21.3 Sample Size Planning for ANOVA
21.3.1 Balanced Designs
21.3.1.1 Single Factor Studies
21.3.1.1.1 Fixed cell means
\[ P(F>f_{(1-\alpha;a-1,N-a)}|\phi) = 1 - \beta \]
where \(\phi\) is the non-centrality parameter (measures how unequal the treatment means \(\mu_i\) are)
\[ \phi = \frac{1}{\sigma}\sqrt{\frac{n}{a}\sum_i (\mu_i - \mu_.)^2} , (n_i \equiv n) \]
and
\[ \mu_. = \frac{\sum \mu_i}{a} \]
To decide on the power probabilities we use the non-central F distribution.
We could use the power table directly when effects are fixed and design is balanced by using minimum range of factor level means for your desired differences
\[ \Delta = \max(\mu_i) - \min(\mu_i) \]
Hence, we need
- \(\alpha\) level
- \(\Delta\)
- \(\sigma\)
- \(\beta\)
Notes:
- When \(\Delta/\sigma\) is small greatly affects sample size, but if \(\Delta/\sigma\) is large.
- Reducing \(\alpha\) or \(\beta\) increases the required sample sizes.
- Error in estimating \(\sigma\) can make a large difference.
21.3.1.2 Multi-factor Studies
The same noncentral \(F\) tables can be used here
For two-factor fixed effect model
Test for interactions:
\[ \begin{aligned} \phi &= \frac{1}{\sigma} \sqrt{\frac{n \sum \sum (\alpha \beta_{ij})^2}{(a-1)(b-1)+1}} = \frac{1}{\sigma} \sqrt{\frac{n \sum \sum (\mu_{ij}- \mu_{i.} - \mu_{.j} + \mu_{..})^2}{(a-1)(b-1)+1}} \\ \upsilon_1 &= (a-1)(b-1) \\ \upsilon_2 &= ab(n-1) \end{aligned} \]
Test for Factor \(A\) main effects:
\[ \begin{aligned} \phi &= \frac{1}{\sigma} \sqrt{\frac{nb \sum \alpha_i^2}{a}} = \frac{1}{\sigma}\sqrt{\frac{nb \sum (\mu_{i.}- \mu_{..})^2}{a}} \\ \upsilon_1 &= a-1 \\ \upsilon_2 &= ab(n-1) \end{aligned} \]
Test for Factor \(B\) main effects:
\[ \begin{aligned} \phi &= \frac{1}{\sigma} \sqrt{\frac{na \sum \beta_j^2}{b}} = \frac{1}{\sigma}\sqrt{\frac{na \sum (\mu_{.j}- \mu_{..})^2}{b}} \\ \upsilon_1 &= b-1 \\ \upsilon_2 &= ab(n-1) \end{aligned} \]
Procedure:
- Specify the minimum range of Factor \(A\) means
- Obtain sample sizes with \(r = a\). The resulting sample size is \(bn\), from which \(n\) can be obtained.
- Repeat the first 2 steps for Factor \(B\) minimum range.
- Choose the greater number of sample size between \(A\) and \(B\).
21.3.2 Randomized Block Experiments
Analogous to completely randomized designs . The power of the F-test for treatment effects for randomized block design uses the same non-centrality parameter as completely randomized design:
\[ \phi = \frac{1}{\sigma} \sqrt{\frac{n}{r} \sum (\mu_i - \mu_.)^2} \]
However, the power level is different from the randomized block design because
- error variance \(\sigma^2\) is different
- df(MSE) is different.