# Statistical Power

Statistical power is the probability Probability Probability is a mathematical tool used to study randomness and provide predictions about the likelihood of something happening. There are several basic rules of probability that can be used to help determine the probability of multiple events happening together, separately, or sequentially. Basics of Probability of detecting an effect when that effect genuinely exists in the population. Other things being equal, a test based on a large sample has more statistical power than a test involving a small sample. There are also ways to increase power without increasing the sample size. Most published studies have low statistical power, which can lead to serious misinterpretation of results.

Last update:

## Prerequisites

In order to comprehend the concept of statistical power, some previous knowledge about descriptive and inferential statistics is recommended.

## Definition and Importance

### Definition

Statistical power (SP) is expressed in 3 different ways:

1. SP is the probability Probability Probability is a mathematical tool used to study randomness and provide predictions about the likelihood of something happening. There are several basic rules of probability that can be used to help determine the probability of multiple events happening together, separately, or sequentially. Basics of Probability of finding significance if the alternative hypothesis is true.
2. SP is the probability Probability Probability is a mathematical tool used to study randomness and provide predictions about the likelihood of something happening. There are several basic rules of probability that can be used to help determine the probability of multiple events happening together, separately, or sequentially. Basics of Probability of correctly rejecting a false null hypothesis, where the null hypothesis is the hypothesis that there is no significant difference between specified populations (e.g., control versus experimental groups).
3. SP = 1 – beta (β), where β = type II (false negative) error, equivalent to 1 – sensitivity. The more powerful a clinical experimental study, the more readily it will detect a treatment effect when it really exists.

### Low statistical power

Fewer than 13% of 31,873 clinical trials published between 1974 and 2017 had adequate SP. A study with low SP means that the test results are questionable and poses potentially serious problems, including:

• A reduced chance of detecting a true, genuine, and meaningful effect in the sample population, which may block the pursuit of further studies
• Lower likelihood that a statistically significant result reflects a true effect (e.g., more false positives)
• Overestimation of the true treatment effect size (ES)
• Low reproducibility
• Possible violation of ethical principles:
• Patients and healthy volunteers continue to participate in research that may be of limited clinical value.
• Unnecessary sacrifice of research animals
• Chaos in the interpretation of small-sample studies using the same methodology but which produce conflicting results

### Too much statistical power: overpowered studies

Studies having too much SP, also called “overpowered studies,” can also be problematic because of the following reasons:

• Can be misleading as they have the potential to show statistical significance and also unimportant/irrelevant clinical differences
• Result in wastage of resources
• May be unethical due to the involvement of human subjects and/or laboratory animals for unnecessary experiments

## Characteristics

Statistical power has relevance only when the null hypothesis can be rejected, and is determined by the following variables:

• Alpha (α)
• Beta (β)
• Population standard deviation (SD)
• Sample size
• Treatment ES

### Alpha

Alpha is the chance of testing positive by a diagnostic test among those without the condition, causing a type I error or a “false positive.”

• Alpha = the chance of rejecting the null hypothesis among those that satisfy the null hypothesis
• Alpha = 1 – specificity = “p-value” = “the significance level”
• Alpha = false positives/(false positives + true negatives)
• At a significance (alpha) level of 0.05, 5% of the samples may show a false significant difference due just to chance.
• Most studies use an alpha cut-off of 5% (0.05).

### Beta

Beta is the chance of testing negative by a diagnostic test among those with the condition, causing a type II error or a “false negative.”

• Beta = the chance of accepting the null hypothesis among those that do not satisfy the null hypothesis
• Beta = 1 – sensitivity
• Beta = false negatives/(false negatives + true positives)
• Beta is directly related to SP of the study (SP = 1 – β).
• At a beta level of 0.2, 20% of the samples may miss a true significant difference.
• Most studies use a beta cut-off of 20% (0.2).
• Unlike alpha, there is a different value of beta for each different mean value of the alternative hypothesis; thus, beta depends on both the cut-off value set by alpha and the mean of the alternative hypothesis.

