Power

DataCamp Vid: Power and sample size

Why do we need power? Makes sure that, when we do reject the null, it was done correctly instead of just due to small data.

Parameter	Meaning
d	Effect size (Cohen's d for means, or h for proportions). Measures standardized difference between 2 groups. Cohen’s effect size conventions (small = 0.2, medium = 0.5, large = 0.8).
power	Probability of correctly rejecting the null. 80% --> "If there is an effect, I want an 80% of detecting it."
sig.level
type = "one.sample"
alternative = "two.sided"

How many subjects do we need?

# Means
library(pwr)
pwr.t.test(d = 0.81,       # Cohen
           #n = ,          # what we're looking for
           power = 0.8,
           sig.level = 0.05,
           type = "one.sample",
           alternative = "two.sided")

# n = 14.3 people per test/control, round up to 15. (Since we testing against some known average, it'll just be 15 people total.)

# Proportions
library(pwr)
pwr.p.test(h = 0.5,        # Cohen
           #n = ,          # what we're looking for
           power = 0.8,
           sig.level = 0.05, 
           alternative = "two.sided")

If we already ran the test, how much power is there?

# Means
library(pwr)
pwr.t.test(d = 0.81, 
           n = 20,
           # power = 0.8,          # what we're looking for
           sig.level = 0.045,
           type = "one.sample",
           alternative = "two.sided")

# power = 0.92. This is larger than 80%, so we're good to go.

# Proportions
library(pwr)
pwr.p.test(h = 0.81,  # Cohen's h for proportions effect size
           n = 20,    # Sample size
           # power = NA,          # Power is what we're calculating
           sig.level = 0.045,    # Significance level
           alternative = "two.sided")   # Two-sided test