Time Invariant Regressor in Nonlinear Panel Model with Fixed Effects 1

Σχετικά έγγραφα
Estimators when the Correlation Coefficient. is Negative

Exam Statistics 6 th September 2017 Solution

CS 1675 Introduction to Machine Learning Lecture 7. Density estimation. Milos Hauskrecht 5329 Sennott Square

Discrete Fourier Transform { } ( ) sin( ) Discrete Sine Transformation. n, n= 0,1,2,, when the function is odd, f (x) = f ( x) L L L N N.

On Hypersurface of Special Finsler Spaces. Admitting Metric Like Tensor Field

Examples of Cost and Production Functions

Markov Processes and Applications

Optimal stopping under nonlinear expectation

Last Lecture. Biostatistics Statistical Inference Lecture 19 Likelihood Ratio Test. Example of Hypothesis Testing.

Lecture 17: Minimum Variance Unbiased (MVUB) Estimators

Multi-dimensional Central Limit Theorem

Minimum density power divergence estimator for diffusion processes

Article Multivariate Extended Gamma Distribution

SUPERPOSITION, MEASUREMENT, NORMALIZATION, EXPECTATION VALUES. Reading: QM course packet Ch 5 up to 5.6

Multi-dimensional Central Limit Theorem

Πανεπιστήµιο Κρήτης - Τµήµα Επιστήµης Υπολογιστών. ΗΥ-570: Στατιστική Επεξεργασία Σήµατος. ιδάσκων : Α. Μουχτάρης. εύτερη Σειρά Ασκήσεων.

Other Test Constructions: Likelihood Ratio & Bayes Tests

Every set of first-order formulas is equivalent to an independent set

6.3 Forecasting ARMA processes

Homework for 1/27 Due 2/5

Solutions: Homework 3

ST5224: Advanced Statistical Theory II

Econ 2110: Fall 2008 Suggested Solutions to Problem Set 8 questions or comments to Dan Fetter 1

A study on generalized absolute summability factors for a triangular matrix

1. For each of the following power series, find the interval of convergence and the radius of convergence:

Solve the difference equation

C.S. 430 Assignment 6, Sample Solutions

HOMEWORK 4 = G. In order to plot the stress versus the stretch we define a normalized stretch:

α & β spatial orbitals in

Statistical Inference I Locally most powerful tests

Matrices and Determinants

One and two particle density matrices for single determinant HF wavefunctions. (1) = φ 2. )β(1) ( ) ) + β(1)β * β. (1)ρ RHF

Markov Processes and Applications

ΝΕΑ ΑΠΟΤΕΛΕΣΜΑΤΑ ΣΧΕΤΙΚΑ ΜΕ ΤΗΝ ΥΠΑΡΞΗ ΕΚΤΙΜΗΤΩΝ ΜΕΓΙΣΤΗΣ ΠΙΘΑΝΟΦΑΝΕΙΑΣ ΓΙΑ ΤΗΝ 3-ΠΑΡΑΜΕΤΡΙΚΗ ΓΑΜΜΑ ΚΑΤΑΝΟΜΗ

The Probabilistic Method - Probabilistic Techniques. Lecture 7: The Janson Inequality

Ordinal Arithmetic: Addition, Multiplication, Exponentiation and Limit

4.6 Autoregressive Moving Average Model ARMA(1,1)

Cytotoxicity of ionic liquids and precursor compounds towards human cell line HeLa

ORDINAL ARITHMETIC JULIAN J. SCHLÖDER

Introduction of Numerical Analysis #03 TAGAMI, Daisuke (IMI, Kyushu University)

derivation of the Laplacian from rectangular to spherical coordinates

3.4 SUM AND DIFFERENCE FORMULAS. NOTE: cos(α+β) cos α + cos β cos(α-β) cos α -cos β

Reminders: linear functions

Math221: HW# 1 solutions

Example Sheet 3 Solutions

Choosing the Number of Moments in Conditional Moment Restriction Models

Diane Hu LDA for Audio Music April 12, 2010

On Inclusion Relation of Absolute Summability

2 Composition. Invertible Mappings

Uniform Convergence of Fourier Series Michael Taylor

Solution Series 9. i=1 x i and i=1 x i.

Prescribing Morse scalar curvatures: subcritical blowing-up solutions

8.1 The Nature of Heteroskedasticity 8.2 Using the Least Squares Estimator 8.3 The Generalized Least Squares Estimator 8.

Fractional Colorings and Zykov Products of graphs

n r f ( n-r ) () x g () r () x (1.1) = Σ g() x = Σ n f < -n+ r> g () r -n + r dx r dx n + ( -n,m) dx -n n+1 1 -n -1 + ( -n,n+1)

The Simply Typed Lambda Calculus

Ψηφιακή Επεξεργασία Εικόνας

Math 446 Homework 3 Solutions. (1). (i): Reverse triangle inequality for metrics: Let (X, d) be a metric space and let x, y, z X.

Homework 3 Solutions

Problem Set 3: Solutions

Nowhere-zero flows Let be a digraph, Abelian group. A Γ-circulation in is a mapping : such that, where, and : tail in X, head in

Studies on Properties and Estimation Problems for Modified Extension of Exponential Distribution

Solutions to Exercise Sheet 5

Concrete Mathematics Exercises from 30 September 2016

Jesse Maassen and Mark Lundstrom Purdue University November 25, 2013

= λ 1 1 e. = λ 1 =12. has the properties e 1. e 3,V(Y

EE512: Error Control Coding

The Equivalence Theorem in Optimal Design

ω ω ω ω ω ω+2 ω ω+2 + ω ω ω ω+2 + ω ω+1 ω ω+2 2 ω ω ω ω ω ω ω ω+1 ω ω2 ω ω2 + ω ω ω2 + ω ω ω ω2 + ω ω+1 ω ω2 + ω ω+1 + ω ω ω ω2 + ω

Chapter 6: Systems of Linear Differential. be continuous functions on the interval

6.1. Dirac Equation. Hamiltonian. Dirac Eq.

Presentation of complex number in Cartesian and polar coordinate system

5.4 The Poisson Distribution.

Theorem 8 Let φ be the most powerful size α test of H

t. Neymann-Fisher factorization: where g and h are non-negative function. = is a sufficient statistic. : t( y)

