Now we relax this assumption and allow that the error variance depends on the independent variables, i.e., heteroskedasticity

Similar documents
Outline. 9. Heteroskedasticity Cross Sectional Analysis. Homoskedastic Case

Lecture 4 Hypothesis Testing

CHAPTER 8 SOLUTIONS TO PROBLEMS

Economics 130. Lecture 4 Simple Linear Regression Continued

Econ Statistical Properties of the OLS estimator. Sanjaya DeSilva

Econ107 Applied Econometrics Topic 3: Classical Model (Studenmund, Chapter 4)

Chapter 3. Two-Variable Regression Model: The Problem of Estimation

Outline. Zero Conditional mean. I. Motivation. 3. Multiple Regression Analysis: Estimation. Read Wooldridge (2013), Chapter 3.

a. (All your answers should be in the letter!

Econ107 Applied Econometrics Topic 9: Heteroskedasticity (Studenmund, Chapter 10)

Econometrics of Panel Data

Chapter 5: Hypothesis Tests, Confidence Intervals & Gauss-Markov Result

F8: Heteroscedasticity

Professor Chris Murray. Midterm Exam

Chapter 7 Generalized and Weighted Least Squares Estimation. In this method, the deviation between the observed and expected values of

Interval Estimation in the Classical Normal Linear Regression Model. 1. Introduction

January Examinations 2015

Chapter 2 - The Simple Linear Regression Model S =0. e i is a random error. S β2 β. This is a minimization problem. Solution is a calculus exercise.

ECONOMICS 351*-A Mid-Term Exam -- Fall Term 2000 Page 1 of 13 pages. QUEEN'S UNIVERSITY AT KINGSTON Department of Economics

If we apply least squares to the transformed data we obtain. which yields the generalized least squares estimator of β, i.e.,

[The following data appear in Wooldridge Q2.3.] The table below contains the ACT score and college GPA for eight college students.

ECONOMETRICS - FINAL EXAM, 3rd YEAR (GECO & GADE)

DO NOT OPEN THE QUESTION PAPER UNTIL INSTRUCTED TO DO SO BY THE CHIEF INVIGILATOR. Introductory Econometrics 1 hour 30 minutes

x i1 =1 for all i (the constant ).

Exam. Econometrics - Exam 1

Tests of Single Linear Coefficient Restrictions: t-tests and F-tests. 1. Basic Rules. 2. Testing Single Linear Coefficient Restrictions

Lecture 6: Introduction to Linear Regression

Biostatistics 360 F&t Tests and Intervals in Regression 1

ECON 351* -- Note 23: Tests for Coefficient Differences: Examples Introduction. Sample data: A random sample of 534 paid employees.

Lecture 9: Linear regression: centering, hypothesis testing, multiple covariates, and confounding

Lecture 9: Linear regression: centering, hypothesis testing, multiple covariates, and confounding

e i is a random error

The Multiple Classical Linear Regression Model (CLRM): Specification and Assumptions. 1. Introduction

Linear Regression Analysis: Terminology and Notation

Resource Allocation and Decision Analysis (ECON 8010) Spring 2014 Foundations of Regression Analysis

where I = (n x n) diagonal identity matrix with diagonal elements = 1 and off-diagonal elements = 0; and σ 2 e = variance of (Y X).

Statistics for Business and Economics

LINEAR REGRESSION ANALYSIS. MODULE IX Lecture Multicollinearity

Systems of Equations (SUR, GMM, and 3SLS)

STAT 3008 Applied Regression Analysis

) is violated, so that V( instead. That is, the variance changes for at least some observations.

LINEAR REGRESSION ANALYSIS. MODULE IX Lecture Multicollinearity

Interpreting Slope Coefficients in Multiple Linear Regression Models: An Example

Question 1 carries a weight of 25%; question 2 carries 20%; question 3 carries 25%; and question 4 carries 30%.

β0 + β1xi and want to estimate the unknown

Chapter 13: Multiple Regression

Properties of Least Squares

since [1-( 0+ 1x1i+ 2x2 i)] [ 0+ 1x1i+ assumed to be a reasonable approximation

Basically, if you have a dummy dependent variable you will be estimating a probability.

