7.4 Graphical Data Exploration: Histograms, Box Plots, & Multi-Vari Charts

7.4 Graphical Data Exploration: Histograms, Box Plots, & Multi-Vari Charts

Core Principle: In the Analyze phase, continuous improvement practitioners must explore continuous variable data graphically before conducting complex hypothesis tests or calculating inferential statistics. Renowned statistician John Tukey emphasized that "numerical summaries focus on expected behavior, but graphical displays expose the unexpected." Histograms display the central tendency, dispersion, and morphology of continuous data distributions, diagnosing skewness, bimodal stream blending, and truncated screening artifacts. Box-and-Whisker Plots (Boxplots) present the Five-Number Summary and enforce Tukey's $1.5 \times \text{IQR}$ rule to mathematically flag process outliers. Complementing these, Multi-Vari Analysis passively decomposes overall process variance into three distinct physical families: Positional (Within-Unit), Cyclical (Piece-to-Piece), and Temporal (Time-to-Time). Finally, Scatter Plots assess bivariate associations while upholding the cardinal Six Sigma rule: correlation does NOT prove causation.

---

The Role of Graphical Data Exploration in Analyze

In DMAIC projects, practitioners collect hundreds or thousands of continuous baseline observations representing cycle times, dimensions, hydraulic pressures, or chemical concentrations. Proceeding immediately to calculate inferential statistics (such as two-sample $t$-tests, ANOVA, or linear regression) without visual exploration is a dangerous practice.

Graphical exploration accomplishes three mandatory diagnostic objectives:

1. Validates Distributional Assumptions: Parametric tests require approximately normal distributions. Graphical plots instantly reveal severe skewness, heavy tails, or extreme outliers that would invalidate parametric math.
2. Detects Hidden Operational Subgroups: Blended datasets (such as day shift vs. night shift, or Cavity 1 vs. Cavity 2) generate deceptive summary statistics. Visual plots expose multimodality and clustering.
3. Isolates the Dominant Source of Variation: Before attempting to control every input parameter, graphical tools isolate whether variance is concentrated within individual units, between successive units, or across operating shifts.

---

Histograms: Visualizing Continuous Data Distributions

A Histogram is the foundational graphical display for continuous variable data. It groups individual continuous measurements into contiguous, non-overlapping intervals called class intervals or bins, plotting the frequency of observations falling within each bin as a vertical bar.

                      Histogram Construction Architecture

     Frequency
        60 ┌───────────────────┬───┐
           │                   │   │
        45 │               ┌───┤   ├──┐
           │               │   │   │  │
        30 │           ┌───┤   │   │  ├──┐
           │           │   │   │   │  │  │
        15 │       ┌───┤   │   │   │  │  ├──┐
           │       │   │   │   │   │  │  │  │
         0 └─┴─────┴───┴───┴───┴───┴──┴──┴──┴─┴──▶ Continuous Scale (mm)
            10.0  10.2  10.4  10.6  10.8  11.0
            [-- Class Interval / Bin Width --]
            Note: Bars touch with NO gaps to signify an unbroken continuum.

Construction Guidelines

• Determining Number of Bins ($k$): Selecting too few bins over-smoothes the data, hiding bimodal peaks; selecting too many bins creates a jagged, empty display. Standard rules of thumb:
  • Square Root Rule: $k = \sqrt{n}$ (For $n = 100$, $k = 10$ bins).
  • Sturges' Rule: $k = 1 + 3.322 \log_{10} n$ (For $n = 100$, $k = 1 + 3.322(2) = 7.64 \approx 8$ bins).
• Determining Bin Width ($W$):

$W = \frac{\text{Maximum Observed Value} - \text{Minimum Observed Value}}{k}$

• Touching Bars Rule: Unlike discrete bar charts (which display deliberate gaps between bars to signify separate, unordered categories), histogram bars touch one another without gaps. This visually reinforces that the horizontal axis represents an unbroken mathematical continuum.

---

Distribution Morphologies & Diagnostic Clues

The visual profile of a histogram reveals critical insights regarding underlying process physics, operational constraints, and special-cause variation:

                      Histogram Distribution Morphologies

     1. NORMAL (SYMMETRIC)           2. POSITIVE (RIGHT) SKEW        3. NEGATIVE (LEFT) SKEW
             ┌───┐                               ┌───┐                           ┌───┐
          ┌──┤   ├──┐                         ┌──┤   │                        │   ├──┐
        ┌─┤  │   │  ├─┐                     ┌─┤  │   ├──┐                  ┌──┤   │  ├─┐
      ┌─┤ │  │   │  │ ├─┐                 ┌─┤ │  │   │  ├──┐            ┌──┤  │   │  │ ├─┐
    ──┴─┴─┴──┴───┴──┴─┴─┴──             ──┴─┴─┴──┴───┴──┴──┴─┴──      ──┴─┴──┴───┴──┴─┴─┴──
       Mean = Median = Mode                  Mode < Median < Mean          Mean < Median < Mode