Higher Order Properties of Bootstrap and Jackknife Bias Corrected Maximum Likelihood Estimators

Example of the Baum-Welch Algorithm

MATH423 String Theory Solutions 4. = 0 τ = f(s). (1) dτ ds = dxµ dτ f (s) (2) dτ 2 [f (s)] 2 + dxµ. dτ f (s) (3)

Econ Spring 2004 Instructor: Prof. Kiefer Solution to Problem set # 5. γ (0)

Μηχανική Μάθηση Hypothesis Testing

Durbin-Levinson recursive method

8.324 Relativistic Quantum Field Theory II

( y) Partial Differential Equations

1 Complete Set of Grassmann States

CHAPTER 25 SOLVING EQUATIONS BY ITERATIVE METHODS

Lecture 2. Soundness and completeness of propositional logic

IIT JEE (2013) (Trigonomtery 1) Solutions

Numerical Analysis FMN011

2. Let H 1 and H 2 be Hilbert spaces and let T : H 1 H 2 be a bounded linear operator. Prove that [T (H 1 )] = N (T ). (6p)

Congruence Classes of Invertible Matrices of Order 3 over F 2

Parametrized Surfaces

SCITECH Volume 13, Issue 2 RESEARCH ORGANISATION Published online: March 29, 2018

ECE Spring Prof. David R. Jackson ECE Dept. Notes 2

Suppose Mr. Bump observes the selling price and sales volume of milk gallons for 10 randomly selected weeks as follows

On Generating Relations of Some Triple. Hypergeometric Functions

Problem Set 9 Solutions. θ + 1. θ 2 + cotθ ( ) sinθ e iφ is an eigenfunction of the ˆ L 2 operator. / θ 2. φ 2. sin 2 θ φ 2. ( ) = e iφ. = e iφ cosθ.

Notes on the Open Economy

Phys460.nb Solution for the t-dependent Schrodinger s equation How did we find the solution? (not required)

Local Approximation with Kernels

6. MAXIMUM LIKELIHOOD ESTIMATION

Transcript:

me Ivarat Regressor Nolear Pael Model wt Fxed ffects Jyog Ha UCLA February 26, 23 I am grateful to Da Ackerberg ad Jerry Hausma for elpful commets.

Abstract s paper geeralzes Hausma ad aylor s (98) tuto, ad develop a metod of dealg wt tme-varat regressor olear pael model wt Þxed effects. e metod requres large umber of observatos per dvdual ( ) as well as a large umber of dvduals ().

Itroducto Pael data allow te possblty of cotrollg for uobserved dvdual specþc effects, wc may be correlated wt observed explaatory varables. I lear models, suc Þxed effects are usually elmated by dfferecg, wc yelds a model free of suc cdetal parameter. A uteded cosequece of dfferecg s tat t also elmates tme-varat regressor, wc reders te coeffcet of te tmevarat regressor udetþed. Hausma ad aylor (98) used a strumetal varables approac to overcome suc problem. I ts paper, I geeralze ter tuto, ad develop a metod of dealg wt tme-varat regressor te olear framework. s metod requres large umber of observatos per dvdual ( ), so ts applcablty s lmted to te case were s large. Because te strumetal varables estmato requres a large umber of dvduals (), I adopt a asymptotc framework were bot ad grow to Þty at te same rate. s result s made possble by recet teccal progress of pael aalyss uder suc alteratve asymptotcs. See, e.g., Arellao (2), Ha ad Kuersteer (22, 23), Ha ad Newey (22), ad Wouterse (22). 2 Prelmares Suppose tat we are gve a set of momet restrctos [g (y t, γ wδ x tθ )] =, =,...,; t =,..., () for some vector-valued fucto g, were y t, x t,adw deote te depedet varable te tt perod, tme-varyg regressor te tt perod, ad tme-varat regressor. Uobserved dvdual specþc effects are summarzed by te scalar varable γ. Our prmary focus s to estmate te coeffcet δ of te tme-varat regressor w we γ s possbly correlated wt w ad x t. We ow dscuss ow te parameters ca be cosstetly estmated. If bot ad grow to Þty at te same rate, θ ca be -cosstetly estmated. Lettg α γ wδ we ca rewrte te model as [g (y t, α x tθ )] =, to wc we ca apply varats of recetly developed metods dscussed Arellao (2), Ha ad Kuersteer (22, 23), Ha ad Newey (22), ad Wouterse (22). erefore, te coceptual callege s to develop a cosstet estmator of δ. For ts purpose, we assume tat te data are..d. over : Codto ({y,y 2,...},{x,x 2,...},z,w, γ ) s..d. over. Suppose for a momet tat we observe α. Also suppose tat we observe a addtoal varable z wt dm (z )=dm(w ) adsuctat It s easy to geeralze te dscusso to te over-detþed case were dm (z ) > dm (w ). Because te prmary purpose of ts paper s detþcato ad cosstet estmato of δ, I focus o te exactly detþed case.

Codto 2 () [z γ ]=;()[z w ] s osgular It s clear tat we ca estmate δ by e δ ( P z w) ( P z α ). It s easy to see tat ts estmator s cosstet for δ as. Hausma ad aylor s (98) tuto was tat e δ would rema cosstet eve f we replace α by a ubased estmate. I te olear cotext, t seems dffcult to come up wt suc a ubased estmator for α. erefore, Hausma ad aylor s (98) metod caot be drectly appled. e basc tuto ts paper s tat, we bot ad grow to Þty at te same rate, we cacomeupwta -cosstet estmator for α,saybα. Because te estmato error becomes very smaller as te sample sze creases, te IV estmator b δ z w z bα wll be cosstet for δ geeral. 3 Cosstet stmato of α Let u (t ; θ, α ) g ( t ; θ, α ) v ( t ; θ, α ) were t (y,...,y,x,...,x ),adweredm (θ) =dm(u) =p ad dm (α) =dm(v) =. We assume tat [g ( t ; θ, α )] =, ad cosder te estmator tat solves = u ³ t ; b θ, bα = t= v ³ t ; bθ, bα t= s dcates tat we are separatg g to two compoets. e Þrst compoet u s used trougout te sample for estmato of θ. e secod compoet v s used oly for te t dvdual to estmate α. s separato was adopted because I wated to explot some recet teccal developmet Ha ad Kuersteer (23) or Ha ad Newey (22). I do ot expect ts separato to be costrag practce. For example, te case of lear models y t = α x tθ ε t (2) wt x t strctly exogeous, we may take u ( t ; θ, α ) = x t (y t α x tθ) v ( t ; θ, α ) = y t α x tθ 2

