Day 4A

Asymptotic Theory

c A. Colin Cameron
Univ. of Calif.- Davis

Advanced Econometrics
Bavarian Graduate Program in Economics
.
Based on A. Colin Cameron and Pravin K. Trivedi (2009, 2010),
Microeconometrics using Stata (MUS), Stata Press.
and A. Colin Cameron and Pravin K. Trivedi (2005),
Microeconometrics: Methods and Applications (MMA), C.U.P.

July 22-26, 2013

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013on A. 1Colin
/ 31Ca
 1. Introduction

1. Introduction

Consider sample mean ȳ = 1

N ∑N
i =1 yi as an estimator of the
population mean µ
I asymptotic theory gives properties of ȳ as N ! ∞.
Make assumptions about the data generating process (dgp)
I yi (µ, σ2 ) i.i.d. (independent and identically distributed)
I =) ȳ (µ, σ2 /N ).

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013on A. 2Colin
/ 31Ca
 1. Introduction

The distribution of ȳ has all mass at the mean of µ as N ! ∞

I Intuition: V[ȳ ] = E[(ȳ µ)2 ] = σ2 /N ! 0.
I Formally ȳ converges in probability (de ned below) to µ
p
I This is denoted ȳ ! µ or plim ȳ = µ.
I Then ȳ is consistent for µ.
I The proof uses a law of large numbers (de ned below) for an average.

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013on A. 3Colin
/ 31Ca
 1. Introduction

Statistical inference based on ȳ with N ! ∞ requires scaling ȳ up

ȳ E[ȳ ] ȳ pµ
I Use standardized statistic z = p = .
V[ȳ ] σ/ N
I z (0, 1), so z may have a nondegenerate distribution.
I A central limit theorem (de ned below) proves z N (0, 1) as
N ! ∞.
d
I Formally z converges in distribution (!, de ned below) to N [0, 1].
p d
I Equivalently N (ȳ µ) ! N [0, σ2 ]. p
I Note: ȳ has been scaled up by the multiple N.

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013on A. 4Colin
/ 31Ca
 1. Introduction

p d
For simplicity, the formal result N (ȳ µ) ! N [0, σ2 ] is often
re-expressed in terms of ȳ
a
I ȳ N [0, σ2 /N ]
a
I means \is asymptotically distributed as"
I this means N is large enough that the normal is a good approximation
I but N is not so large that σ2 /N = 0.

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013on A. 5Colin
/ 31Ca
 1. Introduction

Outline

1 Introduction
2 Sequences of random variables
3 Convergence in probability
4 Laws of large numbers (for averages)
5 Convergence in distribution
6 Central limit theorems (for averages)
7 Some Key Results
8 Simulations for LLN and CLT
9 Appendix: Some Further Asymptotic Results
Appendix: Sampling Schemes
Appendix: OLS under Simple Random Sampling

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013on A. 6Colin
/ 31Ca
 2. Sequences of random variables

2. Sequences of random variables

Recall a sequence of real numbers

I e.g. aN = 2 + 3
N
What happens as N ! ∞?
I mathematical convergence (or divergence)
A sequence of nonstochastic real numbers faN g converges to a if for
any ε > 0, there exists N = N (ε) such that for all N > N ,

jaN aj < ε.

I e.g. aN = 2 + 3/N ! 2 since

jaN aj = j2 + 3/N 2j = j3/N j < ε for all N > N = 3/ε.

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013on A. 7Colin
/ 31Ca
 2. Sequences of random variables

We instead consider a sequence of random variables bN .

I e.g. bN = 1
N ∑N 2
i = 1 xi
I e.g. bN = 1
N ∑N
i =1 xi ui
1
I e.g. bN = b
β= 1
N ∑N 2
i = 1 xi
1
N ∑N
i =1 xi ui

What happens as N ! ∞?
I jbN b j may exceed ε due to randomness, so bN 9 b exactly
I instead use convergence in probability.

