MetricsfinalExamcriptexfmax.

mu
m Likelihood

ML asymptoticproperties Small sample properties

generally focuses on .

are
unknown .
fly 10) =
probability density function for a rand .

vary .

L(Olyl=,flYi → Joint
density of
iid observations It shows the .

probability of observing the whole sample That's why. we try to max

O 's such that the probability of observing the

sample (the likelihood) is
maximized Entire idea UL estimation
.

of .

Solve
for 2 02 ItFind points where tog -
Lik is max
)
.

Define first and second moments as (derivatives

of
LL )

fi= Hi=d4nfffiLo and Si

Hi
:

:
Gradient (score ) vector
Hessian matrix

* Eoff (OOD
-

,
= O : First moment evaluated at true O is zero .

* Var [Si too D= -

Eo [Hilool) : Variance is
negative of expected Hessian .

# I too) Var (9 too ) ) =

(
Eo g Coo ) g Coo ) )
.
'
=
-

ETH ( OOD Information

in
Matrix
Information matrix
Equality
* MLE is consistent ,
asymptotical tornal .
 Estimating Asymptotic Variance
I

t={EoµjoY%§
-

① IM Estimator : -

¥ " , no ,
,
=
{ ICED
I'
observe that we have exp .

② Empirical Hessian

(-23%5692)
"

war µ, no,
IETTeakta
= -
-

,
.

③ Outer Product to PG )
t t

LESis EG Ei]
-

)
-

'

rope (8)
"a =
i
=

.
men
:*
:c
en

÷÷÷÷÷÷÷÷÷÷÷¥¥¥÷÷ :

④ Sandwich Estimator
Ei ]
t
wars (on y to) H to #
-

= H .
.

This estimator is valid , even if the information matrix equality doesn't

hold .

A *It In OLS, we make an

assumption of linear model , but in M LE we don't .

in:÷÷÷::::÷:i:::::::::::÷::::
estimators to satisfy BLUE
properties
Conditional ULE : F- (Ah B) = PLA 1131.743)

en Lkl data ) =
lnflyisxik ) =÷L enflyilxi -1¥
,
sa)
,
enslxik )

- -

Conditional Marginal

f = share of men

::::::::::
LL = noo .

{ en (( t S)
-

.
(tort)
) t n so .

{ en ( t -
f) .
Oz
}
+ not .

{ (f H
en .
-

oil )} thee .

{ en (d. Oil}
(nothing tlnootneh.int
} Nj:Si'Ffs,
= '

Mery Women

,¥nae
1- noo Ln . (t -
Oz ) t neo en Oz t not en (t O , )
. .
- t htt Ln Ot .

-
Employed or
unemployed .

and

Log - likelihood is
decomposed into marginal &
conditional LL
LL

Hypothesis Tests

* Denote constrained estimator

by da and Null hypothesis : Ho .
If Ho is

correct then; { Ho :C (ok 93

 }lN2LK2①
1) L K Test : H E) -
L to a ) should be close te Zero

2) Wald Test : C (8) -

g should be close to Zero -

3) L U Test : E should be close to Be .

*LK LK=-2(Lr-Lu)nX :

XlNat : Since Emre is

normally distributed then CCE ) -

q is also

normally distributed with mean O and variance V. {481-9}

NH E) then ( x An XZ (in general !)
'
xn ,
Cx pl E
-
.
-

lN=[RE-qI[K{AsyHarlEB.K]7txE-e
=
RSS, - K Ssu

While Wald test doesn't explicitly test

for any
alternative
hypothesis it doesn't require strong distributional assumptions
,
.

It requires that 8 is normal distributed

only easy
.
.

*LuiLu=fenff9IItodTµenLf)rx= me

* Pdf of
Normal distribution : fly lol =
(2h02It exp . ftp.fl-XBI)
In C Bane ) maximize the likelihood
order to
find UL E ,
we
log -
.

FOC 's are

given by :
 i:÷÷H::¥÷ ÷¥¥EI
3^uLE=Eo↳=Cx'x5tx and
%hE=
* As we see ,
In LE is
asymptotically efficient ,
but
on¥ is downward

biased in small samples .

