Two sample t-test

Independent sample t-test, two sample t-test
Difference of Two means Hypothesis test 등 같은 의미

Theory

가정

우리는 이미
central limit theorem 문서와
statistical review 문서의 expected value and variance properties 와
mean and variance of the sample mean 문서를 통해서 아래를 알고 있다.
\begin{eqnarray*} \overline{X} & \sim & \left( \mu, \;\; \frac{\sigma^2}{n} \right) \\ & & \text{in other words, } \\ E \left[ \overline{X} \right] & = & \mu \\ Var \left[ \overline{X} \right] & = & \frac{\sigma^2}{n} \\ & & \text {Assuming that X1 and X2 are independent } \\ \overline{X_{1}} & \sim & \left( \mu_{1}, \frac{\sigma^2_{1}}{n_{1}} \right) \\ \overline{X_{2}} & \sim & \left( \mu_{2}, \frac{\sigma^2_{2}}{n_{2}} \right) \\ & & \text{note that } n_{1}, n_{2} \text{ are sample size.} \\ & & \text{and } \\ & & \frac{\sigma^2_{1}}{n_{1}} = Var \left[ \overline{X_{1}} \right] \\ \end{eqnarray*}

두 샘플 평균들의 차이를 모아 놓은 집합의 (distribution of sample mean difference) 성격은 아래와 같을 것이다.
\begin{eqnarray*} E \left[ \overline{X_{1}} - \overline{X_{2}} \right] & = & \mu_{1} - \mu_{2} \;, \;\;\; \text{and} \\ Var \left[ \overline{X_{1}} - \overline{X_{2}} \right] & = & Var \left[ \overline{X_{1}} \right] + Var \left[ \overline{X_{2}} \right] \\ & = & \frac{\sigma^2_{1}}{n_{1}} + \frac{\sigma^2_{2}}{n_{2}} \\ \text{SE}_{\overline{X_{1}} - \overline{X_{2}}} & = & \text{SE}_{\text{diff}} \\ & = & \sqrt { \frac{\sigma^2_{1}}{n_{1}} + \frac{\sigma^2_{2}}{n_{2}} } \\ \\ & & \text{If variance of each population} \text{is unknown,} \\ & & \text{we use sample variances, instead of using } \sigma \text{.} \\ & & \text{If degrees of freedom for each group is different} \\ & & \text{we use the following method to obtain pooled variance, } \; \text{s}^{2}_{\text{p}}\\ \text{s}^{2}_{\text{p}} & = & \dfrac {\text{SS}_{1} + \text{SS}_{2}} {\text{df}_{1} + \text{df}_{2} } \\ & & \text{Hence, } \\ \text{SE}_{\text{diff}} & = & \sqrt {\frac{\text{s}^{2}_{\text{p}}}{n_1} + \frac{\text{s}^{2}_{\text{p}}}{n_2} } \\ \end{eqnarray*} 

rm(list=ls())
rnorm2 <- function(n,mean,sd){ 
  mean+sd*scale(rnorm(n)) 
}
ss <- function(x) {
  sum((x-mean(x))^2)
}

N.p <- 1000000
m.p <- 100
sd.p <- 10

set.seed(101)
p1 <- rnorm2(N.p, m.p, sd.p)
mean(p1)
sd(p1)

p2 <- rnorm2(N.p, m.p+10, sd.p)
mean(p2)
sd(p2)

sz1 <- sz2 <- 50
df1 <- sz1 - 1
df2 <- sz2 - 1
df.tot <- df1 + df2

iter <- 100000
mdiffs <- rep(NA, iter)
means.s1 <- rep(NA, iter)
means.s2 <- rep(NA, iter)
tail(mdiffs)

for (i in 1:iter) {
  # means <- append(means, mean(sample(p1, s.size, replace = T)))
  s1 <- sample(p1, sz1, replace = T)
  s2 <- sample(p2, sz2, replace = T)
  means.s1[i] <- mean(s1)
  means.s2[i] <- mean(s2)
  mdiffs[i] <- mean(s1)-mean(s2)
}
# 정리, 증명에 의한 계산 
mu <- mean(p1) - mean(p2)
# var(x1bar-x2bar) = var(x1bar) + var(x2bar)
# var(x1bar) = var(x1)/n, n = sample size
ms <- var(p1)/sz1 + var(p2)/sz2 
se <- sqrt(ms)

mu
ms
se

# 시뮬레이션에 의한 집합 (distribution)
# mdiffs
m.diff <- mean(mdiffs)
var.diff <- var(mdiffs)
sd.diff <- sd(mdiffs)
m.diff
var.diff
sd.diff

var(means.s1)
var(p1)/sz1
var(means.s2)
var(p2)/sz2
var(means.s1-means.s2)
var(means.s1) + var(means.s2) - 2 * cov(means.s1, means.s2)
# 두 집합이 완전히 독립적일 때 cov = 0 이므로
var(p1)/sz1 + var(p2)/sz2

var.diff <- (var(p1)/sz1) + (var(p2)/sz2)
var.diff
sqrt(var.diff)
se.diff <- sqrt(var.diff)
se.diff

# 이것을 그래프로 그려보면
hist(mdiffs, breaks=50)
abline(v=mean(mdiffs), 
       col="black", lwd=2)

se.diff
one <- qt(1-(.32/2), df=df.tot)
two <- qt(1-(.05/2), df=df.tot)
thr <- qt(1-(.01/2), df=df.tot)

ci68 <- se.diff*one
ci95 <- se.diff*two
ci99 <- se.diff*thr

abline(v=c(m.diff-ci68, m.diff-ci95, m.diff-ci99,
           m.diff+ci68, m.diff+ci95, m.diff+ci99), 
       col=c("green", "blue", "black"), lwd=2)
text(x=m.diff, y=iter/12, 
     labels=paste("mu_1-mu_2"), 
col="black", pos = 4, cex=1.2)

text(x=m.diff-ci95, y=iter/12, 
     labels=paste("-SE_diff"), 
     col="blue", pos = 4, cex=1.2)
text(x=m.diff+ci95, y=iter/12, 
     labels=paste("SE_diff"),
     col="blue", pos = 4, cex=1.2)