Outline. Overview of Issues. Spatial Regression. Luc Anselin

Similar documents
Spatial Econometrics

Spatial Regression. 9. Specification Tests (1) Luc Anselin. Copyright 2017 by Luc Anselin, All Rights Reserved

Panel Data Models. James L. Powell Department of Economics University of California, Berkeley

Spatial Regression. 11. Spatial Two Stage Least Squares. Luc Anselin. Copyright 2017 by Luc Anselin, All Rights Reserved

Econometric Analysis of Cross Section and Panel Data

1 Outline. 1. Motivation. 2. SUR model. 3. Simultaneous equations. 4. Estimation

Spatial Regression. 6. Specification Spatial Heterogeneity. Luc Anselin.

Econometrics. Week 4. Fall Institute of Economic Studies Faculty of Social Sciences Charles University in Prague

Spatial Regression. 13. Spatial Panels (1) Luc Anselin. Copyright 2017 by Luc Anselin, All Rights Reserved

Spatial Econometrics. Wykªad 6: Multi-source spatial models. Andrzej Torój. Institute of Econometrics Department of Applied Econometrics

SPATIAL PANEL ECONOMETRICS

Repeated observations on the same cross-section of individual units. Important advantages relative to pure cross-section data

Testing Random Effects in Two-Way Spatial Panel Data Models

Spatial Regression. 15. Spatial Panels (3) Luc Anselin. Copyright 2017 by Luc Anselin, All Rights Reserved

Econometrics Summary Algebraic and Statistical Preliminaries

Spatial Regression. 14. Spatial Panels (2) Luc Anselin. Copyright 2017 by Luc Anselin, All Rights Reserved

Proceedings of the 8th WSEAS International Conference on APPLIED MATHEMATICS, Tenerife, Spain, December 16-18, 2005 (pp )

GMM Estimation of Spatial Error Autocorrelation with and without Heteroskedasticity

Next, we discuss econometric methods that can be used to estimate panel data models.

RAO s SCORE TEST IN SPATIAL ECONOMETRICS

LECTURE 2 LINEAR REGRESSION MODEL AND OLS

splm: econometric analysis of spatial panel data

Departamento de Economía Universidad de Chile

Linear Regression with Time Series Data

Econometrics of Panel Data

Applied Econometrics (MSc.) Lecture 3 Instrumental Variables

Luc Anselin and Nancy Lozano-Gracia

Econometrics. Week 8. Fall Institute of Economic Studies Faculty of Social Sciences Charles University in Prague

Lecture: Simultaneous Equation Model (Wooldridge s Book Chapter 16)

Intermediate Econometrics

Regression with time series

The regression model with one stochastic regressor (part II)

Topic 10: Panel Data Analysis

Econ 582 Fixed Effects Estimation of Panel Data

Lecture 2: Spatial Models

The exact bias of S 2 in linear panel regressions with spatial autocorrelation SFB 823. Discussion Paper. Christoph Hanck, Walter Krämer

Non-linear panel data modeling

Econ 510 B. Brown Spring 2014 Final Exam Answers

SPATIAL ECONOMETRICS: METHODS AND MODELS

1. The OLS Estimator. 1.1 Population model and notation

Lecture 6: Hypothesis Testing

Econometrics of Panel Data

Recent Advances in the Field of Trade Theory and Policy Analysis Using Micro-Level Data

1 Estimation of Persistent Dynamic Panel Data. Motivation

ON THE NEGATION OF THE UNIFORMITY OF SPACE RESEARCH ANNOUNCEMENT

Economics 536 Lecture 7. Introduction to Specification Testing in Dynamic Econometric Models

1 Introduction to Generalized Least Squares

Discrete time processes

Outline. Nature of the Problem. Nature of the Problem. Basic Econometrics in Transportation. Autocorrelation

Title. Description. var intro Introduction to vector autoregressive models

Panel Data Models. Chapter 5. Financial Econometrics. Michael Hauser WS17/18 1 / 63

