Transcript Stat 502 - Topic #2

Topic 2 – Simple Linear Regression
KKNR Chapters 4 – 7
1
Overview

Regression Models; Scatter Plots


Estimation and Inference in SLR


SAS GPLOT Procedure
SAS REG Procedure
ANOVA Table & Coefficient of
Determination (R2)
2
Simple Linear Regression Model

We take n pairs of observations
(X 1,Y 1 ), (X 2,Y 2 ),..., (X n ,Y n )

The goal is to find a model that best fits with
the data.

Model will be linear in terms of the parameters
(betas). These won’t appear in exponents or
anything unusual.

Allowed to be nonlinear in terms of predictor
variables (we may transform these somewhat freely).
We may also transform the response.
3
Simple Linear Regression Model (2)

Some sample models
Yi   0  1 X i   i
Yi   0  1  log X i    i
log Yi   0  1

X i  i
Notice the betas always function in the same
way, and the analysis will always proceed in
the same way too (after we make whatever
transformations we might need).
4
Simple Linear Regression Model (3)

Key question: How do you decide on the
“best” form for the model?

Always view a scatter plot (use PROC
GPLOT in SAS). Curvature in the plot will
help you determine the need for a
transformation on either X or Y.

Always consider residual plots. Some
patterns in these plots will also indicate the
need for transformation (more on this later).
5
Scatter Plot Approach

If you can look at a scatter plot and the data
“look linear”, then likely no transformation is
necessary. Try not to look for things that are
not there.

If you see curvature, then some
transformation may be appropriate:

Use scientific theory & experience

Try transformations you think may work – look
at scatter plots of the transformed data to
assess whether they do work.
6
Finding the “Best” Model

There is no “absolute” strategy.

Some common mistakes (why are these bad?):

Try several different methods and simply take
the one for which you get the best results
(e.g. highest R2)

Over-fit the model by including lots of extra
terms (e.g. squares, cubes, etc.) in hopes to
get the curve to go through all of the data
points (note that this would be MLR)
7
Collaborative Learning Activity
CLG #2.1-2.3
First, make sure you read enough
to understand the dataset we will
be considering. Then, please try
to answer these questions related
to scatter plots.
8
Scatter Plot Examples (1)
Wi t h
E s t i ma t e d
Re g r e s s i o n
Li ne
1200
1100
1000
900
800
3
4
5
6
S t a t e wi d e
7
Ex p e n d i t u r e s
8
9
10
9
Scatter Plot Examples (2)
Wi t h
E s t i ma t e d
Re g r e s s i o n
Li ne
1200
1100
1000
900
800
0
10
20
30
Pe r c e n t a g e
40
of
El i g i b l e
50
St u d e n t s
60
Tak i ng
70
SAT
80
90
10
Scatter Plot Examples (3)
Wi t h
No n p a r a me t e r i c
30
40
S mo o t h
1200
1100
1000
900
800
0
10
20
Pe r c e n t a g e
of
El i g i b l e
50
St u d e n t s
60
Tak i ng
70
SAT
80
90
11
Scatter Plot Examples (4)
Log
T r a n s f o r me d
Pr e d i c t o r
Wi t h
No n p a r a me t e r i c
S mo o t h
1200
1100
1000
900
800
1
2
L o g - T r a n s f o r me d
3
Pe r c e n t a g e
of
4
El i g i b l e
St u d e n t s
Tak i ng
5
SAT
12
Comments on GPLOT

Utilize SYMBOL, AXIS, and TITLE
statements to make your plots look nice.

ORFONT provides a good symbol-set. You
can also manipulate the COLOR of symbols
in order help the viewer differentiate groups.

Be careful to remember that SAS reuses
these statements, so you will need to
redefine them as necessary.
13
SAS ORFONT
14
Fitting the SLR Model
•
Once we decide on the form of our model,
we need to estimate the parameters that
yield the “best” fit.
•
The arithmetic involved is accomplished with
a computer, but it is useful to have some
understanding of the how the estimates are
calculated.
15
The SLR Model

Whatever the transformations may be, our model is in
the form of a straight line:
Y  0  1 X  

Epsilon represents the inherent variation (or error) in
the model.