were y P t= y t ad x P t= x t. Note tat ts wll result te usual Þxed effects estmator of θ. We assume followg regularty codtos: Codto 3 () Gve tme-varat varables (α,z,w ), (y t,x t ) s..d. over t; () for every, G () (θ, α )=; () for eac η >, f f {(θ,α): (θ,α) (θ,α ) >η} G () (θ, α) >, were bg () (θ, α ) P t= g ( t; θ, α ), G () (θ, α ) [g ( t ; θ, α )]. Codto 4, suc tat ρ, were < ρ <. Codto 5 () e fucto g ( ; θ, α) s cotuous (θ, α) Υ; () e parameter space Υ s compact; () ere exsts a fucto M ( t ) suc tat mm2 g ( t ; θ, α ) m α m2 M ( t) m m 2,...,6 ad sup M ( t ) Q < for some Q>64. Codto 6 () m v ( t ; θ, α ) 2 > ; () lm P I >, were U ( t ; θ, α ) ³ u ( t ; θ, α ) ρ v ( t ; θ, α ), ρ v(t;θ,α ) α u(t;θ,α ) α,adi U(t;θ,α ). We ca sow tat bα are uformly cosstet for α. eorem Uder Codtos 3, 4, 5, ad 6, we ave (bα α )=λ κ, were µ λ vt /2 α t= v t = O p () ad Pr [ κ ]=o (). Here, v t v t ( t, θ, α ). Proof. See Appedx C. 4 IV stmato of δ Recall tat, f α s kow, we could ave estmate δ by e δ z w z α. (3) Because t s ot kow, we ca replace α by bα, ad cosder b δ z w z bα. (4) 3

eorem mples tat we ave ³ bδ δ = z w /2 z γ /2 /2 z λ o p () Note tat, f z λ = P t= z v t does ot ave a zero expectato, t would complcate asymptotc aalyss. We terefore assume tat te strumet z satsþes [z v t ]=. Codto 7 [v t z ]= Codto 7 guaratees tat te error estmato of α does ot complcate te ferece regardg bδ. Suc codto was adopted by Hausma ad aylor (98). We ave v t = ε t te lear model (2), ad Hausma ad aylor s strumet z for suc lear model s requred to satsfy [z ε t ]=. It turs out tat te substtuto of bα for α (3) does ot affect te asymptotc dstrbuto of te resultat estmator (4). It s because /2 /2 P z λ = o p () uder Codtos 3, ad 7. We terefore obta ³ bδ δ = z w /2 z γ o p () Note tat ³ eδ δ = z w /2 z γ o p () as well. s s a tutve result. Note tat b δ s a IV estmator of bα o w.becausebα s a proxy for α, ad because te measuremet error dsappears as, we sould expect tat te asymptotc dstrbuto of ³ bδ δ sould be detcal to tat of ³ eδ δ uder te large asymptotcs. As a cosequece, we obta eorem 2 Assume Codtos, 2, 3, 4, 5, 6, ad 7. Furter assume tat [ z w ] < ad γ 2 z z <. Weteave ³ ³ bδ δ N, ( [z w ]) γ 2 z z ]) ( [w z. 5 Summary ad Future Work I ts paper, I geeralzed Hausma ad aylor s (98) result to olear pael models wt Þxed effects. e result s expected to ave a mplcato for some olear models of socal teractos. ssbecausetemodel() could be uderstood as a model wt some partcular kds of socal effects. Followg te termology troduced by Mask (993), let y t deote te outcome of te tt dvdual of te t group. Also, let x t deote te dvdual level caracterstcs tat operate at te dvdual level oly. If w s equal to te group average [y t ] of y t, te te model cotas edogeous socal effects. If w s equal to te group average [x t ] of x t, te te model cotas exogeous (cotextual) socal effects. Fally γ, wc s ot observed by te ecoometrca, captures te presece of correlated 4

group effects. erefore, te result ts paper may be applcable to some partcular class of models of socal teractos. 2 2 IdetÞcato of δ wt varous kds of socal effects s expected to requre some more fuctoal restrcto, suc as te excluso restrcto cosdered by Graam ad Ha (23), or selecto effects as cosdered by Brock ad Durlauf (2). Nolear models of socal teracto are expected to be plagued by multple equlbra geeral. erefore, t s mportat tat te model () s ot uderstood as te Þrst order codto of te log lkelood. After all, textbook style lkelood may ot ext. e model () sould be uderstood as a set of momet restrctos tat are vald regardless of te potetal multplcty of equlbra.wt multple equlbra, te..d. assumpto Codto may soud puzzlg. Suc puzzle would dsappear f te equlbrum selecto s cosdered as a Þxed effect, altoug t does ot appear te model (). 5

Appedx A Cosstecy Lemma Assume tat W t are d wt [W t ]=ad Wt 2k <. e, ³P 2k t= W t = C(k) k o( k ) for some costat C(k). Proof. By adoptg a argumet te proof of Lemma 5. Lar (992), we ave ³P 2k t= W P t = 2k P C (α,..., α j ) P j= α I jq s= W α s t s, (5) were for eac Þxed j {,..., 2k}, P α exteds over all j-tuples of postve tegers (α,..., α j ) suc tat α... α j =2k ad P I exteds over all ordered j-tuples (t,..., t j ) of tegers suc tat t j. Also, C(α,..., α j ) stads for a bouded costat. Note, tat f j>kte at least oe of te dces Qj α j =. By depedece ad te fact tat [W t ]=tfollows tat s= W t αs s =weever j>k. ³P s sows tat t= W t 2k C (k) k Wt 2k for some costat C(k). Lemma 2 Suppose tat, for eac, {ξ t,t=, 2,...} s a sequece of zero mea..d. radom varables. We assume tat {ξ t,t=, 2, 3} are depedet across. We also assume tat ξ t 6 <. Fally, we assume tat = O ( ). Weteave Pr for every η >. ξ > η = o 2 t t= Proof. Usg Lemma, weobta 6 ξ C t 8 ξ 6 t, t= were C>s a costat. erefore, we ave 2 Pr ξ > η t 2 C8 6 η 6 ξ 6 t or t= 2 Pr ξ > η C t 6 η t= 6 ξ 6 t = o (). 6