Econometrics I KS. Module 2: Multivariate Linear Regression. Alexander Ahammer. This version: April 16, 2018

Switching Regime Estimation

Reading Assignment. Distributed Lag and Autoregressive Models. Chapter 17. Kennedy: Chapters 10 and 13. AREC-ECON 535 Lec G 1

the error term could vary over the observations, in ways that are related

Spatial Regression. 10. Specification Tests (2) Luc Anselin. Copyright 2017 by Luc Anselin, All Rights Reserved

Economics 308: Econometrics Professor Moody

Appendix A: The time series behavior of employment growth

GeoDa and Spatial Regression Modeling

Analyzing spatial autoregressive models using Stata

Freeing up the Classical Assumptions. () Introductory Econometrics: Topic 5 1 / 94

Agricultural and Applied Economics 637 Applied Econometrics II

Christopher Dougherty London School of Economics and Political Science

A SPATIAL ANALYSIS OF A RURAL LAND MARKET USING ALTERNATIVE SPATIAL WEIGHT MATRICES

Heteroskedasticity. Part VII. Heteroskedasticity

Econometrics of Panel Data

Ch.10 Autocorrelated Disturbances (June 15, 2016)

EC327: Advanced Econometrics, Spring 2007

Instrumental Variables, Simultaneous and Systems of Equations

Introduction to Spatial Statistics and Modeling for Regional Analysis

Measuring The Benefits of Air Quality Improvement: A Spatial Hedonic Approach. Chong Won Kim, Tim Phipps, and Luc Anselin

Vector autoregressions, VAR

Short T Panels - Review

LECTURE 11. Introduction to Econometrics. Autocorrelation

Lesson 17: Vector AutoRegressive Models

Panel Data Model (January 9, 2018)

Heteroskedasticity. We now consider the implications of relaxing the assumption that the conditional

Econometrics I. Professor William Greene Stern School of Business Department of Economics 25-1/25. Part 25: Time Series

G. S. Maddala Kajal Lahiri. WILEY A John Wiley and Sons, Ltd., Publication

State-space Model. Eduardo Rossi University of Pavia. November Rossi State-space Model Fin. Econometrics / 53

Single Equation Linear GMM with Serially Correlated Moment Conditions

A Course in Applied Econometrics Lecture 14: Control Functions and Related Methods. Jeff Wooldridge IRP Lectures, UW Madison, August 2008

W-BASED VS LATENT VARIABLES SPATIAL AUTOREGRESSIVE MODELS: EVIDENCE FROM MONTE CARLO SIMULATIONS

A Non-Parametric Approach of Heteroskedasticity Robust Estimation of Vector-Autoregressive (VAR) Models

Linear models. Linear models are computationally convenient and remain widely used in. applied econometric research

7 Introduction to Time Series

7. GENERALIZED LEAST SQUARES (GLS)

A Guide to Modern Econometric:

Prof. Dr. Roland Füss Lecture Series in Applied Econometrics Summer Term Introduction to Time Series Analysis

Statistics: A review. Why statistics?

Econometrics Honor s Exam Review Session. Spring 2012 Eunice Han

Advanced Econometrics I

Least Squares Estimation of a Panel Data Model with Multifactor Error Structure and Endogenous Covariates

Modeling the Ecology of Urban Inequality in Space and Time

Missing dependent variables in panel data models

7 Introduction to Time Series Time Series vs. Cross-Sectional Data Detrending Time Series... 15

LECTURE 10: MORE ON RANDOM PROCESSES

Dynamic Panels. Chapter Introduction Autoregressive Model

When Should We Use Linear Fixed Effects Regression Models for Causal Inference with Longitudinal Data?

