You are on page 1of 64

Panel DataModelswithHeterogeneityandEndogeneity

J eff Wooldridge
MichiganStateUniversity
ProgrammeEvaluationfor PolicyAnalysis
Institutefor Fiscal Studies
J une2012
1. Introduction
2. General SetupandQuantitiesof Interest
3. AssumptionswithNeglectedHeterogeneity
4. ModelswithHeterogeneityandEndogeneity
5. EstimatingSomePopular Models
1
1. Introduction
∙ Whenpanel datamodelscontainunobservedheterogeneityand
omittedtime-varyingvariables, control functionmethodscanbeusedto
account for bothproblems.
∙ Under fairlyweekassumptionscanobtainconsistent, asymptotically
normal estimatorsof averagestructural functions– providedsuitable
instrumentsareavailable.
∙ Other issueswithpanel data: Howtotreat dynamics? Modelswith
laggeddependent variablesarehardtoestimatewhenheterogeneityand
other sourcesof endogeneityarepresent.
2
∙ Approachestohandlingunobservedheterogeneity:
1. Treat asparameterstoestimate. Canworkwell withlargeT but with
small T canhaveincidental parametersproblem. Biasadjustmentsare
availablefor parametersandaveragepartial effects. Usuallyweak
dependenceor evenindependenceisassumedacrossthetime
dimension.
2. Removeheterogeneitytoobtainanestimatingequation. Worksfor
simplelinear modelsandafewnonlinear models(viaconditional MLE
or aquasi-MLE variant). Cannot bedoneingeneral. Also, maynot be
abletoidentifyinterestingpartial effects.
3
∙ CorrelatedRandomEffects: Mundlak/Chamberlain. Requiressome
restrictionsondistributionof heterogeneity, althoughthesecanbe
nonparametric. Appliesgenerally, doesnot imposerestrictionson
dependenceover time, allowsestimationof averagepartial effects. Can
beeasilycombinedwithCF methodsfor endogeneity.
∙ Cantrytoestablishboundsrather thanestimateparametersor APEs.
Chernozhukov, Fernández-Val, Hahn, andNewey(2009) isarecent
example.
4
2. General Setup and Quantities of Interest
∙ Static, unobservedeffectsprobit model for panel datawithanomitted
time-varyingvariabler
it
:
Py
it
 1|x
it
,c
i
,r
it
  x
it
þ  c
i
 r
it
, t  1,...,T. (1)
What arethequantitiesof interest for most purposes?
(i) Theelement of þ, the[
j
. Thesegivethedirectionsof thepartial
effectsof thecovariatesontheresponseprobability. For anytwo
continuouscovariates, theratioof coefficients, [
j
/[
h
, isidentical tothe
ratioof partial effects(andtheratiodoesnot dependonthecovariates
or unobservedheterogeneity, c
i
).
5
(ii) Themagnitudesof thepartial effects. Thesedependnot onlyonthe
valueof thecovariates, sayx
t
, but alsoonthevalueof theunobserved
heterogeneity. Inthecontinuouscovariatecase,
∂Py
t
 1|x
t
,c,r
t

∂x
tj
 [
j
çx
t
þ  c  r
t
. (2)
∙ Questions: (a) Assumingwecanestimateþ, what shouldwedoabout
theunobservablesc,r
t
? (b) If wecanonlyestimateþ up-to-scale, can
westill learnsomethinguseful about magnitudesof partial effects? (c)
What kindsof assumptionsdoweneedtoestimatepartial effects?
6
∙ Let x
it
,y
it
 : t  1,...,T bearandomdrawfromthecrosssection.
Supposeweareinterestedin
Ey
it
|x
it
,c
i
,r
it
  m
t
x
it
,c
i
,r
it
. (3)
c
i
canbeavector of unobservedheterogeneity, r
it
avector of omitted
time-varyingvariables.
∙ Partial effects: if x
tj
iscontinuous, then
0
j
x
t
,c,r
t
 ≡
∂m
t
x
t
,c,r
t

∂x
tj
, (4)
or discretechanges.
7
∙ Howdoweaccount for unobservedc
i
,r
it
? If weknowenough
about thedistributionof c
i
,r
it
 wecaninsert meaningful valuesfor
c,r
t
. For example, if µ
c
 Ec
i
, µ
r
t
 Er
it
 thenwecancompute
thepartial effect at theaverage(PEA),
PEA
j
x
t
  0
j
x
t

c

r
t
. (5)
Of course, weneedtoestimatethefunctionm
t
andµ
c

r
t
. If wecan
estimatethedistributionof c
i
,r
it
, or featuresinadditiontoitsmean,
wecaninsert different quantiles, or acertainnumber of standard
deviationsfromthemean.
8
∙ Alternatively, wecanobtaintheaveragepartial effect (APE) (or
populationaverageeffect) byaveragingacrossthedistributionof c
i
:
APEx
t
  E
c
i
,r
it

0
j
x
t
,c
i
,r
it
. (6)
Thedifferencebetween(5) and(6) canbenontrivial. Insomeleading
cases, (6) isidentifiedwhile(5) isnot. (6) iscloselyrelatedtothe
notionof theaveragestructural function(ASF) (Blundell andPowell
(2003)). TheASF isdefinedas
ASF
t
x
t
  E
c
i
,r
it

