Transcript Chapter 4: Properties of the Least Squares Estimator

Chapter 4: Properties of the Least Squares
Estimator
In this chapter, we will
• Review the formulas for b1 and b2
• Derive their means, variances and probability density
functions
• To do so, we will use the assumptions we made in Chapter 3.
• Do a Monte Carlo simulation to demonstrate the idea of a
sampling distribution.
4.1
4.2
Formulas
1)
2)
3)
4)
T  xt yt   xt  yt
b2 
2
2
T  xt   xt 
b2 
 ( x t  x )( yt  y )
b2 
2
(
x

x
)
 t
b1  y  b2 x
All 4 formulas are equivalent,
will give the same estimate.
Remember the calculations we
did in class.
 ( x t  x ) yt
 ( xt  x )
b2   wt yt
2
where
xt  x
wt 
2
 ( xt  x )
Properties of the Least Squares Estimators
Introduction
From Chapter 3:
• The population parameters 1 and 2 are unknown population
parameters that must be estimated
• The Method of Least Squares gives us formulas for b1 and b2,
the estimators for 1 and 2
In Chapter 4:
• Recognize that these estimators are random variables
We want to know their mean, variance and shape
First, review the assumptions of the model from Ch. 3
4.3
Review of Assumptions
1. Linear Regression Model
y = 1 + 2 x + e
2. Error Term has a mean of zero:
E(e) = 0  E(y) = 1 + 2x
3. Error term has constant variance:
Var(e) = E(e2) = 2
4. Error term is not correlated with itself
(no serial correlation):
Cov(ei,ej) = E(eiej) = 0 ij
5. Data on X are not random and thus are
uncorrelated with the error term:
Cov(X,e) = E(Xe) = 0
6. (Optional) Error term has a normal
distribution. E~N(0, 2)
4.4
The Differences between  and b
About 1 and 2
About b1 and b2
• They are parameters
• They do not have means,
variances or p.d.f.’s
• Their values are unknown
• There is no “formula” for 1 and
2
• They are estimated using b1 and
b2 and a sample of data on X and
Y
• They are estimators
• We use the values of b1 and b2 to
draw inferences about 1 and 2
• These estimators are formulas
that explain how to combine the
data points in a sample of data to
get the best fitting line.
• They are functions of the data.
Because the data constitute a
random sample  b1 and b2 are
random variables (will vary from
sample to sample)
4.5
Estimator vs. Estimate
• An estimate is an actual value for b1 and b2. Plug in the
data values on X and Y to get an estimate of the intercept
(b1) and the slope (b2)
• An estimator is a function that explains how to combine
the data points on X and Y to estimate the intercept and
slope
Estimators
b2 
Estimates
 ( x t  x )( yt  y )
 ( xt  x )
b1  y  b2 x
2
b2= 0.1283
b1= 40.768
4.6
“Sampling” Properties of the Least Squares
Estimator
We need to derive the mean, variance and probability distribution functions
for b1 and b2.
• Why?
 Because b1 and b2 are random variables and they are also our
estimators.
• What makes a good estimator?
 On average, it is correct in its job of estimating a population
parameter
 It has small variance: an estimator that varies a lot from sample to
sample isn’t as good as an estimator that varies less from sample to
sample (all else constant)
 Ideally, it has well-define p.d.f such as the Normal p.d.f
4.7
Monte Carlo Simulation
1) First, assume we know 1 and 2 . Therefore, when we
estimate the parameters using b1 and b2, we will know how
well our estimator estimates.
Define the Truth:
– choose values for 1 and 2 and define the exact
distribution of the error term et
– We will define the truth as:
– Yt = 20 + 0.6Xt + et where et ~ Normal(0, = 3)
– so we have chosen 1 = 20 and 2 = 0.6
4.8
Monte Carlo Simulation (con’t)
2) Create the data
a) the errors: Generate 100 samples of T=26 observations on the error
term et by taking random draws (this ensures independence) from a
normal distribution with mean 0 and standard deviation 3.
 we are forcing the error term to obey the assumptions of Chapter 3:
E(et)=0, constant variance 2 = 9 (homoskedasticity) and independent
(serially uncorrelated).
b) the X data: choose any T=26 values for X, since they should be nonrandom.
c) The Y data: are generated by the model and the error term
Yt = 20 + 0.6Xt + et
Perform this step 100 times for each set of 26 data values.
4.9
Monte Carlo Simulation (con’t)
3) Estimate the model 100 times, each time you use the same set
of values for X and the set of Y values generated using the
errors and the X data:
4) You will have 100 values of b1 and 100 values of b2
5) We would expect
– the average of the b1 values would be close to 20
– the average of the b2 values would be close to 0.6
– The average of the residuals would be 0 and the standard
deviation of the residuals would be close to 3 (variance
close to 9)
4.10
Analytical Derivation of Sampling Properties
1) Find the Mean (a.k.a. expected value) of b2:
To do so, it is easiest to work with the following
form for b2:
where
wt 
b2   wt yt
b2   wt yt   wt ( 1   2 xt  et )
  wt 1   2  wt xt   wt et
 1  wt   2  wt xt   wt et
  2   wt et
E (b2 )  E (  2   wt et )
  2   wt E (et )
 2
xt  x
2
(
x