Least Squares Estimation-Finite-Sample Properties

Transcription:

Spatial Regression Luc Anselin University of Illinois, Urbana-Champaign http://www.spacestat.com Outline Overview of Issues Spatial Regression Specifications Space-Time Models Spatial Latent Variable Models Overview of Issues 1

Motivation Model-Driven new focus on spatial interaction» new economic geography, interacting agents, spatial equity, spatial externalities, etc. Data-Driven use of geo-referenced information» GIS data, data integration Four Elements of Spatial Econometrics Specifying the Structure of Spatial Dependence» which locations/observations interact Testing for the Presence of Spatial Dependence» what type of dependence, what is the alternative Estimating Models with Spatial Dependence» spatial lag, spatial error, higher order Spatial Prediction» interpolation, missing values Spatial Dependence Estimating the Form/Extent of Spatial Interaction substantive spatial dependence spatial lag models: y Correcting for the Effect of Spatial Spill-overs spatial dependence as a nuisance spatial error models: ε 2

Specifying Spatial Dependence Substantive Spatial Dependence lag dependence include Wyas explanatory variable in regression y = ρwy + Xβ + ε Dependence as a Nuisance error dependence non-spherical error variance E[εε ] = Ω» where Ω incorporates dependence structure Interpretation of Spatial Lag True Contagion related to economic-behavioral process only meaningful if areal units appropriate (ecological fallacy) interesting economic interpretation (substantive) Apparent Contagion scale problem, spatial filtering Interpretation of Spatial Error Spill-Over in Ignored Variables poor match process with unit of observation or level of aggregation apparent contagion: regional structural change economic interpretation less interesting nuisance parameter Common in Empirical Practice 3

Cost of Ignoring Spatial Dependence Ignoring Spatial Lag omitted variable problem OLS estimates biased and inconsistent Ignoring Spatial Error efficiency problem OLS still unbiased, but inefficient OLS standard errors and t-tests biased Specification of Spatial Econometric Models Issues to address formal statistical framework» spatial stochastic process what type of spatial dependence» lag or error, conditional or simultaneous extent of autocovariance» spatial weights, distance decay dependence vs. heterogeneity» some forms of dependence induce heterogeneity simple or higher order model Formal Structure for Spatial Dependence Direct Representation covariance between observations a direct function of a distance metric» Cov[y i, y j ] = f(θ, d ij ) -- Cov symmetric and positive definite continuous function of distance, isotropic Spatial Process Models spatial random process {Z i, i =D}» Markov random field form of process determines form of covariance spatial weights to specify spatial interaction 4

Tests for Spatial Dependence Spatial Autocorrelation Tests Moran s I for regression residuals moments-based tests alternative hypothesis is NOT a specific spatial process model Maximum Likelihood Based Tests Lagrange Multiplier/Score tests» based on OLS residuals Wald and Likelihood Ratio tests» based on Maximum Likelihood estimation of spatial model Estimation of Spatial Econometric Models Maximum Likelihood assume normality likelihood function includes Jacobian» numerical problems GMM/GM/IV robust to non-normality not very efficient no/difficult inference in some cases Spatial Regression Specifications 5

Spatial Autoregressive Model (simultaneous - SAR) Specification assume E[y] = 0, w.l.o.g. y = ρwy + u and E[u] = 0 (I - ρw)y = u, or, y = (I - ρw) -1 u Covariance Structure E[yy ] = (I - ρw) -1 E(uu )(I - ρw ) -1» with E[uu ] = σ 2 I [no need to assume gaussian]» E[yy ] = σ 2 [(I - ρw) (I - ρw)] -1 Conditional Autoregressive Model - CAR Specification E[ y i y* ] = µ i + ρσ j w ij ( y j - µ j ), j i Covariance Structure assuming [ y i y* ] gaussian then joint density of y is MVN[µ, Σ] provided constraints on interaction terms are satisfied SAR vs CAR SAR is NOT First Order CAR SAR variance structure (for symmetric W)»[(I -ρw) (I - ρw)] -1 = [I - 2ρW + ρ 2 W 2 ] -1» compare to CAR = [I - ρw] -1 SAR corresponds to conditional model with first AND second order neighbors CAR is NOT a First Order SAR SAR representation of CAR not useful» requires Choleski decomposition of I - ρw. 6

