Chapter 6: Ordinary Least Squares Estimation Procedure –

The Properties
Chapter 6 Outline
• Clint’s Assignment: Assess the Effect of Studying on Quiz Scores
• Review
o Regression Model
o Ordinary Least Squares (OLS) Estimation Procedure
o The Estimates, bConst and bx, Are Random Variables
• Strategy: General Properties and a Specific Application
o Review: Assessing Clint’s Opinion Poll Results
o Preview: Assessing Professor Lord’s Quiz Results
• Standard Ordinary Least Squares (OLS) Premises
• New Equation for the Ordinary Least Squares (OLS) Coefficient
Estimate
• General Properties: Describing the Coefficient Estimate’s Probability
Distribution
o Mean (Center) of the Coefficient Estimate’s Probability
Distribution
o Variance (Spread) of the Coefficient Estimate’s Probability
Distribution
• Estimation Procedures and the Estimate’s Probability Distribution
o Importance of the Mean (Center)
o Importance of the Variance (Spread)
• Reliability of the Coefficient Estimate
o Variance (Spread) of the Error Term’s Probability Distribution
o Sample Size: Number of Observations
o Range of the Explanatory Variable
o Reliability Summary
• Best Linear Unbiased Estimation Procedure (BLUE)
Chapter 6 Prep Questions
1. Run the Distribution of Coefficient Estimates simulation in the Econometrics
Lab by clicking the following link:

[Link to MIT-Lab 6P.1 goes here.]

After completing the lab, fill in the following blanks:

Numerical
Value Of Your Simulation’s
Coefficient Calculations Calculations
 2

Repetition Estimate Mean Variance Mean Variance

1 _______ _______ _______ _______ _______
2 _______ _______ _______ _______ _______
3 _______ _______ _______ _______ _______
NB: You must click the “Next Problem” button to get to the
simulation’s problem 1.
2. Review the arithmetic of means:
a. Mean of a constant times a variable: Mean[cx] = _____________
b. Mean of a constant plus a variable: Mean[c + x] = _____________
c. Mean of the sum of two variables: Mean[x + y] = _____________
3. Review the arithmetic of variances:
a. Variance of a constant times a variable: Var[cx] = _____________
b. Variance of the sum of a variable and a constant: Var[c + x] =
_____________
c. Variance of the sum of two variables: Var[x + y] =
___________________
d. Variance of the sum of two independent variables: Var[x + y] =
___________________
4. Consider an estimate’s probability distribution:
a. Why is the mean (center) of the probability distribution important?
Explain.
b. Why is the variance (spread) of the probability distribution important?
Explain.
Clint’s Assignment: Assess the Effect of Studying on Quiz Scores
Clint’s assignment is to assess the theory that additional studying increases quiz
scores. To do so he must use data from Professor Lord’s first quiz, the number of
minutes studied and the quiz score for each of the three students in the course:

Minutes
Quiz
Studied
Student Score (y)
(x)
1 5 66
2 15 87
3 25 90
Table 6.1: First Quiz Results

Project: Use data from Professor Lord’s first quiz to assess the effect of
studying on quiz scores.
Review
Regression Model
Clint uses the following regression model to complete his assignment:
 3

yt = βConst + βxxt + et where yt = quiz score of student t

xt = minutes studied by student t
et = error term for student t
The Parameters: βConst and βx
βConst and βx are the parameters of the model. Let us review their interpretation:
• βConst reflects the number of points Professor Lord gives students just for
showing up.
• βx reflects the number of additional points earned for each additional
minute of studying.
The Error Term
The error term, et, plays a crucial role in the model. The error term represents
random influences. The mean of the error term’s probability distribution for each
student equals 0:
Mean[e1] = 0 Mean[e2] = 0 Mean[e3] = 0
Consequently, the error terms have no systematic on affect quiz scores.
Sometimes the error term will be positive and sometimes it will be negative, but
after many, many quizzes each student’s error terms will average out to 0. When
the probability distribution of the error term is symmetric, the chances that a
student will score better than “usual” on one quiz equal the chances that the
student will do worse than “usual.”
Ordinary Least Squares (OLS) Estimation Procedure
As a consequence of the error terms (random influences), we can never determine
the actual values of βConst and βx; consequently, Clint has no choice but to
estimate the values. The ordinary least squares (OLS) estimation procedure is the
most commonly used procedure for doing this:
T

∑(y t − y )( xt − x )
bx = t =1
T
bConst = y − bx x
∑ (x − x )
t =1
t
2

Using the results of the first quiz, Clint estimates the values of the coefficient and
constant:
First Quiz Data Ordinary Least Squares (OLS) Estimates: Esty = 63 + 1.2x
Student x y
1 5 66 bConst = Estimated points for showing up = 63
⎯→
2 15 87 bx = Estimated points for each minute studied = 1.2
3 25 90
 4

x1 + x2 + x3 5 + 15 + 25 45 y1 + y2 + y3 66 + 87 + 90 243
x= = = = 15 y= = = = 81
3 3 3 3 3 3

Student yt y yt − y xt x xt − x
1 66 81 −15 5 15 −10
2 87 81 6 15 15 0
3 90 81 9 25 15 10

Student ( yt − y )( xt − x ) ( xt − x ) 2
1 (−15)(−10) = 150 (-10) 2 = 100
2 (6)(0) = 0 (0) 2 = 0
3 (9)(10) = 90 (10) 2 = 100
Sum = 240 Sum = 200
T T

∑(y
t =1
t − y )( xt − x ) = 240 ∑ (x − x )
t =1
t
2
= 200
T

∑(y t − y )( xt − x )
240 6 6
bx = t =1
T
= = = 1.2 bConst = y − bx x = 81 − × 15 = 63
∑ (x − x ) 2 200 5 5
t
t =1

The Estimates, bConst and bx, Are Random Variables

In the previous chapter, we used the Econometrics Lab to show that the estimates
for the constant and coefficient, bConst and bx, are random variables. As a
consequence of the error terms (random influences), we could not determine the
numerical value of the estimates for the constant and coefficient, bConst and bx,
before we conduct the experiment, even if we knew the actual values of the
constant and coefficient, βConst and βx. Furthermore, we can never expect the
estimates to equal the actual values. Consequently, we must assess the reliability
of the estimates. We shall focus on the coefficient estimate:
Estimate Reliability: How reliable is the coefficient estimate calculated from
the results of the first quiz? That is, how confident can Clint be that the
coefficient estimate, 1.2, will be close to the actual value of the coefficient?
Strategy: General Properties and a Specific Application
Review: Assessing Clint’s Opinion Poll Results
Clint faced a similar problem when he polled a sample of the student population
to estimate the fraction of students supporting him. 12 of the 16 randomly selected
students polled, 75 percent, supported Clint thereby suggesting that he was
leading. But we then observed that it was possible for this result to occur even if
 5

