Scribd
Upload
47 views

Classical Multiple Regression

- Classical multiple linear regression models a dependent variable y as a linear function of independent variables (x's) plus some error term ε. - The OLS estimator β is calculated as β = (X'X)-1X'Y and minimizes the sum of squared errors between the observed and predicted y values. - Key assumptions include: linearity, errors are independent and normally distributed with mean 0 and constant variance σ2, and X is fixed with full column rank. Violations can cause statistical issues. - Goodness of fit is assessed using R2, which measures the proportion of variance in y explained by the regression. Adjusted R2 corrects for the number of predictors. -

Uploaded by

13sandip
Copyright
© © All Rights Reserved
Available Formats
Download as DOC, PDF, TXT or read online on Scribd
47 views

Classical Multiple Regression

- Classical multiple linear regression models a dependent variable y as a linear function of independent variables (x's) plus some error term ε. - The OLS estimator β is calculated as β = (X'X)-1X'Y and minimizes the sum of squared errors between the observed and predicted y values. - Key assumptions include: linearity, errors are independent and normally distributed with mean 0 and constant variance σ2, and X is fixed with full column rank. Violations can cause statistical issues. - Goodness of fit is assessed using R2, which measures the proportion of variance in y explained by the regression. Adjusted R2 corrects for the number of predictors. -

Uploaded by

13sandip
Copyright
© © All Rights Reserved
Available Formats
Download as DOC, PDF, TXT or read online on Scribd
You are on page 1/ 5

Classical Multiple Regression

y is a random scalar that is partially explained by x but partial explained by unobserved


things that together are denoted by the random variable . Since is unobserved, we
scale it so that its mean is zero, but it has a variance 2 that must be measured
y=x+ population theory
Typically x includes the variable 1 as the first variable along with p other variables so it is
a p+1-vector. We will use k=p+1. If we have n observations of (y,x), we lay them in a
row and stack them:
1 x 11 x 1p
y1

Y y i X 1 x i1 x ip
n 1
n k

1 x n1 x np
y n

X is called the design matrix as though the research chose the values of x exogenously.
The critical assumption is actually that x is uncorrelated with the error .
Classical Linear Regression Assumptions
1. Y=X+ Linearity
2. E[Y|X]=X or E[|X]=0 explanatory variables are exogenous, independent of errors
3. Var(Y|X)=I errors are iid ( independent and identically distributed)
4. X is fixed
5. X has full column rank: nk and columns of X are not dependent
6. is normally distributed
Problems that might arise with these assumptions
1. wrong regressors, nonlinearity in the parameters, changing parameters
2. biased intercept
3. autocorrelation and heteroskedasticity
4. errors in variables, lagged values, simultaneous equation bias
5. multicollinearity
6. inappropriate tests
Least Squares Estimator
We want to know the latent values of the parameters and so we have to use
the data Y,X to create estimators. Start with . Let the guess of what might be denoted
by the letter b. If Y=Xb+e, then this is really a definition of the resulting residual errors
from a guess b. We want to make these small in a summed squared sense:
min b SSE ee=(Y-Xb)(Y-Xb)=YY-2bXY+bXXb.

SSE
2X ' Y 2X ' Xb 0 , or
b

OLS estimator of

b=(XX)-1XY

Slightly different derivation: Y=Xb+e multiply by X to get XY=XXb+Xe. Since X


and e are independent, the average value of Xe that we might see is approximately zero.
Thus XY=XXb which also gives the OLS estimator of .
1
Interpretation: b (X ' X ) X ' Y

cov(X, Y )

var(X )

cov(X, Y )
var(X )

var(Y )

var(Y )
var(X )

sY
sX

Hence, b is like a correlation between x and y when we do not standardize the scales of
the variables.
The residual vector e is by definition e=Y-Xb or
e=Y-X(XX)-1XY=(I-X(XX)-1X)Y =MY,
where M=I- X(XX)-1X. This matrix M is the centering matrix around the regression
line and is very much like the mean centering matrix H=I-1 1/n=I-1 (11)-11.
Theorem: M is symmetric and idempotent (MM=M), tr(M)=n-k, MX=0.
Given the regression centering matrix M, the sum of squared errors is SSE=ee=
(MY)(MY)=YMY.
Theorem: E[b]= OLS is unbiased.
proof: E[b]=E[(XX)-1XY]= E[(XX)-1X(X+)]= E[(XX)-1XX+(XX)-1X)]
= +(XX)-1XE[]=.
Theorem: var[b]=2(XX)-1.
proof: E[(b-)(b-)]=+(XX)-1X-)+(XX)-1X-)]=
=(XX)-1XX(XX)-1]=(XX)-1XE[]X(XX)-1
=(XX)-1X2IX(XX)-1=2(XX)-1XX(XX)-1=2(XX)-1.
Theorem: Xe=0, estimated errors are orthogonal to the data generating them.
Proof: XMY=(X-XX(XX)-1X)Y=(X-X)Y=0Y=0.
Now consider estimating . SSE=ee=YMY=(X+)M(X+)=
XMX+2MX+M. The first two terms are zero because MX=0. Hence
ee=M=tr(M) (note: the trace of a scalar is trivial)=tr(M). Given this, the
expected value of ee is just tr(ME[])=tr(M=2tr(M)=2(n-k).