Spatial Moving Average SMA Specification y = ρwu + u y = (I + ρw)u Covariance Structure E[yy ] = (I + ρw)e(uu )(I + ρw)» with E[uu ] = σ 2 I [no need to assume gaussian]» E[yy ] = σ 2 [(I - ρw) (I - ρw)] = σ 2 [ I -ρ(w+w ) + ρ 2 WW ] Mixed Regressive Spatial Autoregressive Model Specification y = ρwy + Xβ + u with u as i.i.d. Spatial Filter (I - ρw)y = Xβ + u Reduced Form y = (I - ρw) -1 Xβ + (I - ρw) -1 u E[ y X ] = (I - ρw) -1 Xβ Spatial Multiplier Expansion of Reduced Form for w ij < 1 and ρ < 1 (I - ρw) -1 = I + ρw + ρ 2 W 2 +» Leontief inverse Spatial Multiplier E[ y X ] = [I + ρw + ρ 2 W 2 + ] Xβ» function of X, WX, W 2 X,» first, second, third order neighbors» all locations involved, but with distance decay 7

Higher Order Models Spatial Autoregressive, Moving Average Process (SARMA) of order p, q AR in y y= ρ 1 W 1 y + ρ 2 W 2 y +... + ρ p W p y+ ε MA in error term ε = λ 1 W 1 ξ + λ 2 W 2 ξ + + λ q W q ξ + ξ Special Cases biparametric SAR (Brandsma and Ketellapper)» y = ρ 1 W 1 y + ρ 2 W 2 y + Xβ + u Higher Order Models (continued) Both Lag and Error AR y = ρw 1 y + Xβ + ε ε = λw 2 ε + u or (I - ρw 1 )y = Xβ + (I - λw 2 ) -1 u (I - λw 2 )(I - ρw 1 )y = (I - λw 2 )Xβ + u Identification Problems for W 1 = W 2 y = (λ + ρ)wy - λρw 2 y + Xβ -λwxβ + u for W1 orthogonal to W2, or W1.W2 = 0 y = λw 2 y + ρw 1 y + Xβ - λw 2 Xβ + u Spatial AR Error Process Spatial Autoregressive Error y = Xβ + ε with ε = λwε + ξ Ω = σ 2 [(I - λw) (I - λw)] -1» Ω = σ 2 [ (I + ρw + ρ 2 W 2 + ) (I + ρw + ρ 2 W 2 + )] -1» range: all observations Spatial Common Factor Model (I - λw)y = (I - λw)xβ + η» OLS on spatially filtered variables spatial Durbin model: y = λwy + Xβ - λwxβ + η y = γ 1 Wy + Xγ 2 + WXγ 3 + η with γ 1.γ 2 = - γ 3» spatial common factor constraint 8

Global Spillovers Nature of Interaction (I - ρw) -1 all locations interact» spatial multiplier Model Taxonomy unmodeled effects only: spatial AR error» y = Xβ + (I - λw) -1 u both unmodeled and X: spatial lag» y = (I - ρw) -1 Xβ + (I - ρw) -1 u X only: spatial lag with MA error» y = (I - ρw) -1 Xβ + u = ρwy + Xβ + u - ρwu Local Spillovers Nature of Interaction W (spatial lag): only immediate neighbors interact Model Taxonomy unmodeled effects only: spatial MA error» y = Xβ + u + λwu X only: spatial cross-regressive» y = Xβ + γwxβ + u unmodeled effects and X: MA error with WX» y = Xβ + γwxβ + u + λwu» y = (I + λw)(xβ + u) (common factor) Spatial Dependence and Heteroskedasticity Spatial dependence induces heteroskedasticity MA process diagonal element»1 + λ (w ii + w ii ) + λ 2 [ (WW ) ii ]»(WW ) ii = Σ j w ij 2» depends on number of neighbors» not constant -> heteroskedastic errors AR process similar» also higher order powers 9

