6.11 Graphical Methods for Exploratory Data Analysis

Exploratory Data Analysis (EDA) represents the initial quantitative phase of the Six Sigma Analyze phase. It focuses on summarizing, visualizing, and understanding the inherent structure, distribution, central tendency, dispersion, and potential anomalies in data prior to formal statistical hypothesis testing. EDA provides Black Belts with visual insights into process behavior, helping identify potential root causes and guiding subsequent modeling efforts.

---

Histograms: Distribution & Shape Characterization

A histogram is a column chart that displays the frequency distribution of a continuous variable partitioned into non-overlapping, adjacent intervals (bins). It provides immediate visual feedback regarding process location, spread, symmetry, and underlying distributional shape.

Binning Rules & Construction Parameters

Choosing the correct number of bins ($k$) is critical; too few bins obscure fine structural details, whereas too many bins introduce random noise. Two standard mathematical guidelines determine bin count:

1. Sturges' Rule: $k = 1 + 3.322 \log_{10}(n)$ Where $n$ represents the total sample size.

2. Square Root Rule: $k = \sqrt{n}$

Bin width ($w$) is calculated by dividing the sample range by the bin count: $w = \frac{\text{Maximum} - \text{Minimum}}{k}$

Interpreting Histogram Shapes

---

Stem-and-Leaf Plots

A stem-and-leaf plot is a semi-graphical technique that organizes continuous numeric data to display distribution shape while preserving individual raw data values. Each data point is split into a stem (leading digits representing the bin interval) and a leaf (trailing digit representing individual observations).

Practical Example & Interpretation

Consider a sample of shaft diameter measurements (in mm): $12.1, 12.3, 12.3, 12.7, 13.0, 13.4, 13.8, 14.2$.

• Stem (10s and 1s): 12, 13, 14
• Leaf (0.1s):
  • 12 | 1 3 3 7
  • 13 | 0 4 8
  • 14 | 2

Key Advantage: Unlike histograms, stem-and-leaf plots allow the practitioner to reconstruct the exact dataset directly from the chart, enabling rapid order-statistic computation (median, range, quartiles) without losing granularity.

---

Box Plots & The 5-Number Summary

A box plot (or box-and-whisker plot) provides a compact visual representation of a dataset's central tendency, dispersion, skewness, and outliers based on the 5-number summary:

1. Minimum ($\text{Min}$): Smallest non-outlier observation.
2. First Quartile ($Q_1$): $25^{\text{th}}$ percentile (lower hinge).
3. Median ($Q_2$): $50^{\text{th}}$ percentile (center line).
4. Third Quartile ($Q_3$): $75^{\text{th}}$ percentile (upper hinge).
5. Maximum ($\text{Max}$): Largest non-outlier observation.

Mathematical Formulation of Box Plot Metrics

The Interquartile Range (IQR) measures the middle 50% of the distribution: $\text{IQR} = Q_3 - Q_1$

Outlier Detection Rules (Tukey's Fences)

Outliers are data points that fall outside the calculated inner and outer fences:

• Inner Fences (Mild Outliers): $\text{Lower Inner Fence} = Q_1 - 1.5 \times \text{IQR}$ $\text{Upper Inner Fence} = Q_3 + 1.5 \times \text{IQR}$ Data points outside inner fences but inside outer fences are plotted as individual asterisks ($*$) or open circles ($o$).

• Outer Fences (Extreme Outliers): $\text{Lower Outer Fence} = Q_1 - 3.0 \times \text{IQR}$ $\text{Upper Outer Fence} = Q_3 + 3.0 \times \text{IQR}$ Points beyond outer fences represent severe process anomalies and warrant immediate root-cause investigation.

---

Scatter Plots & Relationship Identification

A scatter plot displays pairs of continuous numeric data ($(X_i, Y_i)$) on a two-dimensional Cartesian plane to evaluate potential relationships between an independent variable ($X$, potential input factor) and a dependent variable ($Y$, process output metric).

Key Scatter Plot Diagnostics

• Direction: Positive association ($Y$ increases as $X$ increases) vs. Negative association ($Y$ decreases as $X$ increases).
• Form: Linear ($Y = \beta_0 + \beta_1 X$) vs. Curvilinear (quadratic, exponential).
• Strength: Clustering tightness around the trend line (quantified by Pearson correlation $r$).
• Heteroscedasticity: Non-constant variance across the range of $X$ (funnel-shaped dispersion indicating changing process variability).
• Clustering: Distinct grouping of data points indicating uncaptured categorical subgrouping variables (e.g., machine operator or ambient humidity).

---

Pareto Charts & The 80/20 Rule

A Pareto chart is a specialized bar chart where categories are arranged in descending order of frequency or cost, combined with a cumulative percentage line graph. It operationalizes Juran's Pareto Principle (the 80/20 rule), which posits that roughly 80% of process problems or costs stem from 20% of the potential defect causes.

Pareto Structure & Dual Y-Axes

• Left Y-Axis: Frequency, defect count, or financial cost ($).
• Right Y-Axis: Cumulative percentage ($0%$ to $100%$).
• X-Axis: Categorical defect types or failure causes arranged from highest to lowest impact.
• Vital Few vs. Useful Many: The left-most bars contributing up to $80%$ of the cumulative line represent the vital few priorities for project resolution.

---

Run Charts & Non-Random Pattern Rules

A run chart plots process performance data chronologically around the sample median or mean. It is used to detect trends, shifts, and non-random instability prior to establishing formal Statistical Process Control (SPC) limits.

Four Standard Non-Random Signals

1. Shift Rule: A run of $\ge 8$ consecutive data points on one side of the median center line.
2. Trend Rule: $\ge 6$ consecutive data points continuously increasing or decreasing.
3. Number of Runs Rule: Too few or too many crossings of the median line relative to expected statistical tables (indicating mixture or clustering).
4. Astronomical Point: A single point dramatically separated from the rest, representing an obvious special cause event.

---

Multi-Vari Charts: Variance Component Analysis

A Multi-Vari chart is a visual tool that decomposes total process variation into three discrete components without requiring complex analysis of variance calculations. It plots sample measurements across three primary categorical axes:

1. Positional Variation (Within-Piece): Variation occurring within a single unit or part (e.g., thickness variations across a single silicon wafer, taper along a machined cylinder, or temperature differences across a furnace bed).
2. Cyclical Variation (Piece-to-Piece / Batch-to-Batch): Variation occurring between consecutive parts produced short-term (e.g., part-to-part variation within a single batch or lot).
3. Temporal Variation (Time-to-Time): Variation occurring over extended time intervals (e.g., shift-to-shift, day-to-day, or setup-to-setup variations caused by tool wear, environmental changes, or operator techniques).

By systematically plotting high-low lines for positional variation, averages for piece-to-piece variation, and time-ordered blocks for temporal variation, Black Belts isolate which variance component dominates total process variability.