The {InteractionPoweR}
package conducts power analyses
for regression models in cross-sectional data sets where the term of
interest is an interaction between two variables, also known as
‘moderation’ analyses. The package includes functions for simulating
data sets, conducting power analyses, and plotting and interpreting
results. Notable package features include (1) the ability to compute
power for interactions between two continuous variables, (2) effect
sizes are all specified as the cross-sectional Pearson’s correlation,
(3) simulations do not assume that the interacting variables are
independent, (4) any variable in the model, including the outcome, can
have anywhere from 2 (i.e., binary) to 20 discrete values, and (5)
analyses can incorporate the effects of reliability and skew, both of
the interacting variables, as well as of the outcome variable.
For more information see our tutorial paper, the package vignette, and the FAQ.
We also have a Shiny app which implements the major functions in a user-friendly point-and-click interface.
Please report bugs, issues, or questions as an Issue on Github.
You can install InteractionPoweR from CRAN with:
install.packages("InteractionPoweR")
You can also install the development version of InteractionPoweR from github with:
install.packages("devtools")
::install_github("dbaranger/InteractionPoweR/") devtools
If you get an error about a corrupt .rdb file, try restarting your R session.
The simplest use-case is when all the input parameters are known. We know the population-level correlation between our predictors (x1 and x2) and our outcome, we have a smallest effect size of interest in mind for our interaction effect size, and our sample size is already set (maybe we are conducting secondary data analysis). Power can be determined with a single command.
First - analytic power, using variable correlations (and reliability, if provided) to estimate how much additional variance is explained by the interaction term.
library(InteractionPoweR)
<-power_interaction_r2(
test_poweralpha = 0.05, # alpha, for the power analysis
N = 350, # sample size
r.x1x2.y = .15, # interaction effect to test (correlation between x1*x2 and y)
r.x1.y = .2, # correlation between x1 and y
r.x2.y = .1, # correlation between x2 and y
r.x1.x2 = .2 # correlation between x1 and x2
)
test_power#> pwr
#> 1 0.8055776
We see that we have 80% power.
We can also use simulations to estimate power. Simulations are particularly useful because they can account for non-normal data, including variable skew, binary variables, and likert variables.
NB In all these examples we use 1000 simulations for
speed (n.iter = 1000
), but for robust results we recommend
10,000 simulations (n.iter = 10000
).
set.seed(581827)
<-power_interaction(
test_powern.iter = 1000, # number of simulations per unique combination of input parameters
alpha = 0.05, # alpha, for the power analysis
N = 350, # sample size
r.x1x2.y = .15, # interaction effect to test (correlation between x1*x2 and y)
r.x1.y = .2, # correlation between x1 and y
r.x2.y = .1, # correlation between x2 and y
r.x1.x2 = .2, # correlation between x1 and x2
k.y = 2, # y is binary (has 2 levels) - analyses are run as logistic regressions
k.x1 = 5, # x1 has 5 levels (is a likert variable)
skew.x1 = .5, # x1 has a skew of 0.5
skew.x2 = 1.5 # x2 is a continuous variable and has a skew of 1.5
)
test_power#> N pwr
#> 1 350 0.763
The simulation estimates 75% power - it’s accuracy will increase with more iterations.
If you use {InteractionPoweR}
in a publication, please
cite our tutorial
paper:
Baranger DAA, Finsaas MC, Goldstein BL, Vize CE, Lynam DR, Olino TM (2022). “Tutorial: Power analyses for interaction effects in cross-sectional regressions.” PsyArxiv. doi: 10.31234/osf.io/5ptd7