📊

Stat Quest: Statistical Power

Jun 28, 2024

Stat Quest: Statistical Power

Introduction

  • Presenter: Josh Starburns
  • Topic: Statistical Power
  • Prerequisites: Familiarity with p-values and normal distribution
  • Note: Check other quests for p-values and normal distribution if not familiar

Example with Distributions

  • Two distributions:
    • Weights of mice on a special diet (left)
    • Weights of mice eating normal food (right)
  • Little overlap => easy to see diet difference
    • Special diet mice weigh less than normal diet mice

Statistical Test and p-values

  • Collecting measurements:
    • Small set from special diet
    • Small set from normal diet
    • Plot and compare means
  • Example Result: p-value = 0.0004
    • p-value < 0.05 => correctly reject null hypothesis (data sets aren't from same distribution)
  • Repetition: High probability that tests will give small p-value, correctly rejecting null hypothesis
  • Exception: Sometimes data may overlap => large p-value (> 0.05) => cannot reject null hypothesis

Concept of Power

  • Power: Probability of correctly rejecting the null hypothesis
    • High probability of small p-value => large power
  • Power Concepts:
    • No difference between diets => power concept doesn't apply
    • Slight difference, with overlap => power applies

Sample Size and Power

  • Small Sample Size:
    • Example: 3 mice on each diet
    • Result: p-value = 0.34 => fail to reject null hypothesis
    • Repetition: Most tests fail to reject
    • Exception: Sometimes no overlap => correctly reject null hypothesis
  • Power Analysis:
    • Increasing number of measurements => increase power
    • Determines how many measurements needed for good power

Summary

  • Power Definition: Probability of correctly rejecting null hypothesis
  • High Power: Low overlap between distributions => high chance of rejecting null hypothesis
  • Low Power: High overlap & small sample size => low chance of rejecting null hypothesis
  • Increasing Power: Collect more measurements
  • Power Analysis: Tells number of measurements needed for good power

Conclusion

  • Subscribe for more content
  • Support options: Patreon, channel membership, original songs, t-shirts, hoodies, donations
  • "Quest on!"