knitr::opts_chunk$set( # collapse = TRUE, comment = "#>" )
When conducting exact ANOVA power analyses with
Superpower it is possible to calculate the power for both the omnibus $F$-tests and planned contrasts or post-hoc comparisons.
It is possible to use this approach to calculate power for standard contrasts, such as pairwise contrasts between cells.
Here I provide a brief overview of how the output of
ANOVA_exact() can be used to perform power analyses for tailored planned contrasts and follow-up tests.
All power analyses for
emmeans-objects are based on the $F$- and $t$-values from the analyses of the dataset simulated by
ANOVA_exact() assuming two-sided testing.
emmeans_power() does not honor adjustments of the testing procedure due to either one-sided testing (including two one-sided tests) or corrections for multiple comparisons via the
adjust option in
As noted below, for the Bonferroni-adjustment this limitation can be overcome by manually adjusting
First, we will set up a 2 $\times$ 3 repeated measures design.
ANOVA_exact() pairwise comparisons of expected marginal means are added by setting
emm = TRUE and
contrast_type = "pairwise" (default).
# Set up a within design with 2 factors, each with 2 and 3 levels design_result <- ANOVA_design( design = "2w*3w", n = 40, mu = c(0.3, 0, 0.5, 0.3, 0, 0), sd = 2, r = 0.8, labelnames = c("condition", "cheerful", "sad", "voice", "human", "robot", "cartoon") ) exact_result <- ANOVA_exact( design_result, alpha_level = 0.05, verbose = FALSE, emm = TRUE, contrast_type = "pairwise" )
The result contains the power calculations for both the omnibus $F$-tests and pairwise post-hoc comparisons.
The output also contains the
emmeans-object on which these power calculations are based.
By manipulating this object it is possible to tailor the power analyses to the contrasts desired for the planned study.
That is, based on the dataset simulated with
ANOVA_exact() we can write out the analysis code for
emmeans-contrasts, just as we would if we were to analyze the empirical data, and use the output to perform the corresponding power analysis.
emmeans reference grid and contrasts are included in the output of
emmeans_power() on the contrasts, we can reproduce the results of the previous power analysis for the pairwise comparisons.
Now, we can manipulate the
emmeans reference grid to perform additional power analyses.
In the following example, we calculate the power for the contrasts between sad and cheerful condition for each voice.
simple_condition_effects <- emmeans( exact_result$emmeans$emmeans, specs = ~ condition | voice ) emmeans_power(pairs(simple_condition_effects))
We may also calculate the power for testing all condition means against an arbitrary constant.
emmeans_power(test(simple_condition_effects, null = 0.5))
Finally, we can calculate the power for custom contrasts between any linear combination of conditions.
custom_contrast <- contrast( exact_result$emmeans$emmeans, list(robot_vs_sad_human = c(0, 0, 0, 1, -0.5, -0.5)) ) emmeans_power(custom_contrast)
emmeans_power() currently ignores adjustments for multiple comparisons, it is possible to calculate the power for Bonferroni-corrected tests by adjusting
n_contrasts <- nrow(as.data.frame(simple_condition_effects)) emmeans_power( pairs(simple_condition_effects), alpha_level = 0.05 / n_contrasts )
Similarly, if we want to calculate power for a one-sided test, we can doubling
emmeans_power( pairs(simple_condition_effects), alpha_level = 2 * 0.05 )
Note, that because power is calculated from the squared $t$-value, power is only calculated correctly if the alternative hypothesis is true in the simulated dataset. That is, the difference of the condition means is consistent with the tested directional hypothesis.
emmeans can perform equivalence, non-superiority, and -inferiority tests,
emmeans_power() can calculate the corresponding power for these tests.
emmeans_power( pairs(simple_condition_effects, side = "equivalence", delta = 0.3) )
Note, that because power is calculated from the squared $t$-value, power is only calculated correctly if the alternative hypothesis is true in the simulated dataset.
That is, the difference between the condition means is consistent with the tested directional hypothesis (smaller than
Another useful application of
emmeans_power() is to joint tests.
Lets assume we plan to test the main effect of voice for each of the two conditions separately using
We can then calculate the power for each Bonferroni-corrected $F$-test as follows.
voice_by_condition <- joint_tests( exact_result$emmeans$emmeans, by = "condition" ) emmeans_power(voice_by_condition, alpha_level = 0.05 / 2)
Any scripts or data that you put into this service are public.
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.