Lecture 3 Specification

Chapter 11: Simple Linear Regression and Correlation

Psychology 282 Lecture #24 Outline Regression Diagnostics: Outliers

Statistics for Managers Using Microsoft Excel/SPSS Chapter 13 The Simple Linear Regression Model and Correlation

Statistics for Economics & Business

Department of Quantitative Methods & Information Systems. Time Series and Their Components QMIS 320. Chapter 6

Module Contact: Dr Susan Long, ECO Copyright of the University of East Anglia Version 1

Dr. Shalabh Department of Mathematics and Statistics Indian Institute of Technology Kanpur

Polynomial Regression Models

Lecture Notes for STATISTICAL METHODS FOR BUSINESS II BMGT 212. Chapters 14, 15 & 16. Professor Ahmadi, Ph.D. Department of Management

4 Analysis of Variance (ANOVA) 5 ANOVA. 5.1 Introduction. 5.2 Fixed Effects ANOVA

MD. LUTFOR RAHMAN 1 AND KALIPADA SEN 2 Abstract

β0 + β1xi. You are interested in estimating the unknown parameters β

T E C O L O T E R E S E A R C H, I N C.

III. Econometric Methodology Regression Analysis

# c i. INFERENCE FOR CONTRASTS (Chapter 4) It's unbiased: Recall: A contrast is a linear combination of effects with coefficients summing to zero:

Chapter 5 Multilevel Models

The Ordinary Least Squares (OLS) Estimator

See Book Chapter 11 2 nd Edition (Chapter 10 1 st Edition)

LINEAR REGRESSION ANALYSIS. MODULE VIII Lecture Indicator Variables

Chapter 8 Indicator Variables

4.3 Poisson Regression

CHAPER 11: HETEROSCEDASTICITY: WHAT HAPPENS WHEN ERROR VARIANCE IS NONCONSTANT?

Heteroskedasticity (Section )

Introduction to Dummy Variable Regressors. 1. An Example of Dummy Variable Regressors

Statistics II Final Exam 26/6/18

Microeconometrics (Markus Frölich) Chapter 10. Linear Models for Panel Data

1. Inference on Regression Parameters a. Finding Mean, s.d and covariance amongst estimates. 2. Confidence Intervals and Working Hotelling Bands

[ ] λ λ λ. Multicollinearity. multicollinearity Ragnar Frisch (1934) perfect exact. collinearity. multicollinearity. exact

Department of Statistics University of Toronto STA305H1S / 1004 HS Design and Analysis of Experiments Term Test - Winter Solution

Comparison of Regression Lines

Durban Watson for Testing the Lack-of-Fit of Polynomial Regression Models without Replications

Tests of Exclusion Restrictions on Regression Coefficients: Formulation and Interpretation

Y = β 0 + β 1 X 1 + β 2 X β k X k + ε

Chapter 14 Simple Linear Regression

REGRESSION ANALYSIS II- MULTICOLLINEARITY

University of California at Berkeley Fall Introductory Applied Econometrics Final examination

CHAPTER 8. Exercise Solutions

Learning Objectives for Chapter 11

Introduction to Regression

Chapter 15 Student Lecture Notes 15-1

Linear regression. Regression Models. Chapter 11 Student Lecture Notes Regression Analysis is the

Chapter 15 - Multiple Regression

Estimation: Part 2. Chapter GREG estimation

ANSWERS. Problem 1. and the moment generating function (mgf) by. defined for any real t. Use this to show that E( U) var( U)

However, since P is a symmetric idempotent matrix, of P are either 0 or 1 [Eigen-values

First Year Examination Department of Statistics, University of Florida

Two-factor model. Statistical Models. Least Squares estimation in LM two-factor model. Rats

Dr. Shalabh Department of Mathematics and Statistics Indian Institute of Technology Kanpur

Midterm Examination. Regression and Forecasting Models

A Comparative Study for Estimation Parameters in Panel Data Model

Transcription:

ECON 48 / WH Hong Heteroskedastcty. Consequences of Heteroskedastcty for OLS Assumpton MLR. 5: Homoskedastcty var ( u x ) = σ Now we relax ths assumpton and allow that the error varance depends on the ndependent varables,.e., heteroskedastcty Consequences of heteroskedastcty [IMPORTANT] OLS estmators are stll unbased and consstent under heteroskedastcty. Recall that we prove unbasedness under MLR. through MLR.4, but not MLR.5. Also, nterpretaton of R-squared s not changed n the presence of heteroskedastcty. σ u R where the uncondtonal error varance, σ Y heteroskedastcty, (whch refers to the condtonal error varance) Heteroskedastcty nvaldates varance formulas for OLS estmators σ u, s unaffected by The usual F-tests and t-test are not vald under heteroskedastcty because the varance formula for OLS estmator s wrong. Under heteroskedastcty, OLS s no longer the best lnear unbased estmator (BLUE); there mght be more effcent lnear estmator.