m
t
x
t
,c
i
,r
it
. (7)
∙ Passingthederivativethroughtheexpectationin(7) givestheAPE.
9
3. Assumptions with Neglected Heterogeneity
Exogeneity of Covariates
∙ Cannot get bywithjust specifyingamodel for thecontemporaneous
conditional distribution, Dy
it
|x
it
,c
i
.
∙ Themost useful definitionof strict exogeneityfor nonlinear panel
datamodelsis
Dy
it
|x
i1
,...,x
iT
,c
i
  Dy
it
|x
it
,c
i
. (8)
Chamberlain(1984) labeled(8) strict exogeneity conditional on the
unobserved effects c
i
. Conditional meanversion:
Ey
it
|x
i1
,...,x
iT
,c
i
  Ey
it
|x
it
,c
i
. (9)
10
∙ Thesequential exogeneity assumptionis
Dy
it
|x
i1
,...,x
it
,c
i
  Dy
it
|x
it
,c
i
. (10)
Muchmoredifficult toallowsequential exogeneityininnonlinear
models. (Most progresshasbeenmadefor laggeddependent variables
or specificfunctional forms, suchasexponential.)
∙ Neither strict nor sequential exogeneityallowsfor contemporaneous
endogeneityof oneor moreelementsof x
it
, where, say, x
itj
iscorrelated
withunobserved, time-varyingunobservablesthat affect y
it
.
11
Conditional Independence
∙ Inlinear models, serial dependenceof idiosyncraticshocksiseasily
dealt with, either by“cluster robust” inferenceor GeneralizedLeast
Squaresextensionsof FixedEffectsandFirst Differencing. With
strictlyexogenouscovariates, serial correlationnever resultsin
inconsistent estimation, evenif improperlymodeled. Thesituationis
different withmost nonlinear modelsestimatedbyMLE.
∙ Conditional independence (CI) (under strict exogeneity):
Dy
i1
,...,y
iT
|x
i
,c
i
 

t1
T
Dy
it
|x
it
,c
i
. (11)
12
∙ Inaparametriccontext, theCI assumptionreducesour taskto
specifyingamodel for Dy
it
|x
it
,c
i
, andthendetermininghowtotreat
theunobservedheterogeneity, c
i
.
∙ Inrandomeffectsandcorrelatedrandomframeworks(next section),
CI playsacritical roleinbeingabletoestimatethe“structural”
parametersandtheparametersinthedistributionof c
i
(andtherefore, in
estimatingPEAs). Inabroadclassof popular models, CI playsno
essential roleinestimatingAPEs.
13
Assumptions about the Unobserved Heterogeneity
Random Effects
∙ Generallystated, thekeyRE assumptionis
Dc
i
|x
i1
,...,x
iT
  Dc
i
. (12)
Under (12), theAPEsareactuallynonparametricallyidentifiedfrom
Ey
it
|x
it
 x
t
. (13)
∙ Insomeleadingcases(RE probit andRE Tobit withheterogeneity
normallydistributed), if wewant PEsfor different valuesof c, wemust
assumemore: strict exogeneity, conditional independence, and(12)
withaparametricdistributionfor Dc
i
.
14
Correlated Random Effects
A CRE frameworkallowsdependencebetweenc
i
andx
i
, but restricted
insomeway. Inaparametricsetting, wespecifyadistributionfor
Dc
i
|x
i1
,...,x
iT
, asinChamberlain(1980,1982), andmuchwork
since. Distributional assumptionsthat leadtosimpleestimation–
homoskedasticnormal withalinear conditional mean— canbe
restrictive.
15
∙ Possibletodropparametricassumptionsandjust assume
Dc
i
|x
i
  Dc
i
|x
̄
i
, (14)
without restrictingDc
i
|x
̄
i
. Altonji andMatzkin(2005, Econometrica).
∙ Other functionsof x
it
: t  1,...,T arepossible.
16
∙ APEsareidentifiedverygenerally. For example, under (14), a
consistent estimateof theaveragestructural functionis
ASFx
t
  N
−1

i1
N
q
t
x
t
,x
̄
i
, (15)
whereq
t
x
it
,x
̄
i
  Ey
it
|x
it
,x
̄
i
.
∙ Needarandomsamplex
̄
i
: i  1,...,N for theaveragingout to
work.
17
Fixed Effects
∙ Thelabel “fixedeffects” isuseddifferentlybydifferent researchers.
Oneview: c
i
, i  1,...,N areparameterstobeestimated. Usuallyleads
toan“incidental parametersproblem.”
∙ Secondmeaningof “fixedeffects”: Dc
i
|x
i
 isunrestrictedandwe
lookfor objectivefunctionsthat donot dependonc
i
but still identify
thepopulationparameters. Leadsto“conditional MLE” if wecanfind
“sufficient statistics” s
i
suchthat
Dy
i1
,...,y
iT
|x
i
,c
i
,s
i
  Dy
i1
,...,y
iT
|x
i
,s
i
. (16)
∙ Conditional Independenceisusuallymaintained.
∙ Keypoint: PEAsandAPEsaregenerallyunidentified.
18
4. Models with Heterogeneity and Endogeneity
∙ Let y
it1
beascalar response, y
it2
avector of endogenousvariables,
z
it1
exogenousvariables, andwehave
Ey
it1
|y
it2
,z
it1
,c
i1
,r
it1
  m
t1
y
it2
,z
it1
,c
i1
,r
it1
 (17)
∙ y
it2
isallowedtobecorrelatedwithr
it1
(aswell aswithc
i1
).
∙ Thevector of exogenousvariablesz
it
: t  1,...,T withz
it1
⊂ z
it
arestrictlyexogenousinthesensethat
Ey
it
|y
it2
,z
i
,c
i1
,r
it1
  Ey
it
|y
it2
,z
it1
,c
i1
,r
it1

Dr
it1
|z
i
,c
i1
  Dr
it1

(18)
(19)
19
∙ Sometimeswecaneliminatec
i
andobtainanequationthat canbe
estimatedbyIV (linear, exponential). Generallynot possible.
∙ NowaCRE approachinvolvesmodelingDc
i1
|z
i
.
∙ Generally, weneedtomodel howy
it2
isrelatedtor
it1
.
∙ Control Functionmethodsareconvenient for allowingboth.
∙ Supposey
it2
isascalar and
y
it2
 m
it2
z
it
,z
̄
i

2
  v
it2
Ev
it2
|z
i
  0
Dr
it1
|v
it2
,z
i
  Dr
it1
|v
it2

(20)
20
∙ Withsuitabletime-variationintheinstruments, theassumptionsin
(20) allowidentificationof theASF if weassumeamodel for
Dc
i1
|z
i
,v
it2