the election was actually a tossup. In view of this, we asked how confident Clint
should be in the results of his single poll. To address this issue, we turned to the
general properties of polling procedures to assess the reliability of the estimate
Clint obtained from his single poll:
General Properties versus One Specific Application
↓ ↓
Clint’s Estimation
Procedure: Apply the polling procedure
Calculate the fraction of ⎯⎯⎯⎯⎯⎯⎯⎯⎯→ once to Clint’s sample of the 16
the 16 randomly selected randomly selected students:
students supporting Clint
⏐ v + v + … + v16 ⏐
↓ EstFrac = 1 2 ↓
16
Before Poll vt = 1 if for Clint After Poll
↓ = 0 if not for Clint ↓
Random Variable: Estimate: Numerical Value
Probability Distribution ↓
⏐ 12 3
EstFrac = = = .75
⏐ 16 4
↓
How reliable is EstFrac?
Mean[ EstFrac ] = p = ActFrac = Actual fraction of the population supporting Clint
p (1 − p ) p (1 − p )
Var[ EstFrac ] = = where T = SampleSize
T 16
↓
Mean and variance describe the center and spread of the estimate’s probability distribution
While we could not determine the numerical value of the estimated
fraction, EstFrac, before the poll was conducted, we could describe its probability
distribution. Using algebra, we derived the general equations for the mean and
variance of the estimated fraction’s, EstFrac’s, probability distribution. Then, we
checked our algebra with a simulation by exploiting the relative frequency
interpretation of probability: after many, many repetitions, the distribution of the
numerical values mirrors the probability distribution for one repetition.
What can we deduce before the poll is
conducted?
↓
General properties of the polling
procedure described by EstFrac‘s
probability distribution.
↓
 6

Probability distribution is described by

its mean (center) and variance (spread).
↓
Use algebra to derive the equations for
the probability distribution’s
mean and variance.
↓ Check the algebra with a
Mean[ EstFrac ] = p simulation by exploiting the
⎯⎯→
p (1 − p ) relative frequency
Var[ EstFrac ] = interpretation of probability.
T
The estimated fraction’s probability distribution allowed us to assess the
reliability of Clint’s poll.
Preview: Assessing Professor Lord’s Quiz Results
Using the ordinary least squares (OLS) estimation procedure we estimated the
value of the coefficient to be 1.2. This estimate is based on a single quiz. The fact
that the coefficient estimate is positive suggests that additional studying increases
quiz scores. But how confident can we be that the coefficient estimate is close to
the actual value? To address the reliability issue we shall focus on the general
properties of the ordinary least squares (OLS) estimation procedure:
General Properties versus One Specific Application
↓ ⏐
⏐
OLS Estimation Procedure: ↓
Estimate βConst and βx by Apply the estimation
finding the bConst and bx that ⎯⎯⎯⎯⎯⎯⎯→ procedure once to the first
minimize the sum of squared quiz’s data:
residuals ⏐
⏐
↓ Model: ↓
Before experiment yt = βConst + βxxt + et After experiment
↓ ↓
Random Variable: Estimate: Numerical Value
Probability Distribution OLS equations: ↓
T
⏐
⏐ ∑ ( yt − y )( xt − x )
bx =
240 6
= = 1.2
⏐ bx = t =1
200 5
⏐ T

⏐ ∑ ( xt − x ) 2
6
bConst = 81 − × 15 = 63
⏐ t =1
5
↓ bConst = y − bx x
Mean[bx] = ?
Var[bx] = ?
 7

Mean and variance describe the center and spread of the estimate’s probability distribution
While we cannot determine the numerical value of the coefficient estimate
before the quiz is given, we can describe its probability distribution. The
probability distribution tells us how likely it is for the coefficient estimate based
on a single quiz to equal each of the possible values. Using algebra, we shall
derive the general equations for the mean and variance of the coefficient
estimate’s probability distribution. Then, we will check our algebra with a
simulation by exploiting the relative frequency interpretation of probability: after
many, many repetitions, the distribution of the numerical values mirrors the
probability distribution for one repetition.
What can we deduce before the poll is
conducted?
↓
General properties of the OLS
estimation procedure described by the
coefficient estimate’s probability
distribution.
↓
Probability distribution is described by
its mean (center) and variance (spread).
↓
Use algebra to derive the equations for Check the algebra with a
the probability distribution’s simulation by exploiting the
⎯⎯→
mean and variance. relative frequency
interpretation of probability.
The coefficient estimate’s probability distribution will allow us to assess the
reliability of the coefficient estimate calculated from Professor Lord’s quiz.
Standard Ordinary Least Squares (OLS) Regression Premises
To derive the equations for the mean and variance of the coefficient estimate’s
probability distribution, we shall apply the standard ordinary least squares (OLS)
regression premises. As we mentioned Chapter 5, these premises make the
analysis as straightforward as possible. In later chapters, we will relax these
premises to study more general cases. In other words, we shall start with the most
straightforward case and then move on to more complex ones later.
• Error Term Equal Variance Premise: The variance of the error term’s
probability distribution for each observation is the same; all the variances
equal Var[e]:
Var[e1] = Var[e2] = … = Var[eT] = Var[e]
 8

• Error Term/Error Term Independence Premise: The error terms are

independent: Cov[ei, ej] = 0.
Knowing the value of the error term from one observation does not
help us predict the value of the error term for any other observation.
• Explanatory Variable/Error Term Independence Premise: The
explanatory variables, the xt’s, and the error terms, the et’s, are not
correlated.
Knowing the value of an observation’s explanatory variable does not
help us predict the value of that observation’s error term.
To keep the algebra manageable, we shall assume that the explanatory variables
are constants in the derivations that follow. This assumption allows us to apply
the arithmetic of means and variances easily. While this simplifies our algebraic
manipulations, it does not affect the validity of our conclusions.
New Equation for the Ordinary Least Squares (OLS) Coefficient
Estimate
In Chapter 5, we derived an equation that expressed the OLS coefficient estimate
in terms of the x’s and y’s:
T

∑(y t − y )( xt − x )
bx = t =1
T

∑ (x − x )
t =1
t
2

It is advantageous to use a different equation to derive the equations for the mean
and variance of the coefficient estimate’s probability distribution, however; we
shall use an equivalent equation that expresses the coefficient estimate in terms of
the x’s, e’s, and βx rather than in terms of the x’s and y’s:1
T

∑ ( x − x )e t t
bx = β x + t =1
T

∑ (x − x )
t =1
t
2

To keep the notation as straightforward as possible, we shall focus on the 3

observation case. The logic for the general case is identical to the logic for the 3
observation case:
( x − x )e1 + ( x2 − x )e2 + ( x3 − x )e3
bx = β x + 1
( x1 − x ) 2 + ( x2 − x ) 2 + ( x3 − x )2
General Properties: Describing the Coefficient Estimate’s Probability
Distribution
Mean (Center) of the Coefficient Estimate’s Probability Distribution
 9

