Transcript Hatfield.Topic 10

Topic 10 - Linear Regression
• Least squares principle
• Hypothesis tests/confidence
intervals/prediction intervals for regression
1
Linear Regression
•
How much should you pay for a house?
•
Would you consider the median or mean sales price in your area
over the past year as a reasonable price?
•
What factors are important in determining a reasonable price?
– Amenities
– Location
– Square footage
•
To determine a price, you might consider a model of the form:
Price = f(square footage) + e
2
Scatter plots
•
To determine the proper functional relationship between two variables,
construct a scatter plot. For example, for the home sales data from Topic
8, what sort of functional relationship exists between Price and SQFT
(square footage)?
3
Simple linear regression
•
The simplest model form to consider is
Yi = b0 + b1Xi + ei
•
Yi is called the dependent variable or response.
•
Xi is called the independent variable or predictor.
•
ei is the random error term which is typically assumed to have a
•
Normal distribution with mean 0 and variance s2.
We also assume that error terms are independent of each other.
4
Least squares criterion
•
If the simple linear model is appropriate then we need to
estimate the values b0 and b1.
•
To determine the line that best fits our data, we choose the line
that minimizes the sum of squared vertical deviations from our
observed points to the line.
•
In other words, we minimize
n
Q   (Yi  b0 b1X i )2
i 1
5
Scatterplot w/regression line and errors
6
7
Formulas for Least Squares Estimates
S xy [( xi x )( yi  y )] [( xi x )( yi  y )]
ˆ
b1 


2
S xx
(
x

x
)
 i
[( xi x )(xi  x )]
(x  x )
2
is the numerator in sample variance; s
2
 ( x  x)

n 1
2
,
so it is the squared variation of the variable X.
Since you can rewrite
estimated slope
2
(
x

x
)
as

[( x x )(x  x )] , then the
i
i
b̂1 is the ratio of the joint variation of X and Y, to the variation
of X alone.
bˆ0  y  bˆ0 x
. Once the slope is identified, it’s simply a matter of
positioning that line within the graph to minimize the squared variability. The
intercept is a mathematical fallout position on the slope.
8
9
10
11
12
Home sales example
•
For the home sales data, what are least squares estimates for the line of
best fit for Price as a function of SQFT?
•
StatCrunch steps:
– Load the data set
– STAT > REGRESSION > SIMPLE LINEAR
– Set your X and Y values
– Click Calculate…you’ll get a result similar to that shown on the next
slide.
13
Typical Regression Output
Simple linear regression results:
Dependent Variable: PRICE
Independent Variable: SQFT
PRICE = 4781.9307 + 61.36668 SQFT
Sample size: 117
R (correlation coefficient) = 0.8448
R-sq = 0.7136788
Estimate of error standard deviation: 20445.117
Parameter estimates:
Parameter Estimate Std. Err. Alternative DF
Intercept
Slope
4781.9307 6285.482
61.36668 3.6245918
T-Stat
≠ 0 115 0.7607898
P-Value
0.4483
≠ 0 115 16.930645 <0.0001
14
Inference
•
Often times, inference for the slope parameter, b1, is most
important.
•
b1 tells us the expected change in Y per unit change in X.
•
If we conclude that b1 equals 0, then we are concluding that
there is no linear relationship between Y and X.
•
If we conclude that b1 equals 0, then it makes no sense to use
our linear model with X to predict Y.
•
bˆ1has a Normal distribution with a mean of b1 and a variance of
n
s e /  ( X i  X )2
2
i 1
15
Hypothesis test for b1
To test H0: b1 = D0, where D0 is the assumed slope for the
hypothesis (frequently 0), use the test statistic;
HA
bˆ1  D0
T 
se
n
 (X
i 1
 X)
2
i
b1 < D0
b1 > D0
b1 ≠ D0
Reject H0 if
T < -ta,n-2
T > ta,n-2
|T| > ta/2,n-2
16
Home sales example
•
For the home sales data , is the linear relationship between Price and
SQFT significant or is there a significant linear relationship between the
two variables?
Simple linear regression results:
Dependent Variable: PRICE
Independent Variable: SQFT
PRICE = 4781.9307 + 61.36668 SQFT
Sample size: 117
R (correlation coefficient) = 0.8448
R-sq = 0.7136788
Estimate of error standard deviation: 20445.117
Parameter estimates:
Parameter Estimate Std. Err. Alternative DF
Intercept
Slope
4781.9307 6285.482
61.36668 3.6245918
T-Stat
≠ 0 115 0.7607898
P-Value
0.4483
≠ 0 115 16.930645 <0.0001
17
Just to link some concepts together
The denominator in the T calculation for the slope estimator
is;
, the standard error of the slope
se
n
2
(
X

