207 lines
7.8 KiB
R
207 lines
7.8 KiB
R
#----------------------------------------------------------#
|
|
#--- TD4 : regressions sur donnees reelles, biais de ------#
|
|
#--- variable omise, tests ------#
|
|
#----------------------------------------------------------#
|
|
|
|
library(haven) # read_dta : lire les fichiers Stata (.dta)
|
|
library(estimatr) # lm_robust : regression + se robustes
|
|
library(car) # linearHypothesis : tests de Wald
|
|
library(lmtest) # coeftest : test sur un coefficient
|
|
library(ggplot2) # graphiques
|
|
|
|
# Parametres ----
|
|
dossier <- "td4/" # dossier des fichiers .dta (a adapter)
|
|
se_type <- "HC0" # se robuste du cours ; "stata" = HC1,
|
|
# chiffres du corrige commun
|
|
|
|
#----------------------------------------------------------#
|
|
#--- Exercice 1 : le sommeil et le tennis -----------------#
|
|
#----------------------------------------------------------#
|
|
|
|
tennis <- read_dta(file.path(dossier, "Tennis-Sleep.dta"))
|
|
|
|
# ---------- Q1 : statistiques descriptives ----------------
|
|
|
|
# (la colonne Date est du texte : on la retire)
|
|
summary(tennis[, -1])
|
|
|
|
# P1 contre P2 : jeux gagnes, sommeil, alcool (moyennes)
|
|
data.frame(joueur = c("P1", "P2"),
|
|
jeux = c(mean(tennis$GamesP1), mean(tennis$GamesP2)),
|
|
sommeil = c(mean(tennis$SleepP1), mean(tennis$SleepP2)),
|
|
alcool = c(mean(tennis$AlcoholP1),
|
|
mean(tennis$AlcoholP2)))
|
|
|
|
# ---------- Q2 : DifgamesP1P2 sur SleepP1 -----------------
|
|
|
|
reg1 <- lm_robust(DifgamesP1P2 ~ SleepP1, data = tennis,
|
|
se_type = se_type)
|
|
summary(reg1)
|
|
|
|
# Memes tests avec la loi normale du cours
|
|
# (summary affiche des p-valeurs de Student : meme conclusion)
|
|
coeftest(reg1, df = Inf)
|
|
|
|
# ---------- Q3 : signe du biais de variable omise ---------
|
|
|
|
# Cours (chap. 1, prop. 4) : beta_court = beta_long + lambda * beta_G
|
|
# lambda = coef. de SleepP1 dans la regression de AlcoholP1 sur SleepP1
|
|
lambda <- coef(lm(AlcoholP1 ~ SleepP1, data = tennis))["SleepP1"]
|
|
c(correlation = cor(tennis$SleepP1, tennis$AlcoholP1),
|
|
lambda = unname(lambda))
|
|
|
|
# ---------- Q4 : on ajoute AlcoholP1 ----------------------
|
|
|
|
reg2 <- lm_robust(DifgamesP1P2 ~ SleepP1 + AlcoholP1,
|
|
data = tennis, se_type = se_type)
|
|
summary(reg2)
|
|
|
|
# La formule du cours est exacte sur l'echantillon
|
|
biais <- lambda * coef(reg2)["AlcoholP1"]
|
|
c(beta_court = unname(coef(reg1)["SleepP1"]),
|
|
beta_long = unname(coef(reg2)["SleepP1"]),
|
|
biais = unname(biais))
|
|
|
|
# ---------- Q5 : SleepP2 et AlcoholP2 ---------------------
|
|
|
|
# Peu correlees avec SleepP1 et AlcoholP1 : pas de biais
|
|
round(cor(tennis[, c("SleepP1", "AlcoholP1",
