β0 + β1xi and want to estimate the unknown

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

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

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.

e i is a random error

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

j) = 1 (note sigma notation) ii. Continuous random variable (e.g. Normal distribution) 1. density function: f ( x) 0 and f ( x) dx = 1

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

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

Chapter 14 Simple Linear Regression

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

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

Statistics for Economics & Business

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

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

Lecture 4 Hypothesis Testing

Statistics for Business and Economics

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

Chapter 11: Simple Linear Regression and Correlation

Econ Statistical Properties of the OLS estimator. Sanjaya DeSilva

Introduction to Regression

Linear Regression Analysis: Terminology and Notation

Biostatistics 360 F&t Tests and Intervals in Regression 1

Economics 130. Lecture 4 Simple Linear Regression Continued

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

Lecture 6: Introduction to Linear Regression

Correlation and Regression. Correlation 9.1. Correlation. Chapter 9

Properties of Least Squares

Lecture 3 Stat102, Spring 2007

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

The Ordinary Least Squares (OLS) Estimator

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

Topic 7: Analysis of Variance

Y = β 0 + β 1 X 1 + β 2 X β k X k + ε

PubH 7405: REGRESSION ANALYSIS. SLR: INFERENCES, Part II

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

17 - LINEAR REGRESSION II

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

Statistics MINITAB - Lab 2

18. SIMPLE LINEAR REGRESSION III

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

LINEAR REGRESSION ANALYSIS. MODULE IX Lecture Multicollinearity

Basic Business Statistics, 10/e

Learning Objectives for Chapter 11

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

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

Comparison of Regression Lines

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

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

28. SIMPLE LINEAR REGRESSION III

Lab 4: Two-level Random Intercept Model

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

Chapter 13: Multiple Regression

Chapter 15 - Multiple Regression

Chapter 4: Regression With One Regressor

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

Limited Dependent Variables

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

Regression Analysis. Regression Analysis

STAT 3008 Applied Regression Analysis

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

Biostatistics. Chapter 11 Simple Linear Correlation and Regression. Jing Li

Correlation and Regression

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

January Examinations 2015

STATISTICS QUESTIONS. Step by Step Solutions.

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

Professor Chris Murray. Midterm Exam

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

The SAS program I used to obtain the analyses for my answers is given below.

Interpreting Slope Coefficients in Multiple Linear Regression Models: An Example

Statistics for Managers Using Microsoft Excel/SPSS Chapter 14 Multiple Regression Models

Lecture 3 Specification

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

STAT 3340 Assignment 1 solutions. 1. Find the equation of the line which passes through the points (1,1) and (4,5).

Problem of Estimation. Ordinary Least Squares (OLS) Ordinary Least Squares Method. Basic Econometrics in Transportation. Bivariate Regression Analysis

Rockefeller College University at Albany

THE ROYAL STATISTICAL SOCIETY 2006 EXAMINATIONS SOLUTIONS HIGHER CERTIFICATE

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

3.1 Expectation of Functions of Several Random Variables. )' be a k-dimensional discrete or continuous random vector, with joint PMF p (, E X E X1 E X

Chapter 8 Indicator Variables

Lecture 3: Probability Distributions

NANYANG TECHNOLOGICAL UNIVERSITY SEMESTER I EXAMINATION MTH352/MH3510 Regression Analysis

2016 Wiley. Study Session 2: Ethical and Professional Standards Application

Here is the rationale: If X and y have a strong positive relationship to one another, then ( x x) will tend to be positive when ( y y)

STAT 511 FINAL EXAM NAME Spring 2001

Negative Binomial Regression

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

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

The Gaussian classifier. Nuno Vasconcelos ECE Department, UCSD

Probability and Random Variable Primer

A Bound for the Relative Bias of the Design Effect

Parametric fractional imputation for missing data analysis. Jae Kwang Kim Survey Working Group Seminar March 29, 2010

Chapter 14 Simple Linear Regression Page 1. Introduction to regression analysis 14-2

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

Statistics II Final Exam 26/6/18

The Geometry of Logit and Probit

A Comparative Study for Estimation Parameters in Panel Data Model

JAB Chain. Long-tail claims development. ASTIN - September 2005 B.Verdier A. Klinger

