Regression: Introduction

Basic idea:
Use data to identify relationships
among variables and use these
relationships to make predictions.
 Linear regression
• Linear dependence: constant rate of increase of one variable
with respect to another (as opposed to, e.g., diminishing retur
ns).
• Regression analysis describes the relationship between two (o
r more) variables.
• Examples:
– Income and educational level
– Demand for electricity and the weather
– Home sales and interest rates
• Our focus:
–Gain some understanding of the mechanics.
• the regression line
• regression error
– Learn how to interpret and use the results.
– Learn how to setup a regression analysis.
 Two main questions:
•Prediction and Forecasting
– Predict home sales for December given the interest rate for this month.
– Use time series data (e.g., sales vs. year) to forecast future performance
(next year sales).
– Predict the selling price of houses in some area.
• Collect data on several houses (# of BR, #BA, sq.ft, lot size, property tax) an
d their selling price.
• Can we use this data to predict the selling price of a specific house?
•Quantifying causality
– Determine factors that relate to the variable to be predicted; e.g., predict
growth for the economy in the next quarter: use past history on quarterl
y growth, index of leading economic indicators, and others.
– Want to determine advertising expenditure and promotion for the 1999
Ford Explorer.
• Sales over a quarter might be influenced by: ads in print, ads in radio, ads in
TV, and other promotions.
 Motivated Example
• Predict the selling prices of houses in the region.
–Intuitively, we should compare the house for which we need a predicted sellin
g price with houses that have sold recently in the same area, of roughly the sa
me size, same style etc.
• Idea: Treat it as a multiple sample problem.
• Unfortunately, the list of houses meeting these criteria may be quite small, or there
may not be a house of exactly the same characteristics.
• Alternative approach: Consider the factors that determine the selling price of a house
in this region.
• Collect recent historical data on selling prices, and a number of char
acteristics about each house sold (size, age, style, etc.).
–Idea: one sample problem
• To predict the selling price of a house without any particular knowledge of the house
, we use the average selling price of all of the houses in the data set.
–Better idea:
• One of the factors that cause houses in the data set to sell for different amounts of m
oney is the fact that houses come in various sizes.
• A preliminary model might posit that the average value per square foot of a new hou
se is $40 and that the average lot sells for $20,000. The predicted selling price of a h
ouse of size X (in square feet) would be: 20,000 + 40X.
• A house of 2,000 square feet would be estimated to sell for 20,000 + 40(2,000) = $10
0,000.
 Motivated Example
•Probability Model:
– We know, however, that this is just an approximation, and the selling pri
ce of this particular house of 2,000 square feet is not likely to be exactly $
100,000.
– Prices for houses of this size may actually range from $50,000 to $150,000
.
– In other words, the deterministic model is not really suitable. We should t
herefore consider a probabilistic model.
•Let Y be the actual selling price of the house. Then
Y = 20,000 + 40x + ,
where  (Greek letter epsilon) represents a random error term
(which might be positive or negative).
– If the error term  is usually small, then we can say the model is a good o
ne.
– The random term, in theory, accounts for all the variables that are not pa
rt of the model (for instance, lot size, neighborhood, etc.).
– The value of  will vary from sale to sale, even if the house size remains c
onstant. That is, houses of the exact same size may sell for different price
s.
 Regression Model
• The variable we are trying to predict (Y) is called the depend
ent (or response) variable.
• The variable x is called the independent (or predictor, or expl
anatory) variable.
• Our model assumes that
E(Y | X = x) = 0 + 1x (the “population line”) (1)
The interpretation is as follows:
–When X (house size) is fixed at a level x, then we assume the mean of
Y (selling price) to be linear around the level x, where 0 is the (unkno
wn) intercept and 1 is the (unknown) slope or incremental change in Y
per unit change in X.
–0 and 1 are not known exactly, but are estimated from sample data. T
heir estimates are denoted b0 and b1.
• A simple regression model: Consider a model with only one in
dependent variable,.
• A multiple regression model: a model with multiple independe
nt variables.
House Number Y: Actual Selling X: House Size (100s ft2)
Price ($1,000s)
1 89.5 20.0
2 79.9 14.8
3 83.1 20.5
Sample
4 56.9 12.5
15 houses
5 66.6 18.0
from the
6 82.5 14.3
region.