|
|
"SleepP2", "AlcoholP2")]), 2)
|
|
|
|
# ---------- Q6 : regression complete ----------------------
|
|
|
|
reg3 <- lm_robust(DifgamesP1P2 ~ SleepP1 + AlcoholP1 +
|
|
SleepP2 + AlcoholP2,
|
|
data = tennis, se_type = se_type)
|
|
summary(reg3)
|
|
coeftest(reg3, df = Inf)
|
|
|
|
# Test de significativite globale : les 4 coefficients sont nuls
|
|
linearHypothesis(reg3, c("SleepP1 = 0", "AlcoholP1 = 0",
|
|
"SleepP2 = 0", "AlcoholP2 = 0"),
|
|
test = "Chisq")
|
|
|
|
# Le coefficient de SleepP1 selon les variables incluses
|
|
modeles <- list("SleepP1 seul" = reg1,
|
|
"+ AlcoholP1" = reg2,
|
|
"+ SleepP2, AlcoholP2" = reg3)
|
|
coef_sleep <- data.frame(
|
|
modele = factor(names(modeles), levels = rev(names(modeles))),
|
|
coef = sapply(modeles, function(m) coef(m)["SleepP1"]),
|
|
se = sapply(modeles, function(m) m$std.error["SleepP1"]))
|
|
|
|
ggplot(coef_sleep, aes(x = coef, y = modele)) +
|
|
geom_vline(xintercept = 0, linetype = "dashed") +
|
|
geom_pointrange(aes(xmin = coef - 1.96 * se,
|
|
xmax = coef + 1.96 * se)) +
|
|
labs(title = expression("Coefficient de " * SleepP1 *
|
|
" et intervalle de confiance a 95 %"),
|
|
x = "Effet d'une heure de sommeil de P1 (en jeux)",
|
|
y = NULL) +
|
|
theme_minimal()
|
|
|
|
# ---------- Q7 : meme effet du sommeil pour les deux ------
|
|
|
|
# H0 : beta1 = - beta3, soit R beta0 = 0 avec R = (0, 1, 0, 1, 0)
|
|
R <- c(0, 1, 0, 1, 0)
|
|
b <- 0
|
|
linearHypothesis(reg3, hypothesis.matrix = R, rhs = b,
|
|
test = "Chisq")
|
|
|
|
# Autre methode : theta = beta1 + beta3 devient un coefficient
|
|
# y = b0 + theta x1 + b2 x2 + g (x1 - x3) + b4 x4, avec g = - beta3
|
|
tennis$DifsleepP1P2 <- tennis$SleepP1 - tennis$SleepP2
|
|
reg3_bis <- lm_robust(DifgamesP1P2 ~ SleepP1 + DifsleepP1P2 +
|
|
AlcoholP1 + AlcoholP2,
|
|
data = tennis, se_type = se_type)
|
|
summary(reg3_bis)
|
|
|
|
# Le test de H0 : theta = 0 est le test de SleepP1 (loi normale)
|
|
coeftest(reg3_bis, df = Inf)["SleepP1", ]
|
|
|
|
#----------------------------------------------------------#
|
|
#--- Exercice 2 : les prix des logements ------------------#
|
|
#----------------------------------------------------------#
|
|
|
|
hprice <- read_dta(file.path(dossier, "hprice1.dta"))
|
|
|
|
# ---------- Q0 : statistiques descriptives ----------------
|
|
|
|
# price et assess en milliers de $, sqrft et lotsize en pieds carres
|
|
summary(hprice[, c("price", "assess", "bdrms", "lotsize", "sqrft")])
|
|
|
|
# ---------- Q1-Q2 : l'evaluation est-elle rationnelle ? ---
|
|
|
|
# price = b0 + b1 assess + u ; H0 : b0 = 0 et b1 = 1
|
|
reg_prix <- lm_robust(price ~ assess, data = hprice,
|
|
se_type = se_type)
|
|
summary(reg_prix)
|
|
|