Model involves two other parameters (unknown, but
fixed in value):

slope (change in y for a one unit change in x)

intercept (value of y for x = 0; usually not
particularly interested in this)
16
Observations

An observation Y at a particular X is a
random variable. So you can think of each
observation as having been drawn from a
normal distribution centered at  0  1 X and
having standard deviation  .

Be careful to remember that these
parameters (represented by greek letters)
are fixed – but can never be known exactly.
We can only estimate them.
17
Graphical Representation
18
Model Assumptions

We make three assumptions on the error
term in our model. A simple statement of
IID
2
these assumptions is that  i ~ N 0,



The assumptions on the errors apply to both
regression and ANOVA and we will be
assuming these throughout the course.

For regression, we also make a 4th
assumption that our model (in this case
linear relationship between X and Y) is
appropriate.
19
Assumptions on Errors

Constance Variance (Homoscedasticity) – the
variance associated to the error is the same for
ANY value of X.

Normality – the errors follow a normal
distribution with a mean of zero.

Independence – the errors (and hence also
the responses) are statistically independent
of each other.
20
Checking Your Assumptions

Reminder: We can never know the exact
values of the errors because we can never
know the true regression equation.

We can (and will) estimate the errors by the
residuals. The residuals can then be used
(mostly in graphical analyses) to assess the
assumptions – giving us some idea of
whether the assumptions of our model are
satisfied. More on this later...
21
Estimation of the “Best” Line

We want to obtain estimates of the
parameters  0 and 1 (remember, we can
never know them exactly).

Notation: Generally, I will use lower case
English letters to represent estimates for
parameters. You may also see hat-notation.
For example, if  is a parameter, ˆ would be
its estimated value from data.

Our estimates will be denoted  b0 , b1  . The
residuals will be denoted by ei .
22
Parameter Estimates



Key Point: Parameters are fixed, but their
estimates are random variables. If we take a
different sample, we’ll get a different estimate.
Thus all of the estimates we compute  b0 , b1 , ei 
will have associated standard errors that we
may also estimate.
The method of least squares is used to obtain
both parameter estimates and standard errors.
This method is desirable because the estimates
are unbiased, minimum variance estimates.
23
Least-squares Method

The least squares method obtains the
estimated regression line that minimizes the
sum of the squared residuals (also called
the SSE or sum of squares error).
2
2
SSE  e 
Y  Yˆ


i

i
i

Another way to think of this is that the least
squares estimates allow us to explain as
much of the variation in Y as we possibly
can using X as a predictor. The SSR (sums
of squares due to our model) is maximized.
24
Least Squares Estimates

The estimates have formulas in terms of the
data:
n
 X i  X Yi  Y 

b1  i 1 n
2
 Xi  X 
i 1
b0  Y  b1 X
ei  Yi  Yˆi  Yi   b0  b1 X i 
s 2  ˆ 2  SSE  n  2 
25
Least Squares Estimates (2)

It is not important to memorize these formulas.
I won’t ask you to calculate a parameter
estimate by hand from the data. We have
computers for this.

What is important will be to understand that,
because the Yi are random variables, and
because all of these estimates depend on the
Yi, the estimates themselves will also be
random variables. Thus we may estimate their
standard errors, develop confidence intervals,
and draw statistical inferences.
26
Inference about the Slope
•
A non-zero slope implies a linear
association between the predictor and
response.
•
In some experimental cases the
relationship may be causal as well.
•
Thus statistical inference for the slope is
quite important.
27
Inference About the Slope

The first thing to remember is that b1 is a
random variable. In order to do inference, we
must first consider that...
  X  X Y  Y 
n

b1 
i
i 1
n
 X
i 1

i
i
X
2
is normally distributed (why?)
The standard error associated to b1 is
s b1 
MSE
MSE
 n  X i  X 2
SSX
i 1
28
Slope Inference (2)

For testing H 0 : 1  k , the statistic
b1  k
T
s b1
has a t-distribution with n – 2 degrees of
freedom when the null hypothesis is true.