To calculate the mean of bx’s probability distribution, review the arithmetic of

means:
• Mean of a constant times a variable: Mean[cx] = c Mean[x];
• Mean of a constant plus a variable: Mean[c + x] = c + Mean[x];
• Mean of the sum of two variables: Mean[x + y] = Mean[x] + Mean[y];
and recall that the error term represents random influences:
• The mean of each error term’s probability distribution is 0:
Mean[e1] = Mean[e2] = Mean[e3] = 0.
Now, we apply algebra to the new equation for the coefficient estimate, bx:
( x1 − x )e1 + ( x2 − x )e2 + ( x3 − x )e3
[
Mean[bx ] = Mean β x +
( x1 − x ) 2 + ( x2 − x ) 2 + ( x3 − x ) 2
]
Applying Mean[c + x] = c + Mean[x]
( x − x )e1 + ( x2 − x )e2 + ( x3 − x )e3
= β x + Mean 1 [
( x1 − x ) 2 + ( x2 − x ) 2 + ( x3 − x ) 2
]
Rewriting the fraction as a product
= β x + Mean [( 1
( x1 − x ) + ( x2 − x ) 2 + ( x3 − x ) 2
2
)(( x − x )e + ( x
1 1 2 − x )e2 + ( x3 − x )e3 )]

Applying Mean[cx] = cMean[x]

= βx + 2
1
( x1 − x ) + ( x2 − x ) 2 + ( x3 − x ) 2
Mean [(( x − x )e + ( x − x )e + ( x − x )e )]
1 1 2 2 3 3

Applying Mean[x + y] = Mean[x] +

Mean[y]
= βx +
1
( x1 − x ) + ( x2 − x ) 2 + ( x3 − x ) 2
2
[
Mean[( x1 − x )e1 ] + Mean[( x2 − x )e2 ] + Mean[( x3 − x )e3 ] ]
Applying Mean[cx] = cMean[x]
= βx +
1
( x1 − x ) + ( x2 − x ) 2 + ( x3 − x ) 2
2
[
( x1 − x )Mean[e1 ] + ( x2 − x )Mean[e2 ] + ( x3 − x )Mean[e3 ] ]
Since Mean[e1] = Mean[e2] =
Mean[e3] = 0
= βx
So, we have shown that
Mean[bx] = βx
 10

Probability Distribution

bx
Mean[bx] = βx
Figure 6.1: Probability Distribution of Coefficient Estimates

Consequently, the ordinary least squares (OLS) estimation procedure for

the value of the coefficient is unbiased. In any one repetition of the experiment,
the mean (center) of the probability distribution equals the actual value of the
coefficient. The estimation procedure does not systematically overestimate or
underestimate the actual coefficient value, βx. If the probability distribution is
symmetric, the chances that the estimate calculated from one quiz will be too high
equal the chances that it will be too low.
Econometrics Lab 6.1: Checking the Equation for the Mean
We can use the Distribution of Coefficient Estimates simulation in our
Econometrics Lab to replicate the quiz many, many times. But in reality, Clint
only has information from one quiz, the first quiz. How then can a simulation be
useful? The relative frequency interpretation of probability provides the answer.
The relative frequency interpretation of probability tells us that the distribution of
the numerical values after many, many repetitions of the experiments mirrors the
probability distribution of one repetition. Consequently, repeating the experiment
many, many times reveals the probability distribution for the one quiz:
Distribution of the
Numerical Values
⏐ After many,
⏐ many
↓ repetitions
Probability Distribution
We can use the simulation to check the algebra we used to derive the equation for
the mean of the coefficient estimate’s probability distribution:
 11

Mean[bx] = βx
If our algebra is correct, the mean (average) of the estimated coefficient values
should equal the actual value of the coefficient, βx, after many, many repetitions.

Figure 6.2: Distribution of Coefficient Estimates Simulation

[Link to MIT-Lab 6.1 goes here.]

Recall that a simulation allows us to do something that we cannot do in the real

world. In the simulation, we can specify the actual values of the constant and
coefficient, βConst and βx. The default setting for the actual coefficient value is 2.
Be certain that the Pause checkbox is checked. Click Start. Record the numerical
value of the coefficient estimate for the first repetition. Click Continue to simulate
the second quiz. Record the value of the coefficient estimate for the second
repetition and calculate the mean and variance of the numerical estimates for the
first two repetitions. Note that your calculations agree with those provided by the
simulation. Click Continue again to simulate the third quiz. Calculate the mean
and variance of the numerical estimates for the first three repetitions. Once again,
note that your calculations and the simulation’s calculations agree. Continue to
click Continue until you are convinced that the simulation is calculating the mean
and variance of the numerical values for the coefficient estimates correctly.
Now, clear the Pause checkbox and click Continue. The simulation no
longer pauses after each repetition. After many, many repetitions, click Stop.
 12

Probability Distribution

bx
2
Figure 6.3: Histogram of Coefficient Value Estimates

Question: What does the mean (average) of the coefficient estimates equal?
Answer: It equals about 2.0.
This lends support to the equation for the mean of the coefficient estimate’s
probability distribution that we just derived. Now, change the actual coefficient
value from 2 to 4. Click Start and then after many, many repetitions, click Stop.
What does the mean (average) of the estimates equal? Next, change the actual
coefficient value to 6 and repeat the process.

Equation: Simulation:
Mean of Mean (Average) of
Coef Estimate Estimated Coef
Actual Prob Dist Simulation Values, bx, from
βx Mean[bx] Repetitions the Experiments
2 2 >1,000,000 ≈2.0
4 4 >1,000,000 ≈4.0
6 6 >1,000,000 ≈6.0
Table 6.2: Distribution of Coefficient Estimate Simulation Results

Conclusion: In all cases, the mean (average) of the estimates for the
coefficient value equals the actual value of the coefficient after many, many
repetitions. The simulations confirm our algebra. The estimation procedure does
not systematically underestimate or overestimate the actual value of the
coefficient. The ordinary least squares (OLS) estimation procedure for the
coefficient value is unbiased.
 13

