Power Analysis and Effect Size Estimation

What is Statistical Power?

Let’s start by going over type I and type II error which occurs in hypothesis testing. To aid in this review I’ll use an example of comparing two samples. I have one sample of students which did not study for a test and another sample of students (from the same population) that took the same test and studied for 2 hours.

  • Null hypothesis (Ho): The mean score of each sample is the same. Said differently, there is no difference in the means between the two samples (i.e. there is no effect present)
  • Alternative hypothesis (Ha): The mean score for the students who studied will be different than the sample of students that did not study. This is representative of a hypothesis with two tails, if we were only looking for a positive effect (scores go up for the study group) it would be a one tail test.

The possible errors that may occur are:

  • Type I (α)- Often referred to as a false positive. Basically this is when you observe an effect that isn’t actually present (Rejecting the null hypothesis when it is actually true).
    • In this example that would mean our results indicate the group that studied had a higher (or lower) score on average than students that didn’t. Only problem is that we would be wrong.
  • Type II (β)- False Negative. This is what happens when we don’t detect a difference (effect) that is actually present.
    • This would mean we fail to reject the null hypothesis (i.e. we believe both groups are the same based on our data), but we are wrong and the study group actually did score higher (or lower) on average.

I’m going to focus on type II error for the moment, as this is the one most associated with power. The statistical power is a measure of our ability to detect an effect, if one is present, and it is defined as 1-β.

Why Do We Need Power?

Our experiment looks at the difference studying has on test scores. Before we perform such an experiment we need to answer a few questions:

  1. What would we consider a practically significant effect? Is an increase (decrease) of 1 point really anything to care about? Or do we only care about a change of 10 points or more?
  2. At what point do we reject the null hypothesis?
  3. How many participants do we need in each group to detect the effect? This would be the effect we determined in question 1.

Calculating the statistical power of the experiment will allow us to answer question 3, however first we must answer the first 2 questions.

Effect Size

Let’s address question 1 first. In order to understand what power really is we need to understand effect size. Effect size is a standardized measure of an effect. Is a change of 10 points on the test when the standard deviation is 2 points any different than a change of 5 points with a standard deviation of 1 point? Effect size allows for the comparison of experiments which may be on different scales.

In our case we are concerned with comparing two means. In the case of a t-test the effect is simply the difference between the sample means divided by the sample standard deviation. The effect size can be calculated as follows assuming equal sample sizes in each group:

So back to question 1, what size effect are we looking for? This is somewhat of a tricky question. If you already know your standard deviation or you can estimate it based on past studies you can calculate the effect size you are looking for. However if this is not the case there are generally accepted guidelines on effect size proposed by Cohen.

Cohen’s d is not the only method for calculating effect sizes. Depending on your data and test you are performing there are various options, some can be found here:

What Is the Acceptable Amount of Type I Error?

Let’s get back to question 2 - “At what point do we reject the null hypothesis?”. This question is basically asking “What level of type I error are we comfortable with?”. I won’t go into much detail on this as the generally consensus is an p-value below an α of 0.05. This translates to “If the probability of the effect seen is due to chance (sampling error) is less than 5%, reject the null hypothesis.”

For further reading on the pitfalls/shortcomings of p-values checkout:

A Slight Detour

Now that we have talked about what effect size is and the generally excepted Type I error, let’s go through some visual examples. The plots below show the regions of overlap between the null hypothesis (Ho) and the alternative hypothesis (Ha). In this example the alternative hypothesis is defined as an effect of 0.1 which can be seen by the mean value of it’s distribution.

The grey area in the plot above is the area where type I error can occur. There is overlap between the tail of the null hypothesis and the alternative hypothesis. The grey area in this plot is the area of the alternative hypothesis that cannot be detected due to overlap with the null hypothesis. In other words we would fail to reject the null hypothesis (i.e. detect a real effect) in this region. The grey area in this plot shows the region we are capable of detecting an effect and rejecting the null hypothesis, i.e. the power.

Selecting the Sample Size for an Experiment

Now on to question 3. Sample size should be determined prior to starting an experiment. To select the proper sample size we use power analysis. A power value of 0.80 (Type II error of 0.2) is considered the required minimum power value when performing an experiment.

First let’s take a look at the relation ship between sample size and power? Below you will see the distributions of the null and alternative hypotheses based on an α of 0.05 and an effect size of 0.1 with varying sample sizes.