7 126.3 27.5
8 79.3 16.5
9 119.9 24.3
10 87.6 20.2
11 112.6 22.0
12 120.8 .019
13 78.5 12.3
14 74.3 14.0
15 74.8 16.7
Averages 88.84 18.17
 Least Squares Estimation
•price<- c(89.5,79.9,83.1,56.9,66.6,82.5,126.3,79.3,119.9,87.6,112.6,120.8,78.5,74.3,74.8)
•size<- c(20.0,14.8,20.5,12.5,18.0,14.3,27.5,16.5,24.3,20.2,22.0,19.0,12.3,14.0,16.7)
•plot(size,price,xlab= “House size (100 sq ft)”,ylab=“Selling price ($1
,000)”,main=“House Size (X) vs Selling Price (Y)”)
 Assumptions
• These data do not form a perfect line. This is not surprising, c
onsidering that our data are random. In other words, if we as
sume equation (1) then our line predicts the mean for any giv
en level x. However, when we actually take a measurement (i
.e., observe the data), we observe:
Yi = 0 + 1Xi + i, for i = 1,2,…, n = 15,
where i is the random error associated with the ith observati
on.
–Since we don't know the true values of 0 and 1, it is clear that we do n
ot observe the actual errors (i) precisely either.
• Assumptions about the Error
–E(i ) = 0 for i = 1, 2,…,n.
–(i ) =  where  is unknown.
–The errors are independent, that is, the error in the ith observation is independ
ent of the error observed in the jth observation.
–The i are normally distributed (with mean 0 and standard deviation ).
 Least Squares Estimation
• Recall 0 and 1 are (unknown) population parameters.
– From the sample data, we will calculate numbers ˆ0 andˆ1 that are e
stimates of the population parameters.
– How should these numbers be chosen? For any choice ofˆ0 and ˆ1 , w
e can write the following prediction equation
ŷ = ˆ0 + ̂1 X.
– The “hat” is used to denote a value estimated from the model, as oppo
sed to one that is actually observed.
– For each house in our sample of 15 we could check to see how well thi
s equation works at predicting the actual selling prices. Define ei to be
the error associated with the ith observation. That is:
ei = yi - (estimated selling price)
These are sometimes called the residuals or simply errors.
ˆ0 ˆ1
• We will pick the values of and that minimize Si ei2, the s
um of the squares of the residuals. This method is often calle
d Least Squares Regression.
 Using the Equation
• Method of Least squares leads to that the intercept is 18.354 a
nd the slope is 3.879.
–How do we predict the selling price of a house of 1,650 square feet?
• Plug in the value 16.50 (1,650 translated to 100s of square feet) in the regres
sion equation and get predicted selling price = 18.354 + 3.879× (16.50) = 82.3
57.
• Translate to a dollar amount, i.e., $82,357. This is the best estimate you hav
e of the selling price of this house, that is, without any further information
about the house (e.g., neighborhood, number of rooms, lot size, age, etc.).
• Analyzing a Regression
• Estimating the Standard Error
–From the assumptions about the error, the magnitude of  should be a
good guide to the accuracy of a prediction.
–The number  is a population parameter, so we cannot know for certa
in what its value is.
–We therefore use an estimate s that is provided in the regression outp
ut under the name “standard error of the estimate” or just “standard er
ror.”
 Making Predictions
• The estimate s is calculated by (SSE/(n-2))1/2.

–The reason why we divide by n - 2 and not n - 1 has to do with the deg
rees of freedom issue.
–The value of s gives us some idea of the standard deviation of the erro
rs if the model is used to estimate selling prices. In addition, we will m
ake use of the normality assumption to help us make assessments of a
prediction.
• Suppose a house occupies 2,000 square feet. How do we pre
dict the selling price?
–prediction interval: This is used if our goal is to determine a 95% confi
dence interval on the actual selling price of the house. A 95% predictio
n interval for the actual selling price is given by
(18.354 + 3.879× 20 )  t(n - 2, 0.025)s = 95.94  28.07.
–confidence interval: This is used if our goal is to determine a 95% confi
dence interval on the mean selling price of all houses of this size (2,000
square feet). (E[Y|X = x])
It is 95,940  t(n - 2, 0.025)s/√n = 95.94  7.25 .
–In the above examples use the t distribution with n - 2 degrees of freed
om. If n - 2  30 then the standard normal distribution can be used inst
 Making Inferences about Coefficients
• To assess the accuracy of the model, it involves determining
whether a particular variable like house size has any effect o
n the selling price.
–Suppose that when a regression line is drawn it produces a horizontal
line. This means the selling price of the house is unaffected by the size
of the house.
