3.4 Histograms & Scatter Diagrams

3.4 Histograms & Scatter Diagrams

While check sheets and Pareto charts excel at capturing and prioritizing discrete attribute defects (such as counts of cracks, burrs, or leaks), quality technicians constantly measure continuous variable data—diameters, lengths, tensile strengths, surface finishes, hardness values, and cycle times. To understand the underlying process behavior, the technician relies on Histograms to evaluate the shape and spread of a single variable, and Scatter Diagrams to explore the relationship between paired variables.

---

Histograms: Architecture and Purpose

A histogram is a vertical column chart that displays the frequency distribution of continuous measurement data grouped into adjacent, non-overlapping intervals called classes or bins.

Histogram vs. Standard Bar Chart

Quality technicians must clearly distinguish between a bar chart and a histogram:

• Bar Chart: Displays categorical (attribute) data (e.g., machine names, defect types, shift numbers). The bars are separated by gaps to emphasize that each category is discrete and independent.
• Histogram: Displays continuous variable data measured along an unbroken scale (e.g., millimeters, pounds per square inch, degrees Celsius). The bars touch one another without gaps, visually communicating that the underlying measurement continuum is uninterrupted.

+-----------------------------------------------------------------------------------+
|                         HISTOGRAM ARCHITECTURE                                    |
+-----------------------------------------------------------------------------------+
|  Frequency (Counts)                                                               |
|   25 |                  +-----+                                                   |
|   20 |            +-----+     |                                                   |
|   15 |            |     |     +-----+                                             |
|   10 |      +-----+     |     |     +-----+                                       |
|    5 |+-----+     |     |     |     |     +-----+                                 |
|    0 +------+-----+-----+-----+-----+-----+-----+-----> Measured Variable (X)     |
|      [10.0-10.2)[10.2-10.4)[10.4-10.6)...           (Class Intervals / Bins)     |
|        Bars touch to signify a continuous variable scale!                         |
+-----------------------------------------------------------------------------------+

---

Constructing a Histogram: Rules, Formulas, and Class Boundaries

Constructing an accurate histogram requires a standardized 5-step mathematical procedure:

Step 1: Collect Data and Determine Sample Size ($n$)

Gather a representative sample of continuous variable measurements from a stable process. For reliable distribution analysis, a minimum of $n \ge 50$ observations is recommended; $n \ge 100$ is preferred.

Step 2: Determine the Number of Classes ($k$)

Selecting too few classes over-aggregates data into an uninformative block; selecting too many creates a jagged, comb-like distribution with numerous empty bins. Two standard mathematical conventions are used:

1. The Square-Root Rule: $k = \sqrt{n}$ (For example, if $n = 64$, $k = \sqrt{64} = 8$ classes; if $n = 100$, $k = \sqrt{100} = 10$ classes.)

2. Sturges' Rule: $k = 1 + 3.322 \log_{10}(n)$ (For $n = 100$: $k = 1 + 3.322(2) = 7.644 \approx 8$ classes.)

Step 3: Calculate the Range ($R$)

Subtract the smallest sample value ($X_{\min}$) from the largest ($X_{\max}$): $R = X_{\max} - X_{\min}$

Step 4: Calculate the Class Width ($H$ or $w$)

Divide the range $R$ by the number of classes $k$: $H = \frac{R}{k}$ Rounding Rule: Round $H$ upward to the nearest convenient measurement unit or gage resolution to ensure that the $k$ classes encompass the entire sample range.

Step 5: Establish Class Boundaries

To prevent ambiguity regarding which bin an observation belongs to, class boundaries should be established at one extra decimal place of precision (or at half-units of the gage measurement resolution). For example, if parts are measured to the nearest $0.001,\text{in}$, class boundaries should end in $0.0005,\text{in}$ (e.g., $1.0005,\text{in}$ to $1.0055,\text{in}$). This guarantees that no physical measurement ever falls directly on a boundary line.