ECON 48 / WH Hong Heteroskedastcty. Heteroskedastcty-Robust Inference after OLS Estmaton Formula for OLS standard errors and related statstcs have been developed that are robust to heteroskedastcty of unknown form To derve those formula, consder the followngs: Smple regresson model case Estmaton model: y = β0 + βx + u We assumes that the frst four Gauss-Markov assumptons hold, but not the last one. The general form of contan heteroskedastcty: ( u x ) = σ var The subscrpt ndcates that the varance of the error depends upon the partcualar value of x. Then, followng the same argument, we can show that var( ˆ ) n ( ) x x σ = β = SSTx. Therefore, the samplng varance can be estmated by: var ( ˆ ) n ( ) x x uˆ = β = SSTx

ECON 48 / WH Hong Heteroskedastcty Multple regresson model case Estmaton model: = β0 + β +... + βk k + y x x u The general form of contan heteroskedastcty: ( u x ) = σ Formula for heteroskedastcty-robust OLS varance s: var ( ˆ j ) n = β = ru ˆ ˆ SSR j j var where rˆj denotes the th resdual from regressng x j on other ndependent varables, and SSRj s the sum of squared resduals from ths regresson. The squared root of ths formula provdes heteroskedastcty-robust OLS standard error, and s also called Whte standard errors. Dscusson on heteroskedastcty-robust OLS standard error All formula are only vald n large samples. Usng the formula, the usual t-test s vald asymptotcally. The usual F-statstc does not work under heteroskedastcty, but heteroskedastcty robust versons are avalable n most software. 3

ECON 48 / WH Hong Heteroskedastcty (Example) Hourly wage equaton Estmated equaton ( ) log wage = 0.8 + 0.0904educ + 0.040exper 0.0007exper (0.05) (0.0075) (0.005) (0.000) [0.07] [0.0078] [0.0050] [0.000] The usual OLS standard errors are presented n the parentheses and the heteroskedastcty-roust standard errors are presents n the squared brackets. The heteroskedastcty-roust standard errors may be larger or smaller than ther nonrobust counterparts. The dfferences are often small n practce. The usual t-test sad that all the ndependent varables are sgnfcant. H β β : 0 0 exper = = exper F = 7.95 vs. F = 7.99 => the null s hghly rejected robust F-statstcs are also often not too dfferent. However, f there s strong heteroskedascty, dfferences may be larger. To be on safe sde, t would be better to always compute robust standard errors. 4

ECON 48 / WH Hong Heteroskedastcty 3. Testng for Heteroskedastcty It may stll be nterestng whether there s heteroskedastcty because then OLS may not be the most effcent lnear estmator anymore () Breusch-Pagan test for heteroskedastcty Ratonale H ( u x ) = σ 0 :var ; Homoskedastcty var( u ) = E( u ) E( u ) = E( u ) x x x x under MLR.4: E ( u x ) = 0 => ( x ) ( ) E u = E u = σ Ths mples that the mean of u must not vary wth x,..., x k. 5

ECON 48 / WH Hong Heteroskedastcty How to test Consder the followng model: ˆ δ0 δ... δk k u = + x + + x + error Regress squared resduals on all explanatory varables Hypothess testng: H0 : δ =... = δ k = 0 The hypothess means that all the explanatory varables do not explan the squared resduals, mplyng homoskedastcty. F-statstc: F = uˆ ( R )/( n k ) uˆ R / k ~ F kn, k where s the R-squared from the regresson of û on all explanatory varables R û LM test (alternatve test statstc) LM = n R χ ~ uˆ k Ths test s only asymptotcally vald. 6

ECON 48 / WH Hong Heteroskedastcty () Whte test for heteroskedastcty How to test Consder the followng model wth 3 ndependent varable case: û = δ + δ x + δ x + δ x + δ x + δ x + δ x + δ x x + δ x x + δ x x + e rror 0 3 3 4 5 6 3 7 8 3 9 3 Regress squared resduals on all explanatory varables, ther squares, and nteractons. Hypothess testng: H0 : δ =... = δ9 = 0 The Whte test detects more general devatons from heteroskedastcty than the Breush- Pagan test LM = n R χ ~ ˆ u 9 Dsadvantage of ths form of the Whte test: (we can also use an F-test for ths hypothess.) Includng all squares and nteractons leads to a large number of estmated parameters (e.g. k = 6 leads to 7 parameters to be estmated) Alternatve form of the Whte test uˆ = δ + δ yˆ + δ yˆ + error 0 Hypothess testng: H : δ = δ = 0, LM = n R χ 0 ~ uˆ 7