–A horizontal line has a slope of 0, so when no linear relationship exist
s between an independent variable and the dependent variable we sho
uld expect to get 1 = 0.
–But of course, we only observe estimate of 1, which might only be “cl
ose” to zero. To systematically determine when 1 might in fact be zer
o, we will make inferences about it using our estimate , specifically, w
e will do hypothesis tests and build confidence intervals.
• Testing 1, we can test any of the following:
–H0 : 1 = 0 versus HA : 1  0 
–H0 : 1  0 versus HA : 1 < 0
–H0 : 1  0 versus HA : 1 > 0
•  In each case, the null hypothesis can  / sreduced
ˆ1  0be to H0: 1 = 0
ˆ
1
 Example
• Can we conclude at the 1% level of significance that the size
of a house is linearly related to its selling price? Test H0 : 1 =
0 versus HA : 1  0
–Note this is a two-sided test, we are interested in whether there is any
relationship at all between price and size.
–Calculate T = (3.879 - 0) / 0.794 = 4.88.
–That is, we are 4.88 standard deviations from 0. So at the 1% level (cor
responding to thresholds  t(13, 0.005) =  3.012), we reject H0.
–There is sufficient evidence to conclude that house size does linearly a
ffect selling price.
• To get a p-value on this we would need to look up 4.88 insid
e the t-table.
–It is 0.00024 or 0.024%; very small indeed.
ˆ1  t( n2,0.025) s ˆ
• A 95% confidence interval for  is given by 1
1
–For this example: It is 3.879  (2.160)(0.794) = 3.879  1.715.
–Using the 15 data points, we are 95% confident that every extra squar
e foot increases the price of the house by anywhere from $21.64 to $55.
94.
 Method III: Measuring the Strength of the Linea
r Relationship
• Consider the following equation:
Yi - Y = (Yˆ - Y ) + ei.
–Squaring both sides and summing over all data points, and after a little
algebra, we get:
i (Yi -Y )2 = i (Yˆ -Y )2 + i ei2, which we usually rewrite as:
SST = SSR + SSE, (2)
where SST = i (Yi - Y )2 , SSR = i (Yˆ -Y )2 and SSE = i ei2.
–Interpretation:
• SST stands for the “total sum of squares” - this is essentially the total variatio
n in the data set, i.e., the total variation of selling prices.
• SSR stands for “sum of squares due to regression” - this is the squared variat
ion around the mean of the estimated selling prices. This is sometimes called
the total variation explained by the regression.
• SSE stands for “sum of squares due to error” - this is simply the sum of the s
quared residuals, and it is the variation in the Y variable that remains unexpl
ained after taking into account the variable X.
–The interpretation of equation (2) is that the total variation in Y (SST) is
made up of two parts: the total variation explained by the regression (S
 Regression Statistics
• Define R2 = SSR/SST = 1- SSE/SST
–The fraction of the total variation explained by the regression.
–R2 is a measure of the explanatory power of the model.
–Multiple-R = (R2)1/2 (in one variable case = |rXY|)
• According to the definition of R2, adding extraneous explanat
ory variables will artificially inflate the R2.
–We must be careful in interpreting this number.
–Introducing extra variables can lead to spurious results and can interfe
re with the proper estimation of slopes for the important variables.
• In order to penalize an excess of variables, we consider the ad
justed R2, which is
adjusted R2 = 1- [SSE/(n-k-1)]/[SST/(n-1)] .
Here n is the number of data and k is the number of explanat
ory variables.
–The adjusted R2 thus divides numerator and denominator by their DF.
 How to determine the value of used cars that custo
mers trade in when purchasing new cars?
• Car dealers across North America use the “Red Book” to help them
determine the value of used cars that their customers trade in when
purchasing new cars.
–The book, which is published monthly, lists average trade-in values for all bas
ic models of North American, Japanese and European cars.
–These averages are determined on the basis of the amounts paid at recent use
d-car auctions.
–The book indicates alternative values of each car model according to its condi
tion and optional features, but it does not inform dealers how the odometer rea
ding affects the trade in value.
• Question: In an experiment to determine whether the odometer read
ing should be included in the Red Book, an interested buyer of used
cars randomly selects ten 3-year-old cars of the same make, conditi
on, and optional features.
–The trade-in value and mileage for each car are shown in the following table.
