Overview
In this vignette, we focus on statistical power and the role of the
effectsize easystats package in power analysis. As
such, we are interested in accomplishing several things with this
vignette:
- Reviewing statistical power and its value in a research task
- Demonstrating the role of the
effectsizepackage in the context of exploring statistical power - Highlighting the ease of calculating and understanding of
statistical power via the easystats ecosystem, and the
effectsizepackage specifically - Encouraging wider adoption of power analysis in applied research
Disclaimer: This vignette is an initial look at power analysis via easystats. There’s much more we could do, so please give us a feedback about what features would you like to see in easystats to make power analysis easier.
What is statistical power and power analysis?
Statistical power allows for the ability to check whether an effect observed from a statistical test actually exists, or that the null hypothesis really can be rejected (or not). Power involves many related concepts including, but not limited to, sample size, estimation, significance threshold, and of course, the effect size.
What is effectsize?
The goal of the effectsize package is to provide
utilities to work with indices of effect size and standardized
parameters, allowing computation and conversion of indices such as
Cohen’s d, r, odds-ratios, among many others. Please explore the breadth
of effect size operations included in the package by visiting the package
docs.
Putting the Pieces Together: Hypothesis Testing
Let’s take a closer looks at the key ingredients involved in statistical power before walking through a simple applied example below.
Statistical test: In research we often start with a statistical test to test expectations or explore data. For example, we might use a t-test to check for differences between two group means. This would help assess whether the difference of means between the groups is likely the same/indistinguishable from zero (null, \(H_0\)) or not (alternative, \(H_A\))
Significance threshold: This is the threshold against which we compare the p-value from our statistical test, which helps determine which hypothesis has the most support (and which we should reject). That is, we need to assess the probability that the result is likely indistinguishable from 0, or whether we have picked up on a likely real difference or result. To this end, if the p-value associated with our test is less than the significance threshold (e.g., \(p < 0.05\)), then, this tells us that the chance of observing the result we observed due to chance alone is extremely low, and very unlikely. In the case of comparing group mean differences, for example, we would have evidence allowing us to “reject the null hypothesis of no difference,” and conclude that there is a greater chance of the group means being significantly different from each other in line with \(H_A\).
Effect size: This is the magnitude of difference. A common way to calculate this is via Cohen’s \(d\), which measures the estimated standardized difference between the means of two populations. There are many other extensions (e.g., correcting for small-sample bias via Hedges’ \(g\)). This is where the
effectsizepackage comes in, which allows for easy calculation of many different effect size metrics.Statistical power: This brings us to statistical power, which can be thought of in many ways, such as the probability that we are correctly observing an effect or group difference, or that we are correctly rejecting the null hypothesis, and so on (see, e.g., (Cohen 1988/2013), (Greene 2000) for more). But regardless of the interpretation, all of these interpretations are all pointing to a common idea: the ability for us to trust the result we get from the hypothesis test, regardless of the test.
Let’s put these pieces together with a simple example. Say we find a “statistically significant” (\(p < 0.05\)) difference between two group means from a two-sample t-test. In this case, we might be tempted to stop and conclude that the signal is sufficiently strong to conclude that the groups are different from each other. But our test could be incorrect for a variety of reasons. Recall, that the p-value is a probability, meaning in part that we could be erroneously rejecting the null hypothesis, or that an insignificant result is insignificant due to a small sample size, and so on.
This is where statistical power comes in.
Statistical power helps us go the next step and more thoroughly assess the probability that the “significant” result we observed is indeed significant, or detect a cause of an insignificant result (e.g., sample size). In general, before beginning a broader analysis, it is a good idea to check for statistical power to ensure that you can trust the results you get from your test(s) downstream, and that your inferences are reliable.
So this is where we focus in this vignette, and pay special attention
to the ease and role of effect size calculation via the
effectsize package from easystats. The following
section walks through a simple applied example to ensure 1) the concepts
surrounding and involved in power are clear and digestible, and 2) that
the role and value of the effectsize package are likewise
clear and digestible. Understanding both of these realities will allow
for more complex extensions and applications to a wide array of research
problems and questions.
Example: Comparing Means of Independent Samples
In addition to relying on the easystats
effectsize package for effect size calculation, we will
also leverage the simple, but excellent pwr package for the
following implementation of power analysis (Champely and Rosario 2017).
