GraphPad Statistics Guide

Key concepts: Statistical Power

Key concepts: Statistical Power

Previous topic Next topic No expanding text in this topic  

Key concepts: Statistical Power

Previous topic Next topic JavaScript is required for expanding text JavaScript is required for the print function Mail us feedback on this topic!  

Definitions of power and beta

Even if the treatment really does affect the outcome, you might not obtain a statistically significant difference in your experiment. Just by chance, your data may yield a P value greater than 0.05 (or whatever value, alpha, you use as your cutoff).
 
Let's assume we are comparing two means with a t test. Assume that the two means truly differ by a particular amount, and that you perform many experiments with the same sample size. Each experiment will have different values (by chance) so each t test will yield different results. In some experiments, the P value will be less than alpha (usually set to 0.05), so you call the results statistically significant. In other experiments, the P value will be greater than alpha, so you will call the difference not statistically significant.

If there really is a difference (of a specified size) between group means, you won't find a statistically significant difference in every experiment. Power is the fraction of experiments that you expect to yield a "statistically significant" P value. If your experimental design has high power, then there is a high chance that your experiment will find a "statistically significant" result if the treatment really works.
 
The variable beta is defined to equal 1.0 minus power (or 100% - power%). If there really is a difference between groups, then beta is the probability that an experiment like yours will yield a "not statistically significant" result.

How much power do I need?

The power is the chance that an experiment will result in a "statistically significant" result given some assumptions. How much power do you need? These guidelines might be useful:

If the power is less than 50% to detect some effect that you think is worth detecting, then the study is really not helpful.

Many investigators choose sample size to obtain a 80% power. This is arbitrary, but commonly used.

Ideally, your choice of acceptable power should depend on the consequence of making a Type II error.

GraphPad StatMate

GraphPad Prism does not compute statistical power or sample size, but the companion program GraphPad StatMate does.