Lemma 3 Suppose tat Codtos 4 ad 5 old. We te ave for all η > tat Pr sup bg () (θ, α) G () (θ, α) η = o (θ,α) Proof. Let η > be gve. We ote tat Pr sup bg () (θ, α) G () (θ, α) η (θ,α) Pr sup bg () (θ, α) G () (θ, α) η. (6) Let ε > be cose suc tat 2ε [M ( t )] < η 3. Dvde Υ to subsets Υ, Υ 2,...,Υ M(ε) suc tat (θ, α) θ, α < ε weever (θ, α) ad θ, α are te same subset. Let (θ j, α j ) deote some pot Υ j for eac j. e, sup bg () (θ, α) G () (θ, α) =sup bg () (θ, α) G () (θ, α), (θ,α) j Υ j ad terefore Pr sup bg M(ε) () (θ, α) G () (θ, α) η Pr sup bg () (θ, α) G () (θ, α) η (7) (θ,α) j= Υ j For (θ, α) Υ j,weave bg () (θ, α) G () (θ, α) bg () (θ j, α j ) G () (θ j, α j ) ε (M ( t ) [M ( t )]) 2ε [M ( t)] (θ,α) t= e, Pr sup bg () (θ, α) G () (θ, α) η Υ j Pr bg () (θ j, α j ) G () (θ j, α j ) η 3 = o 2 Pr (M ( t ) [M ( t )]) η 3ε t= (8) by Lemma 2. Combg (6), (7), (8), ad = O ( ), we obta te desred cocluso. eorem 3 Uder Codtos 4, 5, ad 3, Pr b θ θ η = o for every η >. ³ ³ Proof. It s useful to ote bθ, bα,...,bα solves bg () bθ, bα =for every. I oter words, ³ bθ, bα,...,bα solves m P bg () (θ, α ). Letη be gve, ad let ε f f (θ,α) (θ,α ) >η G() (θ, α). Note tat ε >. Wt probablty equal to o,weave m bg () (θ, α ) m θ θ >η,α,...,α (θ,α) (θ bg () (θ, α ),α ) >η > m (θ,α) (θ,α ) >η > 2 ε > = 7 bg () ³ bθ, bα, G() (θ, α ) 2 ε

were te secod ad fourt ³ equaltes are based o Lemma 3, ad te last equalty follows from te deþto of te estmators bθ, bα,...,bα. We ca terefore coclude tat Pr b θ θ η = o. eorem 4 Uder Codtos 4, 5, ad 3, Pr [ bα α η] =o for every η >. Proof. We Þrst prove tat Pr bg ³ () bθ, α G () (θ, α) η = o () (9) sup α for every η >. Notetat sup bg ³ () bθ, α G () (θ, α) α sup bg ³ ³ ³ () bθ, α G () bθ, α sup G () bθ, α G () (θ, α) α α sup bg () (θ, α) G () (θ, α) [M ( t)] bθ θ. erefore, (θ,α) Pr Pr sup α sup (θ,α) Pr b θ θ = o () bg ³ () bθ, α G () (θ, α) η bg () (θ, α) G () (θ, α) η 2 η 2( [M ( t )]) by Lemma 3 ad eorem 3. We ow get back to te proof of eorem 4. It suffces to prove tat Pr bα α η = o () for every η >. Let η be gve, ³ ad let ε f f {α: α α >η} G () (θ, α ) >. Codto o te evet sup α bg () bθ, α G () (θ, α) 3 o, ε wc as a probablty equal to o by (9). We te ave m α α >η bg ³ () bθ, α > m α α >η ad terefore, bα α η for every. G () (θ, α ) 2 ε > ³ 2 ε > = G() b bθ, bα, 8

B xpaso Lettg U ( t ; θ, α ) u ( t ; θ, α ) ρ v ( t ; θ, α ) µ u (t ; θ, α ) v (t ; θ, α ) ρ α v ( t ; θ, bα (θ)) t= V ( t ; θ, α ) v ( t ; θ, α ) we ca recogze tat b θ s a soluto to = U t= ³ ³ t ; bθ, bα b θ. ³ Let F (F,...,F ) deote te collecto of dstrbuto fuctos. Let bf F b,..., bf, were bf deotes te emprcal dstrbuto fucto for te stratum. DeÞe F (²) F ² ³ F b F for ², /2.ForeacÞxed θ ad ², letα (θ,f (²)) be te soluto to te estmatg equato = V [θ, α (θ,f (²))] df (²), α ad let θ (F (²)) be te soluto to te estmatg equato = U ( t ; θ (F (²)), α (θ (F (²)),F (²))) df (²). By aylor seres expaso, we ave ³ θ F b θ (F )= θ ² () µ 2 θ ²² () µ 3 θ ²²² (e²), 2 6 were θ ² (²) dθ (F (²))/ d², θ ²² (²) d 2 θ (F (²)) ± d² 2,...,ade² s somewere betwee ad /2. We terefore ave ³ ³ θ bf θ (F ) = θ ² () µ 2 θ ²² () r θ ²²² (e²). () 2 6 e last term () ca be sow to be o p () bytesamemetodashaadnewey(22). Let (,²) U ( ; θ (F (²)), α (θ (F (²)),F (²))) () e Þrst order codto may be wrtte as = (,²) df (²) (2) 9