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013on A. 8Colin
/ 31Ca
 3. Convergence in probability

3. Convergence in probability and consistency

Informal de nition: The sequence fbN g converges in probability to b

if for any ε > 0
lim Pr[jbN b j < ε] = 1.
N !∞

Formal de nition: A sequence of random variables fbN g converges in

probability to b if for any ε > 0 and δ > 0, there exists N = N (ε, δ)
such that for all N > N ,

Pr[jbN b j < ε] > 1 δ.

p
We write plim bN = b or bN ! b
I the limit b may be a constant or a random variable.

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013on A. 9Colin
/ 31Ca
 3. Convergence in probability Consistency

Consistency
b
Suppose the sequence bN is an estimator, say bN = β.
b!p b is consistent for β.
I If β β, a constant, then we say β
A simple consistency proof uses convergence in mean square
I that is lim E[(bN b )2 ] = 0
N !∞
ms
I ! implies convergence in probability.
b is used to estimate β
Suppose β
I b β )2 ] = V[ β
E[( β b ] + (bias[ β
b ])2 as MSE = variance + bias2
ms
b ! β if V[ β
b ] ! 0 and bias[ β b ] ! 0 as N ! ∞
I so β
b p
I it follows that β ! β if the variance and bias go to zero as N ! ∞.
p
We use the weaker convergence in probability as βb! β is possible
b
even if the mean and variance of β do not exist.

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013 on A.10Colin
/ 31Ca
 4. Law of large numbers

4. Law of large numbers

Easy way to get probability limit when bN is an average

bN = X̄N = 1
N ∑N
i =1 Xi .

I Xi here is general notation for a random variable. e.g. Xi = xi ui .

Weak Law of Large Numbers (WLLN):
Speci es conditions on the individual terms Xi in X̄N under which
p
(X̄N E[X̄N ]) ! 0.

Khinchine's Theorem (WLLN):

Let fXi g be i.i.d. (independent and identically distributed).
p
If and only if E[Xi ] = µ exists, then (X̄N µ) ! 0.

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013 on A.11Colin
/ 31Ca
 4. Law of large numbers

Other LLN's are Kolmogorov and, for i.n.i.d. data, Markov

I these are given later.
If a LLN can be applied then

plim X̄N = lim E[X̄N ] in general

1 N
= lim N ∑i =1 E[Xi ] equivalently
=µ if Xi i.i.d.

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013 on A.12Colin
/ 31Ca
 5. Convergence in distribution

5. Convergence in distribution

bN has (unknown) cumulative distribution function (cdf) FN .

Like any other function, FN may have a limit function.
Convergence in Distribution:
A sequence of random variables fbN g converges in distribution to a
random variable b if

lim FN = F , where F is the c.d.f. of b

N !∞

at every continuity point of F , where convergence is in the usual

mathematical sense.
d
We write bN ! b, and call F the limit distribution of fbN g.
Basically FN is very complicated and F is simple like N [0, 1].

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013 on A.13Colin
/ 31Ca
 6. Central limit theorems

6. Central limit theorems

Easy way to get limit distribution when bN is an average X̄N .
X̄N has a degenerate limit distribution with all mass at one point
p
since X̄N ! limE[X̄N ] by a LLN.
So rescale X̄N to standardized variate

X̄N E[X̄N ]
bN = ZN = p [0, 1].
V[X̄N ]

Central Limit Theorem (CLT):

A CLT speci es the conditions on the individual terms Xi in X̄N under
which
d
ZN ! N [0, 1].
Lindeberg-Levy CLT:
Let fXi g be i.i.d. with E[Xi ] = µ and V[Xi ] = σ2 .
p d
Then ZN = N (X̄N µ)/σ ! N [0, 1].
c A. Colin Cameron Univ. of Calif.- Davis (Advanced
BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013 on A.14Colin
/ 31Ca
 6. Central limit theorems

Note that
p
ZN = (X̄N E[X̄N ])/ [X̄N ]
Vq in general
= ∑N (Xi N
E[Xi ])/ ∑i =1 V[Xi ] if Xi independent over i
pi =1
= N (X̄N µ)/σ if Xi i.i.d.

The last expression can be rewritten as

X̄N µ d
p ! N [0, 1].
σ/ N
p d
It follows that µ) ! N [0, σ2 ].
N (X̄N
p d
More generally we often nd N (b
β β) ! N [0, V ].
p
I Scale consistent b
β up by N to get a limit distribution.