Moving from first derivatives to 2nd derivatives;

:÷÷÷÷÷÷÷i÷f
÷ """

22µL
202231

Now : Since matrix matrix is evaluated with expectations

information
under true then
parameters ; we
get var -
cover matrix as :

µ
t
?

%)
-

t
)
Remember IM
-

I too ) = estimator for

±
z,

{
I ⑦o )
}
↳ asymptotic variance is
④

For numerical value; replace 02 with JULE (found in Foc )

④ Observe that there is no covariance between Fsu LE and Juke .

 Quasi U LE : When the density of a random variable is mis specified ,
.

We use Q MILE !pseudo U LE .

Information matrix equality doesn't hold
anymore .
In this case ,
we use 4th method ( sandwich estimator ) to

estimate asymptotic variance . However , mis

specified density leads to

inconsistent estimator ( E too )

Al -
Estimator : An estimator which is obtained by maximizing over a

criterion is re -
estimator! extremism estimator .

Qn ( O ) =
In Egly ,
X O
,
) } Average of n
-
sub
functions .

This is it
more
general form of
MLE ,
satisfies properties of U LE .

Examples : OLS , NLS , LAD , M LE .

Maximum Simulated Likelihood (MSL) : used when log - likelihoods include

integrals that are

expectations .

Binary Ply =L l ) pl x ) response probability

choice Mod Is X = is called .

LPM :
Response probability linear in parameters BJ : = 249=11×41 .
We
2 XJ

obtain unbiased and consistent estimates Also;

simply use OLS to
of 13 .

Interpretation
since
y is binary ,
conditional distribution D H Ix) is not normal .

is the same as OLS .

On the other hand :

}
Transform Xp by
→ predicted probabilities range outside [o D '

GC ) . { = link
function)
→
Marginal probabilities are sometimes non sense-
.

⇐ ×
13=951×7!! ?
'

→ LPM is heteros kedastic

by its
range .

logitm.de/:gcz,=eIez=eIIxp=N#
""

nap

tz# =§d where # HEATH -

¥2

Both are similar ,

logistic regression is thicker at the tails .

I robit
Log
1- * →

#
→ it

* ** Due to the non - linear structure the

, parameters in binary response models
cannot be interpreted the same OLS
way as .

* Effect of continuous variable :

214¥
)
=
24130-311241132×212 =
g (Bota X,
+ 132×2)
.

. 132
* Effect of categorical variable :
D= G Bo ( t B, Kk ti) t 132×2) -
G (Bo -113, Get BzXz)
*
Effect of a
Dummy : D= G ( Bot Bit 132×2 ) -

G (Bo t 132×2)

Partial Effect at Average HEA ) :

PEA=T¥ ,
First
get the
mean
of

data and evaluate

function there .

→
Evaluatefor

Average Partial Effect CAPE ) :

[email protected]
*i÷÷:÷÷÷÷:÷÷÷:::
en
(Ip ) =
xp .
If X changer by 1% ,

U L E of Binary choice Mode Is

""
Yi
Phil Xi ) = [ GHi BD .
[I -
G (Xi BD , yi = o
,
t .
Then the
log -
likelihood

lil3)
in
=
Yi .
en [ GaliBD t H Yi ) en ( t
-
.
-
G ki BD =

(
en G 24i ( -
t ) .
XiB ))
LL Thus
for each observation . we sum all LL 's .

In Kp ) = I LilB ) No analytical Soen '

Use numerical ML .

Variables ( E EV)
Endogenous Explanatory
-

* For a linear model,

directly apply 2SLS .

y it = aik teh t
E '

Yet [41*70] yµo ) .

I:))
Yz Zizi 72822
VI
= + +

gf then Yz
endogenous
p#
o is
, .
 Ul = O, .lk t e
, O, = Cov (4th) Nar (vz) = ( PT2) HE = 9th

insert this to the

Then first eq :

[email protected]#ekiD
y
*
Alternatively ,
we could follow two-step approach s .
t

→ OLS
of Yz on Z, I Zz and obtain Iz
→ Probit
of
Y, on Zi , Yz and Iz .
Test whether coefficient of Iz is

zero .