Variance (Spread) of the Coefficient Estimate’s Probability Distribution

Now, we turn our attention to the variance of the coefficient estimate’s probability
distribution. To derive the equation for the variance, begin by reviewing the
arithmetic of variances:
• Variance of a constant times a variable: Var[cx] = c2Var[x].
• Variance of the sum of a variable and a constant: Var[c + x] = Var[x].
• Variance of the sum of two independent variables: Var[x + y] = Var[x] +
Var[y].
Focus on the first two standard ordinary least squares (OLS) premises:
• Error Term Equal Variance Premise: Var[e1] = Var[e2] = Var[e3] = Var[e].
• Error Term/Error Term Independence Premise: The error terms are
independent; that is, Cov[et, ej] = 0.
Recall the new equation for bx:
( x − x )e1 + ( x2 − x )e2 + ( x3 − x )e3
bx = β x + 1
( x1 − x ) 2 + ( x2 − x ) 2 + ( x3 − x )2
Therefore,
( x1 − x )e1 + ( x2 − x )e2 + ( x3 − x )e3
[
Var[bx ] = Var β x +
( x1 − x ) 2 + ( x2 − x ) 2 + ( x3 − x ) 2
]
Applying Var[c + x] = Var[x]
( x − x )e1 + ( x2 − x )e2 + ( x3 − x )e3
= Var 1 [
( x1 − x ) 2 + ( x2 − x ) 2 + ( x3 − x ) 2
]
Rewriting the fraction as a product
= Var [( ( x − x ) + ( x −1 x ) + ( x − x ) )(( x − x )e + ( x − x )e + ( x − x )e )]
2 2 2 1 1 2 2 3 3
1 2 3

Applying Var[cx] = c2Var[x]

=
[( x − x ) 2
1
+ ( x2 − x ) 2 + ( x3 − x ) 2 ]2
Var [(( x − x )e + ( x − x )e + ( x − x )e )]
1 1 2 2 3 3
1

Error Term/Error Term Independence Premise:

Var[x + y] = Var[x] + Var[y]
=
[( x − x ) 2
1
+ ( x2 − x ) 2 + ( x3 − x ) 2 ]2
[Var[( x − x )e ] + Var[( x − x )e ] + Var[( x − x )e ]]
1 1 2 2 3 3
1

Applying Var[cx] = c2Var[x]

=
1
[( x − x ) + ( x − x ) + ( x − x ) ]
2 2
[( x − x ) Var[e ] + ( x − x ) Var[e ] + ( x − x ) Var[e ]]
2 2 1
2
1 2
2
2 3
2
3
1 2 3

Error Term Equal Variance Premise:

Var[e1] = Var[e2] = Var[e3] = Var[e]
 14

=
[( x − x ) 2
1
+ ( x2 − x ) 2 + ( x3 − x ) 2 ]2
[( x − x ) Var[e] + ( x − x ) Var[e] + ( x − x ) Var[e]]
1
2
2
2
3
2

Factoring out the Var[e]

=
1
[( x − x ) + ( x − x ) + ( x − x ) ]
2 2
[( x − x ) + ( x − x ) + ( x − x ) ]Var[e]
2 2 1
2
2
2
3
2

1 2 3

Simplifying
Var[e ]
=
( x1 − x ) + ( x2 − x ) 2 + ( x3 − x ) 2
2

Var[e ]
We can generalize this: Var[bx ] = T

∑ (x − x )
t =1
t
2

The variance of the coefficient estimate’s probability distribution equals the

variance of the error term’s probability distribution divided by the sum of squared
x deviations.
Econometrics Lab 6.2: Checking the Equation for the Variance
We shall now use the Distribution of Coefficient Estimates simulation to check
the equation that we just derived for the variance of the coefficient estimate’s
probability distribution.

Figure 6.4: Distribution of Coefficient Estimates Simulation

[Link to MIT-Lab 6.2 goes here.]

The simulation automatically spreads the x values uniformly between 0 and 30.
We shall continue to consider three observations; accordingly, the x values are 5,
 15

15, and 25. To convince yourself of this, be certain that the Pause checkbox is
checked. Click Start and then Continue a few times to observe that the values of x
are always 5, 15, and 25.
Next, recall the equation we just derived for the variance of the coefficient
estimate’s probability distribution:
Var[e ] Var[e ]
Var[bx ] = T =
( x − x ) + ( x2 − x ) 2 + ( x3 − x ) 2
2

∑ ( xt − x )2 1
t =1
By default, the variance of the error term probability distribution is 500; therefore,
the numerator equals 500. Let us turn our attention to the denominator, the sum of
squared x deviations. We have just observed that the x values are 5, 15, and 25.
Their mean is 15 and their sum of squared deviations from the mean is 200:
x +x +x 5 + 15 + 25 45
x= 1 2 3 = = = 15
3 3 3
Student xt x xt − x ( xt − x ) 2
1 5 15 −10 (−10) 2 = 100
2 15 15 0 (0) 2 = 0
25 15 10 2
3 (10) = 100
Sum = 200
T

∑ (x − x )
t =1
t
2
= 200
That is,
( x1 − x ) 2 + ( x2 − x ) 2 + ( x3 − x )2 = 200
When the variance of the error term’s probability distribution equals 500 and the
sum of squared x deviations equals 200, the variance of the coefficient estimate’s
probability distribution equals 2.50:
Var[e ] Var[e ] 500
Var[bx ] = T = = = 2.50
( x1 − x ) + ( x2 − x ) + ( x3 − x )
2 2 2

∑ ( xt − x ) 2 200
t =1
To show that the simulation confirms this, be certain that the Pause
checkbox is cleared and click Continue. After many, many repetitions click Stop.
Indeed, after many, many repetitions of the experiment the variance of the
numerical values is about 2.50. The simulation confirms the equation we derived
for the variance of the coefficient estimate’s probability distribution.
Estimation Procedures and the Estimate’s Probability Distribution:
Importance of the Mean (Center) and Variance (Spread)
Let us review what we learned about estimation procedures when we studied
Clint’s opinion poll in Chapter 3:
 16

• Importance of the Probability Distribution’s Mean: Formally, an

estimation procedure is unbiased whenever the mean (center) of the
estimate’s probability distribution equals the actual value. The relative
frequency interpretation of probability provides intuition: If the
experiment were repeated many, many times the average of the numerical
values of the estimates will equal the actual value. An unbiased estimation
procedure does not systematically underestimate or overestimate the actual
value. If the probability distribution is symmetric, the chances that the
estimate calculated from one repetition of the experiment will be too high
equal the chances the estimate will be too low.

Estimate
Actual Value
Figure 6.5: Probability Distribution of Estimates − Importance of the Mean

• Importance of the Probability Distribution’s Variance: When the

estimation procedure is unbiased, the variance of the estimate’s probability
distribution’s variance (spread) reveals the estimate’s reliability; the
variance tells us how likely it is that the numerical value of the estimate
calculated from one repetition of the experiment will be close to the actual
value:
 17