     4. BIMODAL (MIXED STREAMS)      5. TRUNCATED (CLIFF-HANGER)     6. COMB / JAGGED
          ┌───┐     ┌───┐                         ┌───┐                ┌───┐   ┌───┐   ┌───┐
        ┌─┤   ├──┐┌─┤   ├──┐                      │   ├──┐             │   ├───┤   ├───┤   │
      ┌─┤ │   │  ├┤ │   │  ├─┐                    │   │  ├──┐          │   │   │   │   │   │
    ──┴─┴─┴───┴──┴┴─┴───┴──┴─┴──                ──┴───┴──┴──┴─┴──      ┴───┴───┴───┴───┴───┴
      Two mixed machines/shifts                 100% sorting/scraping     Measurement rounding error

---

Box-and-Whisker Plots (Boxplots) & Tukey's Outlier Fences

Developed by statistician John Tukey, the Box-and-Whisker Plot (boxplot) provides a compact graphical summary of process location, dispersion, skewness, and outliers based on the Five-Number Summary:

1. Minimum: The smallest observed value within the non-outlier fence boundaries.
2. First Quartile ($Q_1$): 25th percentile (separates lower 25% from upper 75%).
3. Median ($Q_2$): 50th percentile (central process line).
4. Third Quartile ($Q_3$): 75th percentile (separates lower 75% from upper 25%).
5. Maximum: The largest observed value within the non-outlier fence boundaries.

       Lower Outlier                                               Upper Outlier
          *                                                           *   *
   ───────┼─────────────┬──────────────┬──────────────┬───────────────┼──────────
         Fence         Min            Q1            Median           Q3         Fence
     [Q1 - 1.5*IQR]                    ┌──────────────┬──────────────┐      [Q3 + 1.5*IQR]
                       ├───────────────┤              │              ├──────┤
                                       └──────────────┴──────────────┘
                                       ◄──────────── IQR ────────────►
                                       (Middle 50% of Process Data)

Interquartile Range (IQR) & Tukey's Outlier Rule

The Interquartile Range (IQR) quantifies the spread of the central 50% of ordered observations:

$\text{IQR} = Q_3 - Q_1$

Tukey established objective mathematical fences to identify suspected process anomalies:

• Lower Inner Fence (LIF): $\text{LIF} = Q_1 - 1.5 \times \text{IQR}$
• Upper Inner Fence (UIF): $\text{UIF} = Q_3 + 1.5 \times \text{IQR}$

Any data point falling below the LIF ($x_i < \text{LIF}$) or above the UIF ($x_i > \text{UIF}$) is classified as a suspected outlier and plotted individually as an asterisk or dot. Points falling beyond $3.0 \times \text{IQR}$ are classified as extreme outliers.

Comparative Side-by-Side Boxplots

Boxplots are exceptionally powerful when displayed side-by-side to compare multiple subgroups (e.g., 4 operating shifts or 3 raw material vendors). Differences in median lines highlight centering shifts; differences in box height (IQR) highlight dispersion differences; asymmetric whisker lengths reveal subgroup skewness.

---

Multi-Vari Analysis: Decomposing Process Variation

Pioneered by Leonard Seder in 1950, Multi-Vari Analysis is an empirical, non-interventional graphical technique used to isolate the primary source of process variation without complex mathematical modeling. Unlike Design of Experiments (DOE)—which actively manipulates machine setpoints—Multi-Vari analysis is purely passive and observational. The practitioner samples the process as it naturally operates.

Multi-Vari decomposes overall process variance into three primary structural families:

                      The Three Families of Variation

                               TOTAL VARIATION
                                      │
         ┌────────────────────────────┼────────────────────────────┐
         ▼                            ▼                            ▼
     POSITIONAL                    CYCLICAL                     TEMPORAL
   (Within-Unit)                (Piece-to-Piece)             (Time-to-Time)
 • Variation across a single   • Variation between          • Variation across
   physical component or form    consecutive units or         hours, shifts, days,
 • Geometric taper, runout,      production lots              weeks, or seasons
   thickness gradients         • Cavity-to-cavity,          • Machine warmup, tool
 • Form field variances          batch-to-batch, setup        wear, shift handovers

The Three Families of Variation in Detail

1. Positional (Within-Piece / Within-Unit) Variation:

   • Definition: Variation occurring across different physical locations, dimensions, or orientations of the exact same individual unit at a single point in time.
   • Physical Examples: Radial runout across a machined shaft; thickness taper from top to bottom of an injection-molded housing; temperature gradient across an industrial furnace bed; coating thickness between horizontal and vertical faces.
   • Transactional Examples: Error rate variations between Page 1 and Page 5 of a loan application; cycle time discrepancies between Section A and Section B of an insurance claim.
   • Primary Suspects: Fixture clamping distortion, tool misalignment, mold warpage, uneven heating/cooling nozzle flow.