X
)
 i
 s bˆ1
i 1
se
is the
sbˆ
1
is the
and the denominator can be solved for algebraically. It’s the
square root of the numerator in sample variance;
18
Confidence interval for b1
•
A (1-a)100% confidence interval for b1 is
bˆ1  ta /2,n 2se /
n
2
(
X

X
)
 i
i 1
•
For the home sales data , what is a 95% confidence interval for
the expected increase in price for each additional square foot?
•
Note that StatCrunch doesn’t do the CI for you perse, but you
do have all the terms you need….the standard error for the
slope IS given to you by StatCrunch.
19
Confidence interval for mean response
•
Sometimes we want a confidence interval for the average (expected)
value of Y at a given value of X = x*.
•
With the home sales data, suppose a realtor says the average sales
price of a 2000 square foot home is $120,000. Do you believe her?
•
bˆ0  bˆ1x * has a Normal distribution with a mean of b0 + b1x* and
a variance of


*
2
1

(
x

X
)
2

se   n
2
n
(
X

X
)

i


i 1
20
Confidence interval for mean response
•
A (1-a)100% confidence interval for b0 + b1x* is
bˆ0  bˆ1x  ta /2,n 2se
*
1
(x *  X )2
 n
n
 (X i  X )2
i 1
•
With the home sales data, do you believe the realtor’s claim?
•
Note that StatCrunch DOES do this for you.
– Additional steps once you’re in Regression would be to click next,
choose “Predict Y for X”, where X is the X* you’re given in the
problem and then calculate. You’ll get all the basic regression
outputs as well as a 95% CI and a 95% Prediction Interval
(explained later).
21
Prediction interval for a new response
•
Sometimes we want a prediction interval for a new value of Y at a given
value of X = x*.
•
A (1-a)100% prediction interval for Y when X = x* is
*
2
1
(
x

X
)
bˆ0  bˆ1x  ta /2,n 2se 1   n
n
2
(
X

X
)
 i
*
i 1
•
With the home sales data , what is a 95% prediction interval for the
amount you will pay for a 2000 square foot home?
•
See slide on Confidence Interval for StatCrunch procedures.
22
Extrapolation
•
Prediction outside the range of the data is risky and not
appropriate as these predictions can be grossly inaccurate.
This is called extrapolation.
This is supposed to illustrate
why extrapolation is bad. If the
confidence intervals get wider
the farther that you get from the
mean X value, think about what
they’d look like if you picked
some X value that was way out
of range….The resulting CI
would be so wide, it could be
useless.
23
Correlation
•
The correlation coefficient, r, describes the direction and
strength of the straight-line association between two variables.
•
We will use StatCrunch to calculate r and focus on interpretation.
•
If r is negative, then the association is negative. (A car’s value vs. its
age)
•
If r is positive, then the association is positive. (Height vs. weight)
•
r is always between –1 and 1 (-1 < r < 1).
– At –1 or 1, there is a perfect straight line relationship.
– The closer to –1 or 1, the stronger the relationship.
– The closer to 0, the weaker the relationship.
•
Run the correlation by eye applet in StatCrunch.
24
Home sales example
• For the home sales data , consider the correlation
between the variables.
25
Correlation and regression
• As correlation (r) approaches either -1 (negative
relationship) or +1 (positive relationship), it is
visually more apparent that the relationship exists
and the variation around the line is less.
• The square of the correlation, r2, is the proportion
of variation in the value of Y that is explained by the
regression model with X.
• If correlation increases (the data is closer to the
line), there is less variation (error) that is
unexplained by our model.
• Therefore, there is a direct relationship between r
and r2, as correlation increases, so does r2.
26
Association and causation
•
A strong relationship between two variables does not always
mean a change in one variable causes changes in the other.
•
The relationship between two variables is often due to both
variables being influenced by other variables lurking in the
background.
•
The best evidence for causation comes from properly designed
randomized comparative experiments.
27
Does smoking cause lung cancer?
• Unethical to investigate this relationship with a
randomized comparative experiment.
• Observational studies show strong association between
smoking and lung cancer.
• The evidence from several studies show consistent
association between smoking and lung cancer.
– More and longer cigarettes smoked, the more often lung cancer
occurs.
– Smokers with lung cancer usually began smoking before they
developed lung cancer.
• It is plausible that smoking causes lung cancer
• Serves as evidence that smoking causes lung cancer, but
not as strong as evidence from an experiment.
The additional file for Topic 10 contains discussion as well as calculations of key components
28
in regression.