### Relationship between alpha and beta

The relationship between alpha and beta is often depicted in graphs that show:

• 2 normalized populations or distributions:
• A control group
• An experimental group (which may have a statistically significant, different mean)
• H0: the null hypothesis, which states that there is only 1 true mean (of the control group) and that any variation found in the sample/experimental group is only due to normal random variation in distribution
• H1: the alternative hypothesis, which is a statement that directly contradicts the null hypothesis by stating that the actual value of a population parameter is less than or greater than the value stated in the null hypothesis
• Alpha (α): the false-positive area (usually shown as a 2-tailed hypothesis test) Relationship between alpha and beta:This graph is a representation of 2 normalized populations or distributions of a control group (green) and an experimental group (purple) with a statistically significant different mean. H0 is the null hypothesis. α (alpha) = false-positive area, shown in a one-tailed hypothesis test. Note the inverse relationship between alpha and beta. Image by Lecturio. License: CC BY-NC-SA 4.0

This graph is a representation of 2 normalized populations or distributions of a control group (green) and an experimental group (purple) with a statistically significant different mean.

There is an inverse relationship between alpha and beta. If beta is decreased:

• The area of alpha would increase.
• The number of false negatives or type II errors would decrease.
• The number of false positives or type I errors would increase.

The inverse relationship of alpha and beta can also be appreciated in a 2 x 2 contingency table that compares the positive and negative findings of reality versus a study.

### Population standard deviation

Standard deviation is a measure of the amount of variation or dispersion Dispersion Central tendency is a measure of values in a sample that identifies the different central points in the data, often referred to colloquially as "averages." The most common measurements of central tendency are the mean, median, and mode. Identifying the central value allows other values to be compared to it, showing the spread or cluster of the sample, which is known as the dispersion or distribution. Measures of Central Tendency and Dispersion of a set of values relative to the mean.

• Calculated as the square root of the variance, which is the average of the squared differences from the mean.
• The higher the SD, the more patients are needed in a sample to demonstrate a statistically significant difference.

### Sample size

The sample size is the number of observations in a sample.

• A larger sample will better represent the population; thus, the power of the test will naturally increase.
• Most commonly used parameter to increase the power of a study

For a 2 sample, 2-tailed t-test with an alpha level of 0.05, the simple formula below will give an approximate sample size needed to have a statistical power of 80% (beta = 0.2):

$$n = \frac{16s^{2}}{d^{2}}$$

where n = size of each sample, s = standard deviation (assumed to be the same in each group), and d = difference to be detected. The mnemonic, as suggested by the originator of the formula, Robert Lehr, is “16 s-squared over d-squared.” (Note: “s-squared” is also known as the variance).

Examples:

1. Find the approximate number of patients with Stage I hypertension Hypertension Hypertension, or high blood pressure, is a common disease that manifests as elevated systemic arterial pressures. Hypertension is most often asymptomatic and is found incidentally as part of a routine physical examination or during triage for an unrelated medical encounter. Hypertension (systolic 130–139 mm Hg or diastolic 80–89 mm Hg) needed to provide 80% power to detect a 15-mm Hg difference in diastolic BP treatments A and B using a 2-sample, 2-tailed, alpha = 0.05 t-test, given that the expected standard deviation for each group is 15 mm Hg.
The approximate sample size n = “16 s-squared over d-squared” = 16 x 152/152 = 16 x 225/225 = 16 people in each group. Note that one of the “treatments” is usually set as the control (not treated) group.
1. To turn it around: In 2 groups with 16 patients each with Stage I hypertension Hypertension Hypertension, or high blood pressure, is a common disease that manifests as elevated systemic arterial pressures. Hypertension is most often asymptomatic and is found incidentally as part of a routine physical examination or during triage for an unrelated medical encounter. Hypertension, a treatment difference of 15 mm Hg was found after the patients in each group were treated with 2 different treatments.
If alpha = 0.05 and beta = 0.2, was the sample size sufficient to detect a significant difference?
n = 16 x 152/152 = 16 people in each group.  So, yes, the sample size was sufficient.
2. In question 1, what would be the approximate number of patients needed, if the investigator wanted to detect a 7.5-mm Hg difference instead of a 15-mm Hg difference, with all other parameters remaining the same?
n = 16 x 152/7.52 = 16 x 225/56.25 =  64 people in each group
In this last example, note that in order to detect half as big a difference you need 4 times the sample size, which is made evident by the simple formula.

