Interpreting RSM Results for Second-Order Model Development

1. The Goal: Building a Second-Order Model

The data from an RSM experiment allows us to fit a more complex model than the simple linear one from a factorial design. We use a second-order (or quadratic) model. For a process with two factors, $X_1$ and $X_2$, the model looks like this:

$Y={\beta }_0+{\beta }_1{X}_1+{\beta }_2{X}_2+{\beta }_{12}{X}_1{X}_2+{\beta }_{11}{X}_1^2+{\beta }_{22}{X}_2^2+ϵ$

Let's quickly break this down:

• ${\beta }_0$: The intercept, or the baseline response.
• ${\beta }_1{X}_1,{\beta }_2{X}_2$: The familiar linear effects of your factors.
• ${\beta }_{12}{X}_1{X}_2$: The interaction effect between the factors.
• ${\beta }_{11}{X}_1^2,{\beta }_{22}{X}_2^2$: The new and crucial quadratic terms. These terms model the curvature in the response, allowing us to find a peak or valley.

Our main task is to use statistical software to analyze our experimental data and find the estimated values for these coefficients ($\beta$'s), creating a predictive equation for our specific process.

2. The Analysis Workflow: From Data to Model

Analyzing RSM results follows a structured, practical workflow. We will use a video demonstration to walk through these steps in Minitab, a common software in Six Sigma practice. The principles are the same regardless of the software you use.

The video continues the virtual catapult example from a previous lesson, using a Central Composite Design.

Now, let's formalize and expand on the key steps you just saw.

Step 1: Interpret the ANOVA Table

The Analysis of Variance (ANOVA) table is the primary tool for judging your model. It tells you if the model is statistically significant and which terms are contributing to it.

To summarize, when you look at the ANOVA output, you're looking for:

• Model p-value: Should be low (e.g., < 0.05). This confirms that your model is significantly better at predicting the response than just using the average.
• Term p-values: Individual p-values for linear (A, B), interaction (AB), and quadratic (A², B²) terms should be low. High p-values (> 0.10) suggest terms that aren't contributing much and could potentially be removed.
• Lack-of-Fit p-value: This is unique to experiments with "pure error" (i.e., replicated runs, like our center points). You want this p-value to be high (> 0.05). A high p-value means there is no statistical evidence that your model is "missing" something important. It's a good sign that your chosen quadratic model fits the data well.
• R-squared values:
  • R-squared (R²): The percentage of variation in the response explained by the model. Higher is better.
  • Adjusted R²: Adjusts for the number of terms in the model. It's useful for comparing models with different numbers of terms.
  • Predicted R²: Shows how well the model is expected to predict new observations. A large gap between R² and Predicted R² can indicate an over-fit model.

Step 2: Refine the Model

As you saw in the video (LINK from 19:28 to 20:46), a common practice is to remove insignificant terms to create a simpler, more robust model. This process is called model reduction.

The general rule, based on the hierarchy principle, is:

1. Start by evaluating the highest-order terms (quadratics and interactions).
2. If a term is not significant (high p-value) and is not part of a significant higher-order term, it's a candidate for removal. For example, if the interaction A*B is significant, you should keep both A and B in the model, even if their own p-values are high.
3. Remove the least significant term (highest p-value) one at a time and re-run the analysis until only significant terms (and any required lower-order terms) remain.

This refinement results in a final predictive equation that is both statistically sound and easier to interpret.

Step 3: The Final Regression Equation

After refining the model, the software provides the final regression equation. This equation is the mathematical description of your response surface. It's typically provided in two forms:

• Uncoded Units: This equation uses the actual process values (e.g., Temperature = 85 °C, Time = 48 minutes). It is the most practical form for making predictions.
• Coded Units: This equation uses the coded levels (-1, +1, etc.). It is useful for comparing the relative impact of the factors, as the magnitude of the coefficients directly reflects the effect size.

For example, the final (uncoded) equation from the RSM Part-2 video allows you to plug in any Firing Angle and Release Angle within your experimental range and predict the resulting Catapult Distance.

3. Validating the Model with Residuals

A good-looking ANOVA table is not enough. The statistics are only valid if the underlying assumptions of the analysis are met. We check these assumptions by examining the residuals—the differences between the actual observed values from your experiment and the values predicted by your model.

Here’s what to look for in these crucial diagnostic plots:

1. Normal Plot of Residuals: The points should fall roughly along the straight red line. This indicates that the residuals follow a normal distribution, a key assumption. S-curves or major deviations suggest a problem.
2. Residuals vs. Fits (or Predicted vs. Actual): The points should be randomly scattered in a horizontal band around the zero line. Funnel shapes or other patterns indicate non-constant variance, meaning the model predicts better in some areas than others.
3. Residuals vs. Run Order: The points should be randomly scattered. Any trends, shifts, or patterns could indicate that a time-dependent variable (like tool wear, ambient temperature change) affected your results, which violates the assumption of independent errors.
4. Leverage or Cook's Distance: These are influence plots that help you identify outlier data points that may have an unusually large impact on the model. Points that exceed the red limits on these plots should be investigated.

If your residual plots look good (random scatter, points on the line), you can be confident that your model is statistically valid.

4. A Second Example: Semiconductor Manufacturing

To reinforce the workflow, let's look at another practical example from a different industry.

This text-based example perfectly complements the video walkthrough, showing the same logical process of fitting, assessing, refining, and validating a second-order model.

Conclusion

In this lesson, you've learned the critical process of turning raw experimental data into a validated, predictive process model. This is the analytical engine of the "Improve" phase.

Key Takeaways:

• The goal of RSM analysis is to fit a second-order (quadratic) model that captures linear, interaction, and curvature effects.
• The ANOVA table is used to assess the overall model significance, the contribution of individual terms, and the goodness-of-fit (via the Lack-of-Fit test).
• Model refinement involves removing statistically insignificant terms to create a simpler and more robust predictive equation, following the hierarchy principle.
• Residual analysis is a non-negotiable step to validate that the statistical assumptions of the model are met, ensuring your conclusions are reliable. A good model has normally distributed residuals with constant variance.

Preview of the Next Lesson:

You have successfully built and validated a mathematical model of your process. You now hold the key to optimization. In our next lesson, "Use contour plots and optimization tools to identify optimal process settings from a response surface model," we will finally use this equation. You will learn how to create visual tools like contour and 3D surface plots to explore the "lay of the land" and use numerical optimizers to find the exact factor settings that will give you the peak performance you've been looking for.