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

CIS526: Machine Learning Lecture 3 (Sept 16, 2003) Linear Regression. Preparation help: Xiaoying Huang. x 1 θ 1 output... θ M x M

Transcription:

SLR Models Estmaton Those OLS Estmates Estmators (e ante) v. estmates (e post) The Smple Lnear Regresson (SLR) Condtons -4 An Asde: The Populaton Regresson Functon B and B are Lnear Estmators (condtonal on the s) OLS estmators are unbased! (under SLR.-SLR.4) but B s not alone OLS estmators have a varance SLR.5 - Homoskedastcty MSE/RMSE (Goodness of Ft) and Standard Errors OLS estmators are BLUE! (under SLR.-SLR.5) (separate handout) Those OLS Estmates ) Recall the brth of OLS estmates: a) You have a dataset consstng of n observatons of (, ) y: { }, y =,,... n. b) You beleve that ecept for random nose n the data, there s a lnear relatonshp between the s and the y s: y β + β and want to estmate the unknown parameters β (ntercept parameter) and β (slope parameter). c) Adoptng OLS, you estmate β and β by mnmzng sum squared resduals (SSRs): mn SSR = ( y ( )) β + β wrt β and β d) For the gven sample, the OLS estmates of the unknown ntercept and slope parameters are: ) Slope: ˆ ( )( y y) ( ) y y y y β = = = w, where ( ( n ) S ( w = and w = so the w ' s ( n ) S sum to and are proportonal to the square of the -dstance from. y y ) Recall that s the slope of the lne connectng (, y ) to the sample means pont, (, y ). ) Intercept: ˆ β ˆ = y β.

SLR Models Estmaton Estmates (e post) v. Estmators (e ante) ) epost (actual; after the event): After the data set s generated, OLS provdes numerc estmates of the slope and ntercept parameters, for the gven dataset (estmates are numbers, not random varables). a) After we have drawn the sample {, y } ) Slope estmate: ˆ ( )( y y) β =, and ( ) Intercept estmate: ˆ β ˆ = y β., we have the OLS estmates: 3) eante (before the event): But pror to the generaton of the data, the s and y s are random varables, X s and Y s, and OLS provdes a rule for estmatng the OLS coeffcents (for any realzed data set). a) We call these rules estmators. Estmators wll take on dfferent values dependng on the actual drawn sample, the actual 's and y's. Accordngly, estmators are random varables. b) The X ' s and Y ' s are random varables, and OLS provdes us wth slope and ntercept estmators: ( X X)( Y Y) ( X X) Y () Slope estmator: B = = ( X X) ( X X) () Intercept estmator: B = Y BX j j 4) To revew notaton, we have: a) Random varables (upper case letters): X's and Y's b) Data (lower case letters): 's and y's c) True parameters: β and β d) Parameter estmators (random varables; upper case letters): B and B e) Parameter estmates (estmated coeffcents): ˆ β ˆ and β

SLR Models Estmaton The Smple Lnear Regresson (SLR) Condtons -4 5) SLR. Lnear model (DGM): Y = β + βx + U, where X, Y and U are random varables and β and β are unknown parameters to be estmated. a) Ths s sometmes referred to as the Data Generaton Mechansm (DGM), as t descrbes the process by whch the data are assumed to have been generated. b) Notce that X, U and Y are now random varables, reflectng the random nature of the DGM. c) U s the uneplaned (or unobserved) error term (or dsturbance), and captures other factors (ecluded from the model) that eplan Y: ) = β β U Y X ) Put dfferently: U s the part of Y not eplaned/generated by the lnear functon of X. 6) SLR. Random samplng: the sample {(, )} y s a random sample; the th elements n the sample, (, y ), s the realzaton of two random varables ( X, Y ), where Y = β + β X + U. 7) SLR.3 Sample varaton n the ndependent varable: the ' s are not all the same value 8) SLR.4 U has zero condtonal mean: EU ( X= ) = for all a) EU ( X= ) = for any mples that: ) EU ( ) = (U has mean zero) () If the epected value of U s condtonal on any value of, then the overall epected value of U, whch wll be a weghted average of the condtonal epectatons of U, wll also be. () Put dfferently: E( U ) = prob( X = ) E( U X = ) = prob( X = ) = ) Cov( X, U ) = (X and U are uncorrelated) () Cov( X, U) E( ( X µ X)( U µ U) ) E( ( X µ X) U) U = E ( XU ) µ E ( U ) = E ( XU ) = prob( X = ) E( U X = ) = = snce µ = X = prob( X = ) E( U X = ) = prob( X = ) = At tmes we wll consder the X values to be eogenously gven, n whch case the eplanatory varable s not random. Recall that f X and U are ndependent then Cov( X, U ) =. A covarance of, however, does not mply ndependence, but rather than X and U do not move together n much of a lnear way. 3

