Linear Regression Models Under Conditional

Heteroskedasticity and Autocorrelation

Professor Yongmiao Hong

September 22, 2024
Copyright ◎ 2024 by Professor Hong Yongmiao, All rights reserved. Requests for permission should be mailed to: [email protected]
1. 版权归作者洪永淼教授所有；
2. 不得移除作者署名，否则将视为侵权；
3. 对于不遵守此声明或者其他违法使用本文内容者，作者依法保留追究权等。
4. 发现课件错误请联系作者 [email protected]
 CONTENTS

6.1 Motivation
6.2 Framework and Assumptions
6.3 Long-Run Variance-Covariance Matrix
Estimation
6.4 Consistency of the OLS Estimator
5.5 Asymptotic Normality of the OLS Estimator
6.6 Hypothesis Testing
6.7 Cochrane-Ornut Procedure

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Motivation

In many economic applications， 𝜀𝑡 is serially correlated， 𝑋𝑡 𝜀𝑡 is

generally no longer a MDS.

Example 6.1 [Testing a Zero Population Mean]

Suppose the daily stock return 𝑌𝑡 is an stationary ergodic process

with 𝐸 𝑌𝑡 = 𝜇
• the null hypothesis： 𝐇0 : 𝜇 = 0
• the alternative hypothesis： 𝐇𝐴 : 𝜇 ≠ 0
A test for ℍ0 can be based on the sample mean：
𝑌᪄𝑛 = 𝑛−1 ∑𝑛𝑡=1 𝑌𝑡
By a suitable CLT (White (1999))：
𝑑
𝑛𝑌᪄𝑛 → 𝑁(0, 𝑉)
September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Motivation

Example 6.1 [Testing a Zero Population Mean] (Cont.)

Where：
𝑉 ≡ avar 𝑛𝑌᪄𝑛
◼ Because
var 𝑛𝑌᪄𝑛 = 𝑛−1 ∑𝑛𝑡=1 var 𝑌𝑡
+2𝑛−1 ∑𝑛𝑡=2 ∑𝑡−1
𝑗=1 cov 𝑌𝑡 , 𝑌𝑡−𝑗
◼ Serial correlation in {𝑌𝑡 } is expected to affect the asymptotic
variance of 𝑛𝑌᪄𝑛 . avar( 𝑛𝑌᪄𝑛 ) is no longer equal to var(𝑌𝑡 )

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Motivation

Example 6.1 [Testing a Zero Population Mean] (Cont.)

𝑝
• Suppose: 𝑉ƶ → 𝑉, by Slutsky’s theorem,
under 𝐇0 and 𝑛 → ∞
𝑛𝑌᪄𝑛 𝑑
𝑇𝑟 = → 𝑁(0,1)
𝑉ƶ

Eugene Slutsky

◆ Question
Why does serial correlation exist?

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Motivation

Example 6.2 [Unbiasedness Hypothesis]

• Consider the following linear regression model:

𝑆𝑡+𝜏 = 𝛼 + 𝛽𝐹𝑡 (𝜏) + 𝜀𝑡+𝜏
where 𝑆𝑡+𝜏 is the spot foreign exchange rate at time 𝑡 + 𝜏, 𝐹𝑡 (𝜏)
is the forward exchange rate (with maturity τ > 0) at time 𝑡, and
the disturbance 𝜀𝑡+𝜏 is not observable

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Motivation

Example 6.2 [Unbiasedness Hypothesis] (Cont.)

• Unbiasedness Hypothesis:
𝐸 𝑆𝑡+𝜏 ∣ 𝐼𝑡 = 𝐹𝑡 𝜏
where 𝐼𝑡 is the information set available at time t. This implies
𝐇0 : 𝛼 = 0, 𝛽 = 1
𝐸 𝜀𝑡+𝜏 ∣ 𝐼𝑡 = 0, 𝑡 = 1, 2, …
• However, with 𝜏 > 1, 𝐸 𝜀𝑡+𝜏 ∣ 𝐼𝑡 ≠ 0 a.s. for 1 ≤ 𝑗 ≤ 𝜏 − 1
• There exists serial correlation in 𝜀𝑡 up to 𝜏 − 1 lags under 𝐇0 .
This will affect the asymptotic variance of the OLS estimator 𝑛𝛽ƶ

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Motivation

Example 6.3 [Long Horizon Return Predictability]

• Consider a predictive regression:

