Statistics 512: Applied Linear Models

Topic 3
Topic Overview
This topic will cover
thinking in terms of matrices
regression on multiple predictor variables
case study: CS majors
Text Example (NKNW 241)
Chapter 5: Linear Regression in Matrix Form
The SLR Model in Scalar Form
Y
i
=
0
+
1
X
i
+
i
where
i

iid
N(0,
2
)
Consider now writing an equation for each observation:
Y
1
=
0
+
1
X
1
+
1
Y
2
=
0
+
1
X
2
+
2
.
.
.
.
.
.
.
.
.
Y
n
=
0
+
1
X
n
+
n
The SLR Model in Matrix Form
_

_
Y
1
Y
2
.
.
.
Y
n
_

_
=
_

0
+
1
X
1

0
+
1
X
2
.
.
.

0
+
1
X
n
_

_
+
_

2
.
.
.

n
_

_
_

_
Y
1
Y
2
.
.
.
Y
n
_

_
=
_

_
1 X
1
1 X
2
.
.
.
.
.
.
1 X
n
_

_
_

0

1
_
+
_

2
.
.
.

n
_

_
(I will try to use bold symbols for matrices. At rst, I will also indicate the dimensions as
a subscript to the symbol.)
1
X is called the design matrix.
is the vector of parameters
is the error vector
Y is the response vector
The Design Matrix
X
n2
=
_

_
1 X
1
1 X
2
.
.
.
.
.
.
1 X
n
_

_
Vector of Parameters

21
=
_

0

1
_
Vector of Error Terms

n1
=
_

2
.
.
.

n
_

_
Vector of Responses
Y
n1
=
_

_
Y
1
Y
2
.
.
.
Y
n
_

_
Thus,
Y = X +
Y
n1
= X
n2

21
+
n1
2
Variance-Covariance Matrix
In general, for any set of variables U
1
, U
2
, . . . , U
n
, their variance-covariance matrix is dened
to be

2
{U} =
_

2
{U
1
} {U
1
, U
2
} {U
1
, U
n
}
{U
2
, U
1
}
2
{U
2
}
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. {U
n1
, U
n
}
{U
n
, U
1
} {U
n
, U
n1
}
2
{U
n
}
_

_
where
2
{U
i
} is the variance of U
i
, and {U
i
, U
j
} is the covariance of U
i
and U
j
.
When variables are uncorrelated, that means their covariance is 0. The variance-covariance
matrix of uncorrelated variables will be a diagonal matrix, since all the covariances are 0.
Note: Variables that are independent will also be uncorrelated. So when variables are
correlated, they are automatically dependent. However, it is possible to have variables that
are dependent but uncorrelated, since correlation only measures linear dependence. A nice
thing about normally distributed RVs is that they are a convenient special case: if they are
uncorrelated, they are also independent.
Covariance Matrix of

2
{}
nn
= Cov
_

2
.
.
.

n
_

_
=
2
I
nn
=
_

2
0 0
0
2
0
.
.
.
.
.
.
.
.
.
.
.
.
0 0
2
_

_
Covariance Matrix of Y

2
{Y}
nn
= Cov
_

_
Y
1
Y
2
.
.
.
Y
n
_

_
=
2
I
nn
Distributional Assumptions in Matrix Form
N(0,
2
I)
I is an n n identity matrix.
Ones in the diagonal elements specify that the variance of each
i
is 1 times
2
.
Zeros in the o-diagonal elements specify that the covariance between dierent
i
is
zero.
This implies that the correlations are zero.
3
Parameter Estimation
Least Squares
Residuals are = YX. Want to minimize sum of squared residuals.

2
i
= [
1

2

n
]
_

2
.
.
.

n
_

_
=

We want to minimize

= (YX)

(YX), where the prime ()

denotes the transpose

of the matrix (exchange the rows and columns).
We take the derivative with respect to the vector . This is like a quadratic function: think
(YX)
2
.
The derivative works out to 2 times the derivative of (YX)

with respect to .
That is,
d
d
((YX)

(YX)) = 2X

(Y X). We set this equal to 0 (a vector of

zeros), and solve for .
So, 2X

(YX) = 0. Or, X

Y = X

X (the normal equations).