¥t If the endogenous variable is a

binary variable we cannot use

the 2 -

step approach ,
rather we
follow ULE . This is also called

" "
forbidden regression .

Binary choice and Panel Data Models

LPM Plyit =L Hit ) Xit.pt Ci ,

i
,
Ci =

assuming
strict
exogeneity .

L Pll restriction : OE Xit B ft

puts a . t Ci

For probit and

logit models; P Hit = I Hit Ci )
,
=
(Xit 13 tCi)

The partial effects

- of a continuous variable : =
By X ( Xit BtCi)
.

(Depends on unobserved effect Ci),

If Cin N ( o ,
of ) :
then ; PEA#Xt) = By 0C Xt B)
.
and

APE =
31cg 0(Xt Bc)
.
where Bc = 13/11+0240.5
 Random Effects : P (bit = l l X it , Ci ) = (X it 13 tCi)
.
and it assumes;

Strict conditional Cin Nco of)

exogeneity , independence and ,

plyit-tlxits-Ehit.to/hto2Pi# Be %ef.EE?efhusdoIenabYso
,

need to estimate of by
In random effects model we ,
assume no sandwich estimator .

dependence between Ci and Xit .

Correlated Random (CRE)

Effects :
Again assume strict
exogeneity and
We also allow extent
CI .

dependence between Ci and Xit to some

(Mandlak)

ig+aaiyga
Yit*=YtXitptXTgtait Estimation equation :

Fixed Effects CFE ) Strict conditional

:
exogeneity , independence ate

still valid Estimate

. Ci as n
parameters to estimate .

Dynamic Panel Data Models : Add tags of dependent variable .

yit-p.li#itXitl3-cituit&sextrgicetneeiIy9eniotuYgfIodineYu;taig}a .

Due to correlation between Yi t ,

-
I and Ci, OLS and GLS estimators

are upward biased ou ( Yi t

,
-
I ,
Kit Uit
D) YIP
= . Also within
group

deviation ( -
Yi ) doesn't solve the problem either because
, transformed
 lagged dependent variable and error -
term are correlated . The solution

is to take first -

differences ( Arellano Bond)

Dynamic Models (probit) :P Hit = l l Xi Yi, t -

I ,
- - - - Yi ,
o, Ci )
,

= # ( Xit B tf Yi t
.

,
- it Ci )

Yt

¥1# Hpt Mist Dtt ft

t
-

The joint density -

itCi .
-

OIL xp tpyi.t.it Cif

Initial conditions problem due to the
arises
correlation of Yo and Ci

↳ Five Heckman
different solutions available . & IN ookidge Two most

popular ones .

Corner Solution Models :

÷: : :÷*¥+÷
"

i
. .. ....
't
y , if y > O

to HBK)
The
probability the
given by 144=01 )
* at corner is x : =L -

* For y > 0 the

, density is as usual : fusil =

ftp.expf-tz.fY-gXIJ}
(
Yi-fl3-JLilp.rl-f-OIlxkr.DE#o?z.pfY-xf3y.c4i
Therefore the LL :
=L .
0

Solve LL by numerical maximum Thi .

Case When there at the then the LL reduces
#1 is
no observation corners
:
,

It (
to Stan dart LL .

can
be estimated
by OLS tu LE
Regression estimation)

Case # 2 : When all observations are on two corners ,

then it reduces to a

probit model :
-

43,0=[214--0]
log-t-OIHBKDa-b-tff-1.LY
en L .

TT
i =D log -4 OIL
'
)) Thu

IOI (
-
-

X b x BK
)
Coefficient Interpretations :

→ Latent variable .