Dfferetatg repeatedly wt respect to ², weobta = d (,²) df (²) (,²) d (3) d² = d 2 (,²) d² 2 df (²)2 d (,²) d (4) d² = d 3 (,²) d² 3 df (²)3 d 2 (,²) d² 2 d (5) were ³ F b F. B. θ ² () Because d (,²) d² = (,²) (,²) α α α we may rewrte (3) as = µ (,²) (,²) α valuatg at ² =, ad otg tat U =, we obta α θ ² () = We terefore ave θ ² () = I U d (,²) α α (,²) α I U t= N, lm I lm α df (²) Φ lm (,²) d (6) I (7) B.2 α θ ad α ² I te t stratum, α (θ,f (²)) solves te estmatg equato V ( ; θ, α (θ,f (²))) df (²) = (8) Dfferetatg te LHS wt respect to θ ad ², weobta µ V (, θ,²) V (, θ,²) α (θ,f (²)) = df (²) α df (²), µ V (, θ,²) α (θ,f (²)) = df (²) V (, θ,²) d. α

Observe tat α (θ,f (²)) = α (θ,f (²)) = µ V (, θ,²) α µ V (, θ,²) α µ V (, θ,²) df (²) µ df (²) df (²), V (, θ,²) d. quatg tese equatos to zero ad solvg for dervatves of α evaluated at ² =gves µ α θ V V () = α = O (), (9) µ α ² V () = V t = O p (), (2) were α t= α θ α (θ,f ()), α ² α (θ,f ()). B.3 θ ²² () Because d (,²) d² = k= (,²) l= k= (,²) α l α l l= (,²) α l α l we ave d 2 (,²) d² 2 = k= k = l= k= l= k= k = l= l = k= 2 (,²) 2 (,²) α l α l k= l= k= k = l= k= l= l = l= k= 2 (,²) α l 2 (,²) α l α l (,²) α l α l α l 2 α l l= k= k = 2 (,²) α l (,²) 2 θ k 2 α l (,²) α l 2 θ k α l 2 k= 2 (,²) α l α l α l k = (,²) 2 α l α l l= l= l = k= k = l= l= 2 (,²) α l α l l= l = k= (,²) 2 α l α l 2 (,²) α l α l α l 2 (,²) α l α l 2 α l 2 (,²) α l α l α l α l α l α l

valuatg (4) at ² =, ad otg tat [U α ]=,weobta = 2 U k 2 k= k = l= k= k = 2 2 U k α l k α l U l= k= k k= 2 U αl α l α l α l l= l = k= k = 2 2 U αl α l α l α l l= l = k= l= 2 µ U d θ k 2 µ U d k α k= l l= k= 2 µ U αl d α l l= Because 2 U αl α l α l α l l= l = α = 2 U () α α α. α 2 U (p) α α α ³ P ³ t= = V t V α ³ P ³ t= V t V α 2 U () α α. 2 U (p) α α ³ ³ V α V α 2 U α l 2 θ k 2 l = ³ ³ k α l 2 U αl α l α l α l αl P t= V t P t= V t 2 l= µ U αl d α l = 2 t= µ U α V α t= V t 2 k= 2 2 k= l= l = k= l= 2 U α l 2 U α l α l 2 k= l= αl αl α l µ U d µ U α l d αl µ µ µ = O p O p = O p µ µ µ = O p O p = O p µ µ µ θ k = O p O p = O p µ µ µ = O p O p = O p 2

ad 2 l= k= k = k= k = l= l = k= k = 2 U 2 U αl k α l 2 U αl α l α l α l we may wrte ³ P ³ t= I θ ²² () = V t V α ³ P ³ t= V t V α µ U 2 V α α t= µ µ µ = O () O p O p = O p µ µ µ = O () O p O p = O p µ µ µ = O () O p O p = O p 2 U () α α. 2 U (p) α α t= V t ³ ³ V α V α ³ ³ P t= V t P t= V t or Note tat ad = 2 O p µ µ 2 θ ²² () 2 r r = I I t= ( µ trace t= V t µ V α µ U () α ³ P ³ t= V ³ V α t P ³ t= V t t= 2 U () α l α l V α µ U α 2 U () α l α l V α V α V α µ V t= 2 U () α α. 2 U (p) α α t= ³ ³ µ V α V t = α ³ V α P t= V t ³ V P α t= V t V t o p () t= [V V ] ) o p () ( µ V trace α V t ) U V () α o p () It terefore follows tat 2 µ 2 θ ²² () = 2 r I Υ o p () 3

B.4 α θθ, α θ², ad α ²² ³ Secod order dfferetato = α 2 2, 2 2, 2 2 of (8) yelds 2 V (, θ,²) df (²) α µ (θ,f (²)) 2 V (, θ,²) α df (²) µ 2 µ V (, θ,²) α (θ,f (²)) V (, θ,²) 2 α (θ,f (²)) df (²) α df (²) α µ 2 V (, θ,²) α (θ,f (²)) α (θ,f (²)) df (²), ad = = µ 2 µ V (, θ,²) α (θ,f (²)) V (, θ,²) 2 α (θ,f (²)) df (²) df (²) α α µ 2 V (, θ,²) α (θ,f (²)) α (θ,f (²)) α 2 df (²) µ V (, θ,²) V (, θ,²) α (θ,f (²)) d d, α µ V (, θ,²) 2 µ α (θ,f (²)) 2 µ V (, θ,²) α (θ,f (²)) df (²) α 2 α 2 df (²) µ V (, θ,²) α (θ,f (²)) 2 d. α 2 ese tree equaltes caracterzes 2 α (θ,f (²)) Lemma 4 Pr b θ (²) θ η = o () ², 2 α (θ,f (²)),ad 2 α (θ,f (²)). 2 ad Pr ² bα (²) α η = o () for every η >. Proof. Oly te Þrst asserto s proved. e secod asserto ca be proved smlarly. Let η be gve, ad let ε f f {(θ,α): (θ,α) (θ,α ) >η} G() (θ, α) >. Recall tat F (²) F ² ³ bf F, ², We ave ³ g ( ; θ, α (θ)) df (²) = ² G () (θ, α )² G b () (θ, α ) ad ³ g ( ; θ, α (θ)) df (²) G () (θ, α ) ² bg () (θ, α) G () (θ, α) bg () (θ, α) G () (θ, α). 4