---

Interpreting Histogram Distribution Shapes on the Shop Floor

A quality technician diagnoses process health by reading the visual morphology of the histogram. Standard shapes serve as immediate diagnostic signatures:

+-----------------------------------------------------------------------------------+
|                       HISTOGRAM SHAPES & PROCESS DIAGNOSTICS                      |
+-----------------------------------------------------------------------------------+
| 1. NORMAL (BELL-SHAPED)      2. SKEWED (RIGHT/POSITIVE)    3. BIMODAL (TWO PEAKS) |
|          _--_                        _--_                          _--_     _--_  |
|        _-    -_                    _-    -_                      _-    -_ _-    -_|
|       /        \                  /        \--__                /        V        \|
|     Stable random noise          Bounded: Runout, Flatness    Mixed lots, 2 tools |
|                                                                                   |
| 4. TRUNCATED (CLIFF-LIKE)    5. COMB (SERRATED)            6. ISOLATED ISLAND     |
|              |                                                     _--_           |
|           _--|                   _   _   _   _                   _-    -_     _   |
|         _-   |                  | |_| |_| |_| |                 /        \   |_|  |
|     100% Sorting/Screening     Gage rounding, bin bias         Tool chip, outlier |
+-----------------------------------------------------------------------------------+

1. Normal (Bell-Shaped / Gaussian) Distribution

• Visual Shape: Symmetrical, unimodal curve where the mean, median, and mode coincide at the center, with frequencies tapering smoothly toward both tails.
• Diagnostic Meaning: The process is stable, in statistical control, and influenced solely by inherently random, common-cause variation.

2. Positively Skewed (Right-Skewed) Distribution

• Visual Shape: The peak is shifted to the left, with a long, drawn-out tail extending toward higher positive values. In a right-skewed distribution, the mathematical relationship of central tendency is: $\text{Mean} > \text{Median} > \text{Mode}$.
• Diagnostic Meaning: Naturally occurs when the parameter has a rigid physical or geometric lower boundary at zero that cannot be crossed. Typical examples include concentricity, perpendicularity, surface roughness ($R_a$), total indicator reading (TIR), and hole location true position.

3. Negatively Skewed (Left-Skewed) Distribution

• Visual Shape: The peak is shifted to the right, with a long tail extending toward smaller or lower values. The mathematical relationship is: $\text{Mean} < \text{Median} < \text{Mode}$.
• Diagnostic Meaning: Frequently encountered where an upper physical ceiling exists (such as chemical purity percentages bounded at 100%, or process yield metrics).

4. Bimodal (Double-Peaked) Distribution

• Visual Shape: Two distinct frequency peaks separated by a trough.
• Diagnostic Meaning: The classic signature of a mixed population. Two distinct distributions have been blended into a single dataset. Causes include combining parts produced by two different CNC machines, two mold cavities, two different operators, day shift versus night shift, or two distinct raw material heats. The immediate corrective action is stratification—separating the data by source to analyze each distribution independently.

5. Multimodal Distribution

• Visual Shape: Three or more distinct peaks across the measurement spectrum.
• Diagnostic Meaning: Indicates multiple mixed input streams, such as blending parts from an 8-cavity injection mold or a multi-spindle screw machine.

6. Truncated (Cliff-Like / Cut-Off) Distribution

• Visual Shape: A distribution that appears normal across one portion but drops off abruptly to zero on one or both edges, creating a sheer vertical cliff precisely at or near the engineering specification limits (USL or LSL).
• Diagnostic Meaning (ASQ Exam Favorite): This is the definitive indicator of 100% inspection, sorting, or screening. The manufacturing process is inherently non-capable and is producing nonconforming product; however, an inspector or automated gage is sorting the parts and discarding nonconforming units prior to customer shipment. Sorting is expensive, error-prone, and masks process instability.

7. Comb (Serrated / Alternating Tooth) Distribution

• Visual Shape: Alternating tall and short bars resembling the teeth of a comb.
• Diagnostic Meaning: Indicates measurement errors, including gage resolution that is too coarse for the class width, operator rounding bias (e.g., inspectors rounding readings to even numbers or multiples of five), or poorly chosen class boundaries where raw values cluster unevenly.

8. Isolated Peak (Island) Distribution

• Visual Shape: A small, separate cluster of bars isolated away from the main distribution body.
• Diagnostic Meaning: Represents an intermittent special-cause event, such as an inclusion in raw bar stock, a broken cutting tool tip, an incorrect gage zeroing, or setup parts inadvertently mixed into production boxes.