𝑌𝑡+ℎ,ℎ = 𝛽0 + 𝛽1 𝑟𝑡 + 𝛽2 𝑑𝑡 − 𝑝𝑡 + 𝜀𝑡+ℎ,ℎ
where 𝑌𝑡+ℎ,ℎ is the cumulative return over the holding period
from time 𝑡 to time 𝑡 + ℎ, namely:
𝑌𝑡+ℎ,ℎ = ∑ℎ𝑗=1 𝑅𝑡+𝑗
where 𝑅𝑡+𝑗 is an asset return in 𝑡 + 𝑗, 𝑟𝑡 is the short term interest
rate in time 𝑡, 𝑑𝑡 − 𝑝𝑡 = ln(𝐷𝑡 /𝑃𝑡 )
• When h > 1, cov 𝜀𝑡+ℎ , 𝜀𝑡+h−j ≠ 0 for 1 ≤ 𝑗 ≤ ℎ − 1

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Motivation

Example 6.4 [Relationship Between GDP and Money Supply]

• Consider a linear macroeconomic regression model:

𝑌𝑡 = 𝛼 + 𝛽𝑀𝑡 + 𝜀𝑡
𝑌𝑡 : GDP growth rate in time 𝑡
𝑀𝑡 : money supply growth rate in time 𝑡
𝜀𝑡 : an unobservable disturbance such that 𝐸 𝜀𝑡 ∣ 𝑀𝑡 = 0, but
there may exist persistent serial correlation of unknown form
in {𝜀𝑡 }

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Motivation

◆ Question
What happens to the OLS estimator 𝛽መ if the disturbance {𝜀𝑡 }
displays conditional heteroskedasticity (i.e., 𝐸 𝜀𝑡2 ∣ 𝑋𝑡 ≠ 𝜎 2 )
and/or autocorrelation (i.e., cov 𝜀𝑡 , 𝜀𝑡−1 ≠ 0 at least for some j >
0)?

• In particular, we would like to address the following important

issues:
✓ Is the OLS estimator 𝛽መ consistent for 𝛽𝑜 ?
✓ Is 𝛽መ asymptotically most efficient?
✓ Is 𝛽መ after properly scaled, asymptotically normal?
✓ Are the t-test and F-test statistics applicable for large sample
inference?
September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
 CONTENTS

6.1 Motivation
6.2 Framework and Assumptions
6.3 Long-Run Variance-Covariance Matrix
Estimation
6.4 Consistency of the OLS Estimator
6.5 Asymptotic Normality of the OLS Estimator
6.6 Hypothesis Testing
6.7 Cochrane-Ornut Procedure

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Framework and Assumptions

Assumption 6.1.
[Ergodic Stationarity]: The observable stochastic process
{𝑌𝑡 , 𝑋𝑡′ }𝑛𝑡=1 is ergodic stationary, where 𝑌𝑡 is a random variable and 𝑋𝑡
is a 𝐾 × 1 random vector.

Assumption 6.2.
[Linearity]:
𝑌𝑡 = 𝑋𝑡′ 𝛽𝑜 + 𝜀𝑡 ,
where 𝛽𝑜 is a 𝐾 × 1 unknown parameter vector, and 𝜀𝑡 is the
unobservable disturbance.

Assumption 6.3.
[Correct Model Specification]: 𝐸(𝜀𝑡 |𝑋𝑡 ) = 0 with 𝐸(𝜀𝑡2 ) =
𝜎 2 < ∞.

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Framework and Assumptions

Assumption 6.4. [Nonsingularity]: The 𝐾 × 𝐾 matrix

𝑄 = 𝐸(𝑋𝑡 𝑋𝑡′ )
is symmetric, finite and nonsingular.

Assumption 6.5. [Long-Run Variance-Covariance Matrix]

a) For 𝑗 = 0, ±1, … , define the 𝐾 × 𝐾 autocovariance function of
𝑋𝑡 𝜀𝑡
Γ(𝑗) = cov 𝑋𝑡 𝜀𝑡 , 𝑋𝑡−𝑗 𝜀𝑡−𝑗
′
= 𝐸 𝑋𝑡 𝜀𝑡 𝜀𝑡−𝑗 𝑋𝑡−𝑗