SLR Models Estmaton () Notce the connecton to Omtted Varable Bas, whch s drven by correlaton between U and X. An Asde: The Populaton Regresson Functon 9) Under these assumptons (specfcally SLR. and SLR.4): a) β β β β EY ( X= ) = + + EU ( ) = + +, so the epected value of Y gven s EY ( X= ) = β + β. ) Populaton Regresson Functon (PRF): EY ( X= ) = β + β a) Ths traces out the condtonal means (condtonal on the values of X, the partcular values of the RHS varable) of the dependent varable Y. B and B are Lnear Estmators (condtonal on the s) ) The OLS slope estmator, B, s lnear n the Y ' s (condtonal on the s), snce we can epress the estmator as: ( ) ( ) a) B = by, where b = = ( ) ( n ) S j b) Note that condtonal on the s means that we are takng the values as gven, and not as random varables wth values to be determned. ) And the OLS ntercept estmator s also lnear n the Y ' s (condtonal on the s) snce: a) B = Y by = b Y n n 3) So condtonal on the 's, B and B are lnear estmators. 4

SLR Models Estmaton OLS estmators are unbased! (under SLR.-SLR.4) Who saw ths comng? 4) Gven assumptons/condtons SLR.-SLR.4, B and B, the OLS estmators of the ntercept and slope parameters, are unbased estmators (so for each, the epected value of the estmator s the true parameter value). a) We'll be provng ths below. b) The trck n the proof s to assume a partcular set of sample ' s, and to show that for any such sample, the OLS estmators wll be unbased. And snce ths s true for any sample of ' s, t must be true n epectaton. ) Ths s sometmes referred to as the Law of Iterated Epectaton. We wll skp the proof of ths Law but the ntuton s pretty straghtforward, yes? and you saw t n acton above n the proof that EU ( ) = gven SLR.4. 5) Condtonal on the s, the OLS estmators (random varables note the captal B's below) are defned by: a) ( )( Y Y) ( )( Y Y) ( ) Y Y B = = =, ( ) ( n ) S ( n ) S ( Y Y) and so B = w, ( ) where w ( = ( n ) S are non-negatve weghts that sum to, w =. b) B = Y B 6) An nterestng result: a) Assume SLR.-SLR.4. ) Then Y = β + βx + U, and µ Y = β + βµ X (snce EU ( ) = ) ) β β β β Cov( X, Y ) = Cov( X, + X + U ) = Cov( X, X ) + Cov( X, U ) = Cov( X, X ) snce Cov( X, U ) =. Cov( X, Y ) Cov( X, Y ) ) But then β =. and β = µ βµ = µ µ. Cov( X, X ) Y X Y Cov( X, X ) X b) Notce the resemblance to the OLS estmators! 5

SLR Models Estmaton 7) The OLS slope estmator B s unbased! - E( B ) = β Y Y a) Step - E = β: To evaluate the epected value of B condtonal on the s, we Y Y need to determne E, for each. But E( Y ) = β + β + EU ( ) = β + β (gven SLR.4). And snce EY ( s ' ) = EY ( ) = ( β + β) n n = β + β, we have: Y Y ( β + β ) ( β + β) β( ) E = = = β. ( ) ( ) b) Step - ( ' ) E B s = β ( B n an unbased estmator of β, condtonal on the s): Y Y Y Y E B ' s = E w = we = wβ = β weghts sum to w = ). ) Snce ( ) (snce the c) Step 3 - E( B ) = β : And snce E( B s ' ) = β for all s, E( B ) = β. 3 8) The OLS ntercept estmator B s also unbased! - E( B ) a) Step - ( ' ) = β E B s = β ( B n an unbased estmator of β, condtonal on the s): ( ' ) ( ) ( ) E B s = E Y E B = β + β β = β. b) Step - E( B ) = β : And snce E( B s ' ) = β for all s, E( B ) = β. 9) OLS LUE (gven SLR.-SLR.4): So gven the SLR condton -4, the OLS slope and ntercept estmators are lnear unbased estmators (LUEs) of the unknown parameter values! 4 ) Who saw ths comng? Who ever thought that process of mnmzng SSRs would lead to unbased estmators? 3 Ths last step s an applcaton of the Law of Iterated Epectatons. 4 But remember that they are only lnear condtonal on the 's. 6

