Description
1. Using the stopping data set (from package alr4), fit a simple linear regression model with
the stopping distance as the response and speed as the predictor. Answer the following
parts.
In parts (b) through (g), you should draw a specific conclusion and clearly refer to the
diagnostic tool(s) (plots or statistics) you used to draw your conclusion.
(a) [2 pts] Produce the four default diagnostic plots given by R.
(b) [2 pts] Using an appropriate diagnostic plot, check whether the assumed mean
function is appropriate.
(c) [2 pts] Check the assumption of constant variance.
(d) [2 pts] What is the largest (most positive) least-squares residual value? What is the
smallest (most negative) least-squares residual value?
(e) [2 pts] Identify one observation that has the largest leverage value.
(f) [2 pts] Check for outliers. (You should use diagnostic plots — a formal test is not
necessary.)
(g) [2 pts] Check for influential points.
2. [14 pts] Using the drugcost data set (from package alr4), fit a model with COST as the
response and all of the other variables as predictors. Then answer the same parts as in
Problem 1.
3. Using the fuel2001 data set (from package alr4), fit a model with FuelC as the response
and Tax, Drivers, and Income as predictors. Answer the following.
(a) [2 pts] Produce a plot of the standardized residuals ri versus the ordinary (least
squares) residuals ˆei
. (Show R code.)
(b) [2 pts] The points in this plot do not exactly fall on a straight line. Briefly explain
why. [ Hint: What is the formula for the standardized residuals? ]
(c) [2 pts] List the studentized residuals ti (which are used as test statistics in the Mean
Shift Test).
(d) [2 pts] Perform all Mean Shift Tests without Bonferroni adjustment, using α = 0.05.
Which states are identified as outliers?
(e) [2 pts] Perform all Mean Shift Tests with Bonferroni adjustment, using α = 0.05.
Which states are identified as outliers?
4. Using R, produce a grid of 9 normal probability plots (qqnorm) for samples of size n = 50
simulated independently from a geometric distribution with parameter prob equal to 0.4.
(Refer to the last section of the Diagnostics in Linear Regression slides. Use the R function
rgeom to simulate from the geometric distribution.)
1
(a) [2 pts] Display your plots and the R code you used to produce them.
(b) [2 pts] Describe two distinct ways in which these plots tend to differ in appearance
from what you would expect for normally-distributed data.
5. [ GRADUATE SECTION ONLY ] Consider the linear regression model Y = Xβ + e under
the Gauss-Markov conditions and with X having p
0
columns and full column rank.
Consider the leverage values hii from the hat matrix H and the random residuals ˆei (from
ordinary least squares).
(a) [2 pts] Let tr(M) be the sum of all diagonal elements of the square matrix M, called
the trace of M. If the matrix product AB is square, show that
tr(AB) = tr(BA)
(Hint: The ith diagonal element of AB is P
j
aij bji and the jth diagonal element of
BA is P
i
bjiaij .)
(b) [2 pts] Using the previous part, show that
Xn
i=1
hii = tr(H) = p
0
(Hints: Take A = X and B =
XTX
−1XT
. What is the trace of an identity matrix?)
(c) [2 pts] Find an expression for E
eˆ
2
i
in terms of hii and σ
2
.
(d) [2 pts] Using results from the previous parts, show that the usual error variance
estimator ˆσ
2
is unbiased.
Some reminders:
• Unless otherwise stated, all data sets are either automatically available or can be found in
either the alr4 package or the faraway package in R.
• Unless otherwise stated, use a 5% level (α = 0.05) in all tests.