====== 하나의 독립변인이 갖는 고유의 영향력 혹은 설명력 파악하기 ======
[[:multiple regression]] 참조
[[:partial and semipartial correlation]] 참조
<code>
datavar <- read.csv("http://commres.net/wiki/_media/regression01-bankaccount.csv") 
datavar
colnames(datavar) <- c("y", "x1", "x2")
datavar
attach(datavar)

library(ggplot2) # ggplot 사용을 위해서 
# install.packages("ggplot2")

lm.y.x1 <- lm(y~x1, data=datavar)
summary(lm.y.x1)
str(lm.y.x1)
intercept.y <- lm.y.x1$coefficients[1]
slope.x <- lm.y.x1$coefficients[2]
intercept.y
slope.x
ggplot(data=datavar, aes(x1,y)) +
  geom_point(color="blue", size=1.5, pch=2) +
  geom_hline(aes(yintercept=mean(y)), size=1.5, color="red") +
  geom_abline(intercept=intercept.y, slope=slope.x, size=1.5, color="darkgreen")

# recapping
ss.tot <- var(y)*(length(y)-1)
ss.tot
ss.res <- sum(lm.y.x1$residuals^2)
ss.res
ss.reg <- sum((lm.y.x1$fitted.values-mean(y))^2)
ss.reg
ss.tot
ss.res + ss.reg

summary(lm(lm.y.x1$residuals~x1))
summary(lm((lm.y.x1$fitted.values~x1)))
summary(lm((lm.y.x1$fitted.values-mean(y))~x1))
# recapping ends

lm.y.x2 <- lm(y~x2, data=datavar)
summary(lm.y.x2)

intercept.y <- lm.y.x2$coefficients[1]
slope.x <- lm.y.x2$coefficients[2]
intercept.y
slope.x
ggplot(data=datavar, aes(x2,y)) +
  geom_point(color="blue", size=1.5, pch=2) +
  geom_hline(aes(yintercept=mean(y)),size=1.5, color="red") +
  geom_abline(intercept=intercept.y, slope=slope.x, size=1.5, color="darkgreen")


lm.y.x1x2 <- lm(y~x1+x2, data=datavar)
summary(lm.y.x1x2)

##################################
res.lm.y.x2 <- resid(lm.y.x2) # y - x2 

# x2를 제외하기 위한 계산
lm.x1.x2 <- lm(x1~x2, data=datavar)
res.lm.x1.x2 <- resid(lm.x1.x2) # x1 - x2

lm(res.lm.y.x2~res.lm.x1.x2)
summary(lm(res.lm.y.x2~res.lm.x1.x2)) # x2를 y, x1에서 모두 제외한 분석
# or 
cor(res.lm.y.x2, res.lm.x1.x2)
# cor제곱 값은 x2의 설명력을 제어 혹은 제외한 후에 보는 
# y와 x1간의 r 제곱값이다 
cor(res.lm.y.x2, res.lm.x1.x2)^2
summary(lm(res.lm.y.x2~res.lm.x1.x2))$r.squared

############################
# http://commres.net/wiki/partial_and_semipartial_correlation 문서 확인할 것
# install.packages("ppcor")
library(ppcor)
pcor(datavar)

datavar # data 확인
pcor.test(y, x1, x2) # x2제어 하고, y와 x1과의 correlation값을 묻는 펑션
summary(lm(res.y.minus.x2~res.x1.minus.x2))
r.sq <- (summary(lm(res.lm.y.x2~res.lm.x1.x2)))$r.squared
r.sq 
sqrt(r.sq)

# 같은 논리로 semi-partial을 살펴보면 공통 영향력의 크기를 
# 알아낼 수 있다. 
sp.y.x1 <- spcor.test(y, x1, x2)
sp.y.x2 <- spcor.test(y, x2, x1)

b <- sp.y.x1$estimate^2
c <- sp.y.x2$estimate^2
b
c

temp.bc <- b + c
temp.bc
summary(lm.y.x1x2)$r.squared # b+c+d / a+b+c+d
common.area <- summary(lm.y.x1x2)$r.squared - temp.bc
common.area