Space-Time Models Time and Spatial Dependence Example on Single Dimension processes on a line» locations i-1, i, i+1 autoregressive process in time (AR)»y i,t = ρ y i,t-1 + u i,t : y at i and t related to y at i in t-1»with process stable across space» y i-1,t = ρ y i-1,t + u i-1,t : same for y at i-1 (and i+1) autoregressive process on line (SAR)»y i,t = λ (y i-1,t + y i+1,t ) + ε i,t : y at i in t related to i-1,i+1 in t»with process stable over time:» y i,t-1 = λ (y i-1,t-1 + y i+1,t-1 ) + ε i,t-1 : same for y at t-1 Identification Problems Space-Time Dependence substitute y i,t-1 into AR» y i,t = ρ.λ (y i-1,t-1 + y i+1,t-1 ) + [ρ ε i,t-1 + u i,t ] substitute y i-1,t-1, y i+1,t-1 into SAR» y i,t = λ.ρ (y i-1,t-1 + y i+1,t-1 ) + [λ(ε i-1,t-1 + ε i+1,t-1 ) + u i,t ] Identification Problem without further structure, λ and ρ are not separately identifiable from simple space-time AR model» identification problem: ρ.λ Identification Requires Separable Models spatially dependent time proces time dependent spatial process 10

Model Dependence General Framework y = Σ r ρ r W r y + Xβ + ε» y as NT by 1 matrix, stacked by time or region»w r as NT by NT weights matrix expressing all space-time dependencies» different W r to express spatial and/or time dependence time dependence: initial value problem space-time dependence: what weights? Space-Time Dependence Pure space-recursive i at t depends on neighbors at t-1» spatial lag at t-1 is exogenous y it = γ[wy] i,t-1 + f(z) + ε i,t» spatial lag becomes endogenous for space-time error dependence, not for either serial or spatial alone Time-space recursive i at t depends on i at t-1 and neighbors at t-1» serial lag exogenous in absence of serial error dependence y it = ρy i,t-1 + γ[wy] i,t-1 + f(z) + ε i,t» serial lag endogenous for serial error dependence» spatial lag endogenous for space-time error dependence Space-Time Dependence (2) Time-space simultaneous i at t depends on i at t-1 and neighbors at t» spatial lag is endogenous y it = ρy i,t-1 + γ[wy] i,t + f(z) + ε it» implies dependence of y i,t on spatial lag at t-1 The works i at t depends on t-1 and current and past neighbors y it = ρy i,t-1 + λ[wy],i, t + γ[wy] i,,t-1 + f(z) + ε it» identification problems 11

Error Dependence Identification Problem Cov[y it,y js ] 0 for some i,j,t,s Impose Structure groupwise dependence» limit dependence to one dimension»e[ε i,t.ε j,s ] = σ h for all (i,t,j,s) =S h» classical SUR: E[ε i,t.ε j,s ] = σ ij for all t, s» spatial SUR: E[ε i,t.ε j,s ] = σ ts for all i, j parametric dependence» spatial AR or serial AR» identification issues between two dimensions Spatial SUR Model Parameters vary by time, fixed across space y it = x it β t + ε it spatial weights fixed: W N by N Spatial Lag Dependence y it = ρ t [Wy t ] i,t + x it β t + ε it Var[ε it ] = σ 2 t E[ε it ε is ] = σ 2 ts Spatial Error Dependence ε it = λ t [Wε t ] i,t + u it Var[u it ] = σ 2 t E[u it u is ] = σ 2 ts Practical Modeling Strategies To Pool or Not to Pool test constraints on β it = β» Chow test and generalizations test for groupwise heteroskedasticity locational and/or time dummies SUR or Not test on diagonality of cross-equation covariance with fixed β or variable β it Spatial Effects care in fixed effects models» condition if locational dummies lag vs error 12