ECON 48 / WH Hong Heteroskedastcty 4. Weghted Least Squared Estmaton () The Heteroskedastcty s known up to a multplcatve constant (Weghted LS) Assume that ( u ) = σ h( ) x x, h( x ) = h > 0 var The functonal form of the heteroskedastcty,.e., y = β0 + βx +... + βkxk + u y x x k u => = β0 + β +... + βk + h h h h h <=> * * * * y = β x + β x + + β x + u ; Transformed model 0 0... k k * h( x ), s known. Example: Savngs and ncome sav β0 βnc u = + +, ( u nc ) = σ var nc sav nc = β0 + β + u nc nc nc Note that ths regresson model has no constant. * 8

ECON 48 / WH Hong Heteroskedastcty The transformed model s homoskedastc ( ) E u ( x) * u E u σ h x = E x = = = σ h h h Provded that the other Gauss-Markov assumptons hold as well, OLS appled to the transformed model s the best lnear unbased estmator. OLS n the transformed model s called as weghted least squares (WLS) y x x n k mn b0 b... bk = h h h h <=> n = ( ) mn y b bx... b x / h 0 k k Observaton wth a large varance gets a smaller weght n the optmzaton problem Why s WLS more effcent than OLS n the orgnal model? Observaton wth a large varance are less nformatve than observaton wth small varance and therefore should get less weght. WLS s a specal case of generalzed least squares (GLS) 9

ECON 48 / WH Hong Heteroskedastcty () Unknown heteroskedastcty functon (feasble GLS) ( u x) = σ ( δ + δ x + + δ x ) = σ h( x ) var exp 0... k k Assumed general form of heteroskedastcty. That s, h( x ) s unknown and therefore should be estmated before runnng the estmaton model. exp-functon s used to ensure postvty. u = σ exp ( δ + δ x +... + δkxk) v where v s multplcatve error 0 ( ) = + + + + log( ) => log u α0 δx... δkxk e where Steps for estmaton of FGLS. Regress y on all ndependent varables,. Estmates the model, ( ˆ ) => log α0 δ... δk k ( ˆ α0 δ δ ) h ˆ = exp + ˆ x +... + ˆ x k k α = σ + δ. 0 0 x,..., x k, and obtan the resduals, u = ˆ + ˆx + + ˆ x + error, n order to obtan h.. Apply WLS usng hˆ obtaned n the prevous step,.e., estmate the followng model: uˆ ˆ * * * * * y y = β0x0 + βx +... + βkxk + u where * y =, hˆ x * j xj = for j = 0,,.., k, and hˆ u * = u hˆ 0

ECON 48 / WH Hong Heteroskedastcty (Example) Demand for cgarettes Estmaton by OLS ( ) ( ) cgs = 3.64 + 0.880log ncome 0.75log cgprc 0.50educ (4.08) (0.78) (5.773) (0.67) 0.77age 0.0090age.83restaurn (0.60) (0.007) n = 807, R = 0.056 (.) cgs: cgarettes smoked per day; cgprc : cgarettes prce restaurn : smokng restrcton n restaurants p value = 0.000 => reject homoskedastcty Breusch Pagan

ECON 48 / WH Hong Heteroskedastcty Estmaton by FGLS ( ) ( ) cgs = 5.64 +.30log ncome.94log cgprc 0.463educ (7.80) (0.44) (4.46) (0.0) 0.48age 0.0056age 3.46restaurn (0.097) (0.0009) (0.80) n = 807, R = 0.34 Dscusson The ncome elastcty s now statstcally sgnfcant; other coeffcents are also more precsely estmated (wthout changng qualtatve results)

EC ON 48 / WH Hong Heteroskedastcty What f the assumed heter oskedastcty functon s wrong? If the heteroskedastcty functon s msspecfed, WLS s stll unbased and consstent under MLR. through MLR.4, but not effcent However, f we estmate ˆ( ) h x wth a msspecfed functonal form, then FGLS s based and nconsstent. In contrast, OLS s always unbased n the presence of heteroskedastcty. Practcally, OLS estmaton wth robust standard error adjustment would be a safe way. 3