Generally, wecanestimate
Ey
it1
|y
it2
,z
i
,v
it2
  Ey
it1
|y
it2
,z
it1
,z
̄
i
,v
it2
 ≡ g
t1
y
it2
,z
it1
,z
̄
i
,v
it2
 (21)
21
∙ TheASF isnowobtainedbyaveragingout z
̄
i
,v
it2
:
ASFy
t2
,z
t1
  E
z̄
i
,v
it2

g
t1
y
t2
,z
t1
,z
̄
i
,v
it2

∙ Most of thiscanbefullynonparametric(Altonji andMatzkin, 2005;
Blundell andPowell, 2003) althoughsomerestrictionisneededon
Dc
i1
|z
i
,v
it2
, suchas
Dc
i1
|z
i
,v
it2
  Dc
i1
|z
̄
i
,v
it2

∙ WithT sufficientlylargewecanaddother featuresof
z
it
: t  1,...,T toz
̄
i
.
22
5. Estimating Some Popular Models
Linear Model with Endogeneity
∙ Simplest model is
y
it1
 o
1
y
it2
 z
it1
ö
1
 c
i1
 u
it1
≡ x
it1
þ
1
 c
i1
 u
it1
Eu
it1
|z
i
,c
i1
  0
(22)
∙ Thefixedeffects2SLSestimator iscommon. Deviatevariablesfrom
timeaveragestoremovec
i1
thenapplyIV:
ÿ
it1
 ẍ
it1
þ
1
 ü
it1

it
 z
it
− z
̄
i
23
∙ Easytomakeinferencerobust toserial correlationand
heteroskedasticityinu
it1
. (“Cluster-robust inference.”)
∙ Test for (strict) exogeneityof y
it2
:
(i) Estimatethereducedformof y
it2
byusual fixedeffects:
y
it2
 z
it
ö
1
 c
i2
 u
it2
Get theFE residuals,

ü
it2
 ÿ
it2
− z̈
it
ö
̂
1
.
∙ Estimatetheaugment equation
y
it1
 o
1
y
it2
 z
it1
ö
1
 ¡
1

