special cause.

---

Standard Normal Distribution and Z-Score Transformations

Manufacturing processes have different engineering units: inches, millimeters, foot-pounds, micro-inches, or Rockwell C hardness. To calculate defect rates and probabilities across disparate operations, statisticians convert raw dimensions into a unitless standard known as the Standard Normal Distribution ($Z \sim N(0, 1)$), which has a standardized mean of $\mu = 0$ and standard deviation of $\sigma = 1$.

The $Z$-Score Formula

The $Z$-score (standard normal deviate) defines the exact number of standard deviations an individual measurement $X$ lies above or below the mean: $Z = \frac{X - \mu}{\sigma} \quad \text{(Population)} \qquad \text{or} \qquad Z = \frac{X - \bar{X}}{s} \quad \text{(Sample Estimate)}$

• A positive $Z$-score ($Z > 0$) indicates the dimension is above the mean.
• A negative $Z$-score ($Z < 0$) indicates the dimension is below the mean.
• $Z = 0$ indicates the dimension falls exactly on the process mean.

Practical Step-by-Step Defect Calculation

A quality inspector evaluates a CNC turned hydraulic piston whose outer diameter is specified on the drawing as $\varnothing 1.250 \pm 0.004\text{ in}$. This establishes:

• $\text{Upper Specification Limit (USL)} = 1.254\text{ in}$
• $\text{Lower Specification Limit (LSL)} = 1.246\text{ in}$

Quality auditing demonstrates that the process is normally distributed with a mean of $\mu = 1.251\text{ in}$ (shifted $0.001\text{ in}$ high) and a standard deviation of $\sigma = 0.0015\text{ in}$. What is the expected scrap rate exceeding the $USL$?

Step 1: Calculate the Z-score for the USL:
  Z_USL = (USL - μ) / σ
  Z_USL = (1.254 - 1.251) / 0.0015 = +0.003 / 0.0015 = +2.00

Step 2: Look up Z = +2.00 in the Standard Normal Cumulative Probability Table:
  Cumulative Area from -∞ to Z = +2.00 is Φ(2.00) = 0.9772 (97.72% conforming below USL).

Step 3: Calculate the Exceedance Probability (Upper Tail Scrap):
  P(X > USL) = 1.0000 - 0.9772 = 0.0228 or 2.28%.

Step 4: Calculate the Z-score for the LSL:
  Z_LSL = (LSL - μ) / σ
  Z_LSL = (1.246 - 1.251) / 0.0015 = -0.005 / 0.0015 = -3.33

Step 5: Look up Z = -3.33 in the Standard Normal Table:
  P(X < LSL) = Φ(-3.33) ≈ 0.0004 or 0.04%.

Step 6: Total Expected Defect Rate:
  Total Rejection Rate = Upper Scrap + Lower Rework
  Total Rejection Rate = 2.28% + 0.04% = 2.32% (or 23,200 PPM).

---

Skewness and Kurtosis: Deviations from Normality

Real-world manufacturing data does not always form a pristine bell curve. Mechanical constraints, tooling mechanics, and inspector sorting introduce systematic distortions categorized as Skewness and Kurtosis.

Skewness (Asymmetry of the Distribution)

Skewness quantifies the degree of horizontal asymmetry of a distribution about its mean.

+-----------------------------------------------------------------------------+
|                        DISTRIBUTION SKEWNESS COMPARISON                     |
+-----------------------------------------------------------------------------+
|  POSITIVE (RIGHT) SKEW             NORMAL (SYMMETRIC)      NEGATIVE (LEFT) SKEW     |
|          Peak                             Peak                       Peak   |
|         /|                                |  \                       /|\    |
|        / |                               /|   \                     / | \   |
|       /  |                              / |    \                   /  |  \  |
|     .'   |                             /  |     \                 /   |   '.|
|  --'-----+----------->             ---+---+---+---      <--------+----+---'-|
|    Mode Median Mean                 Mean=Median=Mode          Mean Median Mode |
|    (Long tail to RIGHT)              (Skewness = 0)           (Long tail to LEFT) |
+-----------------------------------------------------------------------------+