By Lemma 3, we ave Pr ² sup (θ,α) g ( ; θ, α (θ)) df (²) G () (θ, α) η = o erefore, for every ² wt probablty equal to o,weave m θ θ >η,α,...,α g ( ; θ, α (θ)) df (²) m (θ,α) (θ,α ) >η g ( ; θ, α (θ)) df (²) > m (θ,α) (θ,α ) >η > 2 ε > = We terefore obta tat Pr ² b θ (²) θ η = o. G () (θ, α ) 2 ε g ³ ; b ³ θ (²), α bθ (²) df (²) Lemma 5 Suppose tat, for eac, {ξ t (φ),t=, 2,...} sasequeceofzeromea..d. radom varables dexed by some parameter φ Φ. We assume tat {ξ t (φ),t=, 2, 3} are depedet across. We also assume tat sup φ Φ ξ t (φ) B t for some sequece of radom varables B t tat s..d. across t ad depedet across. Fally, we assume tat B t 64 <, ad = O ( ). We te ave µ Pr ξ t (φ ) > υ = o t= for every υ suc tat υ < 6. Here, {φ } s a arbtrary sequece Φ. Proof. By Markov s equalty, Pr sup ξ φ Φ t (φ ) > υ =Pr sup ξ t= φ Φ t (φ ) > 3 5 υ t= P sup φ Φ t= ξ t (φ ) 64 = 3 5 64 64υ η 64 were te last equalty s based o domated covergece. By Lemma, weave 64 ξ t (φ ) C 32 ξ t (φ ) 64 32 C B t 64 t= for some C. erefore, we ave 2 Pr ξ t (φ ) > υ t= 2 C 32 ³ 64υ 92 5 64υ η = O 32 5, 64 P sup φ Φ t= ξ t (φ ) 64 92 5 64υ η 64, 5

ad Pr t= ξ t (φ ) > υ Pr t= ξ t (φ ) > υ = C 32 92 5 64υ η 64 = o (). Lemma 6 Suppose tat K ( ; θ (²), α (θ (²),²)) s equal to mm2 g ( t ; θ (²), α (θ (²),²)) m α m2 for some m m 2,...,5. e, for ay η >, weave Pr K ( ; θ (²), α (θ (²),²)) df (²) ² [K ( t ; θ, α )] > η = o ad Pr Also, Pr ² ² K ( ; θ (²), α (θ (²),²)) df (²) [K ( t ; θ, α )] > η = o. K ( ; θ (²), α (θ (²),²)) d >C υ = o for some costat C> ad υ < 6. Proof. Note tat we may wrte K ( ; θ (²), α (θ (²),²)) df (²) K ( ; θ (²), α (θ (²),²)) df K ( ; θ (²), α (θ (²),²)) df (²) K ( ; θ (), α (θ (), )) df (²) K ( ; θ (), α (θ (), )) df (²) K ( ; θ (), α (θ (), )) df M( t )(kθ(²) θk α (θ(²),²) α ) d F (²) ² ³ K ( ; θ (), α (θ (), )) d F b F. erefore, we ave K ( ; θ (²), α (θ (²),²)) df (²) [K ( t ; θ, α )] kθ (²) θk [M ( t )] /2 (α (θ (²),²) α ) 2 t= M ( t ) t= [M ( t )] 2 M ( t ) K ( t ; θ (), α (θ (), )) [K ( t ; θ (), α (θ (), ))], 6 t= /2

te RHS of wc ca be bouded by usg Lemmas?? ad 4 absolute value by some η > wt probablty o. Because K ( ; θ (²), α (θ (²),²)) df (²) [K ( t ; θ, α )] θ (²) θ [M ( t )] M ( t ) t= α (θ (²),²) α [M ( t )] M ( t ) t= M ( t ) [M ( t )], we ca boud ² t= K ( ; θ (²), α (θ (²),²)) df (²) [K ( t ; θ, α )] absolute value by some η > wt probablty o. Usg Lemma 5, we ca also sow tat K ( ; θ (²), α (θ (²),²)) d ca be bouded by absolute value by C υ for some costat C> ad υ suc tat υ < 6 wt probablty o. Lemma 7 Pr α θ ² (²) >C Pr α ² (²) >C υ ² = o = o for some costat C> ad υ < 6. Proof. From Appedx B.2, we obta µ µ α (θ,f (²)) V (, θ,²) V (, θ,²) = df (²) df (²), α µ µ α (θ,f (²)) V (, θ,²) = df (²) V (, θ,²) d. α ³ R Usg Lemma 6, we ca see tat V (,θ,²) α df (²) s uformly bouded away from zero wt probablty o. We ca also see tat, wt probablty o, R V (,θ,²) df (²) s uformly bouded by some costat C, ad R V (, θ,²) d s uformly bouded by C υ. 7

Lemma 8 Pr θ ² (²) >C υ = o ² for some costat C> ad υ < 6. Proof. From (6), we ave θ ² µ (,²) (²) = (,²) α α df (²) µ α (,²) df (²) (,²) d α Usg Lemmas 6, ad 7, we ca boud te deomator of θ ² (²) by some C>, ad te umerator by some C υ wt probablty o. Lemma 9 θ Pr α r θ r ² (²) >C θ Pr α r ² ² (²) >C υ Pr (²) >C ³ υ 2 α ²² ² = o = o = o for some costat C> ad υ < 6. Here, αθ rθ r 2 α r r. We smlarly deþe αθ r². Proof. Note tat 2 V (, θ,²) = df (²) α µ (θ,f (²)) 2 V (, θ,²) α df (²) µ 2 µ V (, θ,²) α (θ,f (²)) V (, θ,²) 2 α (θ,f (²)) df (²) α df (²) α µ 2 V (, θ,²) α (θ,f (²)) α (θ,f (²)) df (²), = α 2 µ 2 V (, θ,²) α (θ,f (²)) df (²) α µ V (, θ,²) 2 α (θ,f (²)) df (²) α µ V (, θ,²) V (, θ,²) d α µ 2 V (, θ,²) α 2 α (θ,f (²)) d α (θ,f (²)) df (²) ad µ V (, θ,²) 2 µ α (θ,f (²)) 2 µ V (, θ,²) α (θ,f (²)) = df (²) α 2 α 2 df (²) µ V (, θ,²) α (θ,f (²)) 2 d. α e result te follows by applyg te same argumet as te proof of Lemma 7. 8, α (θ,f (²)) 2