 Data
Odometer Reading(1,000 miles) 59 92 61 72 52 67 88 62 95 83
Trade-in Value ($100s) 37 41 43 39 41 39 35 40 29 33
• Run the regression, with Trade-in Value as the dependent variable
(Y) and Odometer Reading as the independent variable (X). The out
put appears on the following page.
• Regression Statistics
– Multiple R = 0.893, R2 = 0.798, Adjusted R2 = 0.773 Standard Error =
2.178
– Analysis of Variance
df SS MS F Significance F
Regression 1 150.14 150.14 31.64 0.000
Residual 8 37.96 4.74
Total 9 188.10 
– Testing
Coeff. Stnd Error t-Stat P-value
Intercept 56.205 3.535 15.90 0.000
x -0.26682 0.04743 -5.63 0.000
 F and F-significance
• F is a test statistic testing whether the estimated model is mea
ningful; i.e., statistically significant.
–F =MSR/MSE
–A large F or a small p-value (or F-significance) implies that the model i
s significant.
–It is unusual not to reject this null hypothesis.
 Questions
• What does the regression line tell us about the relationship between
the two variables?
• Can we conclude at the 5% significance level that, for all cars of th
e type described in the experiment, higher mileage results in a lowe
r trade-in value?
• Predict with 95% confidence the trade-in value of such a car that ha
s been driven 60,000 miles.
• A large national courier company has a policy of selling its cars wh
en the odometer reading reaches 75,000 miles. The company is abo
ut to sell a large number of 3-year-old cars, each equipped with the
same optional features and in the same condition as the 10 cars desc
ribed in the experiment. The company president would like to know
the cars' mean trade-in price. Determine the 95% confidence interva
l estimate of the expected value of all cars that have been driven 75,
000 miles.
 Salary-budget Example
• A large corporation is concerned about maintaining parity in
salary levels across different divisions.
–As a rough guide, it determines that managers responsible for compara
ble budgets in different divisions should have comparable compensatio
n.
• Data Analysis: The following is a list of salary levels for 20 ma
nagers and the sizes of the budgets they manage.
–salary<- c(59.0,67.4,50.4,83.2,105.6,86.0,74.4,52.2,82.6,59.0,44.8,111.4,122.
4, 82.6,57.0,70.8,54.6,111.0,86.2,79.0)
–budget<- c(3.5,5.0,2.5,6.0,7.5,4.5,6.0,4.0,4.5,5.0,2.5,12.5,9.0,7.5,6.0, 5.0,3.0,
8.5, 7.5, 6.5)
–Salary Y ($1000s)
–Budget X ($100,000s)
 Salary-budget Example
• Want to fit a straight line to this data.
– The slope of this line gives the marginal increase in salary with respect to inc
rease in budget responsibility.
– The regression equation is SALARY = 31.9 + 7.73 BUDGET
– Each additional $100,000 of budget responsibility translates to an expected a
dditional salary of $7,730.
– If we wanted to know the average salary corresponding to a budget of 6.0, w
e get a salary of 31.9 + 7.73(6.0) = 78.28.
• Why is the least squares criterion the correct principle to follow?
• Assumptions Underlying Least Squares
– The errors e ,…, e are independent of the values of X ,…,X .
1 n 1 n
– The errors have expected value zero; i.e., E[ei] = 0.
– All the errors have the same variance: Var[ei] = s2, for all i = 1,…,n.
– The errors are uncorrelated; i.e., Corr[ei, ej] = 0 if i  j.
• The first two assumptions imply that E[Y|X = x] = 0 + 1x.
– Do we necessarily believe that the variability in salary levels among manager
s with large budgets is the same as the variability among managers with small
budgets?
 How do we evaluate and use the regression line?
• Evaluate the explanatory power of a model.
– Without using X, how do we predict Y?
– Determine how much of the variability in Y values is explained by the X.
• Measure variability using sums of squared quantities.
• The ANOVA table.
– ANOVA is short for analysis of variance.
– This table breaks down the total variability into the explained and unexplain
ed parts.
– Total SS (9535.8) measures the total variability in the salary levels.
• Without using x, we will use sample mean to do prediction.
– The Regression SS (6884.7) is the explained variation.
• It measures how much variability is explained by differences in budgets.
– Error SS (2651.1) is the unexplained variation.
• This reflects differences in salary levels that cannot be attributed to differences in
budget responsibilities.
– The explained and unexplained variation sum to the Total SS.
• R-squared: R = SSR/SST = 6884.7/9538.8 = 72:2%