2. Cyclical (Piece-to-Piece / Batch-to-Batch) Variation:

   • Definition: Variation observed between consecutively produced individual units, parts, or batches produced within a short, contiguous time window.
   • Physical Examples: Weight differences between consecutive tablets compressed on a high-speed press; dimensional variance between Cavity 1 and Cavity 8 of a multi-cavity injection mold; lot-to-lot chemical purity differences between consecutive raw material containers.
   • Transactional Examples: Processing time differences between consecutive customer service calls handled by the same representative on the same morning.
   • Primary Suspects: Raw material supplier inconsistency, multi-cavity tooling imbalances, fluctuating hydraulic pressure, inconsistent operator manual loading.

3. Temporal (Time-to-Time / Shift-to-Shift) Variation:

   • Definition: Variation observed over longer operational time horizons (hours, shifts, days, weeks, or seasons).
   • Physical Examples: Morning thermal warmup drift; cutting tool wear over an 8-hour shift; seasonal ambient humidity shifts altering paint cure times; monthly maintenance calibration decay.
   • Transactional Examples: Call center volume spikes on Monday mornings; claim submission surges during end-of-month financial closing.
   • Primary Suspects: Machine thermal expansion, progressive cutting tool wear, chemical bath aging, shift handover communication breakdowns, ambient seasonal changes.

Comparative Overview of Variation Families

---

Multi-Vari Chart Construction & Interpretation

A Multi-Vari Chart is a structured graphical display that plots continuous response measurements against a hierarchical horizontal axis representing the three variation families.

                          Multi-Vari Chart Architecture

   Response (Y)
     ▲                                               BATCH 2 (Afternoon Shift)
     │              BATCH 1 (Morning Shift)          ┌───────────────────────┐
     │             ┌───────────────────────┐         │   •                   │
     │             │                       │         │   │                   │
     │             │       •               │         │   │   •               │
     │             │       │               │         │   •   │               │
     │             │   •   │   •           │         │       │   •           │
     │             │   │   •   │           │         │       •   │           │
     │             │   •       •           │         │           •           │
     │             │                       │         │                       │
     └─────────────┴───────────────────────┴─────────┴───────────────────────┴──▶
                   Unit 1  Unit 2  Unit 3            Unit 4  Unit 5  Unit 6
                   [-- Cyclical (P-to-P) -]          [-- Cyclical (P-to-P) -]
                   [----------- TEMPORAL SHIFT VARIATION (Time) -------------]

Diagnostic Decision Rules

1. Long Vertical Lines, Flat Unit Means $\implies$ Positional Variation Dominates: If the vertical lines connecting measurements on the same part are long, but part-to-part averages and batch-to-batch averages are relatively stable, positional variation is the root cause. Focus on fixturing, clamping pressure, and part geometry.
2. Short Vertical Lines, Erratic Unit Means $\implies$ Cyclical Variation Dominates: If individual parts are uniform internally (short lines), but consecutive parts fluctuate wildly in average value, cyclical variation dominates. Focus on raw material lots, cavity-to-cavity balancing, and machine loading repeatability.
3. Short Lines, Tight Part Clusters, Shifting Batch Averages $\implies$ Temporal Variation Dominates: If parts are internally uniform and consecutive parts match closely, but the entire batch average drifts upward or downward across hours or shifts, temporal variation dominates. Focus on thermal warmup, tool wear, environmental controls, and shift handovers.

---

Scatter Plots: Evaluating Bivariate Associations

Once the dominant family of variation is isolated, the team identifies specific candidate input variables ($X$) within that family and evaluates their relationship to the critical output ($Y$). A Scatter Plot (or scatter diagram) is the foundational visual tool for displaying bivariate continuous data.