ü
it2
 c
i1
 error
it
(23)
byFE anduseacluster-robust test of H
0
: ¡
1
 0.
24
∙ TherandomeffectsIV approachassumesc
i1
isuncorrelatedwithz
i
,
andnominallyimposesserial independenceonu
it1
.
∙ Simplewaytotest thenull whether REIV issufficient. (Robust
Hausmantest comparingREIV andFEIV.)
Estimate
y
it1
 p
1
 x
it1
þ
1
 z
̄
i
ç
1
 a
i1
 u
it1
(24)
byREIV, usinginstruments1,z
it
,z
̄
i
. Theestimator of þ
1
istheFEIV
estimator.
∙ Test H
0
: ç
1
 0, preferablyusingafullyrobust test. A rejectionis
evidencethat theIVsarecorrelatedwithc
i
, andshoulduseFEIV.
25
∙ Other thantherankcondition, thekeyconditionfor FEIV tobe
consistent isthat theinstruments, z
it
, arestrictlyexogenouswith
respect tou
it
. WithT ≥ 3timeperiods, thisiseasilytested– asinthe
usual FE case.
∙ Theaugmentedmodel is
y
it1
 x
it1
þ
1
 z
i,t1
ç
1
 c
i1
 u
it1
, t  1,...,T − 1
andweestimateit byFEIV, usinginstrumentsz
it
,z
i,t1
.
∙ Useafullyrobust Waldtest of H
0
: ç
1
 0. Canbeselectiveabout
whichleadstoinclude.
26
Example: EstimatingaPassenger DemandFunctionfor Air Travel
N  1,149, T  4.
∙ Usesrouteconcentrationfor largest carrier asIV for logfare.
. use ai r f ar e
. * Reduced f or mf or l f ar e; concen i s t he I V.
. xt r eg l f ar e concen l di st l di st sq y98 y99 y00, f e cl ust er ( i d)
( St d. Er r . adj ust ed f or 1149 cl ust er s i n i d)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
| Robust
l f ar e | Coef . St d. Er r . t P| t | [ 95% Conf . I nt er val ]
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
concen | . 168859 . 0494587 3. 41 0. 001 . 0718194 . 2658985
l di st | ( dr opped)
l di st sq | ( dr opped)
y98 | . 0228328 . 004163 5. 48 0. 000 . 0146649 . 0310007
y99 | . 0363819 . 0051275 7. 10 0. 000 . 0263215 . 0464422
y00 | . 0977717 . 0055054 17. 76 0. 000 . 0869698 . 1085735
_cons | 4. 953331 . 0296765 166. 91 0. 000 4. 895104 5. 011557
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
si gma_u | . 43389176
si gma_e | . 10651186
r ho | . 94316439 ( f r act i on of var i ance due t o u_i )
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
27
. xt i vr eg l passen l di st l di st sq y98 y99 y00 ( l f ar e  concen) , r e t het a
G2SLS r andom- ef f ect s I V r egr essi on Number of obs  4596
Gr oup var i abl e: i d Number of gr oups  1149
R- sq: wi t hi n  0. 4075 Obs per gr oup: mi n  4
bet ween  0. 0542 avg  4. 0
over al l  0. 0641 max  4
Wal d chi 2( 6)  231. 10
cor r ( u_i , X)  0 ( assumed) Pr ob  chi 2  0. 0000
t het a  . 91099494
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
l passen | Coef . St d. Er r . z P| z| [ 95% Conf . I nt er val ]
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
l f ar e | - . 5078762 . 229698 - 2. 21 0. 027 - . 958076 - . 0576763
l di st | - 1. 504806 . 6933147 - 2. 17 0. 030 - 2. 863678 - . 1459338
l di st sq | . 1176013 . 0546255 2. 15 0. 031 . 0105373 . 2246652
y98 | . 0307363 . 0086054 3. 57 0. 000 . 0138699 . 0476027
y99 | . 0796548 . 01038 7. 67 0. 000 . 0593104 . 0999992
y00 | . 1325795 . 0229831 5. 77 0. 000 . 0875335 . 1776255
_cons | 13. 29643 2. 626949 5. 06 0. 000 8. 147709 18. 44516
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
si gma_u | . 94920686
si gma_e | . 16964171
r ho | . 96904799 ( f r act i on of var i ance due t o u_i )
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
I nst r ument ed: l f ar e
I nst r ument s: l di st l di st sq y98 y99 y00 concen
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
. * The quasi - t i me- demeani ng par amet er i s qui t e l ar ge: . 911 ( " t het a" ) .
28
. xt i vr eg2 l passen l di st l di st sq y98 y99 y00 ( l f ar e  concen) , f e cl ust er ( i d)
War ni ng - col l i near i t i es det ect ed
Var s dr opped: l di st l di st sq
FI XED EFFECTS ESTI MATI ON
- - - - - - - - - - - - - - - - - - - - - - - -
Number of gr oups  1149 Obs per gr oup: mi n  4
avg  4. 0
max  4
Number of cl ust er s ( i d)  1149 Number of obs  4596
F( 4, 1148)  26. 07
Pr ob  F  0. 0000
Tot al ( cent er ed) SS  128. 0991685 Cent er ed R2  0. 2265
Tot al ( uncent er ed) SS  128. 0991685 Uncent er ed R2  0. 2265
Resi dual SS  99. 0837238 Root MSE  . 1695
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
| Robust
l passen | Coef . St d. Er r . z P| z| [ 95% Conf . I nt er val ]
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
l f ar e | - . 3015761 . 6124127 - 0. 49 0. 622 - 1. 501883 . 8987307
y98 | . 0257147 . 0164094 1. 57 0. 117 - . 0064471 . 0578766
y99 | . 0724166 . 0250971 2. 89 0. 004 . 0232272 . 1216059
y00 | . 1127914 . 0620115 1. 82 0. 069 - . 0087488 . 2343316
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
I nst r ument ed: l f ar e
I ncl uded i nst r ument s: y98 y99 y00
Excl uded i nst r ument s: concen
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
29
. egen concenb  mean( concen) , by( i d)
. xt i vr eg l passen l di st l di st sq y98 y99 y00 concenb ( l f ar e  concen) , r e t het a
G2SLS r andom- ef f ect s I V r egr essi on Number of obs  4596
Gr oup var i abl e: i d Number of gr oups  1149
Wal d chi 2( 7)  218. 80
cor r ( u_i , X)  0 ( assumed) Pr ob  chi 2  0. 0000
t het a  . 90084889
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
l passen | Coef . St d. Er r . z P| z| [ 95% Conf . I nt er val ]
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
l f ar e | - . 3015761 . 2764376 - 1. 09 0. 275 - . 8433838 . 2402316
l di st | - 1. 148781 . 6970189 - 1. 65 0. 099 - 2. 514913 . 2173514
l di st sq | . 0772565 . 0570609 1. 35 0. 176 - . 0345808 . 1890937
y98 | . 0257147 . 0097479 2. 64 0. 008 . 0066092 . 0448203
y99 | . 0724165 . 0119924 6. 04 0. 000 . 0489118 . 0959213
y00 | . 1127914 . 0274377 4. 11 0. 000 . 0590146 . 1665682
concenb | - . 5933022 . 1926313 - 3. 08 0. 002 - . 9708527 - . 2157518
_cons | 12. 0578 2. 735977 4. 41 0. 000 6. 695384 17. 42022
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
si gma_u | . 85125514
si gma_e | . 16964171
r ho | . 96180277 ( f r act i on of var i ance due t o u_i )
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
I nst r ument ed: l f ar e
I nst r ument s: l di st l di st sq y98 y99 y00 concenb concen
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
30
. i vr eg l passen l di st l di st sq y98 y99 y00 concenb ( l f ar e  concen) , cl ust er ( i d)
I nst r ument al var i abl es ( 2SLS) r egr essi on Number of obs  4596
F( 7, 1148)  20. 28
Pr ob  F  0. 0000
R- squar ed  0. 0649
Root MSE  . 85549
( St d. Er r . adj ust ed f or 1149 cl ust er s i n i d)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
| Robust
l passen | Coef . St d. Er r . t P| t | [ 95% Conf . I nt er val ]
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
l f ar e | - . 3015769 . 6131465 - 0. 49 0. 623 - 1. 50459 . 9014366
l di st | - 1. 148781 . 8809895 - 1. 30 0. 193 - 2. 877312 . 5797488
l di st sq | . 0772566 . 0811787 0. 95 0. 341 - . 0820187 . 2365319
y98 | . 0257148 . 0164291 1. 57 0. 118 - . 0065196 . 0579491
y99 | . 0724166 . 0251272 2. 88 0. 004 . 0231163 . 1217169
y00 | . 1127915 . 0620858 1. 82 0. 070 - . 0090228 . 2346058
concenb | - . 5933019 . 2963723 - 2. 00 0. 046 - 1. 174794 - . 0118099
_cons | 12. 05781 4. 360868 2. 77 0. 006 3. 50164 20. 61397
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
I nst r ument ed: l f ar e
I nst r ument s: l di st l di st sq y98 y99 y00 concenb concen
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
31
. * Now t est whet her i nst r ument ( concen) i s st r i ct l y exogenous.
. xt i vr eg2 l passen y98 y99 concen_p1 ( l f ar e  concen) , f e cl ust er ( i d)
FI XED EFFECTS ESTI MATI ON
- - - - - - - - - - - - - - - - - - - - - - - -
Number of gr oups  1149 Obs per gr oup: mi n  3
avg  3. 0
max  3
Number of cl ust er s ( i d)  1149 Number of obs  3447
F( 4, 1148)  33. 41
Pr ob  F  0. 0000
Tot al ( cent er ed) SS  67. 47207834 Cent er ed R2  0. 4474
Tot al ( uncent er ed) SS  67. 47207834 Uncent er ed R2  0. 4474
Resi dual SS  37. 28476721 Root MSE  . 1274
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
| Robust
l passen | Coef . St d. Er r . z P| z| [ 95% Conf . I nt er val ]
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
l f ar e | - . 8520992 . 3211832 - 2. 65 0. 008 - 1. 481607 - . 2225917
y98 | . 0416985 . 0098066 4. 25 0. 000 . 0224778 . 0609192
y99 | . 0948286 . 014545 6. 52 0. 000 . 066321 . 1233363
concen_p1 | . 1555725 . 0814452 1. 91 0. 056 - . 0040571 . 3152021
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
I nst r ument ed: l f ar e
I ncl uded i nst r ument s: y98 y99 concen_p1
Excl uded i nst r ument s: concen
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
32
. * What i f we j ust use f i xed ef f ect s wi t hout I V?
. xt r eg l passen l f ar e y98 y99 y00, f e cl ust er ( i d)
Fi xed- ef f ect s ( wi t hi n) r egr essi on Number of obs  4596
Gr oup var i abl e: i d Number of gr oups  1149
R- sq: wi t hi n  0. 4507 Obs per gr oup: mi n  4
bet ween  0. 0487 avg  4. 0
over al l  0. 0574 max  4
F( 4, 1148)  121. 85
cor r ( u_i , Xb)  - 0. 3249 Pr ob  F  0. 0000
( St d. Er r . adj ust ed f or 1149 cl ust er s i n i d)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
| Robust
l passen | Coef . St d. Er r . t P| t | [ 95% Conf . I nt er val ]
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
l f ar e | - 1. 155039 . 1086574 - 10. 63 0. 000 - 1. 368228 - . 9418496
y98 | . 0464889 . 0049119 9. 46 0. 000 . 0368516 . 0561262
y99 | . 1023612 . 0063141 16. 21 0. 000 . 0899727 . 1147497
y00 | . 1946548 . 0097099 20. 05 0. 000 . 1756036 . 213706
_cons | 11. 81677 . 55126 21. 44 0. 000 10. 73518 12. 89836
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
si gma_u | . 89829067
si gma_e | . 14295339
r ho | . 9753002 ( f r act i on of var i ance due t o u_i )
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
33
. * Test f or mal l y f or endogenei t y of l f ar e i n FE:
. qui ar eg l f ar e concen y98 y99 y00, absor b( i d)
. pr edi ct u2h, r esi d
. xt r eg l passen l f ar e y98 y99 y00 v2h, f e cl ust er ( i d)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
| Robust
l passen | Coef . St d. Er r . t P| t | [ 95% Conf . I nt er val ]
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
l f ar e | - . 301576 . 4829734 - 0. 62 0. 532 - 1. 249185 . 6460335
y98 | . 0257147 . 0131382 1. 96 0. 051 - . 0000628 . 0514923
y99 | . 0724165 . 0197133 3. 67 0. 000 . 0337385 . 1110946
y00 | . 1127914 . 048597 2. 32 0. 020 . 0174425 . 2081403
u2h | - . 8616344 . 5278388 - 1. 63 0. 103 - 1. 897271 . 1740025
_cons | 7. 501007 2. 441322 3. 07 0. 002 2. 711055 12. 29096
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
. * p- val ue i s about . 10, so not st r ong evi dence even t hough FE and
. * FEI V est i mat oest i mat es ar e ui t e di f f er ent .
34
∙ Turnsout that theFE2SLSestimator isrobust torandomcoefficients
onx
it1
, but oneshouldincludeafull set of timedummies.
(Murtazashvili andWooldridge, 2005).
∙ Canmodel randomcoefficientsanduseaCF approach.
y
it1
 c
i1
 x
it1
b
i1
 u
it1
y
it2
 p
2
 z
it
ö
2
 z
̄
i
ç
2
 v
it2
∙ AssumeEc
i1
|z
i
,v
it2
 andEb
i1
|z
i
,v
it2
 arelinear inz
̄
i
,v
it2
 and
Eu
it1
|z
i
,v
it2
 islinear inv
it2
, canshow
Ey
it1
|z
i
,v
it2
  t
1
 x
it1
þ
1
 z
̄
i
ç
1
 ¡
1
v
it2
 z
̄
i
− µ