A two-sided confidence interval for the slope
will be:
b1  tn2,1 2 s b1
29
Slope Inference (3)


Want to determine: Does X help explain Y
through a linear model????
If we reject the null hypothesis H 0 : 1  0,
then we may conclude that there is a
linear association between X and Y


Must have assumptions satisfied.
Key point: Failing to reject does not necessarily
allow us to conclude that X is unimportant

Maybe we need a bigger sample to give better power
30
Slope Inference (4)

Another Key Point: Violations of the model
assumptions may invalidate the significance
test.

In particular, if there is a nonlinear
association or some type of dependence
issue, the SLR model should not be used.

See pages 65 for some pictures illustrating
this.
31
Experimental Control

In some situations you have experimental control
over your predictor variable. Thus you have some
control over the SE for the slope:
MSE
s b1 
SSX

Making SSX large will decrease the SE of your
estimate for the slope. Do this by spreading you
chosen X values further apart.

Increasing n (and hence increasing degrees of
freedom) may also help to decrease the SE.
32
Inference About the Intercept

Hypothesis tests and confidence intervals may
be constructed similar to inference for the
slope. See page 63.

Key point: Unless the observed predictor is
often in the neighborhood of zero, we have no
reason to be interested in the intercept.

In fact, the intercept will usually just be an
artifact of the model. And if the scope of the
model does not include zero, there is no reason
to even worry whether the value of b0 makes
sense.
33
Further Inference

Confidence Intervals for the
Mean Response

Prediction Intervals
34
The Predicted Value

The line describes the mean population
response for each value of the predictor. If
an association exists, then the mean
response depends on the value of X.

The predicted value of Y at a given X = x0 is
Yˆx0  b0  b1 x0

Reminder: Notation may differ some from
the text – I try to keep our notation as simple
as possible.
35
C.I. for the Mean Response or
Prediction Interval?

If you are trying to predict for a group


C.I. for the Mean Response

Example: Trying to predict the average blood pressure for all
40 year olds

Interval is usually narrower
If you are trying to predict for a single observation

Prediction Interval

Example: Trying to predict the blood pressure for a single 40
year old

Interval is usually much wider because of individual variation

Some 40 year olds will have much higher or lower B.P.s
36
*Calculations of SE for Mean
Response

First step is to write in terms of estimates. We also
need to use a small trick to avoid worrying about a
covariance between the two parameter estimates.
 
Var Yˆx0  Var  b0  b1 x0 
 Var Y  b1 x  b1 x0 
 Var Y   x0  x  b1 
37
*Calculations of SE for Mean
Response (2)

We know the variances for Y-bar and b1. And it
turns out that, even though Y-bar is used in the
calculation of b1, the two are still independent. So
the variance of the sum is the sum of the variances:
 
2
ˆ
Var Yx0  Var Y    x0  x  Var  b1 


2
n
2
x

x

 0 
2

SSX
2

x0  x  

2 1
  

SSX 
 n
38
SE for the Mean Response

The mean response is a random variable
since b0 and b1 are random. Hence it will
have a standard error.
 
s Yˆx0

 1  x0  x 2 
 MSE  

n

SSX


It is good to have some understanding of
how this works, so we will look very briefly at
the calculations (you should just try to follow
them, but not worry about memorizing them)
39
Confidence Intervals for Mean
Response

You sometimes want to get confidence
intervals for the mean response. Because
we have estimated both the mean and
variance, the t-distribution applies (n – 2
degrees of freedom). The CI for a given
value of X is:
 
Yˆx0  tn 2,1 / 2 s Yˆx0
40
SE for the Prediction Interval

So the prediction variance is the sum of
these two components:
Var  pred x0   Var Yˆx0   2
 
 1  X  X 2 
0

 2   2


SSX
n




s pred x0

 1  X  X 2 
0

 MSE 1  
SSX
 n



41
Prediction Intervals (1)

Now consider predicting a new observation
at X = x0. Our point estimate for this would
just be the point on the regression line, Yˆx0 .

Our prediction interval will be of the same
form as for the mean response, but with a
different standard error:

Yˆx0  tn2,1 / 2 s pred x0

42
Why the difference for the
Prediction Interval?

The key to understanding the standard error for
prediction is to understand the random components
involved.