SLR Models Estmaton But B s not alone!!! there are an nfnte number of unbased slope estmators (gven SLR.-SLR.4) ) Any weghted average of the slopes of the lnes connectng the dataponts to the samples means wll also be a LUE of the slope parameter: a) Consder the estmator defned by Y Y α X X, where α =. b) Then condtonal on the s: E Y Y Y Y α = αe = αβ = β, snce α =. c) And snce ths s the case for all s, we have an unbased estmator of β. d) Snce we only requre that the α ' s sum to one, we have an nfnte number of unbased slope estmators (as we vary the α ' s ). ) Test your understandng! From before, you know that gven SLR.-SLR.4, E Y β β EY s ' = β β +. Use these results to show that each of the ( ) = + and ( ) followng wll be unbased slope estmators, condtonal on the 's. And accordngly, by the Law of Iterated Epectatons, unbased slope estmators overall: Y Y a) B = X X 's. Answer: E( B s ' ) 5 b).5 Y Y B.5 Y Y = + X X X5 X 5 c).9 Y Y B. Y Y = + X X X5 X 5 d).5 Y Y B.5 Y Y = X X X5 X Y Y B = 5 e) X X5 Y Y Y Y B =.5 +.5 5 3 7 f) X X5 X3 X7 ( ) ( ) β + β β + β β ( = = = β, all ( ) ( ) 3) And so gettng to BLUE (Best Lnear Unbased Estmators) wll be all about fndng the LUE(s?) (amongst the many) wth the mnmum varance. 7

SLR Models Estmaton OLS estmators (B and B ) have varances 4) Varances of the OLS estmators: Yes, because they are estmators, the OLS estmators, B and B, are random varables, wth a jont dstrbuton, means, varances and a covarance. The sample you are workng wth s just one of many possble samples that you could have drawn. 5) An eample: Y = +.5X + U, X Unform[,] and U N(,). a) The followng show the results from, samples, each wth observatons generated by the random process above wth one slope and ntercept estmate per sample. Dstrbuton of OLS ntercept and slope estmates. summ best best Varable Obs Mean Std. Dev. Mn Ma -------------+--------------------------------------------------------- best, -.77699.776-7.5879 5.9449 best,.548.948-5.437995 6.7959 Densty...3.4 - -5 5 b est. Densty...3.4-5 5 b est. 6) The means of the, estmates are qute close to the true parameters values but notce the large varaton n slope and ntercept estmates, drven by the random nature of the DGM and dependng on the partcular sample that you are workng wth. 7) Worth repeatng!!!: Because they are estmators, the OLS estmators B and B are random varables, wth a jont dstrbuton, means, varances, and a covarance. Dfferent samples wll generate dfferent ntercept and slope estmates. Who knows f your sample s representtve? your estmates could n fact be not at all close to the true parameter values. It all depends on your sample. 8