Lemma Pr ² θ ²² (²) >C ³ υ 2 = o for some costat C> ad υ < 6. Proof. e cocluso follows by usg te caracterzato of θ ²² (²) Appedx B.3, ad Lemmas 6, 7, 8, ad 9. Lemma θ Pr α r θ r θ r ² θ Pr α rθ r ² (²) Pr Pr ² θ α r²² (²) >C ² α ²²² ² (²) >C ³ (²) >C >C υ ³ υ 2 υ 3 = o = o = o = o for some costat C> ad υ < 6. Proof. It was see Appedx B.4 tat ad = = µ 2 V (, θ,²) α (θ,f (²)) df (²) r α µ V (, θ,²) 2 µ α (θ,f (²)) 2 V (, θ,²) α (θ,f (²)) α (θ,f (²)) df (²) α r α 2 df (²) r µ V (, θ,²) V (, θ,²) α (θ,f (²)) d d, r α r µ V (, θ,²) 2 µ α (θ,f (²)) 2 µ V (, θ,²) α (θ,f (²)) df (²) α 2 α 2 df (²) µ V (, θ,²) α (θ,f (²)) 2 d. α 2 9

We terefore obta 3 µ V (, θ,²) 3 V (, θ,²) α (θ,f (²)) = df (²) df (²) r r r α r r r µ 2 α (θ,f (²)) 2 V (, θ,²) df (²) α µ (θ,f (²)) 3 V (, θ,²) df (²) r r α r r α r r α µ (θ,f (²)) α (θ,f (²)) 3 V (, θ,²) r r α 2 df (²) r µ 3 µ V (, θ,²) α (θ,f (²)) 3 V (, θ,²) α (θ,f (²)) df (²) α r r r α 2 df (²) r r µ 2 V (, θ,²) 2 α (θ,f (²)) df (²) α r r r µ 2 V (, θ,²) 2 µ α (θ,f (²)) 2 V (, θ,²) α (θ,f (²)) df (²) α r r r α 2 df (²) r µ V (, θ,²) 3 α (θ,f (²)) df (²) α r r r µ 3 V (, θ,²) α (θ,f (²)) α (θ,f (²)) α 2 df (²) r r r µ 3 V (, θ,²) α (θ,f (²)) α (θ,f (²)) α (θ,f (²)) α 3 df (²) r r r µ 2 V (, θ,²) 2 α (θ,f (²)) α (θ,f (²)) α 2 df (²) r r r µ 2 V (, θ,²) α (θ,f (²)) 2 α (θ,f (²)) df (²), r r r α 2 α (θ,f (²)) r 2 α (θ,f (²)) r r 2

wc caracterzes 3 α (θ,f (²)) r r r, = µ 3 µ V (, θ,²) α (θ,f (²)) 3 V (, θ,²) α (θ,f (²)) α (θ,f (²)) df (²) r r α r α 2 df (²) r µ 2 V (, θ,²) 2 α (θ,f (²)) df (²) r α r µ 2 V (, θ,²) 2 µ α (θ,f (²)) 2 V (, θ,²) α (θ,f (²)) 2 α (θ,f (²)) df (²) r α r α 2 df (²) r r µ V (, θ,²) 3 α (θ,f (²)) df (²) α r r µ 3 V (, θ,²) α (θ,f (²)) α (θ,f (²)) r α 2 df (²) r µ 3 V (, θ,²) α (θ,f (²)) α (θ,f (²)) α (θ,f (²)) α 3 df (²) r r µ 2 V (, θ,²) 2 α (θ,f (²)) α (θ,f (²)) α 2 df (²) r r µ 2 V (, θ,²) α (θ,f (²)) 2 α (θ,f (²)) α 2 df (²) r r 2 µ V (, θ,²) 2 V (, θ,²) α (θ,f (²)) d d r r r α r µ 2 µ V (, θ,²) α (θ,f (²)) 2 V (, θ,²) α (θ,f (²)) α (θ,f (²)) d r α r α 2 d r r µ V (, θ,²) 2 α (θ,f (²)) d, α r r wc caracterzes 3 α (θ,f (²)) r r, = µ 2 V (, θ,²) 2 µ α (θ,f (²)) 2 V (, θ,²) df (²) r α 2 α 2 µ V (, θ,²) 3 α (θ,f (²)) df (²) α r 2 µ 3 µ 2 µ V (, θ,²) α (θ,f (²)) 3 V (, θ,²) r α 2 df (²) α 3 µ 2 V (, θ,²) α (θ,f (²)) 2 α (θ,f (²)) 2 α 2 df (²) r µ 2 µ V (, θ,²) α (θ,f (²)) 2 V (, θ,²) 2 d 2 r α α 2 µ V (, θ,²) 2 α (θ,f (²)) 2 d, α r α (θ,f (²)) 2 α (θ,f (²)) df (²) r 2 α (θ,f (²)) df (²) r µ α (θ,f (²)) α (θ,f (²)) α (θ,f (²)) d r 2 2