                         Scatter Plot Correlation Patterns

     STRONG POSITIVE LINEAR           STRONG NEGATIVE LINEAR            CURVILINEAR (NON-LINEAR)
     Y ▲         ..-*                 Y ▲ *..                          Y ▲        ..***..
       │       .*                       │    '*.                         │     .-'       '-.
       │     .*                         │       '*.                      │   .'             '.
       │   .*                           │          '*.                   │  /                 \
       │ .*                             │             '*                 │ *                   *
       └─────────────▶                  └─────────────▶                  └─────────────▶
       0             X                  0             X                  0             X
          r ≈ +0.92                        r ≈ -0.94                        r ≈ 0.00 (Strong Fit!)

     ZERO / NO CORRELATION            MODERATE POSITIVE CORRELATION
     Y ▲    .  *  .                   Y ▲          . *
       │   *  .  *  .                   │       . *  * .
       │  .  *   .   *                  │     *  .  *
       │   *   .  *  .                  │   .  *  .
       │  .  *   .   *                  │ *  .  .
       └─────────────▶                  └─────────────▶
       0             X                  0             X
          r ≈ 0.00                         r ≈ +0.65

The Pearson Correlation Coefficient ($r$)

The strength and direction of a linear relationship on a scatter plot is quantified by the Pearson correlation coefficient ($r$), which ranges strictly between $-1.0$ and $+1.0$:

• $r = +1.0$: Perfect positive linear relationship (as $X$ increases, $Y$ increases proportionally).
• $r = -1.0$: Perfect negative linear relationship (as $X$ increases, $Y$ decreases proportionally).
• $r = 0.0$: Zero linear correlation (points form a random, circular or rectangular cloud).
• $0.7 < |r| < 1.0$: Strong linear association.
• $0.4 < |r| < 0.7$: Moderate linear association.
• $0.0 < |r| < 0.4$: Weak linear association.

CRITICAL EXAM TRAP — Curvilinear Relationships: The Pearson correlation coefficient measures LINEAR association only! If a scatter plot exhibits a perfect U-shaped parabola (as in the curvilinear diagram above), the calculated Pearson $r$ will be approximately $0.00$, despite an absolute, deterministic mathematical relationship between $X$ and $Y$. Always visually inspect the scatter plot before relying solely on numerical correlation statistics.

---

The Cardinal Six Sigma Rule: Correlation Does NOT Prove Causation

A foundational axiom of Six Sigma quality engineering is:

$\mathbf{Correlation \ne Causation}$

Just because two continuous variables demonstrate a statistically significant correlation coefficient ($r = 0.95$) does not prove that changes in $X$ physically cause changes in $Y$. A strong statistical correlation can be produced by three deceptive operational phenomena:

1. Confounding (Lurking) Variables: Both $X$ and $Y$ are simultaneously driven by an unmeasured third variable ($Z$).
   • Classic Example: In a chemical plant, daily output yield ($Y$) and cooling tower evaporation ($X$) show a strong positive correlation ($r = 0.88$). An untrained analyst might assume increasing cooling tower evaporation will increase plant yield. In reality, ambient outside temperature ($Z$) drives both variables simultaneously.
2. Reverse Causality: The presumed output $Y$ is actually driving the input $X$.
3. Spurious Coincidence: Pure chance alignment in historical data. In large observational databases, unrelated variables frequently demonstrate high correlation coefficients purely by coincidence.

The Analytical Rule of Engagement

In the Analyze phase, a scatter plot establishes an empirical association, not definitive proof. Before spending capital or altering production tooling, the Green Belt must verify causation through:

• Direct physical/chemical domain knowledge.
• Controlled confirmation runs.
• Formal Design of Experiments (DOE) or hypothesis testing, where $X$ is actively manipulated while holding lurking variables constant.

---

Critical Exam Traps to Avoid

• Trap 1: Confusing Histograms with Bar Charts — Bar charts display discrete categorical variables with visible gaps between bars. Histograms display continuous numerical variables with touching bars. Claiming a histogram is used for categorical pass/fail counts is an instant exam error.
• Trap 2: The Direction of Skewness — Skewness is always defined by the direction of the long tail, NOT the location of the tall hump! A distribution with a tail stretching out to the right is positively (right) skewed, and its mean is greater than its median.
• Trap 3: Confusing Positional Variation with Cyclical Variation — Positional variation occurs across different physical locations on the same single unit (e.g., taper or thickness gradient). Cyclical variation occurs between separate consecutive units (e.g., Part 1 vs. Part 2).
• Trap 4: Assuming $r = 0$ Means No Relationship Exists — Forgetting that Pearson's $r$ evaluates linear relationships only. Curvilinear or parabolic relationships will yield $r \approx 0$ despite strong deterministic relationships.
• Trap 5: Concluding Causation from a High Correlation Coefficient — Selecting an exam choice that asserts "Because $r = 0.96$, management can conclude that $X$ causes $Y$." Correlation never proves causation without controlled verification.