From * 126 , 13 the effect X Yt not y We interested in

measures
of on . are

effects on
y . ELY IX) : Unconditional expectation

E (Y IX y > o ) Conditional
,
:
expectation

E (y l X ) = 144=01 x) .
O t Ply > ol x ) . E H IX, y > o )

EHlxt-OIXBHJ.EU/x#
-
Rule If =L NN lo, ) then
: t
,
for constant c ;

fIztzm=f¥
Thus '
Enix y >o) , =
xp to

*¥I¥÷÷÷÷¥tI¥h : inverse wins ratio

* The partial effect of conditional expectation is
of y :

µEHlY7=BstBs.dHxd3#=Bs{tHxBkDfxBk¥g
-
Partial Effect Expectation
of Conditional .

The
*
effect is
again not determined by By itself The sign .
of PE =
Sign of

=¥xBlN.Bs }
'

Partial Effect of Unconditional

Expectation .

If discrete (binary ) variable if there

* X is a or are
large jumps in continuous
variables, we
compute partial effects as
differences between EN IX ) for

different values
of x .

** Tobit Model makes an assumption of Normality and homoskedasticity

in the Hn der
lying latent variable In . case these assumptions are violated,

t obit estimates yield inconsistent MLE estimates .

Reduced form =

No
endogenous
vars in RH s

Censoring & Sample Selection

Missing
Censored regression, truncated regression & incidental Data
regression
*

Problems .

In Variables
censored regression → are
reported only to certain level .

Ex : INages are
reported up
to 70.000 E .
If the data are
missing Above
Threshold .
The censored normal regression is :

Yi = Bot XiB t Ui , Hilli ,

Cin N 6,02)

t.
{ Threshold level
Hi = min Yi Ci ,
} .

*
Regressing Wi on Xi would yield incorrect results

*
Regressing Yi on Xi would yield correct results but Yi ,
is unobserved .

This is similar to the

Pdf of a censored
regression is :

{¥
" """ "" &" ""
" t " * " " " """

but their interpretation is

f- (wiki ) =

e -

ozfcizxil) ,
if wi = C"

different .
In censored

regression model, the coefficients can be directly interpreted , Just

as in a linear model .
However ,
U LE estimation relies on normality and homes
tedasticity of
Ui Violation these two leads to inconsistent LULE)
.

of .

In draw the units If outcome

censored regression , we
randomly .
i is
missing
then it is above threshold
oz or
=

Truncated Normal Regression :

Yi =
Bot Xi B t Ui ,
4it Xi NN lo , 02 ) } Yi e Xi observed
if Yi sci Here
-

is not
sample a tandem one .

:÷÷÷÷÷÷÷:÷:::i:::::
Just in censored regression if
normality and homos kedashcity
-

as
, of

the error terms are violated, then MLE estimates are inconsistent .

Consistency of OLS : Choose si = I if part of sample, Si =D otherwise .

Si Yi = Si Xi B t Si Ui

1) E (s XJU ) E (s ) ECXTU ) Sample selection

= . =D : is
independent of
explanatory variables and error term .

2) E ( S U l SXI , S Xz
,
. .
-
) =D :
Exogenous sample selection .

3) Selection depends on
exp .
Kars .
and some other
factors uncorrelated

with the error term .

The Heckman (Incidental Truncation ) Model

Incidental Truncation IN Women in labor market : Woman who do

ages of
:

( but who would be

they ask for it)
not work
offered a
wage if are

never in the sample . Since truncation depends on another var .

(Labor

this called incidental truncation

force participation 1 ,
is .

The main
xp ECU 1×1=0
e
g
:
y = tu ,

The selection
eq : 5- I [ ZY t VZo] ,
X is a strict subset
of Z .

Hyun it: e:D

E-HittiiliisobservedJ-xiptf-E.si/dv
+
Since Mills ratio (7) is correlated with X , linear
inverse a
regression

would
yield biased estimates unless
f
= 0 .

Estimation
of Heckman Model

① Estimate error term : Estimate 7

from a
probit of s on Z
using the

entire
sample PCs l l Z) OI FEY) Then
using probit coefficients
calculate
= =
: .

I 0fEY inverse mills ratio : ;

=

OI Hii )

② Yi = Bo t Xi B t
f . 54 .
I terror , if p .
Tu to ,
there is selection bias .