|
# Test joint : R = I2, b = (0, 1)'
|
|
R <- diag(2)
|
|
b <- c(0, 1)
|
|
linearHypothesis(reg_prix, hypothesis.matrix = R, rhs = b,
|
|
test = "Chisq")
|
|
|
|
# Tests simples (pour une egalite : Chisq = z^2)
|
|
linearHypothesis(reg_prix, "(Intercept) = 0", test = "Chisq")
|
|
linearHypothesis(reg_prix, "assess = 1", test = "Chisq")
|
|
|
|
# Pourquoi la difference ? Les deux estimateurs sont tres correles
|
|
V <- vcov(reg_prix)
|
|
V[1, 2] / sqrt(V[1, 1] * V[2, 2])
|
|
|
|
# Ellipse de confiance a 95 % pour (b0, b1) et rectangle forme
|
|
# par les deux intervalles a 95 % pris separement
|
|
centre <- coef(reg_prix)
|
|
t <- seq(0, 2 * pi, length.out = 200)
|
|
cercle <- cbind(cos(t), sin(t)) * sqrt(qchisq(0.95, df = 2))
|
|
ellipse <- as.data.frame(sweep(cercle %*% chol(V), 2, centre, "+"))
|
|
names(ellipse) <- c("b0", "b1")
|
|
se <- sqrt(diag(V))
|
|
|
|
ggplot(ellipse, aes(x = b0, y = b1)) +
|
|
annotate("rect", fill = "grey85",
|
|
xmin = centre[1] - 1.96 * se[1],
|
|
xmax = centre[1] + 1.96 * se[1],
|
|
ymin = centre[2] - 1.96 * se[2],
|
|
ymax = centre[2] + 1.96 * se[2]) +
|
|
geom_path(linewidth = 1) +
|
|
annotate("point", x = centre[1], y = centre[2], size = 2) +
|
|
annotate("point", x = 0, y = 1, size = 3, colour = "red") +
|
|
annotate("text", x = 0, y = 1.03, colour = "red",
|
|
label = "H0 : (0, 1)") +
|
|
labs(title = expression("Region de confiance a 95 % pour (" *
|
|
beta[0] * ", " * beta[1] * ")"),
|
|
x = expression(beta[0]), y = expression(beta[1])) +
|
|
theme_minimal()
|
|
|
|
# Ecart moyen evaluation - prix : la droite estimee passe par
|
|
# (assess moyen ; price moyen), l'evaluation est trop elevee
|
|
lm_robust(I(assess - price) ~ 1, data = hprice, se_type = se_type)
|
|
|
|
# ---------- Q3 : effet d'une chambre de 150 pieds carres --
|
|
|
|
reg_log <- lm_robust(lprice ~ sqrft + bdrms, data = hprice,
|
|
se_type = se_type)
|
|
summary(reg_log)
|
|
|
|
# Effet = 150 * b_sqrft + b_bdrms (modele log-niveau)
|
|
effet <- sum(c(0, 150, 1) * coef(reg_log))
|
|
c(effet_en_log = effet, effet_en_pct = 100 * (exp(effet) - 1))
|
|
|
|
# ---------- Q4 : regression qui donne directement l'effet -
|
|
|
|
# theta = 150 b1 + b2 donc b2 = theta - 150 b1 :
|
|
# lprice = b0 + b1 (sqrft - 150 bdrms) + theta bdrms + u
|
|
hprice$sqrft150b <- hprice$sqrft - 150 * hprice$bdrms
|
|
reg_theta <- lm_robust(lprice ~ sqrft150b + bdrms, data = hprice,
|
|
se_type = se_type)
|
|
summary(reg_theta)
|
|
|
|
# Intervalle de confiance a 95 % (loi normale du cours)
|
|
theta_hat <- coef(reg_theta)["bdrms"]
|
|
se_theta <- reg_theta$std.error["bdrms"]
|
|
c(borne_basse = unname(theta_hat - 1.96 * se_theta),
|
|
borne_haute = unname(theta_hat + 1.96 * se_theta))
|