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013 on A.15Colin
/ 31Ca
 6. Central limit theorems Multivariate central limit theorem

Multivariate central limit theorem

Consider vector XN with mean µN and variance VN

XN [ µ N , VN ] .

Rescale XN to standardized variate

ZN = VN 1/2 (XN µN ) [0, I].

Central Limit Theorem (CLT):

A CLT speci es the conditions on the individual terms Xi in XN
under which
d
ZN ! N [0, I].

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013 on A.16Colin
/ 31Ca
 6. Central limit theorems Multivariate central limit theorem

Often lim NVN is nite nonzero

I for example if Xi (µ, Σ) then VN = V[XN ] = N 1 Σ, so NVN = Σ.
d
Then VN 1/2 (XN µN ) ! N [0, I] implies
p d 1
N ( XN µN ) ! N [0, lim N VN ]
p
I scaling the average XN by a multiple N gives a limit distribution
with a nite nonzero variance.

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013 on A.17Colin
/ 31Ca
 7. Some Key Results

7. Some Key Results

Probability Continuity and Continuous Mapping Theorems
Let bN be a vector of random variables, and g ( ) be a continuous
real-valued function. Then
p p
bN ! b, a constant ) g (bN ) ! g (b) Probability Continuity
d d
bN ! b ) g ( bN ) ! g ( b ) Continuous Mapping

Transformation Theorem:
d p
If aN ! a (a random variable) and bN ! b (a constant), then
d
(i ) aN + bN ! a + b
d
(ii ) aN bN ! ab
d
(iii ) aN /bN ! a/b, provided Pr[b = 0] = 0.

I We use especially a matrix version of (ii).

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013 on A.18Colin
/ 31Ca
 7. Some Key Results

Product Limit Normal Rule:

For vector aN and matrix HN , if
d
aN ! N [µ, A]
p
HN ! H, where H is positive de nite

then
d
HN aN ! N [Hµ, HAH0 ].
Leading example is OLS:
p 1 0 1
b
N (β β0 ) = ( X X) 1
p (X0 u )
N N
d 1 1 10
! N [A 0, A BA ].

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013 on A.19Colin
/ 31Ca
 8. Simulations For LLN and CLT

8. Simulations for LLN and CLT

Uniform on (0, 1) has mean 0 and variance 1/12.
Sample average of N uniforms has mean 0 and variance (1/12)/N.

. * Draw from uniform with population mean 0.5

. * Demonstrate LLN by finding average for a very large sample
. * Demonstrate CLT by simulating to obtain many averages
.
. * Small sample: sample mean differs from population mean
. set obs 30
obs was 0, now 30

. set seed 10101

. quietly generate x = runiform()

. mean x

Mean estimation Number of obs = 30

Mean Std. Err. [95% Conf. Interval]

x .5459987 .0511899 .4413036 .6506939

For N = 30: x̄ = 0.546 di ers appreciably from µ = 0.500.

c A. Colin Cameron Univ. of Calif.- Davis (Advanced
BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013 on A.20Colin
/ 31Ca
 8. Simulations For LLN and CLT

. * Consistency: Large sample: sample mean is very close to population mean

. clear all

. set obs 100000

obs was 0, now 100000

. set seed 10101

. quietly generate x = runiform()

. mean x

Mean estimation Number of obs = 100000

Mean Std. Err. [95% Conf. Interval]

x .4988239 .0009133 .4970339 .5006138

For N = 100, 000: x̄ = 0.499 is very close to µ = 0.500.