SLR Models Estmaton SLR.5 Homoskedastcty 8) SLR.5: Homoskedastcty (constant condtonal varance of the error term) a) To derve the varances of the estmators, we make one addtonal assumpton: SLR.5: Var( U X ) σ = = for all b) SLR.5 holds f U s ndependent of X, so that Var( U X = ) = Var( U ) = σ. c) Heteroskedastcty: the condtonal varances are not all the same. Eample: Newton real estate sales prces and lot szes (heteroscedastcty) Source SS df MS Number of obs = 84 -------------+---------------------------------- F(, 8) = 69.4 Model.374e+3.374e+3 Prob > F =. Resdual 5.4e+3 8.7873e+ R-squared =.97 -------------+---------------------------------- Adj R-squared =.943 Total 6.776e+3 83.8e+ Root MSE = 4.e+5 ------------------------------------------------------------------------------ prce Coef. Std. Err. t P> t [95% Conf. Interval] -------------+---------------------------------------------------------------- lotsze 4.99 5.7549 8.3. 3.393 5.98 _cons 37448.4 6384.9 6.. 5348.8 49578 ------------------------------------------------------------------------------.e+6.e+6 3.e+6 5 5 5 LotSze Prce Ftted values Resduals - 5 5 5 LotSze predcteds v. actuals resduals v. lot sze 9

SLR Models Estmaton A smple test for heteroskedastcty: Regress the squared resduals on the RHS varable. Are there more complcated test? Of course!. predct resd, res. gen resd=resd^. reg resd lotsze Source SS df MS Number of obs = 84 -------------+---------------------------------- F(, 8) = 43.3 Model 6.3833e+4 6.3833e+4 Prob > F =. Resdual 4.89e+5 8.4833e+3 R-squared =.34 -------------+---------------------------------- Adj R-squared =.93 Total 4.8e+5 83.736e+3 Root MSE = 3.9e+ ------------------------------------------------------------------------------ resd Coef. Std. Err. t P> t [95% Conf. Interval] -------------+---------------------------------------------------------------- lotsze 3.3e+7 46344 6.56..e+7 3.94e+7 _cons -.57e+ 5.59e+ -.8.5 -.67e+ -4.73e+ ------------------------------------------------------------------------------ Varance of the OLS Estmators (assumng SLR.-SLR.5) 9) If SLR.5 holds, n addton to SLR.-SLR.4, then we have the followng varances of the : OLS estmators, condtonal on the partcular sample of { } a) b) Var( B ) = Var( B ) σ ( j σ = n ( j and StdDev B sd B Var B ( ) = ( ) = ( ) = 3) Here s the proof for the varance of the slope estmator: σ ( ( a) From above we have B = ( Y = α Y, where α ( n ) S =. ( n ) S b) Snce the Y ' s are ndependent, the varance of the sum s the sum of the varances, and Var( B ) = Var( αy ) = αvar( Y ) = ασ = σ α c) But ( α = ( n ) S. and so

SLR Models Estmaton ( j [( n ) S ] [( n ) S ] ( ) ( n ) S α = = = = ( n ) S ( ) σ σ d) Therefore: Var( B ) = = ( ) ( n ) S j j ) Note that ths ncreases wth ncreases n the error varance, σ, and wth decreases n the varaton of the ndependent varable. Makes sense? ) Where does ths varance come from? The estmator s always just the OLS estmator, so all of the varaton s comng from the dfferent possble data samples generated by the DGM. (See the Ecel smulatons.) MSE/RMSE (Goodness of Ft) and Standard Errors 3) Mean Squared Error (MSE): Typcally, we don t know the actual value of the varance σ. SSR But we can estmate t wth the: ˆ σ = = MSE. n a) Recall that MSE was one of our Goodness of Ft metrcs n OLS/SLR Assessment. b) Under SLR.-SLR.5 and condtonal on the s, MSE = σˆ wll be an unbased estmator of the varance, σ, of the error term U (the homoskedastc error). 5 MSE 3) s an unbased estmator of Var( B ) ( n ) S a) Snce ˆ σ = MSE s an unbased estmator of σ and snce MSE MSE = n S ( ) ( ) s an unbased estmator of Var( B ). Var( B ) = σ ( 33) RMSE: The standard error of the regresson, sometme called the Root MSE (or RMSE), s SSR the square root of ths: ˆ σ = = MSE = RMSE. n a) If we have RMSE = ˆ σ, then we can calculate the standard error of B : ˆ σ RMSE RMSE ( ) = ( ) = = =. S n ) StdErr B se B ( ) ( ) b) Ths s an estmate of StdDev( B) = sd( B), and s useful for constructng confdence ntervals for, and testng hypotheses about, β, the true slope parameter n the DGM. And now you see where that std. err. formula came from n OLS/SLR Assessment!, 5 For the proof, see the tet.