Normal Equations
X

Y = (X

X)
Solving this equation for gives the least squares solution for b =
_
b
0
b
1
_
.
Multiply on the left by the inverse of the matrix X

X. (Notice that the matrix X

X is a
2 2 square matrix for SLR.)
b = (X

X)
1
X

Y
REMEMBER THIS.
Reality Break:
This is just to convince you that we have done nothing new nor magic all we are doing is
writing the same old formulas for b
0
and b
1
in matrix format. Do NOT worry if you cannot
reproduce the following algebra, but you SHOULD try to follow it so that you believe me
that this is really not a new formula.
Recall in Topic 1, we had
b
1
=

(X
i

X)(Y
i

Y )

(X
i

X)
2

SS
XY
SS
X
b
0
=

Y b
1

X
4
Now lets look at the pieces of the new formula:
X

X =
_
1 1 1
X
1
X
2
X
n
_
_

_
1 X
1
1 X
2
.
.
.
.
.
.
1 X
n
_

_
=
_
n

X
i

X
i

X
2
i
_
(X

X)
1
=
1
n

X
2
i
(

X
i
)
2
_
X
2
i

X
i

X
i
n
_
=
1
nSS
X
_
X
2
i

X
i

X
i
n
_
X

Y =
_
1 1 1
X
1
X
2
X
n
_
_

_
Y
1
Y
2
.
.
.
Y
n
_

_
=
_
Y
i

X
i
Y
i
_
Plug these into the equation for b:
b = (X

X)
1
X

Y =
1
nSS
X
_
X
2
i

X
i

X
i
n
_ _
Y
i

X
i
Y
i
_
=
1
nSS
X
_
(

X
2
i
)(

Y
i
) (

X
i
)(

X
i
Y
i
)
(

X
i
)(

Y
i
) + n

X
i
Y
i
_
=
1
SS
X
_

Y (

X
2
i
)

X

X
i
Y
i

X
i
Y
i
n

X

Y
_
=
1
SS
X
_

Y (

X
2
i
)

Y (n

X
2
) +

X(n

X

Y )

X

X
i
Y
i
SP
XY
_
=
1
SS
X
_

Y SS
X
SP
XY

X
SP
XY
_
=
_

Y
SP
XY
SS
X

X
SP
XY
SS
X
_
=
_
b
0
b
1
_
,
where
SS
X
=

X
2
i
n

X
2
=

(X
i

X)
2
SP
XY
=

X
i
Y
i
n

X

Y =

(X
i

X)(Y
i

Y )
All we have done is to write the same old formulas for b
0
and b
1
in a fancy new format.
See NKNW page 200 for details. Why have we bothered to do this? The cool part is that
the same approach works for multiple regression. All we do is make X and b into bigger
matrices, and use exactly the same formula.
Other Quantities in Matrix Form
Fitted Values

Y =
_

Y
1

Y
2
.
.
.

Y
n
_

_
=
_

_
b
0
+ b
1
X
1
b
0
+ b
1
X
2
.
.
.
b
0
+ b
1
X
n
_

_
=
_

_
1 X
1
1 X
2
.
.
.
.
.
.
1 X
n
_

_
_
b
0
b
1
_
= Xb
5
Hat Matrix

Y = Xb

Y = X(X

X)
1
X

Y = HY
where H = X(X

X)
1
X

. We call this the hat matrix because is turns Y s into

Y s.
Estimated Covariance Matrix of b
This matrix b is a linear combination of the elements of Y.
These estimates are normal if Y is normal.
These estimates will be approximately normal in general.
A Useful Multivariate Theorem
Suppose U N(, ), a multivariate normal vector, and V = c +DU, a linear
transformation of U where c is a vector and D is a matrix. Then V N(c +
D, DD

).
Recall: b = (X

X)
1
X

Y = [(X

X)
1
X

] Y and Y N(X,
2
I).
Now apply theorem to b using
U = Y, = X, =
2
I
V = b, c = 0, and D = (X

X)
1
X

The theorem tells us the vector b is normally distributed with mean

(X

X)
1
(X

X) =
and covariance matrix

2
_
(X

X)
1
X

_
I
_
(X

X)
1
X

=
2
(X

X)
1
(X

X)
_
(X

X)
1
_

=
2
(X

X)
1
using the fact that both X

X and its inverse are symmetric, so ((X

X)
1
)

