Assignment 1

Fahim Muntasir_182-001

2024-10-31

# Loading required packages

library(tidyverse)
library(alr4)
library(ggplot2)

Answer to the question no. 2.3.1

UBSprices %>%
mutate(log_rice2003 = log(rice2003), log_rice2009 = log(rice2009)) %>%
ggplot(mapping = aes(x = log_rice2003, y = log_rice2009)) +
geom_point(shape = 1) +
coord_cartesian(xlim = c(1.5, 4.5), ylim = c(1.5, 4.5)) +
scale_x_continuous(breaks = seq(1.5, 4.5, by = 0.5)) +
scale_y_continuous(breaks = seq(1.5, 4.5, by = 0.5)) +
geom_abline(slope = 1, intercept = 0, lwd = 0.9, linetype = "solid") +
geom_smooth(method = "lm", se = FALSE, linetype = "dashed", color = "bl
annotate("text", x = 4.2, y = 1.5, label = "Dashed line : OLS,
Solid line : y = x",
color = "red", size = 3)

1
 4.5

4.0

3.5
log_rice2009

3.0

2.5

2.0

Dashed line : OLS,

1.5
Solid line : y = x
1.5 2.0 2.5 3.0 3.5 4.0 4.5
log_rice2003

[From the continuation of problem 2.2] We can see that the from the scatter
plot of y = rice2009 versus x = rice2003, th points are densely situated around
the point (10,10) but otherwise the points are dispersed. Moreover, there are
some potential leverage points and outliers too. Also in this case a linear
relationship is not evident.
When we use log transformation, we see that the scatter plot becomes a
lot more evenly dispersed and it also eliminates unusual values. Using log
scales, the relationship between the transformed variables become a lot more
evident, visually.
We know, simple linear regression fits a straight line between the independent
and dependent variables. As using log scale makes the scatter plot more linear
and more evenly dispersed, using log scales is preferable in this case.

2
Answer to the question no. 2.3.2

Here the original model is E(y | x) = γ0 .xβ1 . According to the question, taking
log on both sides and assuming log(E(y|x)) ≈ E(log(y|x), we can rewrite the
model as
E(log(y|x) ≈ log(γ0 ) + β1 .log(x),
where
β0 = log(γ0 ).

1. Interpretation of β0 : As β0 = log(γ0 ) , if γ0 = 1, then β0 = 1. In

the original model, γ0 is the baseline or scaling factor of the expected
value of y when x is 1. Thus β0 represents the expected value of log(y)
i.e the transformed dependent variable when x = 1 i.e the transformed
independent variable log(x) is equal to 0 .
2. Interpretation of β1 : From the original model :
E(y|x) = γ0 .xβ1 .........(i)
E(y|X = nx) = γ0 .(nx)β1 .........(ii)
Dividing (ii) by (i), we get,
E(y|nx)/E(y|x) = nβ1

So, if we use the original model to interpret β1 , it gives the proportional

change in y (on an average) with respect to x. Specifically, β1 indicates
percentage change in y associated with 1% change in x. If β1 = 1, then if x
changes to 2x i.e. 100% increase in x will cause the expected value of y to be
twice as much or 100% increase.

3
From the transformed model :
E(log(y)|x) = log(γ0 ) + β1 .log(x).........(iii)
E(log(y)|nx) = log(γ0 ) + β1 .log(nx).........(iv)
Subtracting (iv) from (iii), we get,
E(log(y)|2x) − E(log(y)|x) = β1 .log(n)

If we use the transformed model to interpret β1 , it gives the linear change in

log(y) (on an average) with respect to log(x). Specifically, β1 indicates linear
change in log(y) associated with 1% change in log(x). If β1 = 1, then if log(x)
changes to log(2x) i.e. 100% increase in x will cause the expected value of y
to increase by log(2).

• If β1 > 1, response variable increases at a faster rate than the predictor

variable.
• If β1 < 1, response variable increases at a slower rate than the predictor
variable.
• If β1 = 1, response variable increases at the same rate than the predictor
variable.

4
Answer to the question no. 2.16.1

UN11_log <- UN11 %>%

mutate(log_fertility = log(fertility), log_ppgdp = log(ppgdp))

slr <- function(x, y){

y <- y
x <- x
n <- length(x)
mean_y <- mean(y)
mean_x <- mean(x)
sxx <- sum((x - mean_x)ˆ2)
sxy <- sum((x - mean_x) * (y - mean_y))
syy <- sum((y - mean_y)ˆ2)
beta_1 <- sxy / sxx
beta_0 <- mean_y - (beta_1 * mean_x)
rss <- syy - ( sxyˆ2 / sxx )
est_error_var <- rss / (n - 2)
est_beta1_var <- est_error_var / sxx
est_beta0_var <- est_error_var * ((1 / n) + (((mean_x)ˆ2) /sxx))
R_squared <- 1 - (rss / syy )
residual <- (y - beta_0 - beta_1*x)
return(list("n" = n,
"Mean_x" = mean_x, "Mean_y" = mean_y,
"SXX" = sxx, "SYY" = syy, "SXY" = sxy,
"Intercept" = beta_0, "Slope" = beta_1,
"Error_Variance" = est_error_var,
"Beta_0_Variance" = est_beta0_var,
"Beta_1_Variance" = est_beta1_var,
"Coefficient_of_determination" = R_squared,
"Residuals" = residual))
}

result <- slr(y = UN11_log$log_fertility, x = UN11_log$log_ppgdp)