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013 on A.21Colin
/ 31Ca
 8. Simulations For LLN and CLT

. * Central limit theorem

. * Write program to obtain sample mean for one sample of size numobs (= 30)
. program onesample, rclass
1. args numobs
2. drop _all
3. quietly set obs `numobs'
4. generate x = runiform()
5. summarize x
6. return scalar meanforonesample = r(mean)
7. end

. * Run this program 10,000 times to get 10,000 sample means

. quietly simulate xbar = r(meanforonesample), seed(10101) reps(10000) nodots: ///
> onesample 30

. * Summarize the 10,000 sample means and draw histogram

. summarize xbar

Variable Obs Mean Std. Dev. Min Max

xbar 10000 .4995835 .0533809 .3008736 .6990562

. histogram xbar, normal xtitle("xbar from many samples")

(bin=40, start=.30087364, width=.00995456)

For S = 10, 000 simulations each with sample size N = 30

x̄1 , x̄2 , .... , x̄10000 has distribution with mean
p 0.4996 to µ = 0.500
p close p
and standard deviation 0.0534, close to σ/ N = 1/12/ 30 = 0.0527.
c A. Colin Cameron Univ. of Calif.- Davis (Advanced
BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013 on A.22Colin
/ 31Ca
 8. Simulations For LLN and CLT
p p p
z = (x̄ µ)/(σ/ N ) = (x̄ 0.5)/( 1/12/ 30) = (x̄ 0.5)/0.0527.
Histogram and kernel density estimate for z1 , z2 , .... , z10000 .

.4
.3
Density
.2 .1
0

-4 -2 0 2 4
z from many samples

This is standard normal, as predicted by the CLT.

c A. Colin Cameron Univ. of Calif.- Davis (Advanced
BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013 on A.23Colin
/ 31Ca
 9. Appendix Some Further Asymptotic Results

9. Appendix: Some Further Asymptotic Results

Alternative modes of convergence of bN to b

I Mean square: lim E[(bN b )2 ] = 0.
N !∞
I Chebychev's inequality: Pr[(Z µ)2 > k ] σ2 /k, for any k > 0.
I Almost sure: Prfω j lim bN (ω ) = b (ω )g = 1.
I These imply convergence in probability.
Consequences:
p d
I bN ! b implies bN ! b.
I The reverse is generally not true, unless b is a constant.
I For vector r.v.'s de ne FN and F to be cdf's of vectors bN and b.

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013 on A.24Colin
/ 31Ca
 9. Appendix Some Further Asymptotic Results

Strong Law of Large Numbers (LLN):

as
I The convergence is instead almost surely (!).
Kolmogorov SLLN:
Let fXi g be i.i.d. If and only if E[Xi ] = µ exists and E[jXi j] < ∞,
as
then (X̄N E[X̄N ]) ! 0.
as
I Compared to Khinchine's Theorem ! requires E[jXi j] < ∞.
Markov SLLN:
Let fXi g be i.n.i.d. with E[Xi ] = µi .
If ∑i∞=1 (E[jXi µi j1+δ ]/i 1+δ ) < ∞, for some δ > 0, then
as
(X̄N E[X̄N ]) ! 0.
I Relaxes i.i.d. assumption at expense of requiring existence of (1 + δ)th
absolute moment.
I Easiest to set δ = 1, so need variance.

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013 on A.25Colin
/ 31Ca
 9. Appendix Some Further Asymptotic Results

Liapounov CLT:
Let fXi g be independent with E[Xi ] = µi and V[Xi ] = σ2i .
(2+δ)/2
If lim ∑N
i =1 E[jXi µ i j2+ δ ] / ∑ N 2
i =1 σ i = 0, for some
d
choice of δ > 0, then ZN ! N [0, 1].
I The Liapounov CLT relaxes i.i.d. assumption but needs existence of
(2 + δ)th absolute moment.
Cramer-Wold Device:
d d
If λ0 bN ! N [ , ] for all λ 6= 0 then bN ! N [ , ].
I So prove a multivariate CLT by proving a scalar CLT for λ0 bN .

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013 on A.26Colin
/ 31Ca
 9. Appendix Sampling schemes

9. Appendix: Sampling schemes

Simple Random Sampling (SRS)

I Randomly draw (yi , xi ) from the population with equal probabilities.
I Then xi i.i.d. So xi ui i.i.d. (if errors ui are i.i.d.), and xi2 i.i.d.
I Can use Khinchine's LLN and Lindeberg-Levy CLT.
Fixed regressors
I Experiment where xi are xed and we observe the resulting random yi .
I Then xi xed, ui i.i.d. implies xi ui i.n.i.d. and xi2 nonstochastic.
I Need to use Markov LLN and Liapounov CLT.
Exogenous Strati ed Sampling
I Oversample some values of x and undersample others.
I Then xi i.n.i.d., so xi ui i.n.i.d. and xi2 i.n.i.d.
I Need to use Markov LLN and Liapounov CLT.