= (X

X)
1
.
Next we will use this framework to do multiple regression where we have more than one
explanatory variable (i.e., add another column to the design matrix and additional beta
parameters).
Multiple Regression
Data for Multiple Regression
Y
i
is the response variable (as usual)
6
X
i,1
, X
i,2
, . . . , X
i,p1
are the p 1 explanatory variables for cases i = 1 to n.
Example In Homework #1 you considered modeling GPA as a function of entrance
exam score. But we could also consider intelligence test scores and high school GPA
as potential predictors. This would be 3 variables, so p = 4.
Potential problem to remember!!! These predictor variables are likely to be themselves
correlated. We always want to be careful of using variables that are themselves strongly
correlated as predictors together in the same model.
The Multiple Regression Model
Y
i
=
0
+
1
X
i,1
+
2
X
i,2
+ . . . +
p1
X
i,p1
+
i
for i = 1, 2, . . . , n
where
Y
i
is the value of the response variable for the ith case.

i

iid
N(0,
2
) (exactly as before!)

0
is the intercept (think multidimensionally).

1
,
2
, . . . ,
p1
are the regression coecients for the explanatory variables.
X
i,k
is the value of the kth explanatory variable for the ith case.
Parameters as usual include all of the s as well as
2
. These need to be estimated
from the data.
Interesting Special Cases
Polynomial model:
Y
i
=
0
+
1
X
i
+
2
X
2
i
+ . . . +
p1
X
p1
i
+
i
Xs can be indicator or dummy variables with X = 0 or 1 (or any other two distinct
numbers) as possible values (e.g. ANOVA model). Interactions between explanatory
variables are then expressed as a product of the Xs:
Y
i
=
0
+
1
X
i,1
+
2
X
i,2
+
3
X
i,1
X
i,2
+
i
7
Model in Matrix Form
Y
n1
= X
np

p1
+
n1
N(0,
2
I
nn
)
Y N(X,
2
I)
Design Matrix X:
X =
_

_
1 X
1,1
X
1,2
X
1,p1
1 X
2,1
X
2,2
X
2,p1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1 X
n,1
X
n,2
X
n,p1
_

_
Coecient matrix :
=
_

1
.
.
.

p1
_

_
Parameter Estimation
Least Squares
Find b to minimize SSE = (YXb)

(YXb)
Obtain normal equations as before: X

Xb = X

Y
Least Squares Solution
b = (X

X)
1
X

Y
Fitted (predicted) values for the mean of Y are

Y = Xb = X(X

X)
1
X

Y = HY,
where H = X(X

X)
1
X

.
Residuals
e = Y

Y = YHY = (I H)Y
Notice that the matrices H and (I H) have two special properties. They are
Symmetric: H = H

and (I H)

= (I H).
Idempotent: H
2
= H and (I H)(I H) = (I H)
8
Covariance Matrix of Residuals
Cov(e) =
2
(I H)(I H)

=
2
(I H)
V ar(e
i
) =
2
(1 h
i,i
),
where h
i,i
is the ith diagonal element of H.
Note: h
i,i
= X

i
(X

X)
1
X
i
where X

i
= [1 X
i,1
X
i,p1
].
Residuals e
i
are usually somewhat correlated: cov(e
i
, e
j
) =
2
h
i,j
; this is not unexpected,
since they sum to 0.
Estimation of
Since we have estimated p parameters, SSE = e

e has df
E
= n p. The estimate for
2
is
the usual estimate:
s
2
=
e

e
n p
=
(YXb)

(YXb)
n p
=
SSE
df
E
= MSE
s =

s
2
= Root MSE
Distribution of b
We know that b = (X

X)
1
X

Y. The only RV involved is Y , so the distribution of b is

based on the distribution of Y.
Since Y N(X,
2
I), and using the multivariate theorem from earlier (if you like, go
through the details on your own), we have
E(b) =
_
(X

X)
1
X

_
X =

2
{b} = Cov(b) =
2
(X

X)
1
Since
2
is estimated by the MSE s
2
,
2
{b} is estimated by s
2
(X

X)
1
.
ANOVA Table
Sources of variation are
Model (SAS) or Regression (NKNW)
Error (Residual)
Total
SS and df add as before
SSM + SSE = SST
df
M
+ df
E
= df
Total
but their values are dierent from SLR.
9
Sum of Squares
SSM =