The Two Effects Model One Cross-Section Common Regression Coefficient in each time period t, a cross-section y it = x it β + u it Error Components u it = µ i + ν it Spatial Autoregressive Errors ν it = λ Σ j w ij ν jt + ε it or ν t = (I - λw) -1 ε t = B -1 ε t Matrix Form y t = X t β + µ + B -1 ε t The Two Effects Model All Cross Sections Stacked Equations y NTx1 = X NTxk β kx1 + (ι T I N ) µ Nx1 + (I T B -1 NxN) ε NTx1 T cross sections of N observations» ι T is T by 1 vector of ones»i N(T) is N by N (T by T) identity matrix» Kronecker products yield NT by NT matrices Error Variance Var[uu ] E[uu ] = (ι T ι T I N )σ 2 µ + [I T (B B) -1 ]σ 2»with E[µµ ] = σ 2 µ I N and E[εε ] = σ2 I NT Ω = σ 2 Ψ = σ 2 [(J T I N ) φ + (I T (B B) -1 )]»with J T = ι T ι T (a matrix of ones), φ===σ 2 µ / σ2 Det (Ψ) and Ψ -1 using some special matrix properties Ψ = (B B) -1 + T φ I N B -2(T-1) Ψ -1 = J* T [(B B) -1 + T φ I N ] -1 + E T (B B)»with J* T = (1/T)J T and E T = I T -J* T 13

Spatial Latent Variable Models Spatial Latent Variable Models Latent Variable Structure y i * = x i β + ε i»y i * is unobservable, x i β is index function, ε i random error observables»prob[y i * > 0] = Prob[ x i β + ε i > 0]» requires specification of marginal probabilities Spatial Autocorrelation lag: spatial dependence in y i *»Cov[y i *y j *] 0, for i-j neighbors error: spatial dependence in ε i»cov[ε i ε j ] 0, for i-j neighbors Substantive and Nuisance Spatial Dependence Substantive: Lag Dependence y i * = ρ Σ j w ij y j * + x i β + ε i latent y i * function of latent values at neighbors Interaction between underlying propensity not the same as observed y i = revealed decisions Nuisance: Error Dependence y i * = x i β + ε i with ε i = λ Σ j w ij ε j + u i (u i iid) randomness ε i joint dependent with errors at neighbors unobservables have some spatial structure 14

Spatial Lag Probit Model Simultaneous Simultaneous Model y* are jointly determined y* = (I - ρw) -1 Xβ + (I - ρw) -1 ε»withε i ~ N(0,1)»u = (I -ρw) -1 ε and u ~ MVN(0,[(I -ρw) (I -ρw)] -1 ) no longer independent nor homoskedastic» u i is marginal of MVN integrate out N-1 dimensions standardize by location-specific variance condition for y i = 1 or y* i 0» x i β + ρ[wxβ] i + ρ 2 [W 2 Xβ] I + + u i 0»Prob[y i = 1] = Prob [ u i < G(X,β,ρ) ] depends on all x i Spatial Lag Probit Model Conditional Conditional Model not y* but observed y i y* = ρ Σ j w ij E[y* j X] + x i β + ε i» E[y* j X] exogenous?» if so, set z i = Σ j w ij E[ y* j X] = Σ j w ij y j average of observed 1 for neighbors y* = ρz i +x i β + ε i» treat as standard probit» requires much larger N to compensate for loss in information = coding approach only» conditional perspective not same for inference different interpretation, suitable for interpolation, but NOT for explaining complete spatial pattern Spatial Error Probit Model Error Distribution y i * = x i β + ε i with ε i = λ Σ i w ij ε j +u i with u i ~ N(0, 1): ε ~MVN (0,[(I - ρw) (I - ρw)] -1 ) Characteristics multivariate, not univariate normal heteroskedastic» standard probit inconsistent with heteroskedasticity» Var[ε i ] = ω ii, with ω ii = diagonal of [(I - ρw) (I - ρw)] -1» no analytical expression for ω ii P[ ε i < x i β ] is marginal of N-dimensional MVN 15