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013 on A.27Colin
/ 31Ca
 9. Appendix Sampling schemes

These three di erent sampling schemes ultimately lead to similar

asymptotic results.
Microeconometrics often use survey data obtained by strati ed
sampling.
The simplest results assume simple random sampling.
Big problems arise if the strati ed sampling is Instead
Endogenous Strati ed Sampling
I Oversample some values of y and undersample others.
I Estimators can be inconsistent.
I Leading examples are Tobit and selection models.

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013 on A.28Colin
/ 31Ca
 9. Appendix OLS under simple random sampling

9. Appendix: OLS under simple random sampling

Scalar regressor: yi = βxi + ui .
SRS: (xi , yi ) i.i.d. with xi i.i.d. with mean µx & ui i.i.d. with mean 0.
I 1. As xi ui are i.i.d. apply Khinchine's Theorem.
p
Then N 1 ∑i xi ui ! E[xu ] = E[x ] E[u ] = 0.
I 2. As xi2 are i.i.d. apply Khinchine's Theorem.
p
Then N 1 ∑i xi2 ! E[x 2 ] which we assume exists.
I 3. The probability limit is obtained by combining
1 N
N ∑i =1 xi ui
plim b
β = β + plim 1 N 2
N ∑i =1 xi

plim N1 ∑Ni =1 xi ui
= β+
plim N1 ∑N 2
i =1 xi

= β+ 0 = β,
E[x 2 ]

using probability limit continuity (plim[aN /bN ] = a/b if b 6= 0).

c A. Colin Cameron Univ. of Calif.- Davis (Advanced
BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013 on A.29Colin
/ 31Ca
 9. Appendix OLS under simple random sampling

SRS: assume xi i.i.d. with mean µx and second moment E[x 2 ]

and assume ui i.i.d. with mean 0 and variance σ2 .
Then xi ui are i.i.d. with mean E[xu ] = E[x ] E[u ] = 0, and
V[xu ] = E[(xu 0)2 ] = E[x 2 u 2 ] = E[x 2 ]E[u 2 ] = σ2 E[x 2 ].
I 1. Lindeberg-Levy CLT for N 1
∑N
i =1 xi ui yields
p 1 N p1 ∑N
i =1 xi ui
N
N
p∑i =2 1 xi 2ui 0
= p
N d
! N [0, 1].
σ E[x ] σ 2 E[x 2 ]

I 2. Convert to p1
N
∑N
i =1 xi ui

p p1 ∑N
i =1 xi ui
p1 ∑N
i =1 xi ui = σ 2 E[x 2 ] p
N
N σ 2 E[x 2 ]

d p
! σ 2 E[x 2 ] N [0, 1]

d
! N [0, σ2 E[x 2 ]]
using product limit normal rule.
c A. Colin Cameron Univ. of Calif.- Davis (Advanced
BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013 on A.30Colin
/ 31Ca
 9. Appendix OLS under simple random sampling

I 3. The limit distribution is obtained by combining

p p1 ∑N
i =1 xi ui
N (b
β β) = N
1
N ∑N 2
i =1 xi

d N [0,σ2 E[x 2 ]]
! 1 N 2
plim N ∑i =1 xi

d N [0,σ2 E[x 2 ]]
! E[x 2 ]
h i
d 1
! N 0, σ2 E[x 2 ]

using plim N 1 ∑N 2 2
i =1 xi =E[x ] from consistency proof
and the product normal limit rule
d d p
(or aN bN ! a b if aN ! a and bN ! b ).

c A. Colin Cameron Univ. of Calif.- Davis (Advanced

BGPEEconometrics
Course: Asymptotic
BavarianTheory
Graduate Program in Economics
July 22-26,
. Based
2013 on A.31Colin
/ 31Ca