# common.area가 맞는지 확인 
# b + common.area는 lm(y~x1,data=datavar)에서의 r.square 값이어야 한다
summary(lm.y.x1)$r.squared
b + common.area
# 마찬가지로 c + common.area should be the same as
# summary(lm.y.x2)$r.squared
summary(lm.y.x2)$r.squared
c + common.area
</code>

====== 기울기 b에 대한 통계학적인 테스트 ======
[[:multiple regression]]에서 앞 쪽에 기울기에 대한 테스트 부분
regression 문서에서 [[:regression#eg_3_simple_regressionadjusted_r_squared_slope_test]] 부분

<code>
################################################
# standard error of coefficient의 의미 
# b의 사선을 중심으로 에러의 분포가 정규분포를 
# 이룬다고 할 때, 그 때의 standard error 값.
# 따라서, fitted.value +- 2 * se 값이 에러의
# 범위를 (range) 알려주는 방법.
# 단, 2 = t critical value로 대체해야 
################################################

summary(lm.y.x1x2)
mean(y)
mean(x1)
mean(x2)
intercept <- lm.y.x1x2$coefficients[1]
b1 <- lm.y.x1x2$coefficients[2]
b2 <- lm.y.x1x2$coefficients[3]
intercept
b1
b2

ggplot(datavar,aes(x = x1, y = y)) + geom_point() + geom_smooth(method = "lm")
ggplot(datavar,aes(x = x2, y = y)) + geom_point() + geom_smooth(method = "lm")

summary(lm.y.x1)
a <- lm.y.x1$coefficients[1]
b <- lm.y.x1$coefficients[2]
# y hat = a + b x1
# x1 = 200 ?
p <- as.data.frame(200)
p
colnames(p) <- "x1"
p

predict(lm.y.x1, newdata=p)


lm.y.x1$coefficients
pred <- predict(lm.y.x1,newdata=p)
summary(lm.y.x1)
t.critical <- qt(.975,9)
t.critical

(summary(lm.y.x1))$coefficients 
# coefficient 값 중 독립변인의 se값은 $coefficient[4]
# 절편의 se는 [3]

se.x1 <- (summary(lm.y.x1))$coefficients[4]
se.x1
pred.range.down <- pred - (t.critical*se.x1)
pred.range.up <- pred + (t.critical*se.x1)
pred.range.down
pred.range.up

summary(lm.y.x2)
a <- lm.y.x2$coefficients[1]
b <- lm.y.x2$coefficients[2]
a
b

# y hat = a + b x2
# x2 = 5 명일때?
p <- as.data.frame(5)
colnames(p) <- "x2"

lm.y.x2$coefficients
pred <- predict(lm.y.x2,newdata=p)
pred

t.critical <- qt(.975,9)
t.critical
se.x2 <- (summary(lm.y.x2))$coefficients[4]
se.x2

lim <- t.critical*se.x2

pred.range.down <- pred-lim 
pred.range.up <- pred+lim
pred.range.down
pred.range.up
# or 
pred.range <- pred + c(-lim, lim)
pred.range


# Note that the range depends upon the se value
summary(lm.y.x1)
summary(lm.y.x2)
summary(lm.y.x1x2)


dvar <- read.csv("http://commres.net/wiki/_media/elemapi2_.csv", sep = "\t", fileEncoding="UTF-8-BOM")

mod1 <- lm(api00 ~ ell + acs_k3 + avg_ed + meals, data=dvar)
mod2 <- lm(api00 ~ ell + avg_ed + meals + acs_k3, data=dvar)
mod3 <- lm(api00 ~ meals + acs_k3 + avg_ed + ell, data=dvar)
summary(mod1)
summary(mod2)
summary(mod3)

anova(mod1)
anova(mod2)
anova(mod3)
</code>