(1) The regular variance associated to getting a
predicted value (same as the CI mean response
error)

(2) The individual error for a single observation

It basically gives us an extra σ2 piece

Think of the NORMAL DISTRIBUTION centered
around the regression line (see slide 18)
43
Multiple Confidence Intervals

We did intervals for 40 year olds, both group and
individual, what about 30? 35?

Getting multiple CI’s presents a similar problem to
multiple hypothesis tests. We would expect one
errant CI for every 20 CI’s that we obtain at 95%
confidence. Thus some adjustment may need to be
made.

Bonferroni is too conservative here, because these
CI’s are actually dependent and it is possible to
take advantage of this.
44
Confidence Bands

The solution is to change our critical value. Instead
of using T, we use a critical value related to the F
distribution:
W  2 F2,n2,1

This allows us to produce CI’s for the mean
response at any and all possible values for the
predictor variable. Hence we may also use this to
draw confidence bands around the regression line.

Useful trick: For significance level 0.05, the value of
W is approximately 0.6 more than the value of T.
This is slightly conservative, but simplifies
computation.
45
Interpolation vs. Extrapolation

Interpolation (x0 within the domain of the
observed X’s) is generally ok if the
assumptions are satisfied.

Extrapolation (x0 outside the domain of the
observed X’s) is usually a bad idea.

No assurance that linearity continues outside
the observed domain.

Example – Height regressed on age in
children.
46
Key Concepts

The standard error formulas for the slope,
the regression line, and prediction are
related – your goal should be to understand
these relationships.

You should also be able to construct CI’s
and do hypothesis tests.



Point Estimates
Critical Values (know how to look these up)
Standard Errors (generally would not be
asked to compute these, just use them)
47
SAS Review
proc reg data=sat;
model score=expend / clb clm cli;
id state expend;
output out=fit r=res p=pred;
48
Collaborative Learning Activity
Please complete problems 2.4
(constructing CI’s) and 2.5
(interpreting regression output)
on the handout.
49
Output: ANOVA Table
Source
Model
Error
Total
DF
1
48
49
Root MSE
Dependent Mean
Coeff Var
Sum of
Squares
39722
234586
274308
69.90851
965.92000
7.23751
Mean
Square
39722
4887.2
R-Square
Adj R-Sq
F Value
8.13
Pr > F
0.0064
0.1448
0.1270
50
Output: Parameter Estimates
Parameter Estimates
Variable
Intercept
expend
Variable
Intercept
expend
DF
Parameter
Estimate
Standard
Error
t Value
Pr > |t|
1
1
1089.29372
-20.89217
44.38995
7.32821
24.54
-2.85
<.0001
0.0064
DF
1
1
95% Confidence Limits
1000.04174
-35.62652
1178.54569
-6.15782
51
Output: Output Statistics
Output Statistics
Obs state
expend
9 Louisian 4.761
10 Minnesot
6
11 Missouri 5.383
12 Nebraska 5.935
Obs state
expend
9 Louisian 4.761
10 Minnesot
6
11 Missouri 5.383
12 Nebraska 5.935
Dependent Predicted
Std Error
Variable
Value
Mean Predict
1021
989.8
12.9637
1085
963.9
9.9109
1045
976.8
10.6015
1050
965.3
9.8890
95% CL Predict
846.9
1133
822.0
1106
834.7
1119
823.3
1107
95% CL
963.8
944.0
955.5
945.4
Mean
1016
983.9
998.1
985.2
Residual
31.1739
121.0593
68.1689
84.7013
52
ANOVA Table
ANOVA stands for analysis of variance.
We use an ANOVA table in regression to
organize our estimates of different
components of variation. It is important to
understand how this works for SLR since
we will use ANOVA tables for MLR and
ANOVA procedures as well.
53
ANOVA Table


Table consists of variance estimates used to
assess the following two questions:

Is there an linear association between the
response and predictor(s)?