### Treatment effect size

Effect size is the standardized mean difference between 2 groups, which is exactly equivalent to the “Z-score” of a standard normal distribution.

• If the difference between the 2 treatments is small, more patients will be required to detect a difference.
• Other situations that have ESs:
• The correlation between 2 variables
• The regression coefficient in a regression calculation
• The risk (“baseline incidence”) of a particular event (e.g., stroke)

Calculating ES with Cohen’s d:

Cohen’s d is the most common (but imperfect) method to calculate ES. Cohen’s d = the estimated difference in the means/(pooled estimated standard deviations), where:

$${SD = \sqrt{\frac{(SD1^{2} + SD2^{2})}{2}}}$$

If the SDs are equal in each group, then d = mean differences/SD. For example, if the difference is 150 and the SD is 50, then d = 150/50 = 3, which is a large ES.

Interpretation of Cohen’s d:

• Small ES: If d = 0.2, the score or value of the average subject in the experimental group is 0.2 SDs above the value of the average subject in the control group, exceeding the values of 58% of the control group.
• Medium ES: If d = 0.5, the value of the average subject in the experimental group is 0.5 SDs above the value of the average subject in the control group, exceeding the values of 69% of the control group.
• Large ES: If d = 0.8, the value of the average subject is 0.8 SDs above the value of the average subject in the control group, exceeding the values of 79% of the control group.

### Summary of characteristics

In summary, the SP will tend to be greater when:

• ES (difference between groups) is large.
• Sample size is large.
• SDs of the populations are small.
• Alpha significance level is higher (e.g., 0.05 rather than 0.01).
• Beta cut-off is low (e.g., 0.1 versus 0.2).
• 1-tailed test rather than a 2-tailed test is used.
• However, the directional hypothesis cannot detect a difference that is in the opposite direction.
• This test is rarely used.

## Common Mistakes

• Rejecting a null hypothesis (e.g., there is a significant difference) without considering the practical/clinical significance of the study finding
• Accepting a null hypothesis (e.g., not rejecting a null hypothesis) when a result is NOT statistically significant, without taking power into account
• Being convinced by a research study with low power
• Neglecting to perform a power analysis/sample size calculation
• Not correcting for multiple inference when calculating power:
• Multiple inference is the process of performing more than 1 statistical inference test on the same data set.
• Running multiple tests on the same data set at the same stage of analysis increases the chance of obtaining at least 1 invalid result.
• Using standardized ESs (e.g., the small, medium, and large ESs of Cohen’s d) rather than considering the details of the experimental design itself. An experiment can sometimes have a small Cohen rating but be a better experiment.
• Confusing retrospective power (calculated after data are collected) and prospective power

## Analysis

A power analysis answers 2 big questions:

1. How much SP is considered adequate?
2. How big of a sample size will be needed?

### How much statistical power is considered adequate?

The traditional minimum level of power is 80% (or 0.80), just as the arbitrary 5% (or 0.05) value is the traditional minimum alpha cut-off to set the p-value at 0.05.

• An 80% power level means there is a 20% probability Probability Probability is a mathematical tool used to study randomness and provide predictions about the likelihood of something happening. There are several basic rules of probability that can be used to help determine the probability of multiple events happening together, separately, or sequentially. Basics of Probability of encountering a type II error (false negative).
• This 20% acceptable level of having type II errors is 4 times higher than the 5% likelihood of encountering a type I error (false positive) for the standard value for the significance level.
• Type I errors are generally considered worse than type II errors.