Y
i

Y )
2
SSE =

(Y
i

Y
i
)
2
SSTO =

(Y
i

Y )
2
Degrees of Freedom
df
M
= p 1
df
E
= n p
df
Total
= n 1
The total degrees have not changed from SLR, but the model df has increased from 1 to
p1, i.e., the number of X variables. Correspondingly, the error df has decreased from n2
to n p.
Mean Squares
MSM =
SSM
df
M
=

Y
i

Y )
2
p 1
MSE =
SSE
df
E
=

(Y
i

Y
i
)
2
n p
MST =
SSTO
df
Total
=

(Y
i

Y )
2
n 1
ANOVA Table
Source df SS MSE F
Model df
M
= p 1 SSM MSM
MSM
MSE
Error df
E
= n p SSE MSE
Total df
T
= n 1 SST
F-test
H
0
:
1
=
2
= . . . =
p1
= 0 (all regression coecients are zero)
H
A
:
k
= 0, for at least one k = 1, . . . , p 1; at least of the s is non-zero (or, not all the
s are zero).
F = MSM/MSE
Under H
0
, F F
p1,np
Reject H
0
if F is larger than critical value; if using SAS, reject H
0
if p-value < = 0.05 .
10
What do we conclude?
If H
0
is rejected, we conclude that at least one of the regression coecients is non-zero;
hence at least one of the X variables is useful in predicting Y . (Doesnt say which one(s)
though). If H
0
is not rejected, then we cannot conclude that any of the X variables is useful
in predicting Y .
p-value of F-test
The p-value for the F signicance test tell us one of the following:
there is no evidence to conclude that any of our explanatory variables can help us to
model the response variable using this kind of model (p 0.05).
one or more of the explanatory variables in our model is potentially useful for predict-
ing the response in a linear model (p 0.05).
R
2
The squared multiple regression correlation (R
2
) gives the proportion of variation in the
response variable explained by the explanatory variables.
It is sometimes called the coecient of multiple determination (NKNW, page 230).
R
2
= SSM/SST (the proportion of variation explained by the model)
R
2
= 1 (SSE/SST) (1 the proportion not explained by the model)
F and R
2
are related:
F =
R
2
/(p 1)
(1 R
2
)/(n p)
Inference for Individual Regression Coecients
Condence Interval for
k
We know that b N(,
2
(X

X)
1
)
Dene
s
2
{b}
pp
= MSE (X

X)
1
s
2
{b
k
} =
_
s
2
{b}

k,k
, the kth diagonal element
CI for
k
: b
k
t
c
s{b
k
}, where t
c
= t
np
(0.975).
Signicance Test for
k
H
0
:
k
= 0
Same test statistic t

