10.1 Correlation & Scatter Plots

Correlation & Scatter Plots

Quick Answer: A scatter plot graphs paired coordinates $(X_i, Y_i)$ on Cartesian axes to visually uncover relationships between process inputs ($X$) and outputs ($Y$). The Pearson correlation coefficient ($r$) quantifies linear association on a scale from $-1.0$ (perfect negative) to $+1.0$ (perfect positive), with $0.0$ indicating no linear link. Crucially, correlation demonstrates numerical association, never causation; unmeasured lurking variables frequently generate misleading statistical links that require controlled experimentation to verify. Independent CSSYB study guide by OpenExamPrep.

Bivariate Data Analysis in the Analyze Phase

In the DMAIC roadmap, the Analyze phase transitions an improvement team from describing process performance to discovering the fundamental root causes of variation and defects. While the Measure phase relies on univariate statistics (mean, standard deviation, histograms) to baseline individual metrics, univariate analysis cannot explain why an outcome fluctuates.

To isolate root causes, Six Sigma practitioners evaluate bivariate data—paired measurements of two characteristics collected from the same process units. Bivariate analysis investigates the Six Sigma transfer function:

$$Y = f(X) + \epsilon$$

Where $Y$ is the Key Process Output Variable (KPOV) or customer CTQ metric, $X$ represents a candidate Key Process Input Variable (KPIV) or operating parameter, and $\epsilon$ represents common-cause random variation. By analyzing bivariate data, Yellow Belts separate vital process drivers from inconsequential noise variables.

---

Scatter Plots: Visualizing Bivariate Relationships

The primary graphical tool for exploring bivariate relationships is the scatter plot. A scatter plot plots paired observations $(X_i, Y_i)$ on Cartesian coordinates, placing the independent input variable on the horizontal $X$-axis and the dependent response on the vertical $Y$-axis.

Practitioners examine four core visual features on a scatter plot:

1. Form: The overall geometric shape of the data points.
   • Linear: Points cluster along a straight path where $Y$ changes at a constant rate relative to $X$.
   • Curvilinear (Non-linear): Points trace a curved trajectory, such as a quadratic parabola or exponential curve. For example, chemical yield often rises with temperature up to an optimum point before dropping due to thermal degradation.
2. Direction: The trajectory of the relationship.
   • Positive Association: As $X$ increases, $Y$ increases (upward slope from lower-left to upper-right).
   • Negative Association: As $X$ increases, $Y$ decreases (downward slope from upper-left to lower-right).
   • Zero / No Association: Points form a diffuse, circular cloud with no discernible direction.
3. Strength: The degree of dispersion around the central trend. Points tightly clustered around a line indicate a strong relationship, whereas widely dispersed points indicate a weak relationship.
4. Outliers and Leverage: Scatter plots immediately expose isolated observations that depart from the general pattern. Extreme observations near the ends of the $X$-axis exert high leverage and can distort statistical metrics.

---

Pearson Product-Moment Correlation Coefficient ($r$)

Because visual interpretation of scatter plots can be subjective, Six Sigma uses the Pearson product-moment correlation coefficient ($r$) to quantify the strength and direction of linear association between two continuous variables:

$$r = \frac{\sum_{i=1}^{n} (X_i - \bar{X})(Y_i - \bar{Y})}{\sqrt{\sum_{i=1}^{n} (X_i - \bar{X})^2 \sum_{i=1}^{n} (Y_i - \bar{Y})^2}}$$

Interpretation Scale and Bounds

The correlation coefficient is bounded strictly between $-1.0$ and $+1.0$:

$$-1.0 \le r \le +1.0$$

The algebraic sign reflects direction ($+$ for positive slope, $-$ for negative slope), while the absolute magnitude ($|r|$) indicates strength:

Critical Assumptions and Boundaries

Yellow Belts must verify four key conditions when applying Pearson $r$:

• Continuous Data: Both variables must be continuous interval or ratio measurements.
• Linearity: Pearson $r$ evaluates only linear association. For a strong U-shaped non-linear relationship, $r$ may equal zero because positive and negative slopes cancel out.
• Bivariate Normality: Variables should approximate normal distributions.
• Outlier Vulnerability: A single wild outlier can artificially inflate or suppress $r$.

---

Critical Caveat: Correlation vs. Causation

The cardinal rule of bivariate analysis is that correlation does not equal causation. Calculating a high correlation coefficient ($r = 0.92$) proves that two metrics move together mathematically, but does not prove that changing $X$ causes $Y$ to change.

Misinterpreting correlation as causation frequently leads to flawed process changes due to:

1. Confounding and Lurking Variables

A lurking variable ($Z$) is an unmeasured third factor that simultaneously influences both $X$ and $Y$.

Example: A manufacturing plant finds a strong positive correlation ($r = 0.85$) between monthly electricity usage ($X$) and total product defects ($Y$). Cutting power will not eliminate defects: the lurking variable is monthly production volume ($Z$). High production volume increases electricity consumption and naturally yields more total defects.

2. Reverse Causality and Coincidence

Causality may flow in reverse ($Y$ causes $X$), or two unrelated metrics may correlate purely by chance or shared historical trends (spurious correlation).

Establishing True Causation

To prove causality in the Analyze phase, teams must:

1. Validate Physical Mechanisms: Verify that a known engineering, physical, or operational law explains the relationship.
2. Conduct Multi-Vari Studies: Stratify data across shifts, machines, and lots to verify the relationship persists across subgroups.
3. Execute Controlled Experiments (DOE): Intentionally manipulate input $X$ under controlled conditions while holding all other factors constant to observe the response in $Y$.