5
So, using the user defined function, using log_fertility as dependent vari-
able and log_ppgdp as independent variable and applying the simple linear
regression, we get,

Intercept = 2.6655073
Slope = −0.2071498

6
Answer to the question no. 2.16.2

UN11_log %>%
ggplot(mapping = aes(x = log_ppgdp, y = log_fertility)) +
geom_point() +
geom_abline(intercept = result$Intercept, slope = result$Slope,
lwd = 0.9, color = "orange")

2.0

1.5
log_fertility

1.0

0.5

6 8 10
log_ppgdp

7
Answer to the question no. 2.16.3

Here, we wish to test the hypotheses,

H0 : β1 = 0
H1 : β1 < 0
Let α = 0.05

beta_1_null <- 0
alpha <- 0.05
t_cal <- (result$Slope - beta_1_null) / (sqrt(result$Beta_1_Variance))
t_tab <- qt(p = 1 - alpha, df = result$n - 2, lower.tail = TRUE)

if_else(abs(t_cal) > t_tab,

"Reject the Null Hypothesis",
"Do not Reject the Null Hypothesis")

## [1] "Reject the Null Hypothesis"

So, we can conclude with 95% confidence that, the slope is negative.

8
Answer to the question no. 2.16.4

result$Coefficient_of_determination

## [1] 0.525985

Interpretation: As the “Coefficient of determination”, R2 = 0.5259, we can

say 52.59% of the total variation in the response variable i.e log_fertility can
be explained/captured by the explanatory/predictor variable i.e log_ppgdp.

9
Answer to the question no. 2.16.5

x_pred <- log(1000)

pred_y <- function(x_pred, beta0, beta1) {

y_pred <- beta0 + (beta1*x_pred)
return (y_pred)
}

y_predicted <- pred_y (x_pred = x_pred,

beta0 = result$Intercept,
beta1 = result$Slope)

var_pred <- result$Error_Variance +

(result$Error_Variance) * ((1 / result$n) + (x_pred - result$M

se_pred <- sqrt(var_pred)

a <- y_predicted - qt(p = 1 - 0.05/2, df = result$n - 2 )

b <- y_predicted + qt(p = 1 - 0.05/2, df = result$n - 2 )
pred_intrvl_transform <- c(a, b)
pred_intrvl <- c(exp(a), exp(b))

If ppgdp = 1000,
˜
log_f ertility * = 1.2345673.
˜
sepred(log_f ertility * |x∗ = 1000) = 0.308653
˜
P rediction interval(log_f ertility * |x∗ = 1000) = (−0.7375117, 3.2066463)
˜
P rediction interval(f ertility * |x∗ = 1000) = (0.4783026, 24.6961248)

10
Answer to the question no. 2.16.6

UN11$region <- as.character(UN11$region)

#1
max_fertility_index <- which.max(UN11$fertility)
max_fertility <- c("Locality" = UN11$region[max_fertility_index],
"Fertility" = UN11$fertility[max_fertility_index])

#2
min_fertility_index <- which.min(UN11$fertility)
min_fertility <- c("Locality" = UN11$region[min_fertility_index],
"Fertility" = UN11$fertility[min_fertility_index])

#3
UN11_log <- UN11_log %>%
mutate("Residuals" = result$Residuals)

ordered_residual <- sort(UN11_log$Residuals, decreasing = TRUE)

#Two localities with largest positive residual :

high_resid <- rbind.data.frame(UN11_log[which(UN11_log$Residuals == orde
UN11_log[which(UN11_log$Residuals == orde

#Two localities with negative positive residual :

low_resid <- rbind.data.frame(UN11_log[which(UN11_log$Residuals == ordere
UN11_log[which(UN11_log$Residuals == ordere

11
(1): Locality with highest value of fertility is :

max_fertility[["Locality"]]

## [1] "Africa"

(2): Locality with highest value of fertility is :

min_fertility[["Locality"]]

## [1] "Europe"

(3): Two localities with largest positive residual (in log scales) is

high_resid["region"]

## region
## Equatorial Guinea Africa
## Angola Africa

Two localities with largest negative residual (in log scales) is

low_resid["region"]

## region
## Bosnia and Herzegovina Europe
## Moldova Europe

12