 ⊗ x
it1
æ
1
 v
it2
x
it1
ç
1
(25)
35
(1) Regressy
it2
on1, z
it
, z
̄
i
andobtainresidualsv̂
it2
.
(2) Regress
y
it1
on1, x
it1
, z
̄
i
, v̂
it2
, z
̄
i
− z
̄
 ⊗ x
it1
, v̂
it2
x
it1
∙ Probablyincludetimedummiesinbothstages.
36
Binary and Fractional Response
∙ Unobservedeffects(UE) “probit” model – exogenousvariables. For a
binaryor fractional y
it
,
Ey
it
|x
it
,c
i
  x
it
þ  c
i
, t  1,...,T. (26)
Assumestrict exogeneity(conditional onc
i
) andChamberlain-Mundlak
device:
c
i
 ç  x
̄
i
ç  a
i
, a
i
|x
i
~Normal(0,o
a
2
. (27)
37
∙ Inbinaryresponsecaseunder serial independence, all parametersare
identifiedandMLE (Stata: xtprobit) canbeused. J ust addthetime
averagesx
̄
i
asanadditional set of regressors. Thenj
̂
c
 ç
̂
 x
̄
ç
̂
and
o
̂
c
2
≡ ç
̂

N
−1

i1
N
x
̄
i
− x
̄


x
̄
i
− x
̄
 ç
̂
 o
̂
a
2
. CanevaluatePEsat, say,
j
̂
c
 ko
̂
c
.
∙ Onlyunder restrictiveassumptionsdoesc
i
haveanunconditional
normal distribution, althoughit becomesmorereasonableasT gets
large.
∙ Simpletotest H
0
: ç  0 asnull that c
i
, x
̄
i
areindependent.
38
∙ TheAPEsareidentifiedfromtheASF, estimatedas
ASFx
t
  N
−1

i1
N
x
t
þ
̂
a
 ç
̂
a
 x
̄
i
ç
̂
a
 (28)
where, for example, þ
̂
a
 þ
̂
/1 o
̂
a
2