Figure 6.6: Probability Distribution of Estimates − Importance of the Variance

When the estimation procedure is unbiased, the variance of the estimate’s

probability distribution determines reliability.
o As the variance decreases, the probability distribution becomes
more tightly cropped around the actual value making it more likely
for the estimate to be close to the actual value.
o On the other hand, as the variance increases, the probability
distribution becomes less tightly cropped around the actual value
making it less likely for the estimate to be close to the actual value.
Reliability of the Coefficient Estimate
We shall focus on the variance of the coefficient estimate’s probability
distribution to explain what influences its reliability. We shall consider three
factors:
• Variance of the error term’s probability distribution.
• Sample size.
• Range of the x’s.
Estimate Reliability and the Variance of the Error Term’s Probability
Distribution
What is our intuition here? The error term represents the random influences. It is
the error term that introduces uncertainty into the mix. As the variance of the error
term’s probability distribution increases, uncertainty increases; consequently, the
available information becomes less reliable. As the variance of the error term’s
probability distribution increases, we would expect the coefficient estimate to
become less reliable. On the other hand, as the variance of the error term’s
 18

probability distribution decreases, the available information becomes more

reliable, and we would expect the coefficient estimate to become more reliable.
To justify this intuition, recall the equation for the variance of the
coefficient estimate’s probability distribution:
Var[e ] Var[e ]
Var[bx ] = T =
( x1 − x ) + ( x2 − x ) 2 + ( x3 − x ) 2
2

∑ t(
t =1
x − x ) 2

The variance of the coefficient estimate’s probability distribution is directly

proportional to the variance of the error term’s probability distribution.
Econometrics Lab 6.3: Variance of the Error Term’s Probability Distribution

[Link to MIT-Lab 6.3 goes here.]

We shall use the Distribution of Coefficient Estimates simulation to confirm the

role played by the variance of the error term’s probability distribution. To do so,
check the From-To checkbox. Two lists now appear: a From list and a To list.
Initially, 1.0 is selected in the From list and 3.0 in the To list. Consequently, the
simulation will report the percent of repetitions in which the coefficient estimate
falls between 1.0 and 3.0. Since the default value for the actual coefficient, βx,
equals 2.0, the simulation reports on the percent of repetitions in which the
coefficient estimate falls within 1.0 of the actual value. The simulation reports the
percent of repetitions in which the coefficient estimate is “close to” the actual
value where “close to” is considered to be within 1.0.
By default, the variance of the error term’s probability distribution equals
500 and the sample size equals 3. Recall that the sum of the squared x deviations
equals 200 and therefore the variance of the coefficient estimate’s probability
distribution equals 2.50:
Var[e ] Var[e ] 500
Var[bx ] = T = = = 2.50
( x1 − x ) + ( x2 − x ) + ( x3 − x )
2 2 2

∑ ( xt − x ) 2 200
t =1
Be certain that the Pause checkbox is cleared. Click Start and then after many,
many repetitions, click Stop. As Table 6.3 reports, the coefficient estimate lies
within 1.0 of the actual coefficient value in 47.3 percent of the repetitions.
Now, reduce the variance of the error term’s probability distribution from
500 to 50. The variance of the coefficient estimate’s probability distribution now
equals .25:
Var[e ] Var[e ] 50 1
Var[bx ] = T = = = = .25
( x1 − x ) + ( x2 − x ) + ( x3 − x )
2 2 2

∑ ( xt − x ) 2 200 4
t =1
 19

Click Start and then after many, many repetitions, click Stop. The histogram of
the coefficient estimates is now more closely cropped around the actual value,
2.0. The percent of repetitions in which the coefficient estimate lies within 1.0 of
the actual coefficient value rises from 47.3 percent to 95.5 percent.

Simulations:
Probability Estimated Coefficient Values, bx
Actual Distribution Percent
Values Sample x x Equations: Mean Between
βx Var[e] Size Min Max Mean[bx] Var[bx] (Average) Variance 1.0 and 3.0
2 500 3 0 30 2.0 2.50 ≈2.0 ≈2.50 ≈47.3%
2 50 3 0 30 2.0 .25 ≈2.0 ≈.25 ≈95.5%
Table 6.3: Distribution of Coefficient Estimate Simulation Reliability Results

Why is this important? The variance measures the spread of the probability
distribution. This is important when the estimation procedure is unbiased. As the
variance decreases, the probability distribution becomes more closely cropped
around the actual coefficient value and the chances that the coefficient estimate
obtained from one quiz will lie close to the actual value increases. The simulation
confirms this; after many, many repetitions the percent of repetitions in which the
coefficient estimate lies between 1.0 and 3.0 increases from 47.3 percent to 95.5
percent. Consequently, as the error term’s variance decreases, we can expect the
estimate from one quiz to be more reliable. As the variance of the error term’s
probability distribution decreases, the estimate is more likely to be “close to” the
actual value. This is consistent with our intuition, is it not?
Estimate Reliability and the Sample Size
Next, we shall investigate the effect of the sample size, the number of
observations, used to calculate the estimate. Increase the sample size from 3 to 5.
What does our intuition suggest? As we increase the number of observations, we
will have more information. With more information the estimate should become
more reliable; that is, with more information the variance of the coefficient
estimate’s probability distribution should decrease. Using the equation, let us now
calculate the variance of the coefficient estimate’s probability distribution when
there are 5 observations. With 5 observations the x values are spread uniformly at
3, 9, 15, 21, and 27; the mean (average) of the x’s, x , equals 15 and the sum of
the squared x deviations equals 360:
x + x + x + x + x 3 + 9 + 15 + 21 + 27 75
x= 1 2 3 4 5 = = = 15
5 3 5

Student xt x xt − x ( xt − x ) 2
 20

1 3 15 −12 (−12) 2 = 144

2 9 15 −6 (−6) 2 = 36
3 15 15 0 (0) 2 0
4 21 15 6 (6) 2 36
5 27 15 12 (12) 2 = 144
Sum = 360
T

∑ (x − x )
t =1
t
2
= 360

Applying the equation for the value of the coefficient estimate’s probability
distribution:
Var[e ] Var[e ]
Var[bx ] = T =
( x1 − x ) + ( x2 − x ) + ( x3 − x ) 2 + ( x4 − x ) 2 + ( x5 − x ) 2
2 2

∑ t( x − x
t =1
) 2

50
=
(3 − 15) + (9 − 15) + (15 − 15) 2 + (21 − 15) 2 + (27 − 15) 2
2 2

50 50 50
= = = = .1388… ≈ .14
(−12) + (6) + (0) + (6) + (12) 144 + 36 + 0 + 36 + 144 360
2 2 2 2 2

The variance of the coefficient estimate’s probability distribution falls from .25 to
.14. The smaller variance suggests that the coefficient estimate will be more
reliable.
Econometrics Lab 6.4: Sample Size
Are our intuition and calculations supported by the simulation? In fact, the answer
is yes.

[Link to MIT-Lab 6.4 goes here.]