x
)
 t
4.11
Analytical Derivation of Sampling Properties (Con’t)
The expected value of b2 is the parameter that it is estimating.
This property is called unbiasedness.
Both b1 and b2 are unbiased estimators
E(b2 )  2
E(b1)  1
(Proof omitted)
2) Find the Variance of b2 
We see that the variance of
b2 depends on the variance
of the errors and the amount
of variation in the X data.
b2   2   wt et
Var (b2 )  Var (  2   wt et )
  wt2Var (et )
2

2
(
x

x
)
 t
4.12
Analytical Derivation of Sampling Properties (Con’t)
3) The Shape of the distribution of b2
To see the shape of the distribution it is
best to use this formula for b2:
b2   wt yt
Assumption 6:
et~Normal(0,2)
 yt~Normal(1 + 2x,2)
 b2~Normal(2, 2/(xt-x)2)
Note: If we don’t
want to make assp.
6, we can appeal to
the central limit
theorem to show that
b1 and b2 have a
normal p.d.f.
4.13
Recap
E (b2 )   2
2
Var (b2 ) 
2
(
x

x
)
 t
E (b1 )  1
Var (b1 ) 
 2  xt2
T  ( xt  x ) 2
x
Cov (b1, b2 )  
2
 ( xt  x )
2
4.14
4.15
About the variance formulas:
We want our estimators to be precise. High
precision means low variance. The variances of b1
and b2 are affected by:
• The variance of the error term. If the error term has
small variance, our estimator will be more precise.
• The amount of variation in the X data. If there is
lots of variation in the X data, our estimator will be
more precise. (see graph, p.75)
• The size of the sample (T). The larger the sample,
the more precise the estimator.
Gauss Markov Theorem
This is a theorem that tells us that the least squares estimator
is the best one available:
“Under the assumptions 1-5 (the 6th assumption isn’t
needed for the theorem to be true) of the linear regression
model, the least squares estimators b1 and b2 have the
smallest variance of all linear and unbiased estimators of
1 and 2. They are the BLUE (Best, linear, unbiased,
estimator).”
4.16
About the Gauss-Markov Theorem
1. The estimators b1 and b2 are “best” when compared to similar estimators,
those that are linear and unbiased. The Theorem does not say that b1
and b2 are the best of all possible estimators.
2. The estimators b1 and b2 are best within their class because they have the
minimum variance.
3. In order for the Gauss-Markov Theorem to hold, the assumptions 1) – 5)
must be true. If any of the assumptions 1-5 are not true, then b1 and b2
are not the best linear unbiased estimators of 1 and 2.
4. The Gauss-Markov Theorem does not depend on the assumption of
normality
5. The Gauss-Markov theorem applies to the least squares estimators. It
does not apply to the least squares estimates from a single sample.
4.17
Estimating Variances
Recall that the parameters of the model are:
1 the intercept
2 the slope
2 the variance of the error term (et)
Our estimators b1 and b2 will estimate 1 and 2 respectively.
We now need to estimate 2.
We never see or measure the error term (et) but we do calculate
residuals as ^et = (yt – ^yt). = (yt – b1 – b2xt).
We measure the amount of variation in these residuals and use
this as an estimate of the amount of variation in the error
term.
4.18
Estimating Variances (con’t)
2
ˆ
e
ˆ 2   t
T 2
Vaˆr (b2 ) 
We then use this estimated variance
to estimate the variances for b1 and b2:
ˆ 2
 se(b2 )  Vaˆr (b2 )
 ( xt  x )
ˆ 2  xt2
Vaˆr (b1 ) 
 se(b1 ) 
2
T  ( xt  x )
2
x
Coˆv(b1, b2 )  ˆ
2
(
x

x
)
 t
2
Vaˆr (b1 )
4.19