Galton peas (nonconstant variance and weighted least squares)

• Load the galton data.
• Fit an ordinary least squares (OLS) simple linear regression model of Progeny vs Parent.
• Fit a weighted least squares (WLS) model using weights = \(1/{SD^2}\).
• Create a scatterplot of the data with a regression line for each model.

galton <- read.table("~/path-to-data/galton.txt", header=T)
attach(galton)

model.1 <- lm(Progeny ~ Parent)
summary(model.1)
#             Estimate Std. Error t value Pr(>|t|)    
# (Intercept) 0.127029   0.006993  18.164 9.29e-06 ***
# Parent      0.210000   0.038614   5.438  0.00285 ** 

model.2 <- lm(Progeny ~ Parent, weights=1/SD^2)
summary(model.2)
#             Estimate Std. Error t value Pr(>|t|)    
# (Intercept) 0.127964   0.006811  18.787 7.87e-06 ***
# Parent      0.204801   0.038155   5.368  0.00302 ** 

plot(x=Parent, y=Progeny, ylim=c(0.158,0.174),
     panel.last = c(lines(sort(Parent), fitted(model.1)[order(Parent)], col="blue"),
                    lines(sort(Parent), fitted(model.2)[order(Parent)], col="red")))
legend("topleft", col=c("blue","red"), lty=1,
       inset=0.02, legend=c("OLS", "WLS"))

detach(galton)

Computer-assisted learning (nonconstant variance and weighted least squares)

• Load the ca_learning data.
• Create a scatterplot of the data.
• Fit an OLS model.
• Plot the OLS residuals vs num.responses.
• Plot the absolute OLS residuals vs num.responses.
• Calculate fitted values from a regression of absolute residuals vs num.responses.
• Fit a WLS model using weights = \(1/{(\text{fitted values})^2}\).
• Create a scatterplot of the data with a regression line for each model.
• Plot the WLS standardized residuals vs num.responses.

ca_learning <- read.table("~/path-to-data/ca_learning_new.txt", header=T)
attach(ca_learning)

plot(x=num.responses, y=cost)

model.1 <- lm(cost ~ num.responses)
summary(model.1)
#               Estimate Std. Error t value Pr(>|t|)    
# (Intercept)    19.4727     5.5162   3.530  0.00545 ** 
# num.responses   3.2689     0.3651   8.955 4.33e-06 ***
# ---
# Residual standard error: 4.598 on 10 degrees of freedom
# Multiple R-squared:  0.8891,  Adjusted R-squared:  0.878 
# F-statistic: 80.19 on 1 and 10 DF,  p-value: 4.33e-06

plot(num.responses, residuals(model.1))
plot(num.responses, abs(residuals(model.1)))

wts <- 1/fitted(lm(abs(residuals(model.1)) ~ num.responses))^2

model.2 <- lm(cost ~ num.responses, weights=wts)
summary(model.2)
#               Estimate Std. Error t value Pr(>|t|)    
# (Intercept)    17.3006     4.8277   3.584  0.00498 ** 
# num.responses   3.4211     0.3703   9.238 3.27e-06 ***
# ---
# Residual standard error: 1.159 on 10 degrees of freedom
# Multiple R-squared:  0.8951,  Adjusted R-squared:  0.8846 
# F-statistic: 85.35 on 1 and 10 DF,  p-value: 3.269e-06

plot(x=num.responses, y=cost, ylim=c(50,95),
     panel.last = c(lines(sort(num.responses), fitted(model.1)[order(num.responses)], col="blue"),
                    lines(sort(num.responses), fitted(model.2)[order(num.responses)], col="red")))
legend("topleft", col=c("blue","red"), lty=1,
       inset=0.02, legend=c("OLS", "WLS"))

plot(num.responses, rstandard(model.2))

detach(ca_learning)

Market share (nonconstant variance and weighted least squares)

• Load the marketshare data.
• Fit an OLS model.
• Plot the OLS residuals vs fitted values with points marked by Discount.
• Use the tapply function to calculate the residual variance for Discount=0 and Discount=1.
• Fit a WLS model using weights = 1/variance for Discount=0 and Discount=1.
• Plot the WLS standardized residuals vs fitted values.

marketshare <- read.table("~/path-to-data/marketshare.txt", header=T)
attach(marketshare)

model.1 <- lm(MarketShare ~ Price + P1 + P2)
summary(model.1)
  #             Estimate Std. Error t value Pr(>|t|)    
# (Intercept)  3.19592    0.35616   8.973 3.00e-10 ***
# Price       -0.33358    0.15226  -2.191   0.0359 *  
# P1           0.30808    0.06412   4.804 3.51e-05 ***
# P2           0.48431    0.05541   8.740 5.49e-10 ***

plot(fitted(model.1), residuals(model.1), col=Discount+1)
vars <- tapply(residuals(model.1), Discount, var)
#          0          1 
# 0.01052324 0.02680546 

wts <- Discount/vars[2] + (1-Discount)/vars[1]

model.2 <- lm(MarketShare ~ Price + P1 + P2, weights=wts)
summary(model.2)
#             Estimate Std. Error t value Pr(>|t|)    
# (Intercept)  3.17437    0.35671   8.899 3.63e-10 ***
# Price       -0.32432    0.15291  -2.121   0.0418 *  
# P1           0.30834    0.06575   4.689 4.89e-05 ***
# P2           0.48419    0.05422   8.930 3.35e-10 ***

plot(fitted(model.2), rstandard(model.2), col=Discount+1)

detach(marketshare)

Home price (nonconstant variance and weighted least squares)

• Load the realestate data.
• Calculate log transformations of the variables.
• Fit an OLS model.
• Plot the OLS residuals vs fitted values.
• Calculate fitted values from a regression of absolute residuals vs fitted values.
• Fit a WLS model using weights = \(1/{(\text{fitted values})^2}\).
• Plot the WLS standardized residuals vs fitted values.

realestate <- read.table("~/path-to-data/realestate.txt", header=T)
attach(realestate)

logY <- log(SalePrice)
logX1 <- log(SqFeet)
logX2 <- log(Lot)

model.1 <- lm(logY ~ logX1 + logX2)
summary(model.1)
#             Estimate Std. Error t value Pr(>|t|)    
# (Intercept)  4.25485    0.07353  57.864  < 2e-16 ***
# logX1        1.22141    0.03371  36.234  < 2e-16 ***
# logX2        0.10595    0.02394   4.425 1.18e-05 ***

plot(fitted(model.1), residuals(model.1))

wts <- 1/fitted(lm(abs(residuals(model.1)) ~ fitted(model.1)))^2

model.2 <- lm(logY ~ logX1 + logX2, weights=wts)
summary(model.2)
#             Estimate Std. Error t value Pr(>|t|)    
# (Intercept)  4.35189    0.06330  68.755  < 2e-16 ***
# logX1        1.20150    0.03332  36.065  < 2e-16 ***
# logX2        0.07924    0.02152   3.682 0.000255 ***

plot(fitted(model.2), rstandard(model.2))

detach(realestate)