= b
k
/s{b
k
}
Still use df
E
which now is equal to n p
p-value computed from t
np
distribution.
This tests the signicance of a variable given that the other variables are already in the model
(i.e., tted last). Unlike in SLR, the t-tests for are dierent from the F-test.
11
Multiple Regression Case Study
Example: Study of CS Students
Problem: Computer science majors at Purdue have a large drop-out rate.
Potential Solution: Can we nd predictors of success? Predictors must be available at time
of entry into program.
Data Available
Grade point average (GPA) after three semesters (Y
i
, the response variable)
Five potential predictors (p = 6)
X
1
= High school math grades (HSM)
X
2
= High school science grades (HSS)
X
3
= High school English grades (HSE)
X
4
= SAT Math (SATM)
X
5
= SAT Verbal (SATV)
Gender (1 = male, 2 = female) (we will ignore this one right now, since it is not a
continuous variable).
We have n = 224 observations, so if all ve variables are included, the design matrix X is
224 6. The SAS program used to generate output for this is cs.sas.
Look at the individual variables
Our rst goal should be to take a look at the variables to see...
Is there anything that sticks out as unusual for any of the variables?
How are these variables related to each other (pairwise)? If two predictor variables are
strongly correlated, we wouldnt want to use them in the same model!
We do this by looking at statistics and plots.
data cs;
infile H:\System\Desktop\csdata.dat;
input id gpa hsm hss hse satm satv genderm1;
Descriptive Statistics: proc means
proc means data=cs maxdec=2;
var gpa hsm hss hse satm satv;
The option maxdec = 2 sets the number of decimal places in the output to 2 (just showing
you how).
12
Output from proc means
The MEANS Procedure
Variable N Mean Std Dev Minimum Maximum
---------------------------------------------------------------------------------
gpa 224 2.64 0.78 0.12 4.00
hsm 224 8.32 1.64 2.00 10.00
hss 224 8.09 1.70 3.00 10.00
hse 224 8.09 1.51 3.00 10.00
satm 224 595.29 86.40 300.00 800.00
satv 224 504.55 92.61 285.00 760.00
---------------------------------------------------------------------------------
Descriptive Statistics
Note that proc univariate also provides lots of other information, not shown.
proc univariate data=cs noprint;
var gpa hsm hss hse satm satv;
histogram gpa hsm hss hse satm satv /normal;
Figure 1: Graph of GPA (left) and High School Math (right)
Figure 2: Graph of High School Science (left) and High School English (right)
NOTE: If you want the plots (e.g., histogram, qqplot) and not the copious output from proc
univariate, use a noprint statement
13
Figure 3: Graph of SAT Math (left) and SAT Verbal (right)
proc univariate data = cs noprint;
histogram gpa / normal;
Interactive Data Analysis
Read in the dataset as usual
From the menu bar, select
Solutions -> analysis -> interactive data analysis
Obtain SAS/Insight window
Open library work
Click on Data Set CS and click open.
Getting a Scatter Plot Matrix
(CTRL) Click on GPA, SATM, SATV
Go to menu Analyze
Choose option Scatterplot(Y X)
You can, while in this window, use Edit -> Copy to copy this plot to another program such as
Word: (See Figure 4.)
This graph once you get used to it can be useful in getting an overall feel for the
relationships among the variables. Try other variables and some other options from the
Analyze menu to see what happens.
Correlations
SAS will give us the r (correlation) value between pairs of random variables in a data set
using proc corr.
proc corr data=cs;
var hsm hss hse;
14
Figure 4: Scatterplot Matrix
hsm hss hse
hsm 1.00000 0.57569 0.44689
<.0001 <.0001
hss 0.57569 1.00000 0.57937
<.0001 <.0001
hse 0.44689 0.57937 1.00000
<.0001 <.0001
To get rid of those p-values, which make it dicult to read, use a noprob statement. Here
are also a few dierent ways to call proc corr:
proc corr data=cs noprob;
var satm satv;
satm satv
satm 1.00000 0.46394
satv 0.46394 1.00000
proc corr data=cs noprob;
var hsm hss hse;
with satm satv;
hsm hss hse
satm 0.45351 0.24048 0.10828
satv 0.22112 0.26170 0.24371
15
proc corr data=cs noprob;
var hsm hss hse satm satv;
with gpa;
hsm hss hse satm satv
gpa 0.43650 0.32943 0.28900 0.25171 0.11449
Notice that not only do the Xs correlate with Y (this is good), but the Xs correlate with
each other (this is bad). That means that some of the Xs may be redundant in predicting
Y .
Use High School Grades to Predict GPA
proc reg data=cs;
model gpa=hsm hss hse;
Analysis of Variance
Sum of Mean
Source DF Squares Square F Value Pr > F
Model 3 27.71233 9.23744 18.86 <.0001
Error 220 107.75046 0.48977
Corrected Total 223 135.46279
Root MSE 0.69984 R-Square 0.2046
Dependent Mean 2.63522 Adj R-Sq 0.1937
Coeff Var 26.55711
Parameter Estimates
Parameter Standard
Variable DF Estimate Error t Value Pr > |t|
Intercept 1 0.58988 0.29424 2.00 0.0462
hsm 1 0.16857 0.03549 4.75 <.0001
hss 1 0.03432 0.03756 0.91 0.3619
hse 1 0.04510 0.03870 1.17 0.2451
Remove HSS
proc reg data=cs;
model gpa=hsm hse;
R-Square 0.2016 Adj R-Sq 0.1943
Parameter Standard
Variable DF Estimate Error t Value Pr > |t|
Intercept 1 0.62423 0.29172 2.14 0.0335
hsm 1 0.18265 0.03196 5.72 <.0001
hse 1 0.06067 0.03473 1.75 0.0820
Rerun with HSM only
proc reg data=cs;
model gpa=hsm;
16
R-Square 0.1905 Adj R-Sq 0.1869
Parameter Standard
Variable DF Estimate Error t Value Pr > |t|