Note that the sample size has increased from 3 to 5. Click Start and then after
many, many repetitions click Stop:

Simulations:
Probability Estimated Coefficient Values, bx
Actual Distribution Percent
Values Sample x x Equations: Mean Between
βx Var[e] Size Min Max Mean[bx] Var[bx] (Average) Variance 1.0 and 3.0
2 500 3 0 30 2.0 2.50 ≈2.0 ≈2.50 ≈47.3%
2 50 3 0 30 2.0 .25 ≈2.0 ≈.25 ≈95.5%
2 50 5 0 30 2.0 .14 ≈2.0 ≈.14 ≈99.3%
 21

Table 6.4: Distribution of Coefficient Estimate Simulation Reliability Results

After many, many repetitions the percent of repetitions in which the coefficient
estimate lies between 1.0 and 3.0 increases from 95.5 percent to 99.3 percent. As
the sample size increases, we can expect the estimate from one quiz to be more
reliable. As the sample size increases, the estimate is more likely to be “close to”
the actual value.
Estimate Reliability and the Range of x’s
Let us again begin by appealing to our intuition. As the range of x’s becomes
smaller, we are basing our estimates on less variation in the x’s, less diversity;
accordingly, we are basing our estimates on less information. As the range
becomes smaller, the estimate should become less reliable and consequently, the
variance of the coefficient estimate’s probability distribution should increase. To
confirm this, increase the minimum value of x from 0 to 10 and decrease the
maximum value from 30 to 20. The five x values are now spread uniformly
between 10 and 20 at 11, 13, 15, 17, and 19; the mean (average) of the x’s, x ,
equals 15 and the sum of the squared x deviations equals 40:
x + x + x + x + x 11 + 13 + 15 + 17 + 19 75
x= 1 2 3 4 5 = = = 15
5 3 5
Student xt x xt − x ( xt − x ) 2
1 11 15 −4 (−4) 2 = 16
2 13 15 −2 (−2) =2 4
15 15 0 2 0
3 (0)
17 15 2 2 4
4 (2)
19 15 4 2 = 16
5 (6)
Sum = 40
T

∑ (x − x )
t =1
t
2
= 40
Applying the equation for the value of the coefficient estimate’s probability
distribution:
Var[e ] Var[e ]
Var[bx ] = T =
( x1 − x ) + ( x2 − x ) + ( x3 − x ) 2 + ( x4 − x ) 2 + ( x5 − x ) 2
2 2

∑ t( x − x
t =1
) 2

50
=
(11 − 15) + (13 − 15) + (15 − 15) 2 + (17 − 15) 2 + (19 − 15) 2
2 2

50 50 50 5
= = = = = 1.25
(−4) + (2) + (0) + (2) + (4) 16 + 4 + 0 + 4 + 16 40 4
2 2 2 2 2
 22

The variance of the coefficient estimate’s probability distribution increases from

about .14 to 1.25.
Econometrics Lab 6.5: Range of x’s
Our next lab confirms our intuition.

[Link to MIT-Lab 6.5 goes here.]

After changing the minimum value of x to 10 and the maximum value to 20, click
the Start and then after many, many repetitions click Stop.

Simulations:
Probability Estimated Coefficient Values, bx
Actual Distribution Percent
Values Sample x x Equations: Mean Between
βx Var[e] Size Min Max Mean[ bx ] Var[bx ] (Average) Variance 1.0 and 3.0
2 500 3 0 30 2.0 2.50 ≈2.0 ≈2.50 ≈47.3%
2 50 3 0 30 2.0 .25 ≈2.0 ≈.25 ≈95.5%
2 50 5 0 30 2.0 .14 ≈2.0 ≈.14 ≈99.3%
2 50 5 10 20 2.0 1.25 ≈2.0 ≈1.25 ≈62.8%
Table 6.5: Distribution of Coefficient Estimate Simulation Reliability Results

After many, many repetitions the percent of repetitions in which the coefficient
estimate lies between 1.0 and 3.0 decreases from 99.3 percent to 62.8 percent. An
estimate from one repetition will be less reliable. As the range of the x’s
decreases, the estimate is less likely to be “close to” the actual value.
Reliability Summary
Our simulation results illustrate relationships between information, the variance of
the coefficient estimate’s probability distribution, and the reliability of an
estimate:
More and/or more reliable Less and/or less reliable
information. information.
↓ ↓
Variance of coefficient Variance of coefficient
estimate’s probability estimate’s probability
distribution smaller. distribution larger.
↓ ↓
Estimate more reliable; more Estimate less reliable; less
likely the estimate is “close to” likely the estimate is “close to”
the actual value. the actual value.
Best Linear Unbiased Estimation Procedure (BLUE)
 23

In Chapter 5 we introduced the mechanics of the ordinary least squares (OLS)

estimation procedure and in this chapter we analyzed the procedure’s properties.
Why have we devoted so much attention to this particular estimation procedure?
The reason is straightforward. When the standard ordinary least squares (OLS)
premises are satisfied, no other linear estimation procedure produces more
reliable estimates. In other words, the ordinary least squares (OLS) estimation
procedure is the best linear unbiased estimation procedure (BLUE). Let us
now explain this more carefully.
If an estimation procedure is the best linear unbiased estimation procedure
(BLUE), it must exhibit three properties:
• The estimate must be a linear function of the dependent variable, the yt’s.
• The estimation procedure must be unbiased; that is, the mean of the
estimate’s probability distribution must equal the actual value.
• No other linear unbiased estimation procedure can be more reliable; that
is, the variance of the estimate’s probability distribution when using any
other linear unbiased estimation procedure cannot be less than the variance
when the best linear unbiased estimation procedure is used.
The Gauss-Markov theorem proves that the ordinary least squares (OLS)
estimation procedure is the best linear unbiased estimation procedure.2 We shall
illustrate the theorem by describing two other linear unbiased estimation
procedures that while unbiased, are not as reliable as the ordinary least squares
(OLS) estimation procedure. Please note that while we would never use either of
these estimation procedures to do serious analysis, they are useful pedagogical
tools. They allow us to illustrate what we mean by the best linear unbiased
estimation procedure.
Two New Estimation Procedures
We shall now consider the Any Two and the Min-Max estimation procedures:
• Any Two Estimation Procedure: Choose any two points on the scatter
diagram; draw a straight line through the points. The coefficient estimate
equals the slope of this line.
 24

y Any Two

x
3 9 15 21 27
Figure 6.7: Any Two Estimation Procedure

• Min-Max Estimation Procedure: Choose two specific points on the

scatter diagram; the point with the smallest value of x and the point with
the largest value of x; draw a straight line through the two points. The
coefficient estimate equals the slope of this line.

y Min-Max

x
3 9 15 21 27
Figure 6.8: Min-Max Estimation Procedure

Econometrics Lab 6.6: Comparing the Ordinary Least Squares (OLS), Any Two,
and Min-Max Estimation Procedures
We shall now use the BLUE simulation in our Econometrics Lab to justify our
emphasis on the ordinary least squares (OLS) estimation procedure.

[Link to MIT-Lab 6.6 goes here.]