Scatter Diagrams: Bivariate Correlation and Relationship Analysis

A scatter diagram (or scatter plot) is a graphical technique designed to evaluate the potential relationship between two paired, continuous numerical variables plotted on Cartesian coordinates:

• Independent Variable ($X$): Plotted along the horizontal X-axis. This represents the suspect input, process parameter, machine setting, or environmental condition (e.g., cutting tool temperature, furnace dwell time, spindle RPM).
• Dependent Variable ($Y$): Plotted along the vertical Y-axis. This represents the resulting quality response, performance characteristic, or defect dimension (e.g., surface roughness, weld penetration depth, part shrinkage).

+-----------------------------------------------------------------------------------+
|                         SCATTER DIAGRAM CORRELATION PATTERNS                      |
+-----------------------------------------------------------------------------------+
| STRONG POSITIVE (r ~ +1)     STRONG NEGATIVE (r ~ -1)      NO CORRELATION (r ~ 0) |
| Y                            Y                             Y                      |
|  |        *                   |  *                          |    *   *   *        |
|  |      *                     |    *                        |  *   *   *   *      |
|  |    *                       |      *                      |    *   *   *        |
|  |  *                         |        *                    |  *   *   *   *      |
|  +------------ X              +------------ X               +------------ X       |
|  As X rises, Y rises          As X rises, Y falls           Points form round cloud|
|                                                                                   |
| CURVILINEAR / NON-LINEAR RELATIONSHIP (r ~ 0, but strong deterministic link!)     |
| Y |          *   *                                                                |
|   |       *         *         * Parabolic response: Optimum operating point       |
|   |     *             *         exists at the apex! Linear r is misleading!       |
|   +----------------------- X                                                      |
+-----------------------------------------------------------------------------------+

Interpreting Correlation Patterns

Technicians evaluate scatter patterns to assess the strength and direction of association:

1. Strong Positive Linear Correlation ($r \approx +1$): As $X$ increases, $Y$ increases proportionally in a narrow, well-defined band. (Example: Sintering temperature vs. tensile yield strength.)
2. Strong Negative Linear Correlation ($r \approx -1$): As $X$ increases, $Y$ decreases proportionally in a narrow band. (Example: Grinding wheel speed vs. surface roughness $R_a$.)
3. Moderate / Weak Correlation: Data points exhibit a general directional tilt but form a wide, diffuse cloud, indicating that while $X$ influences $Y$, other unmeasured sources of process variation are simultaneously active.
4. Zero / No Correlation ($r \approx 0$): Data points form a circular or formless cloud. Changes in $X$ have no linear association with changes in $Y$.
5. Curvilinear (Non-Linear) Correlation: Data points follow a distinct curved profile (such as a parabola, U-shape, or asymptotic curve). Quality technicians must recognize that the Pearson correlation coefficient ($r$) measures only linear association; a scatter diagram can display an absolute deterministic parabolic relationship while yielding a linear correlation coefficient of $r = 0.0$!

---

The Critical Quality Rule: Correlation vs. Causation

The most pervasive trap on the ASQ CQT exam is assuming that statistical correlation proves physical causation.

[!CAUTION] Correlation Does NOT Equal Causation: A high correlation coefficient ($r = 0.94$) between an input variable $X$ and an output dimension $Y$ demonstrates that the two variables move together; it does not prove that $X$ causes $Y$. Both variables may be co-dependent on an unmeasured third factor—known as a confounding or lurking variable.

Real Shop-Floor Example of a Confounding Variable

A quality technician at an automotive stamping plant plots ambient plant relative humidity ($X$) against stamping dimensional springback error ($Y$) and discovers a strong positive correlation ($r = 0.88$). The technician hastily concludes that humid air causes sheet metal springback.

A rigorous metallurgical investigation proves that relative humidity has zero physical effect on cold-rolled steel. The true root cause was a confounding variable: during humid summer months, the shop-floor ambient temperature rose by 25°F. The heat lowered the viscosity of the hydraulic stamping oil, reducing die clamping tonnage by 14%, which caused sheet metal springback. The correlation with humidity was purely coincidental, driven by seasonal ambient temperature.

To establish definitive causation, the quality technician must move beyond observational scatter plots and execute Design of Experiments (DOE), where suspect factors are deliberately altered under controlled conditions while holding all other process variables constant.