Intercept 1 0.90768 0.24355 3.73 0.0002
hsm 1 0.20760 0.02872 7.23 <.0001
The last two models (HSM only and HSM, HSE) appear to be pretty good. Notice that R
2
go down a little with the deletion of HSE , so HSE does provide a little information. This is
a judgment call: do you prefer a slightly better-tting model, or a simpler one?
Now look at SAT scores
proc reg data=cs;
model gpa=satm satv;
R-Square 0.0634 Adj R-Sq 0.0549
Parameter Standard
Variable DF Estimate Error t Value Pr > |t|
Intercept 1 1.28868 0.37604 3.43 0.0007
satm 1 0.00228 0.00066291 3.44 0.0007
satv 1 -0.00002456 0.00061847 -0.04 0.9684
Here we see that SATM is a signicant explanatory variable, but the overall t of the model
is poor. SATV does not appear to be useful at all.
Now try our two most promising candidates: HSM and SATM
proc reg data=cs;
model gpa=hsm satm;
R-Square 0.1942 Adj R-Sq 0.1869
Parameter Standard
Variable DF Estimate Error t Value Pr > |t|
Intercept 1 0.66574 0.34349 1.94 0.0539
hsm 1 0.19300 0.03222 5.99 <.0001
satm 1 0.00061047 0.00061117 1.00 0.3190
In the presence of HSM, SATM does not appear to be at all useful. Note that adjusted R
2
is the same as with HSM only, and that the p-value for SATM is now no longer signicant.
General Linear Test Approach: HS and SATs
proc reg data=cs;
model gpa=satm satv hsm hss hse;
*Do general linear test;
* test H0: beta1 = beta2 = 0;
sat: test satm, satv;
* test H0: beta3=beta4=beta5=0;
hs: test hsm, hss, hse;
17
R-Square 0.2115 Adj R-Sq 0.1934
Parameter Standard
Variable DF Estimate Error t Value Pr > |t|
Intercept 1 0.32672 0.40000 0.82 0.4149
satm 1 0.00094359 0.00068566 1.38 0.1702
satv 1 -0.00040785 0.00059189 -0.69 0.4915
hsm 1 0.14596 0.03926 3.72 0.0003
hss 1 0.03591 0.03780 0.95 0.3432
hse 1 0.05529 0.03957 1.40 0.1637
The rst test statement tests the full vs reduced models (H
0
: no satm or satv, H
a
: full
model)
Test sat Results for Dependent Variable gpa
Mean
Source DF Square F Value Pr > F
Numerator 2 0.46566 0.95 0.3882
Denominator 218 0.49000
We do not reject the H
0
that
1
=
2
= 0. Probably okay to throw out SAT scores.
The second test statement tests the full vs reduced models (H
0
: no hsm, hss, or hse,
H
a
: all three in the model)
Test hs Results for Dependent Variable gpa
Mean
Source DF Square F Value Pr > F
Numerator 3 6.68660 13.65 <.0001
Denominator 218 0.49000
We reject the H
0
that
3
=
4
=
5
= 0. CANNOT throw out all high school grades.
Can use the test statement to test any set of coecients equal to zero. Related to extra
sums of squares (later).
Best Model?
Likely the one with just HSM. Could argue that HSM and HSE is marginally better. Well
discuss comparison methods in Chapters 7 and 8.
Key ideas from case study
First, look at graphical and numerical summaries for one variable at a time.
Then, look at relationships between pairs of variables with graphical and numerical
summaries.
Use plots and correlations to understand relationships.
The relationship between a response variable and an explanatory variable depends on
what other explanatory variables are in the model.
18
A variable can be a signicant (p < 0.05) predictor alone and not signicant (p > 0.05)
when other Xs are in the model.
Regression coecients, standard errors, and the results of signicance tests depend on
what other explanatory variables are in the model.
Signicance tests (p-values) do not tell the whole story.
Squared multiple correlations (R
2
) give the proportion of variation in the response
variable explained by the explanatory variables, and can give a dierent view; see also
adjusted R
2
.
You can fully understand the theory in terms of Y = X + .
To eectively use this methodology in practice, you need to understand how the data
were collected, the nature of the variables, and how they relate to each other.
Other things that should be considered
Diagnostics: Do the models meet the assumptions? You might take one of our nal
two potential models and check these out on your own.
Condence intervals/bands, Prediction intervals/bands?
Example II (NKNW page 241)
Dwaine Studios, Inc. operates portrait studios in n = 21 cities of medium size.
Program used to generate output for condence intervals for means and prediction intervals
is nknw241.sas.
Y
i
is sales in city i
X
1
: population aged 16 and under
X
2
: per capita disposable income
Y
i
=
0
+
1
X
i,1
+
2
X
i,2
+
i
Read in the data
data a1;
infile H:\System\Desktop\CH06FI05.DAT;
input young income sales;
proc print data=a1;
19
Obs young income sales
1 68.5 16.7 174.4
2 45.2 16.8 164.4
3 91.3 18.2 244.2
4 47.8 16.3 154.6
...
proc reg data=a1;
model sales=young income
Sum of Mean
Source DF Squares Square F Value Pr > F
Model 2 24015 12008 99.10 <.0001
Error 18 2180.92741 121.16263
Corrected Total 20 26196
Root MSE 11.00739 R-Square 0.9167
Dependent Mean 181.90476 Adj R-Sq 0.9075
Coeff Var 6.05118
Parameter Standard
Variable DF Estimate Error t Value Pr > |t|
Intercept 1 -68.85707 60.01695 -1.15 0.2663
young 1 1.45456 0.21178 6.87 <.0001
income 1 9.36550 4.06396 2.30 0.0333
clb option: Condence Intervals for the s
proc reg data=a1;
model sales=young income/clb;
95% Confidence Limits
Intercept -194.94801 57.23387
young 1.00962 1.89950
income 0.82744 17.90356
Estimation of E(Y
h
)
X
h
is now a vector of values. (Y
h
is still just a number.)
(1, X
h,1
, X
h,2
, . . . , X
h,p1
)