In the previous visualizations I mentioned the distributions have means of zero (null) and 0.1 (alternative), but you may be wondering what the standard deviation for the distributions is. The standard deviation is the standard error:

The standard error is driven down as the sample size goes up as can be seen in the examples below. You can see the distributions become narrower and separate making a clearer decision boundary with less overlap. This has the effect of increasing our ability to detect an effect if present - in other words our power increases.

N=100
N=1000
N=2000

Example: What Sample Size Do I Need?

First step is to identify the sample size necessary to meet our requirements:

  • Detects an effect size of 0.5
  • Power = 0.80
  • α = 0.05
  • two-sided t-test

To estimate sample size I’ve used the pwr package in R.

library(ggplot2)
library(pwr)

t_test_pwr <- pwr.t.test(d = 0.5,
                         power = 0.80, sig.level = 0.05,
                         alternative = 'two.sided')
print(t_test_pwr)
Two-sample t test power calculation

         n = 63.76561
         d = 0.5
 sig.level = 0.05
     power = 0.8
alternative = two.sided

NOTE: n is number in *each* group

We can also plot the power vs. the sample size.

plot(t_test_pwr)

I wrote a small function to create random samples of an experiment.

perform_t_test <- function(n,mu_a, mu_b, s){
  group_a <- rnorm(n, mean = mu_a, sd = s)
  group_b <- rnorm(n, mean = mu_b, sd = s)
  t.test(group_a, group_b)$p.value
}

So now that we know the minimum sample size to meet our requirements, let’s test it out. First we will try an experiment with a sample size smaller than our power analysis stated.

Let’s start with an underpowered experiment. From the power analysis we know the sample size should be 64 or higher to allow us to detect a medium effect. So let’s start with two groups with 25 samples each (underpowered) and an effect present of 0.5. We perform the experiment 1000 times.

n <- 25
mu_a <- 80
mu_b <- 75
s <- 10

cohen_d <- (mu_a - mu_b)/s

effect_detected_percent <- mean(
  replicate(1000, perform_t_test(n, mu_a, mu_b, s)) < 0.05)
print(cohen_d)
[1] 0.5
print(effect_detected_percent)
[1] 0.41

The results show the null hypothesis is only rejected 41% of the time assuming we are using an α of 0.05.

Now let’s try it with a sample size which is greater than the required sample size. We find the null hypothesis is rejected in 88.5% of the experiments.

n <- 80
effect_detected_percent <- mean(
  replicate(1000, perform_t_test(n, mu_a, mu_b, s)) < 0.05)
print(effect_detected_percent)
[1] 0.885

What can happen if I perform an experiment that is overpowered? In the case of a trial that uses animals or humans there are ethical concerns in addition to the added cost and waste of resources. In addition to these issues, an overpowered experiment is more likely to detect extremely small effects as significant. Take the example below. Say my two groups have mean test scores of 80 and 80.5, but the standard deviation of the groups is 25 points instead of 10.

n <- 8000
mu_a <- 80
mu_b <- 80.5
s <- 25

cohen_d <- abs(mu_a - mu_b)/s
effect_detected_percent <- mean(
  replicate(1000, perform_t_test(n, mu_a, mu_b, s)) < 0.05)

print(cohen_d)
[1] 0.02
print(effect_detected_percent)
[1] 0.23

As a result the effect is only 0.02, but the null hypothesis is rejected in 23% of the experiments. This means 23% of the time we would find a significant result that would cause us to reject the null hypothesis, but knowing the data a priori, we know that the effect is negligible and there is most likely no real difference between the two groups.

If we use the same means and standard deviation and adjust to a smaller sample size, 80, we see the null hypothesis is only rejected 5.6% of the time.

n <- 80
mu_a <- 80
mu_b <- 80.5
s <- 25

cohen_d <- abs(mu_a - mu_b)/s
effect_detected_percent <- mean(
  replicate(1000, perform_t_test(n, mu_a, mu_b, s)) < 0.05)

print(cohen_d)
[1] 0.02
print(effect_detected_percent)
[1] 0.056

If you are interested in further reading into overpowered studies checkout The Power of “P”: On Overpowered Clinical Trials and “Positive” Results by Howard S. Hochster, MD