Then ∑∞𝑗=−∞ ∥ Γ(𝑗) ∥< ∞ , ∥ 𝐴 ∥= ∑ 𝐾 𝐾

𝑖=1 𝑗=1 𝐴(𝑖,𝑗)
∑ for any
𝐾 × 𝐾 matrix, and the long-run variance-covariance matrix
𝑉 = ∑∞
𝑗=−∞ Γ(𝑗)

is positive definite
September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Framework and Assumptions

Assumption 6.5. (Cont.)

b) The conditional expectation
𝑞.𝑚.
𝐸 𝑋𝑡 𝜀𝑡 ∣ 𝑋𝑡−𝑗 𝜀𝑡−𝑗 , 𝑋𝑡−𝑗−1 𝜀𝑡−𝑗−1 , … → 0 as 𝑗 → ∞.
1/2
c) ∑∞
𝑗=0 𝐸 𝑟𝑗′ 𝑟𝑗 <∞

𝑟𝑗 = 𝐸 𝑋𝑡 𝜀𝑡 ∣ 𝑋𝑡−𝑗 𝜀𝑡−𝑗 , 𝑋𝑡−𝑗−1 𝜀𝑡−𝑗−1 , …

−𝐸 𝑋𝑡 𝜀𝑡 ∣ 𝑋𝑡−𝑗−1 𝜀𝑡−𝑗−1 , 𝑋𝑡−𝑗−2 𝜀𝑡−𝑗−2 , …
.

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
 CONTENTS

6.1 Motivation
6.2 Framework and Assumptions
6.3 Long-Run Variance-Covariance Matrix
Estimation
6.4 Consistency of the OLS Estimator
6.5Asymptotic Normality of the OLS Estimator
6.6 Hypothesis Testing
6.7 Cochrane-Ornut Procedure

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Long-Run Variance-Covariance Matrix Estimation

◆ Question
Why are we interested in the long-run variance-covariance matrix
𝑉 = ∑∞ 𝑗=−∞ Γ(𝑗)
መ we have
• For the OLS estimator 𝛽,
𝑛 𝛽ƶ − 𝛽𝑜 = 𝑄ƶ −1 𝑛−1/2 ∑𝑛𝑡=1 𝑋𝑡 𝜀𝑡 .
• Suppose a suitable CLT holds for {𝑋𝑡 𝜀𝑡 } in the present context.
That is, suppose
𝑛 𝑑
−1/2
𝑛 ∑𝑡=1 𝑋𝑡 𝜀𝑡 → 𝑁(0, 𝑉)
• As n→∞, where V is an asymptotic variance-covariance matrix,
namely
𝑉 ≡ avar 𝑛−1/2 ∑𝑛𝑡=1 𝑋𝑡 𝜀𝑡 = lim var 𝑛−1/2 ∑𝑛𝑡=1 𝑋𝑡 𝜀𝑡
𝑛→∞

𝑑
• By Slutsky’s theorem, we have 𝑛 𝛽ƶ − 𝛽𝑜 → 𝑁 0, 𝑄 −1 𝑉𝑄−1
September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Long-Run Variance-Covariance Matrix Estimation

◼ Proof
• Put 𝑔𝑡 = 𝑋𝑡 𝜀𝑡 , note that 𝐸(𝑔𝑡 ) = 0 given E(𝜀𝑡 |𝑋𝑡 ) = 0 ,
for some lag order j > 0, Γ 𝑗 = cov(𝑔𝑡 , 𝑔𝑡−𝑗 ) ≠ 0
 −1/2 n   −1/2 n 
var  n  X t  t  = var  n  gt 
 t =1   t =1 
 
 n
 n

= E  n −1/2  gt  n −1/2  g s  
 t =1  s =1 
 

 E ( g g  )
n n
−1
=n t s
t =1 s =1

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Long-Run Variance-Covariance Matrix Estimation

◼ Proof (Cont.)
t −1 n −1

 E ( g g  ) + n  E ( g g  ) + n   E ( g g  )
n n n
−1 −1 −1
=n t t t s t s
t =1 t = 2 s =1 t =1 s =t +1
n −1 −1 n+ j
= n −1  E ( gt gt ) +  n −1  E ( g t g t− j ) + n −1  E ( g t g t− j )
n n

t =1 j =1 t = j +1

j =− ( n −1) t =1
n −1
= 
j =− ( n −1)
(1− | j | / n)( j )


→  ( j ), as n → 
j =−
by dominated convergence. Therefore, we have 𝑉 = ∑∞
𝑗=−∞ Γ(𝑗)

✓ when {𝑔𝑡 } is an MDS, we have 𝑉 = Γ(0)