= X

h
: this is row h of the design matrix. We want a point
estimate and a condence interval for the subpopulation mean corresponding to the set of
explanatory variables X
h
.
Theory for E(Y )
h
E(Y
h
) =
j
= X

h
= X

h
b
s
2
{
h
} = X

h
s
2
{b}X
h
= s
2
X

h
(X

X)
1
X
h
95%CI :
h
s{
h
}t
np
(0.975)
20
clm option: Condence Intervals for the Mean
proc reg data=a1;
model sales=young income/clm;
id young income;
Dep Var Predicted Std Error
Obs young income sales Value Mean Predict 95% CL Mean
1 68.5 16.7 174.4000 187.1841 3.8409 179.1146 195.2536
2 45.2 16.8 164.4000 154.2294 3.5558 146.7591 161.6998
3 91.3 18.2 244.2000 234.3963 4.5882 224.7569 244.0358
4 47.8 16.3 154.6000 153.3285 3.2331 146.5361 160.1210
Prediction of Y
h(new)
Predict a new observation Y
h
with X values equal to X
h
.
We want a prediction of Y
h
based on a set of predictor values with an interval that expresses
the uncertainty in our prediction. As in SLR, this interval is centered at

Y
h
and is wider
than the interval for the mean.
Theory for Y
h
Y
h
= X

h
+

Y
h
=
h
= X

s
2
{pred} = Var(

Y
h
+ ) = Var(

Y
h
) + Var()
= s
2
_
1 +X

h
(X

X)
1
X
h
_
CI for Y
h(new)
:

Y
h
s{pred}t
np
(0.975)
cli option: Condence Interval for an Individual Observation
proc reg data=a1;
model sales=young income/cli;
id young income;
Dep Var Predicted Std Error
Obs young income sales Value Mean Predict 95% CL Predict
1 68.5 16.7 174.4000 187.1841 3.8409 162.6910 211.6772
2 45.2 16.8 164.4000 154.2294 3.5558 129.9271 178.5317
3 91.3 18.2 244.2000 234.3963 4.5882 209.3421 259.4506
4 47.8 16.3 154.6000 153.3285 3.2331 129.2260 177.4311
Diagnostics
Look at the distribution of each variable to gain insight.
Look at the relationship between pairs of variables. proc corr is useful here. BUT note that
relationships between variables can be more complicated than just pairwise: correlations are
NOT the entire story.
21
Plot the residuals vs...
the predicted/tted values
each explanatory variable
time (if available)
Are the relationships linear?
Look at Y vs each X
i
May have to transform some Xs.
Are the residuals approximately normal?
Look at a histogram
Normal quantile plot
Is the variance constant?
Check all the residual plots.
Remedies
Similar remedies to simple regression (but more complicated to decide, though).
Additionally may eliminate some of the Xs (this is call variable selection).
Transformations such as Box-Cox
Analyze with/without outliers
More detail in NKNW Ch 9 and 10
22