How “strong” is that linear association?
We need to start by understanding the
different components of variation for a single
data point.
54
Components of Variation
55
Combining Over All Data

We might look at the total deviation as
follows:

 
Yi  Y  Yˆi  Y  Yi  Yˆi


But we cannot simply add deviations across
data points. Why? Options?
56
Combining Over All Data (2)

Squared deviations are chosen because it
turns out that they can be used to estimate
variances. It also (conveniently) turns out
that:
ˆ
ˆ
Y

Y

Y

Y

Y

Y


  


n
i 1
n
2
i
SSTOT
i 1

n
2
i
SS R
i 1

i
2
i
SS E
57
Sums of Squares

The total sums of squares (SST) represents
the total available variation that could be
explained by the predictor (that not already
explained by Ybar).

We break this into the two components:

Model/Regression sums of squares (SSR) is
the part that is explained by the predictor.

Error sums of squares (SSE) is the part that
is still left unexplained.
58
Degrees of Freedom

Each SS has an associated degrees of freedom.

For simple linear regression,



DFT = n – 1

DFR = 1

DFE = n – 2
Always have DFT  DFR  DFE
In general, you lose one degree of freedom for each
parameter you estimate. Since we estimate Ybar
before we start, dfTOT is n – 1.
59
Degrees of Freedom (2)

It is important to understand how DF are
assigned with the models that we will be
discussing. Some key principles:

DF Total is always 1 less than the number of
observations.

You should next determine DF for the model.
For regression, each continuous variable
requires a slope estimate and takes 1 DF.

Lastly, the error DF is determined by
subtraction (avoid memorization of formulas).
60
Mean Squares

A mean square is a SS divided by its
associated degrees of freedom.

These are the actual variance estimates:
SST
2
sY  MST 
(population variance)
n 1
SSE
2
s  MSE 
(error variance)
n2
SS R
2
s  MSR 
under null hyp. H 0 : 1  0
1
61
F-tests


Because both MSR and MSE estimate 
under the null hypothesis, we may utilize
their ratio in order to test whether there is a
linear association.
2
If the null hypothesis is true, the statistic
F
MSR
MSE
will have an F distribution with 1 and n – 2 DF.

Note: To get your DF for the F-test, simply use
the DF for the associated mean squares.
62
Relationship of F to T

For SLR the F test is identical to the t-test
for the slope as in fact:

F  ˆ1 S ˆ

1

2
T
2
Additionally you will find that in terms of
2
critical values, F1,v  tv .
63
Example
Source
Model
Error
Total
DF
1
48
49
Root MSE
Dependent Mean
Coeff Var
Variable
Intercept
expend
Variable
Sum of
Squares
39722
234586
274308
Mean
Square
39722
4887.2
69.90851
R-Square
965.92000
Adj R-Sq
7.23751
Parameter Estimates
F Value
8.13
Pr > F
0.0064
0.1448
0.1270
DF
Parameter
Estimate
Standard
Error
t Value
Pr > |t|
1
1
1089.29372
-20.89217
44.38995
7.32821
24.54
-2.85
<.0001
0.0064
DF
95% Confidence Limits
64
Example (2)

For SLR, the F-test in ANOVA Table is
exactly the same as the test for zero slope.


Note 8.13 = (-2.85)2.
Caution: When we get into multiple
regression, if the F-test has a small p-value
this is a good start, but not the end! In
multiple regression, the F-test may be thought
of as a test for “model significance”. But it
doesn’t tell us which variable(s) are important
and which are not.
65
Other Statistics from REG

R-square and Adjusted R-square help us to
assess the “strength” of the linear
relationship.

The coefficient of variation is calculated
using
CV  100 MSE Y
It measures the variation as a percentage of
the mean.
66
Coefficient of Determination
The coefficient of determination
(R2) gives us some idea as to
the strength of the regression
relationship.
67
Coefficient of Determination (R2)

Reflects the variation in Y that is explained
by the regression relationship as a
percentage of the total:
SS R
SS E
R 
 1
SSTOT
SSTOT
2

With perfect linear association, SSE will be
zero and R2 will be 1.

If no linear association, SSE will be the
same as SST and R2 will be 0.
68
Common Misconceptions

Steeper slope means bigger R2. This is not true.
In fact R2 has nothing to do with the magnitude of
the slope for our regression line.