It would be much better to have a 90% power level. Although it takes more resources, keep in mind that it would take even more to rerun the study later.

### How big of a sample size will be needed?

• A sample large enough to detect an effect of practical scientific importance to ensure a probability Probability Probability is a mathematical tool used to study randomness and provide predictions about the likelihood of something happening. There are several basic rules of probability that can be used to help determine the probability of multiple events happening together, separately, or sequentially. Basics of Probability high enough such that a false null hypothesis is rejected
• Power analysis must be performed before an experiment is started.
• One cannot continue to add subjects to a completed experiment that had a nearly significant p-value.
• This practice is frowned upon and constitutes what is called “p-hacking” or “data-dredging.”
• Calculating the proper sample size for an independent samples t-test:
• Estimate (by pilot study or historical data) the population means of the 2 groups or the difference between the means, which should be the smallest ES that is of scientific interest.
• Estimate (by pilot study or historical data) the population SDs of the 2 groups.
• Decide what levels of alpha (e.g., 0.05) and beta (e.g., 0.2) are desired.
• Submit these values (alpha, beta, the 2 estimated means, and the estimated pooled SD) into a reputable online SP calculator to obtain the sample size.
• The calculations are somewhat complex and are always done using a computer.
• An approximate sample size can be calculated by using the formula n =16s2/d2, as discussed above.

## Examples of Statistical Power Calculations and Analysis

### Scenario 1

A trial of a new fertilizer called “Grow-A-Lot” was given to a tomato farmer to determine if more tomatoes are produced per plant with the new fertilizer compared to plants not fertilized. The farmer picked 200 tomato seeds from a bucket of his usual planting seeds and divided them into 2 groups:

• A group of 100 seeds that did not receive fertilizer (the control group)
• A group of 100 seeds that received fertilizer (the experimental group)

The null hypothesis is that both groups of plants would produce the same number of tomatoes per plant, whereas the alternative hypothesis would be that the plants receiving the fertilizer would produce a different number of tomatoes.

Trial 1 with large sample sizes:

The fertilized group produced an average of twice the number of tomatoes (300) as the control group (150). There is also a small amount of overlap, since some plants in the control group overperformed the others in their group, whereas some plants in the experimental group underperformed. Just a glance at the graph is convincing enough to note that there is an obvious difference, but a t-test was performed to confirm that the difference was statistically significant, with a very small p-value. Trial 1, scenario 1: Graph demonstrating the use of Grow-A-Lot with large effect size, large SDs, means of 150 vs. 300, SD of 50 in each, and large sample sizes of 100Green: control groupBlue: experimental groupThe 100 plants in the control group resulted in a mean of 150 tomatoes per plant. While the 100 plants which received fertilizer produced significantly more tomatoes with a mean of 300 tomatoes per plant. The result is significant at a p-value of < 0.05, so the null hypothesis is rejected.SD: standard deviation Image by Lecturio. License: CC BY-NC-SA 4.0

Even if the experiment is repeated 1000 times, it would be extremely unlikely that the farmer would happen to pick a different set of seeds at random from the overlap region in order to yield a different result. The large size effect alone gives this trial a large amount of SP because it would be extremely unlikely that repeat sampling would produce a different result.

Trial 2 with small sample sizes:

The experiment would retain large SP with many fewer subject seeds as well, and almost all the t-tests would correctly give a significant (small) p-value. Trial 2, scenario 1: Graph demonstrating the use of Grow-A-Lot with large effect size, large SDs, means of 150 versus 300, SD of 50 in each, and small sample sizes of 30Green: control groupBlue: experimental groupA significant difference can still be seen between the groups both by eyeballing the graph and by statistical testing because of the large effect size. The result is significant at p < 0.05, so the null hypothesis is rejected.SD: standard deviation Image by Lecturio. License: CC BY-NC-SA 4.0

### Scenario 2

A different fertilizer is used (fertilizer “Grow-A-Little”) that has a much lesser effect, producing an average of only 10 extra tomatoes per plant. There will be a larger overlap of tomato production per plant between the experimental and the control groups, which can be detected only by using larger sample sizes.