wc caracterzes 3 α (θ,f (²)) r,ad 2 µ 2 V (, θ,²) α (θ,f (²)) 2 α (θ,f (²)) = α 2 df (²) 2 µ V (, θ,²) 3 α (θ,f (²)) df (²) α 3 µ 3 µ 3 V (, θ,²) α (θ,f (²)) α 3 df (²) µ 2 V (, θ,²) α (θ,f (²)) 2 α (θ,f (²)) 2 df (²) 2 α 2 µ 2 V (, θ,²) 2 α 2 d µ α (θ,f (²)) µ V (, θ,²) µ 2 V (, θ,²) α 2 2 µ V (, θ,²) 2 α α d 2 α (θ,f (²)) 2 µ 2 α (θ,f (²)) d d 2 α (θ,f (²)) 2, wc caracterzes 3 α (θ,f (²)) 3. Ispectg tese dervatves ad applyg Lemmas 6, 7, ad 9, we obta te desred result. Lemma 2 Pr ² θ ²²² (²) >C ³ υ 3 = o for some costat C> ad υ < 6. Proof. From (5), we ave = d 3 (,²) d² 3 df (²)3 d 2 (,²) d² 2 d CombgLemmas6,7,8,9,ad, we ca boud P R ³ d 2 (,²) d² d 2 by C υ 3 wt probablty o. It was see Appedx B.3 tat te r-t compoet d2 (r) (,²) d² of d2 (,²) 2 d² s equal 2 to d 2 (r) (,²) d² 2 = (,²) 2 (r) (,²) (,²) (r) (,²) 2 θ 2 2 (r) (,²) α µ (r) (,²) α 2 (r) (,²) α α α 2 (r) (,²) 2 α α (r) (,²) α 2 (r) µ (,²) α α 2 (r) (,²) α α α 2 (r) µ 2 (,²) α α 2 2 (r) µ (,²) α α α 2 2 α (r) (,²) 2 α α (r) (,²) α 2 θ α 2 2 (r) µ (,²) α α 2 (r) µ 2 (,²) α 2 α 2. α 2 R d 3 (,²) Usg Lemmas 6, 7, 8, 9, ad aga, we ca coclude tat P d² df 3 (²) s equal to ³ P R (,²) df (²) 3 θ plus terms tat ca all be bouded by P R d 2 (,²) 3 d² d 2 by C 22 ³ υ 3

wt probablty o ³ P. Because R (,²) df (²) s bouded away from by Lemma 6, we obta te desred cocluso. Lemma 3 ³ bθ θ = θ ² () 2 Proof. Follows from equato () ad Lemma 2. µ 2 θ ²² () o p () C Proof of eorem Let bα (²) be suc tat = V ³ ; b θ (²), bα (²) df (²) (Note slgt dfferece of deþto. I prevous sectos, bα (θ,²) was uderstood to be a soluto to R V ( ; θ, α (θ,f (²))) df (²) =.) Usg te same argumets as earler, we are lookg for te expaso bα (²) α = bα ² () for some ², /2. Let 2 bα²² v (,²) V (θ (F (²)), α (F (²))). ( ²) e Þrst order codto may be wrtte as = v (,²) df (²) Dfferetatg repeatedly wt respect to ², weobta dv (,²) = df (²) v (,²) d (2) d² d 2 v (,²) dv (,²) = d² 2 df (²)2 d (22) d² were ³ bf F. Because dv (,²) d² = v (,²) v (,²) α we may rewrte (2) as µ v (,²) = v (,²) α valuatg at ² =we obta bα ² () = ( [V α ]) µ α α df (²) P t= v (, θ, α ) V θ v (,²) d θ ² () (23) 23

were θ ² () s deþed(7). It also follows tat µ µ bα ² v (,²) v (,²) (²) = df (²) α df (²) θ ² (²) Next, cosder d 2 v (,²) d² 2 = ² [,/ ] v (,²) 2 v (,²) ( α ) 2 v (,²) 2 θ () 2 2 v (,²) α µ 2 α v (,²) 2 α α () 2 α v (,²) d. (24) suc tat bα ²² (²) s caracterzed by = θ ² (²) v (,²) df (²) θ ² v (,²) (²) df (²) θ ²² (²) v (,²) 2 df (²) θ ² (²) bα ² 2 α (²) v (,²) ( α ) 2 df (²)(bα ² (²))2 v (,²) df (²) bα ²² v (,²) (²) α d θ ² v (,²) (²) d ² bα (²) (25) α We ow sow tat Pr (bα α ) > υ = o, for ay υ < 6. For ts purpose, t suffces to prove Pr sup bα ² ² [,/ (²) > υ = o, ] Pr bα ² () > υ = o, Pr sup bα ²² (²) > ³ υ 2 = o. I order to prove te Þrst clam, we ote tat Pr sup θ ² (²) υ = o ² [,/ ] from Lemma (8). By Lemma (6), we also ave µ Pr sup v (,²) v (,²) ² [,/ ] df (²) > η = o, Pr sup v (,²) v (,²) df (²) α α > η = o. ² [,/ ] By Lemma (6) aga, t follows tat Pr sup v (,²) d > / υ = o. ² [,/ ] 24

s proves te result for bα ² (²) as well as bα (). Forbα ²² (²), te result follows from represetato (25) as well as Lemmas (6), (8), ad (). We ow prove eorem. For ts purpose, we combe bα (²) = α bα ² () 2 bα²² ( ²), bα ² () = ( [V α ]) µ θ ² () = P t= v (, θ, α ) V θ I t= U θ ² () ad obta µ (bα α ) = ( [V α ]) P t= v (, θ, α ) ( [V α ]) V θ I 2 bα²² ( ²), from wc te cocluso follows. t= U 25

Refereces [] Arellao, M. (2): Dscrete Coces wt Pael Data, Ivestgacoes coómcas Lecture, V Smposo de Aálss coómco, Bellaterra. [2] Brock, W. ad S. Durlauf (2): Iteractos-based Models, J. Heckma &. Leamer, ds. Hadbook of coometrcs 5, pp. 3297-338. Amsterdam: Nort-Hollad. [3] Graam, B. S., ad J. Ha (23): IdetÞcato ad stmato of te Lear--Meas Model of Socal Iteractos, upublsed mauscrpt. [4] Ha, J. ad G. Kuersteer (22): Asymptotcally Ubased Iferece for a Dyamc Pael Model wt Fxed ffects We Bot ad are Large, coometrca, 7, 639-657. [5] Ha, J. ad G. Kuersteer (23): Bas Reducto for Dyamc Nolear Pael Models wt Fxed ffects, upublsed mauscrpt. [6] Ha, J. ad W.K. Newey (22): Jackkfe ad Aalytcal Bas Reducto for Nolear Pael Models, upublsed mauscrpt. [7] Hausma, J., ad W. aylor (98), Pael Data ad Uobservable Idvdual ffects, coometrca 49, pp. 377 398. [8] Mask, C. (993): IdetÞcato of edogeous socal effects: te reßecto problem, Revew of coomc Studes 6, pp. 53-542. [9] Wouterse,. (22): Robustess agast Icdetal Parameters, upublsed mauscrpt. 26