1/2
.
∙ For binaryor fractional response, APEsareidentifiedwithout the
conditional serial independenceassumption. UsepooledBernoulli
quasi-MLE (Stata: glm) or generalizedestimatingequations(Stata:
xtgee) toestimatescaledcoefficientsbasedon
Ey
it
|x
i
  x
it
þ
a
 ç
a
 x
̄
i
ç
a
. (29)
(Timedummieshavebeensupressedfor simplicity.)
39
∙ A moreradical suggestion, but inthespirit of Altonji andMatzkin
(2005), istojust useaflexiblemodel for Ey
it
|x
it
,x
̄
i
 directly, say,
Ey
it
|x
it
,x
̄
i
  0
t
 x
it
þ  x
̄
i
y 
x
̄
i
⊗ x
̄
i
ö  x
it
⊗ x
̄
i
q.
(30)
J ust averageout over x
̄
i
toget APEs.
∙ If wehaveabinaryresponse, start with
Py
it
 1|x
it
,c
i
  x
it
þ  c
i
, (31)
andassumeCI, wecanestimateþ byFE logit without restricting
Dc
i
|x
i
.
40
∙ Inanynonlinear model usingtheMundlakassumption
Dc
i
|x
i
  Dc
i
|x
̄
i
, if T ≥ 3canincludeleadvalues, w
i,t1
, tosimply
test strict exogeneity.
∙ Example: MarriedWomen’sLabor ForceParticipation: N  5,663,
T  5(four-monthintervals).
∙ Followingresultsincludeafull set of timeperioddummies(not
reported).
∙ TheAPEsaredirectlycomparableacrossmodels, andcanbe
comparedwiththelinear model coefficients.
41
LFP (1) (2) (3) (4) (5)
Model Linear Probit CRE Probit CRE Probit FE Logit
Est. Method FE PooledMLE PooledMLE MLE MLE
Coef. Coef. APE Coef. APE Coef. APE Coef.
kids −.0389 −.199 −.0660 −.117 −.0389 −.317 −.0403 −.644
.0092 .015 .0048 .027 .0085 .062 .0104 .125
lhinc −.0089 −.211 −.0701 −.029 −.0095 −.078 −.0099 −.184
.0046 .024 .0079 .014 .0048 .041 .0055 .083
kids — — — −.086 — −.210 — —
— — — .031 — .071 — —
lhinc — — — −.250 — −.646 — —
— — — .035 — .079 — —
42
Probit with Endogenous Explanatory Variables
∙ Represent endogeneityasanomitted, time-varyingvariable, in
additiontounobservedheterogeneity:
Py
it1
 1|y
it2
,z
i
,c
i1
,v
it1
  Py
it1
 1|y
it2
,z
it1
,c
i1
,r
it1

 x
it1
þ
1
 c
i1
 r
it1

∙ Elementsof z
it
areassumedstrictlyexogenous, andwehaveat least
oneexclusionrestriction: z
it
 z
it1
,z
it2
.
43
∙ PapkeandWooldridge(2008, J ournal of Econometrics): Usea
Chamberlain-Mundlakapproach, but onlyrelatingtheheterogeneityto
all strictlyexogenousvariables:
c
i1
 ç
1
 z
̄
i
ç
1
 a
i1
, Da
i1
|z
i
  Da
i1
.
∙ EvenbeforewespecifyDa
i1
, thisisrestrictivebecauseit assumes,
inparticular, Ec
i
|z
i
 islinear inz
̄
i
andthat Varc
i
|z
i
 isconstant.
Usingnonparametricscanget bywithless, suchas
Dc
i1
|z
i
  Dc
i1
|z
̄
i
.
44
∙ Onlyneed
Ey
it1
|y
it2
,z
i
,c
i1
,v
it1
  x
it1
þ
1
 c
i1
 v
it1
, (32)
soappliestofractional response.
∙ Needtoobtainanestimatingequation. First, notethat
Ey
it1
|y
it2
,z
i
,a
i1
,r
it1
  x
it1
þ
1
 ç
1
 z
̄
i
ç
1
 a
i1
 r
it1

≡ x
it1
þ
1
 ç
1
 z
̄
i
ç
1
 v
it1
. (33)
45
∙ Assumealinear reducedformfor y
it2
:
y
it2
 ç
2
 z
it
o
2
 z
̄
i
ç
2
 v
it2
,t  1,...,T
Dv
it2
|z
i
  Dv
it2

(34)
(andwemight allowfor time-varyingcoefficients).
∙ Next, assume
v
it1
|z
i
,v
it2
  Normalp
1
v
it2
,k
1
2
,t  1,...,T.
[Easytoallowp
1
tochangeover time; just havetimedummiesinteract
withv
it2
.]
∙ Assumptionseffectivelyruleout discretenessiny
it2
.
46
∙ Write
v
it1
 p
1
v
it2
 e
it1
wheree
it1
isindependent of z
i
,v
it2
 (and, therefore, of y
it2
) and
normallydistributed. Again, usingastandardmixingpropertyof the
normal distribution,
Ey
it1
|y
it2
,z
i
,v
it2
  x
it1
þ
k1
 ç
k1
 z
̄
i
ç
k1
 p
k1
v
it2
 (35)
wherethe“k” denotesdivisionby1 k
1
2

1/2
.
∙ Identificationcomesoff of theexclusionof thetime-varying
exogenousvariablesz
it2
.
47
∙ Twostepprocedure(PapkeandWooldridge, 2008):
(1) Estimatethereducedformfor y
it2
(pooledor for eacht
separately). Obtaintheresiduals, v̂
it2
.
(2) Usetheprobit QMLE toestimateþ
k1

k1

k1
andp
k1
.
∙ Howdoweinterpret thescaledestimates? Theygivedirectionsof
effects. Conveniently, theyalsoindextheAPEs. For giveny
2
andz
1
,
averageout z
̄
i
andv̂
it2
(for eacht):
o
̂
k1
 N
−1