Theorem: s2ee/(n-k)=YMY/(n-k) is an unbiased estimator of 2 and s2(XX)-1 is an


unbiased estimator of var[b]. s is called the standard error of the estimate
Gauss-Markov Theorem: The OLS estimator b=(XX)-1XY is BLUE (the Best Linear
Unbiased Estimator) of .
proof: Let a be another estimator of . For linearity a=AY. For unbiased,
E[AY]=E[AX+A]=AX =, so AX=I. That is a=+A.
var(a)=E[(AY-)(AY-)]=E[AA]=2AA.
Define D=A-(XX)-1X, then
var(a)=2((XX)-1X+D)((XX)-1X+D)=2{(XX)-1+DD+(XX)-1XD+DX(XX)-1}.
But DX=AX-(XX)-1XX=I-I=0. Hence var(a)= 2(XX)-1+2DD. The first term is the
variance of b and the second term is a positive definite matrix so var[a]>var[b].
Note: apply this with just an intercept and it implies that x is the BLUE of
How good is the fit? YHY is a measure of the spread in values of y and is called the sum
of squares Total. The regression can reduce the unknown elements to just the sum of
squared Errors, ee. The amount of sum of squares that the regression explains is the
difference: SST-SSE=SSR. R2 is a common measure of performance (also called the
coefficient of determination:
R2

SSR
SSE
e' e
Y ' MY
1
1
1
SST
SST
Y' HY
Y ' HY .

Note: since b minimizes ee, it also maximizes R2.


R2 always goes up if you add a new variable (since we could always set the
coefficient of that variable to zero, using it optimally always reduces error). But it can
reduce the variance matrix (XX) and hence increase the variance of the estimators.
Adjusted R2 corrects for the number of independent variables:
R2 1

n 1
(1 R 2 )
nk
.

Adding a variable with a t-stat >1.0 will increase adjusted R2. Notice: not t-stat>1.96.

Normal Distribution in Regression


Suppose Y|X ~ N(X, 2I). Note: up to now the only statistical assumption that we made
is is iid and independent of X. Now we layer on normality. The likelihood of observing
Y is just the pdf for multivariate normal:
L(Y|X,2I)=(22)-n/2 exp[-(Y-X)(Y-X)/2].
L or equivalently max ln(L) L .
Maximum Likelihood Estimation (MLE): max
,
,
L
0 X ' (Y X) / 2

MLE (X ' X) 1 X' Y

Note: MLE=b from OLS.


L
n 1 1 1
0

(Y X)' (Y X)
2
2 2 2 4

2 MLE e' e / n .
Note: divide by n not n-k. Hence MLE of 2 is biased.
Theorem: Y|X ~ N(X, 2I) then b~N(,2(XX)-1),
independent.

nk 2
s ~ 2n k , and b and s2 are
2

Confidence Intervals
Joint:
(b-)(XX)(b-) ks2Fk,n-k()
One at a time:
bi SE(bi) tn-k(/2)
Simultaneous:
bi SE(bi) kFk , n k ()
Hypothesis Testing
Ho: R=r, where R qk and r q1 for linear restrictions on k1.
Let a be the OLS estimators of using the above q restrictions: a min ee s.t. Ra=r. Let
b be the unconstrained OLS estimators of . Likelihood ratio LR = La/Lb. Define
1
1
LR = -2ln(LR)=2ln(Lb)-2ln(La)= 2 (Y Xb)' (Y Xb) 2 ( Y Xa )' (Y Xa ) . If we

replace 2 with s2=ee/(n-k), then


LR

(Y - Xa)' (Y - Xa) - e' e / q ~ F

q ,n k .
e' e /( n k )
Hence we can test the restrictions R=r by running the regression with the constraints and
unconstrained and computing

(SSE constr SSE unconstr ) / q


SSE unconst /(n k )

and compare to critical value


Fq,n-k().
There are two other tests that are sometimes done: Lagrange Multiplier test and Wald test.
See graph below. The Wald test is a 2 test of whether Runcontr-r is different from zero.
The Lagrange multiplier test is a 2 test of whether the slope of the likelihood is zero at
the constrained value contr. It has the advantage over the LR test of not requiring unconstr
from being estimated.

?
= 0 Lagrange Multiplier Test

Likelihood
?
Ratio 0 =
Test

R=r

?
= 0 Wald Test
constr

unconstr

You might also like