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Long-Run Variance-Covariance Matrix Estimation

Definition 6.1.
[Spectral Density Matrix]: Suppose {𝑔𝑡 = 𝑋𝑡 𝜀𝑡 } is a K×1 weakly
stationary process with 𝐸 𝑔𝑡 = 0 and autocovariance function
′
Γ 𝑗 ≡ cov(𝑔𝑡 , 𝑔𝑡−𝑗 ) = 𝐸(𝑔𝑡 , 𝑔𝑡−𝑗 ), which is a K × K matrix.
Suppose
∑∞
𝑗=−∞ ∥ Γ 𝑗 ∥< ∞

The Fourier transform of the autocovariance function Γ 𝑗 exists

and is given by
1
𝐻(𝜔) = ∑∞ Γ(𝑗)exp(−𝐢𝑗𝜔), 𝜔 ∈ [−𝜋, 𝜋]
2𝜋 𝑗=−∞

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Long-Run Variance-Covariance Matrix Estimation

• The inverse Fourier transform of the spectral density matrix is

𝜋
Γ(𝑗) = ∫−𝜋 𝐻(𝜔)𝑒 𝐢𝑗𝜔 𝑑𝜔
• when 𝜔 = 0, we have the long-run variance-covariance matrix
𝑉 = 2𝜋𝐻(0) = ∑∞
𝑗=−∞ Γ(𝑗)

◆ Question
How to estimate the long-run variance-covariance matrix 𝑉?

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Long-Run Variance-Covariance Matrix Estimation

• We first consider a naive estimator

𝑉ƶ = ∑𝑛−1 ƶ
𝑗=−(𝑛−1) Γ 𝑗

where the sample autocovariance function

 −1 n
n  X t et X t− j et − j , j = 0,1, , n − 1,
 t = j +1
ˆ ( j ) =  n
n −1

 t =1− j
X t et X t− j et − j , j = −1, −2, , −(n − 1)

However,𝑉ƶ𝑉ƶ isisnot
➢ However, notconsistent
consistentfor
for𝑉𝑉
• There are too many estimated terms in the summation over lag
orders in V. 𝐵𝑖𝑎𝑠[∑𝑛−1 ƶ
𝑗=−(𝑛−1) 𝛤 𝑗 ] → 0 as 𝑛 → ∞ , but
𝑉𝑎𝑟[∑𝑛−1 ƶ
𝑗=−(𝑛−1) 𝛤 𝑗 ] will never go to zero.
September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Long-Run Variance-Covariance Matrix Estimation

𝑝
• So, we consider the following truncated sum 𝑉ƶ = ∑𝑗=−𝑝 Γƶ 𝑗
where p is a positive integer. If p is fixed, we expect
𝑝
𝑝
𝑉ƶ → ∑𝑗=−𝑝 Γ(𝑗) ≠ 2𝜋𝐻(0) = 𝑉
we should let 𝑝 = 𝑝𝑛 → ∞ as 𝑛 → ∞
• However, we cannot let p grow as fast as the sample size n. To
𝑝
ƶ
ensure 𝑉 → 𝑉, using a truncated variance estimator
𝑝𝑛
𝑉ƶ = ∑𝑗=−𝑝𝑛
ƶ
Γ(𝑗)
ƶ ∝ 𝑝𝑛+1 /𝑛 → 0,
where 𝑝𝑛 → ∞, 𝑝𝑛 /𝑛 → 0, 𝑉𝑎𝑟(𝑉)
ƶ =∑ 𝑛 ƶ 𝑝 ∞ ƶ ƶ
𝐵𝑖𝑎𝑠(𝑉) 𝑗=−𝑝𝑛 Γ(𝑗) − ∑ 𝑗=−∞ Γ(𝑗) = −∑|𝑗| >𝑝𝑛 Γ(𝑗) → 0.

✓ Example: 𝑝𝑛 = 𝑛1/3
September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Long-Run Variance-Covariance Matrix Estimation

➢ However, it may not be PSD for all n

• To ensure that it is always PSD, use a weighted average estimator
ƶ𝑉 = ∑𝑝𝑛 𝑘 𝑗/𝑝𝑛
ƶ
Γ(𝑗)
𝑗=−𝑝𝑛
where the weighting function 𝑘 · is called a kernel function.
✓ Example: Bartlett kernel
𝑘(𝑧) = (1 − |𝑧|)𝟏(|𝑧| ≤ 1)
✓ Newey and West (1987) first used this kernel function to
estimate V in econometrics.