i1
N
ço
̂
k1
y
t2
 z
t1
ö
̂
k1
 ç
̂
k1
 z
̄
i
ç
̂
k1
 p̂
k1

it2
 .
48
∙ Application: Effectsof SpendingonTest PassRates
∙ N  501school districts, T  7timeperiods.
∙ Oncepre-policyspendingiscontrolledfor, instrument spendingwith
the“foundationgrant.”
∙ Initial spendingtakestheplaceof thetimeaverageof IVs.
49
. * Fi r st , l i near model :
. i vr eg mat h4 l unch al unch l enr ol l al enr ol l y96- y01 l exppp94 l e94y96- l e94y01
( l avgr exp  l f ound l f ndy96- l f ndy01) , cl ust er ( di st i d)
I nst r ument al var i abl es ( 2SLS) r egr essi on Number of obs  3507
F( 18, 500)  107. 05
Pr ob  F  0. 0000
R- squar ed  0. 4134
Root MSE  . 11635
( St d. Er r . adj ust ed f or 501 cl ust er s i n di st i d)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
| Robust
mat h4 | Coef . St d. Er r . t P| t | [ 95% Conf . I nt er val ]
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
l avgr exp | . 5545247 . 2205466 2. 51 0. 012 . 1212123 . 987837
l unch | - . 0621991 . 0742948 - 0. 84 0. 403 - . 2081675 . 0837693
al unch | - . 4207815 . 0758344 - 5. 55 0. 000 - . 5697749 - . 2717882
l enr ol l | . 0463616 . 0696215 0. 67 0. 506 - . 0904253 . 1831484
al enr ol l | - . 049052 . 070249 - 0. 70 0. 485 - . 1870716 . 0889676
y96 | - 1. 085453 . 2736479 - 3. 97 0. 000 - 1. 623095 - . 5478119
. . .
y01 | - . 704579 . 7310773 - 0. 96 0. 336 - 2. 140941 . 7317831
l exppp94 | - . 4343213 . 2189488 - 1. 98 0. 048 - . 8644944 - . 0041482
l e94y96 | . 1253255 . 0318181 3. 94 0. 000 . 0628119 . 1878392
. . .
l e94y01 | . 0865874 . 0816732 1. 06 0. 290 - . 0738776 . 2470524
_cons | - . 334823 . 2593105 - 1. 29 0. 197 - . 8442955 . 1746496
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
I nst r ument ed: l avgr exp
I nst r ument s: l unch al unch l enr ol l al enr ol l y96 y97 y98 y99 y00 y01
l exppp94 l e94y96 l e94y97 l e94y98 l e94y99 l e94y00 l e94y01
l f ound l f ndy96 l f ndy97 l f ndy98 l f ndy99 l f ndy00 l f ndy01
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
50
. * Get r educed f or mr esi dual s f or f r act i onal pr obi t :
. r eg l avgr exp l f ound l f ndy96- l f ndy01 l unch al unch l enr ol l al enr ol l y96- y01
l exppp94 l e94y96- l e94y01, cl ust er ( di st i d)
Li near r egr essi on Number of obs  3507
F( 24, 500)  1174. 57
Pr ob  F  0. 0000
R- squar ed  0. 9327
Root MSE  . 03987
( St d. Er r . adj ust ed f or 501 cl ust er s i n di st i d)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
| Robust
l avgr exp | Coef . St d. Er r . t P| t | [ 95% Conf . I nt er val ]
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
l f ound | . 2447063 . 0417034 5. 87 0. 000 . 1627709 . 3266417
l f ndy96 | . 0053951 . 0254713 0. 21 0. 832 - . 044649 . 0554391
l f ndy97 | - . 0059551 . 0401705 - 0. 15 0. 882 - . 0848789 . 0729687
l f ndy98 | . 0045356 . 0510673 0. 09 0. 929 - . 0957972 . 1048685
l f ndy99 | . 0920788 . 0493854 1. 86 0. 063 - . 0049497 . 1891074
l f ndy00 | . 1364484 . 0490355 2. 78 0. 006 . 0401074 . 2327894
l f ndy01 | . 2364039 . 0555885 4. 25 0. 000 . 127188 . 3456198
. . .
_cons | . 1632959 . 0996687 1. 64 0. 102 - . 0325251 . 359117
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
. pr edi ct v2hat , r esi d
( 1503 mi ssi ng val ues gener at ed)
51
. gl mmat h4 l avgr exp v2hat l unch al unch l enr ol l al enr ol l y96- y01 l exppp94
l e94y96- l e94y01, f a( bi n) l i nk( pr obi t ) cl ust er ( di st i d)
not e: mat h4 has non- i nt eger val ues
Gener al i zed l i near model s No. of obs  3507
Opt i mi zat i on : ML Resi dual df  3487
Scal e par amet er  1
Devi ance  236. 0659249 ( 1/ df ) Devi ance  . 0676989
Pear son  223. 3709371 ( 1/ df ) Pear son  . 0640582
Var i ance f unct i on: V( u)  u*( 1- u/ 1) [ Bi nomi al ]
Li nk f unct i on : g( u)  i nvnor m( u) [ Pr obi t ]
( St d. Er r . adj ust ed f or 501 cl ust er s i n di st i d)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
| Robust
mat h4 | Coef . St d. Er r . z P| z| [ 95% Conf . I nt er val ]
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
l avgr exp | 1. 731039 . 6541194 2. 65 0. 008 . 4489886 3. 013089
v2hat | - 1. 378126 . 720843 - 1. 91 0. 056 - 2. 790952 . 0347007
l unch | - . 2980214 . 2125498 - 1. 40 0. 161 - . 7146114 . 1185686
al unch | - 1. 114775 . 2188037 - 5. 09 0. 000 - 1. 543623 - . 685928
l enr ol l | . 2856761 . 197511 1. 45 0. 148 - . 1014383 . 6727905
al enr ol l | - . 2909903 . 1988745 - 1. 46 0. 143 - . 6807771 . 0987966
. . .
_cons | - 2. 455592 . 7329693 - 3. 35 0. 001 - 3. 892185 - 1. 018998
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
52
. mar gef f
Aver age par t i al ef f ect s af t er gl m
y  Pr ( mat h4)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
var i abl e | Coef . St d. Er r . z P| z| [ 95% Conf . I nt er val ]
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
l avgr exp | . 5830163 . 2203345 2. 65 0. 008 . 1511686 1. 014864
v2hat | - . 4641533 . 242971 - 1. 91 0. 056 - . 9403678 . 0120611
l unch | - . 1003741 . 0716361 - 1. 40 0. 161 - . 2407782 . 04003
al unch | - . 3754579 . 0734083 - 5. 11 0. 000 - . 5193355 - . 2315803
l enr ol l | . 0962161 . 0665257 1. 45 0. 148 - . 0341719 . 2266041
al enr ol l | - . 0980059 . 0669786 - 1. 46 0. 143 - . 2292817 . 0332698
. . .
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
. * These st andar d er r or s do not account f or t he f i r st - st age est i mat i on.
. * Can use t he panel boot st r ap. Mi ght al so l ook f or par t i al ef f ect s at
. * di f f er ent par t s of t he spendi ng di st r i but i on.
53
Count and Other Multiplicative Models
∙ Conditional meanwithmultiplicativeheterogeneity:
Ey
it
|x
it
,c
i
  c
i
expx
it
þ (36)
wherec
i
≥ 0. Under strict exogeneityinthemean,
Ey
it
|x
i1
,...,x
iT
,c
i
  Ey
it
|x
it
,c
i
, (37)
the“fixedeffects” Poissonestimator isattractive: it doesnot restrict
Dy
it
|x
i
,c
i
, Dc
i
|x
i
, or serial dependence.
54
∙ TheFE Poissonestimator istheconditional MLE derivedunder a
Poissonandconditional independenceassumptions. It isoneof therare
careswheretreatingthec
i
asparameterstoestimategivesaconsistent
estimator of þ.
∙ TheFE Poissonestimator isfullyrobust toanydistributional failure
andserial correlation. y
it
doesnot evenhavetobeisnot acount
variable! Fullyrobust inferenceiseasy(xt pqml inStata).
55
∙ For endogeneitytherearecontrol functionandGMM approaches,
withtheformer beingmoreconvenient but imposingmorerestrictions.
∙ CF usessameapproachasbefore.
∙ Start withanomittedvariablesformulation:
Ey
it1
|y
it2
,z
i
,c
i1
,r
it1
  expx
it1
þ
1
 c
i1
 r
it1
. (38)
∙ Thez
it
 – includingtheexcludedinstruments– areassumedtobe
strictlyexogenoushere.
56
∙ If y
it2
is(roughly) continuouswemight specify
y
it2
 ç
2
 z
it
v
2
 z
̄
i
ç
2
 v
it2
.
∙ Alsowrite
c
i1
 ç
1
 z
̄
i
ç
1
 a
i1
sothat
Ey
it1
|y
it2
,z
i
,v
it1
  expç
1
 x
it1
þ
1
 z
̄
i
ç
1
 v
it1
,
wherev
it1
 a
i1
 r
it1
.
57
∙ Reasonable(but not completelygeneral) toassumev
it1
,v
i2
 is
independent of z
i
.
∙ If wespecifyEexpv
it1
|v
it2
  expp
1
 ¡
1
v
it2
 (aswouldbetrue
under joint normality), weobtaintheestimatingequation
Ey
it1
|y
it2
,z
i
,v
it2
  expk
1
 x
it1
þ
1
 z
̄
i
ç
1
 ¡
1
v
it2
. (39)
58
∙ Nowapplyasimpletwo-stepmethod. (1) Obtaintheresidualsv̂
it2
fromthepooledOLSestimationy
it2
on1, z
it
, z
̄
i
acrosst andi. (2) Usea
pooledQMLE (perhapsthePoissonor NegBinII) toestimatethe
exponential function, wherez
̄
i
,v̂
it2
 areexplanatoryvariablesalong
withx
it1
. (Asusual, afullyset of timeperioddummiesisagoodidea
inthefirst andsecondsteps).
∙ Notethat y
it2
isnot strictlyexogenousintheestimatingequation. and
soGLS-typemethodsaccount for serial correlationshouldnot beused.
GMM withcarefullyconstructedmomentscouldbe.
59
∙ EstimatingtheASF isstraightforward:
ASF
t
y
t2
,z
t1
  N
−1

