Day 16 Review: Applied split-plot
July 2nd, 2025
16.1 Announcements
- Planning to miss >2 classes in July? survey
- Watch last week’s classes (especially days 3+4)
- Homework 2 due today.
- Homework 3 is posted and due next Friday (July 11).
16.2 Background
Yesterday, we designed an experiment with an RCBD to figure out the best temperature to bake the muffins. Check out the original recipe here.
Yesterday we agreed on the following model:
\[y_{ik} = \mu + T_i + b_k + \varepsilon_{ik},\]
\[b_k \sim N(0, \sigma_b^2), \\ \varepsilon_{ik} \sim N(0, \sigma_{\varepsilon}^2),\]
and design:
|
|
|
Dry ingredients:
- 1 1/2 cups all-purpose flour (7 1/2 ounces)
- 1 1/2 teaspoons baking powder
- 1/4 teaspoon baking soda
- 1/2 teaspoon salt
- 1/4 teaspoon ground cinnamon
- 1/4 teaspoon ground nutmeg
- 2/3 cup finely chopped pecans, toasted
Wet ingredients:
- 1/2 cup light brown sugar (3 3/4 ounces)
- 1/4 cup granulated sugar (2 ounces)
- 2 large eggs
- 1 stick unsalted butter, browned and cooled (4 ounces)
- 1 1/2 cups mashed bananas, from very ripe bananas that have been peeled and mashed well with a fork (12 3/4 ounces or 361 gr.)
- 2 tablespoons full-fat sour cream
- 1 teaspoon pure vanilla extract
16.3 What is the best design?
16.3.1 Option A: another RCBD
Corresponding model:
\[y_{ik} = \mu + T_i + B_j + (TB)_{ij} + b_k + \varepsilon_{ik},\]
\[b_k \sim N(0, \sigma_b^2), \\ \varepsilon_{ik} \sim N(0, \sigma_{\varepsilon}^2).\]
|
|
|
16.3.2 Option B: split-plot design
In the model:
- Cooking days are blocks.
- Oven are “mini-blocks” for banana recipe.
Statistical model:
\[y_{ik} = \mu + T_i + B_j + (TB)_{ij} + b_k + w_{i(k)} + \varepsilon_{ik},\]
\[b_k \sim N(0, \sigma_b^2), \\ w_{i(k)} \sim N(0, \sigma^2_w) , \\ \varepsilon_{ik} \sim N(0, \sigma_{\varepsilon}^2).\]
ANOVA table for this split-plot design
|
|
|
16.4 Treatment means and confidence intervals for the split-plot design
The treatment mean for the \(i\)th temperature and \(j\)th banana level is \(\mu_{ij} = \mu + T_i + B_j +(TB)_{ij}\). That mean won’t change under different design structures. What may change is the confidence interval around the mean difference.
- First, recall the formula for a CI: \(\theta \pm t_{df, \frac{\alpha}{2}} \cdot se(\hat{\theta})\)
## [1] 2.776445
- For example, the CI for the differences between means for 300F and 400F \(\mu_{1 \cdot} - \mu_{2 \cdot}\) is \((\mu_{1 \cdot} - \mu_{2 \cdot}) \pm 2.78 \cdot se(\widehat{\mu_{1 \cdot} - \mu_{2 \cdot}})\)
- \(se(\widehat{\mu_{1 \cdot} - \mu_{2 \cdot}}) = \sqrt{\frac{2 (\sigma^2_{\varepsilon} + b \cdot \sigma^2_w)}{b \cdot r}}\)
\[\mu_i \pm 2.78 \cdot \sqrt{\frac{2 (\sigma^2_{\varepsilon} + b \cdot \sigma^2_w)}{b \cdot r}}\]
## [1] 2.446912
- The CI for the differences between means for normal and high banana \(\mu_{\cdot 1} - \mu_{\cdot 2}\) is \((\mu_{\cdot 1} - \mu_{\cdot 2}) \pm 2.44 \cdot se(\widehat{\mu_{\cdot 1} - \mu_{\cdot 2}})\)
- \(se(\widehat{\mu_{\cdot 1} - \mu_{\cdot 2}}) = \sqrt{\frac{2 \sigma^2_{\varepsilon}}{t \cdot r}}\)
\[\mu_i \pm 2.44 \cdot \sqrt{\frac{2 \sigma^2_{\varepsilon}}{r}}\]