Whitney Newey Kenneth West

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Long-Run Variance-Covariance Matrix Estimation

➢ In fact, we can consider a more general

variance estimator for V :
𝑉ƶ = ∑𝑛−1 ƶ
𝑗=1−𝑛 𝑘 𝑗/𝑝𝑛 Γ(𝑗),
where 𝑘(·) may have unbounded
support. In fact, any kernel 𝑘: ℝ →
−1,1 can be used under certain
conditions
✓ Example: Quadratic-Spectral kernel
that k(·) has unbounded support.
Andrews (1991) used it to estimate V
3 sin(𝜋𝑧)
𝑘(𝑧) = − cos(𝜋𝑧) , −∞ < 𝑧 < ∞
(𝜋𝑧)2 𝜋𝑧 Donald Andrews

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Long-Run Variance-Covariance Matrix Estimation

➢ In practice, it is most important to determine an appropriate

smoothing parameter 𝑝𝑛 .
✓ Andrews (1991) and Newey & West (1994) discuss data-
driven methods to choose 𝑝𝑛 .

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
 CONTENTS

6.1 Motivation
6.2 Framework and Assumptions
6.3 Long-Run Variance-Covariance Matrix
Estimation
6.4 Consistency of the OLS Estimator
6.5 Asymptotic Normality of the OLS Estimator
6.6 Hypothesis Testing
6.7 Cochrane-Ornut Procedure

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Consistency of the OLS Estimator

Theorem 6.1.
Suppose Assumptions 6.1 to 6.5(a) hold. Then
𝑝
𝛽ƶ → 𝛽𝑜 as 𝑛 → ∞
◼ Proof:
𝛽ƶ − 𝛽𝑜 = 𝑄ƶ −1 𝑛−1 ∑𝑛𝑡=1 𝑋𝑡 𝜀𝑡
• By Assumptions 6.1, 6.2 and 6.4 and WLLN for an ergodic
stationary process, we have
𝑝 𝑝
ƶ ƶ
𝑄 → 𝑄 and 𝑄 → 𝑄−1
−1

• Similarly , by Assumptions 6.1 𝑝

to 6.3 and 6.5(a), we have
𝑛−1 ∑𝑛𝑡=1 𝑋𝑡 𝜀𝑡 → 𝐸 𝑋𝑡 𝜀𝑡 = 0
• Using WLLN for an ergodic stationary process, where 𝐸 𝑋𝑡 𝜀𝑡 =
0 given Assumption 6.2 𝐸 𝜀𝑡 |𝑋𝑡 = 0 and the law of iterated
expectations.
September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
 CONTENTS

6.1 Motivation
6.2 Framework and Assumptions
6.3 Long-Run Variance-Covariance Matrix
Estimation
6.4 Consistency of the OLS Estimator
6.5 Asymptotic Normality of the OLS Estimator
6.6 Hypothesis Testing
6.7 Cochrane-Ornut Procedure

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Asymptotic Normality of the OLS Estimator

Lemma 6.1.
[CLT for a Zero Mean Ergodic Stationary Process (White 1984,
Theorem 5.15)]: Suppose {𝑍𝑡 } is an ergodic stationary process with:
(1) 𝐸 𝑍𝑡 = 0;
(2) 𝑉 = ∑∞
𝑗=−∞ Γ(𝑗) is finite and nonsingular,
′
where Γ 𝑗 = 𝐸 𝑍𝑡 𝑍𝑡−𝑗 ;
𝑞.𝑚.
(3) 𝐸 𝑍𝑡 ∣ 𝑍𝑡−𝑗 , 𝑍𝑡−𝑗−1 , … → 0, as 𝑗 → ∞;
1/2
(4) ∑∞
𝑗=0 𝐸 𝑟𝑗′ 𝑟𝑗 < ∞. Then as n→∞,
𝑑
𝑛1/2 𝑍᪄𝑛 = 𝑛−1/2 ∑𝑛𝑡=1 𝑍𝑡 → 𝑁(0, 𝑉).

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Asymptotic Normality of the OLS Estimator

Theorem 6.2.
[Asymptotic Normality]: Suppose Assumptions 6.1 to 6.5 hold.