First, let’s fit a simple two sample t-test using the mtcars data to
explore mean MPG for both transmission groups (AM).
t <- t.test(mpg ~ am, data = mtcars)There are many power tests supported by pwr for
different contexts, and we encourage you to take a look and select the
appropriate one for your application. For present purposes of
calculating statistical power for our t-test, we will rely on the
pwr.t2n.test() function. Here’s the basic anatomy:
pwr.t2n.test(
n1 = ..., n2 = ...,
d = ...,
sig.level = ...,
power = ...,
alternative = ...
)But, before we can get to the power part, we need to collect a few ingredients first, as we can see above. The ingredients we need include:
-
d: effect size -
n1andn2: sample size (for each sample) -
sig.level: significance threshold (e.g.,0.05) -
alternative: direction of the t-test (e.g., greater, lesser, two.sided)
(By omitting the power argument, we are implying that we
want the function to estimate that value for us.)
Calculate Effect Size
Given the simplicity of this example and the prevalence of Cohen’s
\(d\), we will rely on this effect size
index here. We have three ways of easily calculating Cohen’s \(d\) via effectsize.
Approach 1: effectsize()
The first approach is the simplest. As previously hinted at, there is
a vast literature on different effect size calculations for different
applications. So, if you don’t want to track down a specific one, or are
unaware of options, you can simply pass the statistical test object to
effectsize(), and either select the type, or
leave it blank for “cohens_d”, which is the default option.
Note, when using the formula interface to
t.test(), this method (currently) only gives an approximate
effect size. So for this first simple approach, we update our test
(t_alt) and then make a call to
effectsize().
t_alt <- t.test(mtcars$mpg[mtcars$am == 0], mtcars$mpg[mtcars$am == 1])
effectsize(t_alt, type = "cohens_d")Note, users can easily store the value and/or CIs as you’d
like via, e.g.,
cohens_d <- effectsize(t, type = "cohens_d")[[1]].
Approach 2: cohens_d()
Alternatively, if you knew the index one you wanted to use, you could
simply call the associated function directly. For present purposes, we
picked Cohen’s \(d\), so we would call
cohens_d(). But there are many other indices supported by
effectsize. For example, see here
for options for standardized differences. Or see here
for options for contingency tables. Or see here
for options for comparing multiple groups, and so on.
In our simple case here with a t-test, users are encouraged to use
effectsize() when working with htest objects
to ensure proper estimation. Therefore, with this second approach of
using the “named” function, cohens_d, users should pass the
data directly to the function instead of the htest object
(e.g., cohens_d(t)).
cohens_d(mpg ~ am, data = mtcars)Approach 3: t_to_d()
When the original underlying data is not available, you may get a warning message like:
Warning: … Returning an approximate effect size using t_to_d()
In these cases, the default behavior of effectsize is to
make a back-up call to t_to_d() (or which ever conversion
function is appropriate based on the input). This step makes the
calculation from the t-test to Cohen’s \(d\). Given the prevalence of calculating
effect sizes for different applications and the many effect size indices
available for different contexts, we have anticipated this and baked in
this conversion “fail safe” in the architecture of
effectsize by detecting the input and making the
appropriate conversion. There are many conversions available in the
package. Take a look here.
This can also be done directly by the user using the
t_to_d() function:
t_to_d(
t = t$statistic,
df_error = t$parameter
)
#> d | 95% CI
#> ----------------------
#> -1.76 | [-2.82, -0.67]Statistical Power
Now we are ready to calculate the statistical power of our t-test given that we have collected the essential ingredients.
For the present application, the effect size obtained from
cohens_d() (or any of the three approaches previously
described) can be passed to the d argument.
(result <- cohens_d(mpg ~ am, data = mtcars))
#> Cohen's d | 95% CI
#> --------------------------
#> -1.48 | [-2.27, -0.67]
#>
#> - Estimated using pooled SD.
(Ns <- table(mtcars$am))
#>
#> 0 1
#> 19 13
pwr.t2n.test(
n1 = Ns[1], n2 = Ns[2],
d = result[["Cohens_d"]],
sig.level = 0.05,
alternative = "two.sided"
)
#>
#> t test power calculation
#>
#> n1 = 19
#> n2 = 13
#> d = 1.478
#> sig.level = 0.05
#> power = 0.9779
#> alternative = two.sidedThe results tell us that we are sufficiently powered, with a very
high power for each group, 0.999 and
0.990.
Notice, though, if you were to change the group sample sizes to
something very small, say n = c(2, 2), then you would get a
much lower power, suggesting that your sample size is too small to
detect any reliable signal or to be able to trust your results.