Trial 3 with large sample sizes and large SDs:

The sample sizes are sufficiently large to counteract the small ES, making the difference statistically significant at a p-value < 0.05. Note that, even though the difference is statistically significant, this small difference may not be of practical or relevant significance to the farmer. Trial 3, scenario 2: Graph demonstrating the use of Grow-A-Little with small effect size, large SDs, means of 150 versus 165, SD of 50 in each, and large sample sizes of 100Green: control groupBlue: experimental group. The difference is statistically significant at a p-value of < 0.05 because the sample sizes were sufficiently large to counteract the small effect size. Note that, even though it is statistically significant, the small difference may not be of practical or relevant significance to the farmer.SD: standard deviation Image by Lecturio. License: CC BY-NC-SA 4.0

Trial 4 with small sample sizes and large SDs:

Due to small sample sizes, no statistically significant difference is found at a p-value < 0.05. So, the null hypothesis cannot be rejected because the trial did not have a large enough effect or sample size. Trial 4, scenario 2: Graph demonstrating the use of Grow-A-Little with small effect size, large SDs, means of 150 versus 165, SD of 50 in each, and small sample sizes of 30Green: control groupBlue: experimental groupNo statistically significant difference is revealed at a p-value of < 0.05, so the null hypothesis can not be rejected because the trial did not have a large enough effect size or large enough sample sizes.SD: standard deviation Image by Lecturio. License: CC BY-NC-SA 4.0

Trial 5 with small sample sizes and small SDs:

Due to a small SD, the difference is statistically significant at a p-value of 0.05. The SD is usually a fixed parameter in a population and cannot be changed, but the same result can be effectively obtained by increasing the sample size. The increase diminishes the impact of a large but fixed value of SD, allowing the detection of smaller differences between the groups tested. Trial 5, scenario 2: Graph demonstrating the use of Grow-A-Little with small effect size, small SDs of 10 in each, means of 150 versus 165, and small sample sizes of 30Green: control groupBlue: experimental groupThe difference is now statistically significant at a p-value of 0.05 because of the small SDs.SD: standard deviation Image by Lecturio. License: CC BY-NC-SA 4.0

## Practice Questions

Investigators involved in designing a randomized clinical trial chose a sample size that would have 90% power of detecting a 20% difference between the control and experimental groups, with a significance level (2-sided) of 5%.

### Question 1

If in actuality, there is no difference in the means, what is the chance that the study will find a statistically significant difference? What is this error called?

Answer: A type I (false positive) error. This is just a terminology question and is typical of the type of question present on the board exams, with the power inserted as a distractor. Refer to the colored graph above: if there is no difference between the 2 groups, then there would just be 1 bell curve, with the alpha cut-off describing the false positives; thus, the chance of finding a statistically significant difference is 5%, creating a type I (false positive) error, because any subject having a value in the alpha area belongs to the same population.

### Question 2

Does power increase/decrease/not change if beta is decreased?

Answer: Power increases if beta is decreased, as the power = 1 – beta. Refer to the colored graph.

### Question 3

Does power increase/decrease/not change if alpha is increased?

Answer: Power increases if alpha is increased, which increases the probability Probability Probability is a mathematical tool used to study randomness and provide predictions about the likelihood of something happening. There are several basic rules of probability that can be used to help determine the probability of multiple events happening together, separately, or sequentially. Basics of Probability of false positives; thus, increasing alpha is not a favored way of increasing power. Refer to the colored graph to see the relationship between alpha and power. In a board exam, a 2 x 2 contingency table of reality/truth versus study/test results is often used to frame this question. It is important to understand how to calculate type I and type II errors.