Then as 𝑛 → ∞,
𝑑
𝑛 𝛽ƶ − 𝛽𝑜 → 𝑁 0, 𝑄−1 𝑉𝑄−1 ,

where 𝑉 = ∑∞
𝑗=−∞ Γ(𝑗)

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Asymptotic Normality of the OLS Estimator

◼ Proof:
• Because
𝑛 𝛽ƶ − 𝛽𝑜 = 𝑄ƶ −1 𝑛−1/2 ∑𝑛𝑡=1 𝑋𝑡 𝜀𝑡 .
• By Assumptions 6.1 to 6.3 and 6.5 and CLT for an ergodic
stationary process, we have
𝑛 𝑑
𝑛 −1/2 ∑𝑡=1 𝑋𝑡 𝜀𝑡 → 𝑁(0, 𝑉),
𝑝 𝑝
• 𝑄ƶ → 𝑄 and ƶ
𝑄 → 𝑄−1 by Assumption 6.4
−1 and WLLN for an
ergodic stationary process. We then have by Slutsky’s theorem
𝑑
𝑛 𝛽ƶ − 𝛽𝑜 → 𝑁 0, 𝑄−1 𝑉𝑄 −1

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
 CONTENTS

6.1 Motivation
6.2 Framework and Assumptions
6.3 Long-Run Variance-Covariance Matrix
Estimation
6.4 Consistency of the OLS Estimator
6.5 Asymptotic Normality of the OLS Estimator
6.6 Hypothesis Testing
6.7 Cochrane-Ornut Procedure

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Hypothesis Testing

➢ Consider testing the null hypothesis

𝐇0 : 𝑅𝛽𝑜 = 𝑟
where R is a nonstochastic 𝐽 × 𝐾 matrix with full rank, r is a 𝐽 ×
1 nonstochastic vector, and 𝐽 ≤ 𝐾
◼ Corollary 6.1. Suppose Assumptions 6.1 to 6.5 hold. Then under 𝐇0 ,
as 𝑛 → ∞,
𝑑
𝑛(𝑅𝛽ƶ − 𝑟) → 𝑁 0, 𝑅𝑄−1 𝑉𝑄 −1 𝑅′ .
𝑝
◼ Assumption 6.6. 𝑉ƶ → 𝑉
In some special scenarios, we may have Γ(𝑗) = 𝟎 for all 𝑗 > 𝑝𝑜 ,
where 𝑝𝑜 is a fixed lag order. In these cases, we can use the
ƶ 𝑝𝑜 ƶ
following variance estimator 𝑉 = ∑ Γ(𝑗) 𝑗=−𝑝𝑜

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Hypothesis Testing

Theorem 6.3.
[Robust t-Test and Wald Test]: Under Assumptions 6.1 to 6.6 and
𝐇0 , we have as 𝑛 → ∞,
✓ When 𝐽 = 1, we define a robust t-test statistic

𝑛 𝑅𝛽ƶ − 𝑟 𝑑
𝑇𝑟 = → 𝑁(0,1)
𝑅𝑄ƶ −1 𝑉ƶ 𝑄ƶ −1 𝑅′
✓ when 𝐽 ≥ 1, we define a robust Wald test statistic
ƶ −1 ′ −1
′ ′ −1 ƶ
𝑊𝑟 = 𝑛(𝑅𝛽 − 𝑟) 𝑅 𝑿 𝑿/𝑛 𝑉 𝑿 𝑿/𝑛 𝑅 ′
(𝑅𝛽ƶ − 𝑟)
𝑑
→ 𝜒𝐽2

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
 CONTENTS

6.1 Motivation
6.2 Framework and Assumptions
6.3 Long-Run Variance-Covariance Matrix
Estimation
6.4 Consistency of the OLS Estimator
6.5 Asymptotic Normality of the OLS Estimator
6.6 Hypothesis Testing
6.7 Cochrane-Ornut Procedure

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Cochrane-Ornut Procedure

• Long-run variance estimators are necessary for statistical inference

of the OLS estimation in a linear regression model when there
exists serial correlation of unknown form.
• If serial correlation in the regression disturbance has of a known
form up to some unknown parameters, then simpler statistical
inference procedures are possible.
✓ Example: Ornut-Cochrane procedure.