 25

By default, the sample size equals 5 and the variance of the error term’s
probability distribution equals 500. The From-To values are specified as 1.0 and
3.0:

Sample Size = 5 Simulations:

Actual Estimated Coefficient Values, bx
Estimation Values Mean Percent Between
Procedure β x Var[e] (Average) Variance 1.0 and 3.0
OLS 2.0 500 ≈2.0 ≈1.4 ≈60.4%
Any Two 2.0 50 ≈2.0 ≈14.0 ≈29.0%
Min-Max 2.0 50 ≈2.0 ≈1.7 ≈55.2%
Table 6.6: BLUE Simulation Results

Initially, the ordinary least squares (OLS) estimation procedure is specified. Be

certain that the Pause checkbox is cleared. Click Start and then after many, many
repetitions click Stop. For the OLS estimation procedure, the average of the
estimated coefficient values equals about 2.0 and the variance 1.4. 60.4 percent of
the estimates lie with 1.0 of the actual value. Next, select the Any Two estimation
procedure instead of OLS. Click Start and then after many, many repetitions click
Stop. For the Any Two estimation procedure, the average of the estimated
coefficient values equals about 2.0 and the variance 14.0; 29.0 percent of the
estimates lie within 1.0 of the actual value. Repeat the process one last time after
selecting the Min-Max estimation procedure; the average equals about 2.0 and the
variance 1.7; 55.2 percent of the estimates lie with 1.0 of the actual value.
Let us summarize:
• In all three cases, the average of the coefficient estimates equal 2.0, the
actual value; after many, many repetitions the mean (average) of the
estimates equals the actual value. Consequently, all three estimation
procedures for the coefficient value appear to be unbiased.
• The variance of the coefficient estimate’s probability distribution is
smallest when the ordinary least squares (OLS) estimation procedure is
used. Consequently, the ordinary least squares (OLS) estimation procedure
produces the most reliable estimates.
What we have just observed can be generalized. When the standard ordinary least
squares (OLS) regression premises are met, the ordinary least squares (OLS)
estimation procedure is the best linear unbiased estimation procedure because no
other linear unbiased estimation procedure produces estimates that are more
reliable.
 26

Appendix 6.1: New Equation for the OLS Coefficient Estimate

Begin by recalling the expression for bx that we derived previously in Chapter 5:
T

∑(y t − y )( xt − x )
bx = t =1
T

∑ (x − x )
t =1
t
2

bx is expressed in terms of the x’s and y’s. We wish to express bx in terms of the
x’s, e’s, and βx.
Strategy: Focus on the numerator of the expression for bx and substitute for the
y’s to express the numerator in terms of the x’s, e’s, and βx. As we shall shortly
show, once we do this, our goal will be achieved.
T
We begin with the numerator, ∑(y
t =1
t − y )( xt − x ) , and substitute βConst + βxxt + et

for yt:
T T

∑ ( yt − y )( xt − x ) = ∑ ( βConst + β x xt + et − y )( xt − x )
t =1 t =1
Rearranging terms.
T
= ∑ ( β Const − y + β x xt + et )( xt − x )
t =1

Adding and subtracting βx x .

T
= ∑ ( β Const + β x x − y + β x xt − β x x + et )( xt − x )
t =1
Simplifying.
T
= ∑ [( βConst + β x x − y ) + β x ( xt − x ) + et )]( xt − x )
t =1
Splitting the summation into three parts.
T T T
= ∑ ( β Const + β x x − y )( xt − x ) + ∑ β x ( xt − x ) 2 + ∑ ( xt − x )et
t =1 t =1 t =1

Simplifying the first and second terms.

T T T
= ( β Const + β x x − y )∑ ( xt − x ) + β x ∑ ( xt − x ) 2 + ∑ ( xt − x )et
t =1 t =1 t =1
 27

T T
Now, focus on the first term, ( β Const + β x x − y )∑ ( xt − x ) . What does ∑ (x − x )
t
t =1 t =1
equal?
T T T

∑ ( xt − x ) =
t =1
∑ xt −
t =1
∑x t =1
T
Replacing ∑t =1
x with Tx .
T
= ∑x
t =1
t − Tx
T

∑x t
Since x = t =1
.
T
T

T ∑x t
= ∑ xt − T
t =1
t =1

T
Simplifying.
T T
= ∑x
t =1
t − ∑x
t =1
t

=0
T
Next, return to the expression for the numerator, ∑(y
t =1
t − y )( xt − x ) :
T T T T

∑ ( yt − y )( xt − x ) = (βConst + β x x − y )∑ ( xt − x ) + β x ∑ ( xt − x )2 +
t =1 t =1 t =1
∑ ( x − x )e
t =1
t t

T
↓ ∑ (x − x ) = 0
t =1
t

T T
= 0 + β x ∑ ( xt − x ) 2 + ∑ ( x − x )e
t t
t =1 t =1

Therefore,
T T T

∑(y
t =1
t − y )( xt − x ) = β x ∑ ( xt − x ) 2 + ∑ ( xt − x )et
t =1 t =1
Last, apply this to the equation we derived for bx in Chapter 5:
 28

∑(y t − y )( xt − x )
bx = t =1
T

∑ (x − x )
t =1
t
2

Substituting for the numerator.

T T
β x ∑ ( xt − x ) 2 + ∑ ( xt − x )et
= t =1
T
t =1

∑ (x − x )
t =1
t
2

Splitting the single fraction into two.

T T
β x ∑ ( xt − x )2 ∑ ( x − x )e
t t
= T
t =1
+ t =1
T

∑ (x − x )
t =1
t
2
∑ (x − x )
t =1
t
2

Simplifying the first term.

T

∑ ( x − x )e
t t
= βx + t =1
T

∑ (x − x )
t =1
t
2

We have now expressed bx in terms of the x’s, e’s, and βx.

 29

Appendix 6.2: Gauss-Markov Theorem

Gauss-Markov Theorem: When the standard ordinary least squares (OLS)
premises are satisfied, the ordinary least squared (OLS) estimation procedure is
the best linear unbiased estimation procedure.
Proof: Let
bxOLS = Ordinary least squares (OLS) estimate
First, note that bxOLS is a linear function of the y’s:3
T

∑(y t − y )( xt − x )
b OLS
x = t =1
T

∑ (x − x )
i =1
i
2

OLS
Let w equal the ordinary least squares (OLS) “linear weights”; more
t
specifically,
T
(x − x )
bxOLS = ∑ wtOLS ( yt − y ) where wtOLS = T t
t =1
∑ ( xi − x )2 i =1
OLS
Now, let us derive two properties of w t :
T
• ∑w
t =1
OLS
t =0
T
• ∑w
t =1
OLS
t ( xt − x ) = 1
T
First, ∑w
t =1
OLS
t = 0:
T T
( xt − x )
∑ wtOLS = ∑ T
t =1 t =1
∑ (x − x )
i =1
i
2

Placing the summation in the numerator.

T

∑ (x − x ) t
= t =1
T

∑ (x − x )
i =1
i
2

Splitting the summations in the numerator.