1. Zero Skewness (Symmetric): A normal distribution has a skewness coefficient of $0$. The mean, median, and mode are equal.
2. Positive Skewness (Right-Skewed): The long tail of the distribution stretches toward the right (higher positive values). The bulk of the data is concentrated at lower values.
   • Mathematical Relationship: $\mathbf{\text{Mode} < \text{Median} < \text{Mean}}$. The mean is pulled farthest toward the extended right tail.
   • Shop Metrology Applications: Features bounded by a physical zero stop—such as surface roughness ($Ra$), true position GD&T callouts, total indicator reading (TIR) runout, perpendicularity, and flatness—are inherently positively skewed because a dimension cannot be less than zero but can experience sporadic upward spikes.
3. Negative Skewness (Left-Skewed): The long tail stretches toward the left (lower values). The bulk of data clusters near the upper limit.
   • Mathematical Relationship: $\mathbf{\text{Mean} < \text{Median} < \text{Mode}}$. The mean is pulled to the left.
   • Shop Metrology Applications: Common when parts undergo sorting inspection that screens out undersized scrap, or in grinding operations where operators intentionally target the high material limit ($MMC$) to leave stock for polishing.

Kurtosis (Peakedness and Tail Weight)

Kurtosis measures the "tailedness" or peak sharpness of a probability distribution relative to a normal distribution:

• Mesokurtic: A standard normal distribution with kurtosis equal to $3$ (or excess kurtosis of $0$).
• Leptokurtic (Excess Kurtosis $> 0$): Sharply peaked center with heavy, fat tails. Indicates a high concentration of parts near nominal, but with a higher-than-expected probability of extreme, sporadic outliers.
• Platykurtic (Excess Kurtosis $< 0$): Broad, flat-topped peak with light, thin tails. Often occurs when blending parts produced across several slightly different machine setups or when tool wear moves uniformly across a tolerance band.

---

Real Shop Inspection Scenarios & Common Exam Traps

• Real Shop Scenario — The Bimodal Hole Location Mystery: A quality inspector on a vertical machining center plots a histogram of 100 center-to-center hole distances. The specification is $4.000 \pm 0.005\text{ in}$. Instead of a bell curve, the histogram reveals two distinct humps centered at $3.998\text{ in}$ and $4.003\text{ in}$. The calculated mean is $4.0005\text{ in}$—suggesting nominal capability. However, the process is generating 4% scrap at both tails. Metrological Diagnosis: The inspector investigates the fixture setup and discovers a dual-clamping pallet where parts are machined two at a time. Pallet Station A is misaligned by $-0.002\text{ in}$, while Pallet Station B is misaligned by $+0.003\text{ in}$. The combined data mimics a wide distribution, but separating the data by pallet reveals two perfectly capable, tightly grouped normal distributions. Mixing independent streams masked the true cause.

• Exam Trap: Bessel's Correction ($n$ vs. $n - 1$): ASQ examination questions frequently ask candidates to calculate sample standard deviation from a small raw dataset ($n = 4$ or $5$). A standard distractor choice will calculate the variance dividing by $n$. Always verify: if the question refers to a sample, divide by $n - 1$; only divide by $N$ if the question explicitly states you have measured the entire population.

• Exam Trap: Direction of Skewness vs. Mean/Median Positioning: A classic exam question asks: "In a skewed-right distribution of hole perpendicularity measurements, what is the correct relative ordering of central tendency metrics?" Many candidates confuse positive skew with the physical peak. Remember: the name of the skew corresponds to the direction of the long, thin tail, and the mean is always pulled closest to the tail. Therefore, for right-skewed data: $\text{Mode} < \text{Median} < \text{Mean}$.

• Exam Trap: Bilateral Specification Defect Calculation: When calculating expected defect rates from $Z$-scores, candidates often calculate the defect percentage exceeding the Upper Specification Limit and forget to check the Lower Specification Limit (or vice versa). Unless the process is located at least $4\sigma$ or $5\sigma$ away from the opposite limit, defects can occur at both tails. Always check both $Z_{\text{USL}}$ and $Z_{\text{LSL}}$.