i1
N
expk
̂
1
 x
t1
þ
̂
1
 z
̄
i
ç
̂
1
 ¡
̂
1

it2
;
that is, weaverageout z
̄
i
,v̂
it2
.
∙ Test thenull of contemporaneousexogeneityof y
it2
byusingafully
robust t statisticonv̂
it2
.
∙ Canallowmoreflexibilitybyiteractingz
̄
i
,v̂
it2
 withx
it1
, or evenjust
year dummies.
60
∙ A GMM approach– whichslightlyextendsWindmeijer (2002) –
modifiesthemoment conditionsunder asequential exogeneity
assumptiononinstrumentsandappliestomodelswithlagged
dependent variables.
∙ Writethemodel as
y
it
 c
i
expx
it
þr
it
Er
it
|z
it
,...,z
i1
,c
i
  1,
(40)
(41)
whichcontainsthecaseof sequentiallyexogenousregressorsasa
special case(z
it
 x
it
).
61
∙ Nowstart withthetransformation
y
it
expx
it
þ

y
i,t1
expx
i,t1
þ
 c
i
r
it
− r
i,t1
. (42)
∙ Caneasilyshowthat
Ec
i
r
it
− r
i,t1
|z
it
,...,z
i1
  0, t  1,...,T − 1.
62
∙ Usingthemoment conditions
E
y
it
expx
it
þ

y
i,t1
expx
i,t1
þ
|z
it
,...,z
i1
 0, t  1,...,T − 1 (43)
generallycausescomputational problems. For example, if x
itj
≥ 0for
somej andall i andt – for example, if x
itj
isatimedummy– thenthe
moment conditionscanbemadearbitarilyclosetozerobychoosing[
j
larger andlarger.
∙ Windmeijer (2002, EconomicsLetters) suggestedmultiplying
throughbyexpµ
x
þ whereµ
x
 T
−1

r1
T
Ex
ir
.
63
∙ So, themodifiedmoment conditionsare
E
y
it
expx
it
− µ
x
þ

y
i,t1
expx
i,t1
− µ
x
þ
|z
it
,...,z
i1
 0. (44)
∙ Asapractical matter, replaceµ
x
withtheoverall sampleaverage,
x
̄
 NT
−1

i1
N

r1
T
x
ir
. (45)
∙ Thedeviatedvariables, x
it
− x
̄
, will alwaystakeonpositiveand
negativevalues, andthisseemstosolvetheGMM computational
problem.
64