 30

T T

∑ xt − ∑ x
= t =1
T
t =1

∑ (x − x )
i =1
i
2

Since there are T −x terms.

T

∑ x − Tx t
= t =1
T

∑ (x − x )
i =1
i
2

∑x t
Since x = t =1
.
T
T

T ∑x t

∑ xt − T t =1
T
= t =1
T

∑ (x − x )
i =1
i
2

Simplifying.
T T

∑x −∑x t t
= t =1
T
t =1

∑ (x − x )
i =1
i
2

Since the numerator equals 0.

=0
T
Second, ∑w
t =1
OLS
t = 0:
T T
( xt − x )
∑ wtOLS ( xt − x ) = ∑ T
( xt − x )
t =1 t =1
∑ (x − x )
i =1
i
2

Simplifying.
T
( xt − x ) 2
=∑ T
t =1
∑ (x − x )
i =1
i
2

Placing the summation in the numerator.

 31

∑ (x − x )
t
2

= t =1
T

∑ (x − x )
i =1
i
2

Since the numerator and denominator are

equal.
=1
Next, consider a new linear estimation procedure whose weights are wtOLS + wt' .
Only when each wt' equals 0 will this procedure to identical to the ordinary least
squares (OLS) estimation procedure. Let bx' equal the coefficient estimate
calculated using this new linear estimation procedure:
T
bx' = ∑ ( wtOLS + wt' ) yt
t =1
Now, let us perform a little algebra:
T
bx' = ∑ ( wtOLS + wt' ) yt
t =1

Since yt = βConst + βxxt + et

T
= ∑ ( wtOLS + wt' )( β Const + β x xt + et )
t =1
Multiplying through.
T T T
= ∑ (w
t =1
OLS
t + wt' ) β Const + ∑ (w
t =1
OLS
t + wt' ) β x xt + ∑ (w
t =1
OLS
t + wt' )et

Factoring out βConst from the first term and

βx from the second.
T T T
= βConst ∑ ( wtOLS + wt' ) + β x ∑ ( wtOLS + wt' ) xt + ∑ (w OLS
t + wt' )et
t =1 t =1 t =1

Again, simplifying the first two terms.

T T T T T
= βConst ∑ wtOLS + β Const ∑ wt' + β x ∑ wtOLS xt + β x ∑ wt' xt + ∑ (w OLS
t + wt' )et
t =1 t =1 t =1 t =1 t =1

T T
Since ∑w
t =1
OLS
t t x = 0 and ∑w
t =1
OLS
t x = 1.
t
 32

T T T
= 0 + β Const ∑ w '
t + βx + βx ∑ w x + '
t t ∑ (w OLS
t + wt' )et
t =1 t =1 t =1

Therefore,
T T T
bx' = β Const ∑ wt' + β x + β x ∑ wt' xt + ∑ ( wtOLS + wt' )et
t =1 t =1 t =1

Now, calculate the mean of the new estimate’s probability distribution, Mean[bt′]:
T T T
Mean[b'x ] = Mean[β Const ∑ wt' + β x + β x ∑ wt' xt + ∑ (w OLS
t + wt' )et ]
t =1 t =1 t =1

Since Mean[c + x] = c + Mean[x]

T T T
= βConst ∑ wt' + βx + β x ∑ wt' xt + Mean[∑ ( wtOLS + wt' )et ]
t =1 t =1 t =1

Focusing on the last term, Mean[cx] =

cMean[x]
T T T
= βConst ∑ wt' + βx + β x ∑ wt' xt + ∑ (w OLS
t + wt' )Mean[et ]
t =1 t =1 t =1

Focusing on the last term, since the error

terms represents random influences,
Mean[et] = 0
T T
= βConst ∑ wt' + βx + β x ∑ wt' xt
t =1 t =1
The new linear estimation procedure must be unbiased:
Mean[bx' ] = β x
Therefore,
T T

∑ wt' = 0
t =1
and ∑w x
t =1
'
t t =0

Next, calculate the variance of bt′:

T T T
Var[b'x ] = Var[βConst ∑ wt' + βx + β x ∑ wt' xt + ∑ (w OLS
t + wt' )et ]
t =1 t =1 t =1

Since Var[c + x] = Var[x].

T
= Var[∑ ( wtOLS + wt' )et ]
t =1
 33

Since the error terms are independent,

covariances equal 0: Var[x + y] = Var[x] +
Var[y].
T
= ∑ Var[( wtOLS + wt' )et ]
t =1

Since Var[cx] = c2Var[x].

T
= ∑ ( wtOLS + wt' )2 Var[et ]
t =1
The variance of each error term’s
probability distribution is identical, Var[e].
T
= ∑ ( wtOLS + wt' ) 2 Var[e]
t =1
Factoring out Var[e].
T
= Var[e ]∑ ( wtOLS + wt' )2
t =1
Expanding the squared terms.
T
= Var[e ]∑ ( wtOLS ) 2 + 2 wtOLS wt' + ( wt' )2
t =1
Splitting up the summation.
T T T
= Var[e ] (∑ (w OLS 2
t ) + 2∑ wtOLS wt' + ∑ ( wt' ) 2 )
t =1 t =1 t =1
T
Now, focus on the cross product terms, ∑wt =1
OLS
t wt' :
T T
( xt − x )
∑w OLS
t wt' = ∑ T
wt'
t =1 t =1
∑ (x − x )
i =1
i
2

Placing the summation in the numerator.

T

∑ ( x − x )w t
'
t
= t =1
T

∑ (x − x )
i =1
i
2

Splitting the summations in the numerator.

T T

∑ ( x w − ∑ xw )
t
'
t
'
t
= t =1
T
t =1

∑ (x − x )
i =1
i
2
 34

Factor out −x from the second term in the

numerator.
T T

∑ x w − x∑wt
'
t
'
t
= t =1
T
t =1

∑ (x − x)
i =1
i
2

T T
Since ∑x w
t =1
t
'
t = 0 and ∑w
t =1
'
t = 0.

0−0
= T

∑ (x − x)
i =1
i
2

Since the numerator equals 0.

=0
Therefore,
T T T
Var[b'x ] = Var[e ] (∑ (w OLS 2
t ) + 2∑ wtOLS wt' + ∑ ( wt' ) 2 )
t =1 t =1 t =1
T
Since ∑w
t =1
OLS
t wt' = 0 .
T T
= Var[e ] (∑ (w OLS 2
t ) + ∑ ( wt' ) 2 )
t =1 t =1
The variance of the estimate’s probability distribution is minimized whenever
each wt′ equals 0, whenever the estimation procedure is the ordinary least squares
(OLS) estimation procedure.

1
Appendix 6.1 appearing at the end of this chapter shows how we can derive the
second equation for the coefficient estimate, bx, from the first.
2
The proof appears at the end of this chapter in Appendix 6.2.
3
To reduce potential confusion, the summation index in the denominator has been
changed from t to i.