### Question 4

Does power increase/decrease/not change if the difference between the mean of the experimental group and that of the control group increases?

Answer: Power increases by increasing the mean difference, which is another way of increasing the ES as there is less overlap between the 2 distributions. See the colored graph.

### Question 5

Does beta increase/decrease/not change if the difference between the mean of the experimental group and that of the control group increases?

Answer: Beta decreases if the mean difference increases, as there is less overlap between the 2 populations. See the colored graph.

## References

1. Peirce, C.S. (1878). Illustrations of the Logic of Science VI. Popular Science Monthly. Vol. 13, August 1878. Retrieved on March 1, 2021, from https://en.wikisource.org/w/index.php?oldid=3592335
2. Clinical tools and calculators for medical professionals—ClinCalc. Retrieved March 19, 2021, from https://clincalc.com/
3. Power/sample size calculator. Retrieved March 20, 2021, from https://www.stat.ubc.ca/~rollin/stats/ssize/n2.html
4. Interactive statistical calculation pages. Retrieved March 20, 2021, from https://statpages.info/#Power
5. Statistical power calculator using average values. SPH Analytics. Retrieved March 20, 2021, from https://www.sphanalytics.com/statistical-power-calculator-using-average-values/
6. Otte, W.M., Tijdink, J.K., Weerheim, P.L., Lamberink, H.J., Vinkers, C.H. (2018). Adequate statistical power in clinical trials is associated with the combination of a male first author and a female last author. eLife, 7:e34412. https://doi.org/10.7554/eLife.34412
7. Bland, M. (2015). An Introduction to Medical Statistics. 4th ed., pp. 295–304.
8. Ellis, P.D. (2010). The Essential Guide to Effect Sizes. Statistical Power, Meta-Analysis, and the Interpretation of Research Results. Pp. 46–86.
9. Walters, S.J., Campbell, M.J., Machin, D. (2020). Medical Statistics, A Textbook for the Health Sciences. 5th ed, pp. 40–48, 99–133.
10. Citrome, L., Ketter, T.A. (2013). When does a difference make a difference? Interpretation of number needed to treat Number needed to treat The number needed to treat (NNT) is the number of patients that are needed to treat to prevent 1 additional adverse outcome (e.g., stroke, death). For example, if a drug has an NNT of 10, it means 10 people must be treated with the drug to prevent 1 additional adverse outcome. Number Needed to Treat, number needed to harm, and likelihood to be helped or harmed. International Journal of Clinical Practice, 67(5):407–411. https://doi.org/https://doi.org/10.1111/ijcp.12142
11. Smith, M.K. (2012). Common mistakes involving power. Retrieved March 21, 2021, from https://web.ma.utexas.edu/users/mks/statmistakes/PowerMistakes.html
12. Ioannidis, J.P., Greenland, S., Hlatky, M.A., et al. (2014). Increasing value and reducing waste in research design, conduct, and analysis. Lancet, 383(9912):166–175.
13. Coe, R. (2002). It’s the effect size, stupid: What effect size is and why it is important. https://www.leeds.ac.uk/educol/documents/00002182.htm
14. Allen, J.C. (2011). Sample size calculation for two independent groups: A useful rule of thumb. Proceedings of Singapore Healthcare, 20(2):138–140. https://doi.org/10.1177/201010581102000213
15. Lehr, R. (1992). Sixteen S-squared over D-squared: A relation for crude sample size estimates. Statistics in Medicine, 11(8), 1099–1102. https://doi.org/https://doi.org/10.1002/sim.4780110811

## Learn even more with Lecturio: Complement your med school studies with Lecturio’s all-in-one study companion, delivered with evidence-based learning strategies.

## Study on the Go Lecturio Medical complements your studies with evidence-based learning strategies, video lectures, quiz questions, and more – all combined in one easy-to-use resource.

## Esta página está disponible en Español.

🍪 Lecturio is using cookies to improve your user experience. By continuing use of our service you agree upon our Data Privacy Statement. 