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Cochrane-Ornut Procedure

• Consider a linear regression model with serially correlated

disturbances:
𝑌𝑡 = 𝑋𝑡′ 𝛽𝑜 + 𝜀𝑡
where E(𝜀𝑡 |𝑋𝑡 ) = 0 but {𝜀𝑡 } follows an AR(p) process
𝑝
𝜀𝑡 = ∑𝑗=1 𝛼𝑗 𝜀𝑡−𝑗 + 𝑣𝑡 , 𝑣𝑡 ∼ IID 0, 𝜎 2
• The OLS estimator 𝛽መ is consistent for 𝛽𝑜 given E(𝑋𝑡 𝜀𝑡 ) = 0, but its
asymptotic variance depends on serial correlation in {𝜀𝑡 }. We can
consider the following transformed linear regression model
′
𝑝 𝑝 𝑝
𝑌𝑡 − ∑𝑗=1 𝛼𝑗 𝑌𝑡−𝑗 = 𝑋𝑡 − ∑𝑗=1 𝛼𝑗 𝑋𝑡−𝑗 𝛽𝑜 + 𝜀𝑡 − ∑𝑗=1 𝛼𝑗 𝜀𝑡−𝑗
′
𝑝
= 𝑋𝑡 − ∑𝑗=1 𝛼𝑗 𝑋𝑡−𝑗 𝛽𝑜 + 𝑣𝑡 .

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Cochrane-Ornut Procedure

• We can write it as follows:

𝑌𝑡∗ = 𝑋𝑡∗′ 𝛽𝑜 + 𝑣𝑡
where
𝑝
𝑌𝑡∗ = 𝑌𝑡 − ∑𝑗=1 𝛼𝑗 𝑌𝑡−𝑗
𝑝
𝑋𝑡∗ = 𝑋𝑡 − ∑𝑗=1 𝛼𝑗 𝑋𝑡−𝑗
• The OLS estimator 𝛽෨ of this transformed regression will be
consistent for 𝛽𝑜 and 𝑛 𝛽ƿ − 𝛽𝑜 will be asymptotically normal:
𝑑
𝑛 𝛽ƿ − 𝛽𝑜 → 𝑁 0, 𝜎𝑣2 𝑄𝑋−1∗ 𝑋 ∗
where 𝑄𝑋 ∗𝑋 ∗ = 𝐸 𝑋𝑡∗ 𝑋𝑡∗′ . The OLS estimator 𝛽෨ is a GLS estimator
discussed in Chapter 3. It is BLUE.

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Cochrane-Ornut Procedure

• However, the GLS estimator 𝛽෨ is infeasible. As a solution, one can

use an adaptive feasible two-step GLS procedure:
◼ Step 1: Regress
𝑌𝑡 = 𝑋𝑡′ 𝛽𝑜 + 𝜀𝑡 , 𝑡 = 1, … , 𝑛
namely, regress 𝑌𝑡 on 𝑋𝑡 and obtain the estimated OLS residual
ƶ Then run an AR(p) model
𝑒𝑡 = 𝑌𝑡 − 𝑋𝑡′ 𝛽.
𝑝
𝑒𝑡 = ∑𝑗=1 𝛼𝑗 𝑒𝑡−𝑗 + 𝑣ƿ 𝑡 , 𝑡 = 𝑝 + 1, … , 𝑛
𝑝
and obtain the OLS estimators 𝛼ƶ𝑗 𝑗=1

◼ Step 2: Regress the transformed model

𝑌ƶ𝑡∗ = 𝑋ƶ 𝑡∗′ 𝛽𝑜 + 𝑣𝑡∗ , 𝑡 = 𝑝 + 1, … , 𝑛

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
Cochrane-Ornut Procedure

• The sampling error resulting from the first step estimation has no
impact on the asymptotic properties of the OLS estimator in the
second step.
• The asymptotic variance estimator of 𝛽ƿ𝑎 is given by
𝑠ƶ𝑣2 𝑄ƶ 𝑋−1∗ 𝑋 ∗
where
1 ∗2 ƶ ∗ ∗
1 𝑛
𝑠ƶ𝑣2 = ∑𝑡=1 𝑣ƶ 𝑡 , 𝑄𝑋 𝑋 = ∑𝑡=1 𝑋ƶ 𝑡∗ 𝑋ƶ 𝑡∗′
𝑛
𝑛−𝐾 𝑛
with
𝑣ƶ 𝑡 = 𝑌ƶ𝑡∗ − 𝑋ƶ 𝑡∗′ 𝛽ƿ𝑎

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation
 Thank You !

September 22,
ADVANCED ECONOMETRICS Linear Regression Models Under Conditional Heteroskedasticity and Autocorrelation