The larger the value of R2, the better the model.
This is also not true. R2 says nothing about
appropriateness of model (see page 98).

R2 could be 0 but there could be a non-linear
association between X & Y

R2 could be near 1 while a curvilinear model
would be more appropriate (scatterplots will
generally reveal this)
69
The Correlation Coefficient (r)

Takes the sign of the slope and, for SLR, is
simply the square root of R2.

Dimensionless – ranges between -1 and 1.

Symmetric – interchanging X & Y will not
change the correlation between them.
70
SAT Example: Interpret R2

Simple interpretation: 14.8% of the variation
in SAT scores is explained by the
expenditures.

Reality Check: (1) though significant, this
is not a very strong relationship and (2) the
slope parameter is negative, suggesting that
increasing expenditures is associated with a
decrease in the average score!
72
Adjusted R2

Uses the mean squares to adjust (penalize)
for the number of parameters in the model
SSE /  n  p 
MS E
R  1
 1
SST /  n  1
MSTOT
2
a

We’ll discuss this more in multiple regression
as it really isn’t important for SLR.
73
Regression Diagnostics
Check Your Assumptions!!!
74
Regression Diagnostics

Assumptions

Correct Model (linearity)

Independent Observations

Normally Distributed Errors

Constant Variance

Checking these generally involves PLOTS of the
residuals and predicted values.

Residual = observed – predicted
ei  Yi  Yˆi
75
Regression Diagnostics (2)

Key Point: Most assumption checks may be
done visually by looking at various plots.
They may also be done using statistical
tests. Looking at plots is generally
easier!

So the general formula is to check the plots,
and if you still have questions then perhaps
consider the statistical tests.
76
Checking Normality



Histogram or Box-plot of Residuals

Is the histogram bell-shaped?

Is the box-plot symmetric?
Normal Probability / QQ plot

This is the method we would normally use.
Ordered residuals are plotted against
cumulative normal probabilities and the result
should be approximately linear.

PROC UNIVARIATE: QQPLOT statement
Shapiro-Wilks or Kolmogorov-Smirnov Test
77
SAS Code: QQPlot
proc univariate data=fit noprint;
var res;
title 'Normal Probability Plot';
qqplot res / normal(l=1 mu=est sigma=est);
78
Constancy of Variance


Plot the residuals against fitted (predicted)
values

Check to see if size of residual is somehow
associated with predicted value.

Megaphone shapes are indicative of a
violation.
Bartlett’s or Levene’s Test

Statistical tests are generally sensitive to
violations of normality and cannot be used if
the normality assumption is not met.
79
Plot: Residuals vs. Fitted Values
proc gplot data=fit;
plot res*pred /vref=0;
80
Checking Independence

This is the hardest assumption to check.

One check on this assumption is to simply think
of how the data are collected. Ask the question:
Is there anything in the collection of data that
could lead to dependent responses?

Plot the residuals over time (if applicable). Is
there a “drift” or other pattern as trials proceed?

Durbin-Watson Test
81
Other Issues

Linearity Assumption: A nonlinear pattern
in the residuals vs. predicted values plot
suggests that we need to revise our
assumption of a linear parametric
relationship between X and Y.

Outliers: These will show up in rather
obvious ways on the various plots.
82
When the assumptions are violated...

Discarding data is almost always the wrong
thing to do. Some things you can do are...


Consider transformations of the data.

Transformations of the response variable
[e.g. Log(Y)] often help with normality and/or
constancy of variance issues.

Transformations of the predictor variable(s)
may solve nonlinearity issues
Lastly, we may consider other more complex
models.
83
When outliers are present...

Some formal tests exist to classify outliers
(we’ll talk about them later)

Investigate – don’t eliminate

Lacking a very good reason (e.g. experimenter
made error in recording the data) you should
never be throwing an outlier away.

One good thing to do is to try to figure out how
much effect the outlier has on your various
estimates (we’ll also learn how to do this later)
84
Collaborative Learning Activity
Please discuss problem #2.6 from
the handout.
85
Questions?
86
Upcoming in Topic 3...
Multiple Regression Analysis
Related Reading: Chapter 8
87