Lecture 1 – Introduction (Reference – Chapter 1, Hayashi) This final part of your core training in econometrics will introduce you to a variety of topics that are relevant for regression analysis involving time series data. Time series data arise in virtually every area of applied economic research, making up nearly all of the data that we study in macroeconomics and financial economics. More comprehensive and advanced treatments of time series analysis are provided in Economics 674 and Statistics 551. Begin by recalling the assumptions that define the classical linear regression model. The Classical Normal Linear Regression Model: Let Xi = [x1i … xki], i = 1,…,n be a collection of kdimensional random vectors X = the nxk random matrix whose i-th row is Xi. Assume that the random variables y1,…,yn are related to X1,…,Xn according to: • A.1. (Linearity) yi = β1x1i + … + βkxki + εi , i = 1,…,n where β1,…, βk are constants • A.2. (Strict Exogeneity) E[εi │X ] = 0, i = 1,…,n • A.3. (No Multicollinearity) Prob(rank(X) = k) = 1 • A.4. (Spherical Disturbances) A.4.1.(Conditional Homoskedasticity) E[εi2 │X ] = σ2 > 0 , i = 1,…,n A.4.2. (Serially Uncorrelated Errors) E[εiεj │X ] = 0, i,j= 1,…,n, i≠j A.4.3. (Normally Distributed Errors), εi│X ~ N(0, σ2) , i = 1,…,n Under these assumptions,the ˆ OLS estimator of β, , • is unbiased • exists and is unique • has variance σ2(X’X)-1 • is conditionally normally distributed • is the BUE of β • is the MLE of β In addition, ˆ ) / se(ˆ ) ( ti = i ~ t(n-k) i i where se(ˆi ) ˆ 2 [( X ' X ) 1 ]ii , ˆ 2 SSR /( n k ) if Rβ = r, where R is qxk and r is qx1 then F= 2 ˆ ˆ -1 -1 ˆ {(R -r)’[R(X’X) R’] (R -)/q}/ ~ F(q,n-k) That is, the “OLS” t-statistic is drawn from a tdistribution and the “OLS” F-statistic is drawn from an F-distribution, providing the basis for the application of classical hypothesis test procedures and the construction of confidence intervals/regions. We will assume throughout this course that assumptions A.1 and A.3 hold. However, we will relax assumptions A.2 and A.4. Note: With regard to A.1, nonlinear models, e.g., threshold models, play an increasingly important role in the analysis of economic time series. Nonlinear time series models are considered in Economics 674. The assumptions that the errors are serially uncorrelated and conditionally homoskedastic are often implausible in regressions with time series data. Instead, it is typically the case that the errors are positively autocorrelated and it is often the case, particularly in applications with financial time series data, that they are conditionally heteroskedastic. Suppose we relax A.4.1 and/or A.4.2 but retain all of the other assumptions. We replace these assumptions with: E(εε’) = σ2Ω, Ω a p.d. nxn matrix How do the properities of the OLS estimator and the t and F statistics change? Under these assumptions the OLS estimator of β, • is unbiased ---- YES • exists and is unique ---- YES • has variance σ2(X’X)-1--- NO • is conditionally normally distributed --- YES, if the ε’s are normally distributed • is the BUE of β ---- NO • is the MLE of β ---- NO In addition, ti = (ˆi i ) / se(ˆi ) ~ t(n-k) ---- NO where se(ˆi ) ˆ 2 [( X ' X ) 1 ]ii 2 , ˆ SSR /( n k ) and if Rβ = r, where R is qxk and r is qx1 then F= {(R ˆ -r)’[R(X’X)-1R’]-1(R ˆ -r)/q}/ ˆ 2 ~ F(q,n-k) --- NO The OLS estimator is still unbiased, but it does not have any optimality properties, and inference based on the OLS t and F statistics as though they were drawn from t and F distributions is not valid. Suppose, however, that the matrix Ω is known. Then the GLS estimator of β, ̂GLS , • is unbiased • exists and is unique • has variance σ2(X’ Ω-1 X)-1 • is conditionally normally distributed • is the BUE of β • is the MLE of β In addition, ti = (ˆGLS , i i ) / se(ˆGLS , i ) ~ t(n-k) where se(ˆGLS , i ) ˆ 2 [( X ' 1 X ) 1 ]ii , ˆ 2 SSR /( n k ) and if Rβ = r, where R is qxk and r is qx1 then F= 2 -1 -1 -1 ˆ (R ̂ -r)’[R(X’Ω X) R’] (R ̂ -r)/(q/ ) ~ F(q,n-k) GLS GLS That is, the “GLS” t-statistic is drawn from a t-distribution and the “GLS” Fstatistic is drawn from an F-distribution, providing the basis for the application of classical hypothesis test procedures and the construction of confidence intervals/regions. But, what if Ω is unknown, which would be the typical situation in practice? We could replace it with an estimate, ̂ , to construct the FGLS estimator of β, ̂ F GLS , and the corresponding t and F statistics. What can we say about the properties of the FGLS estimator? • is unbiased --- NO • exists and is unique --- NO (depends on how we estimate Ω) • has variance σ2(X’ Ω -1X)-1 --- NO • is conditionally normally distributed -- NO • is the BUE of β --- NO • is the MLE of β --- NO In addition, ti = ( ˆ F GLS , i i ) / se( ˆ F GLS , i ) ~ t(n-k)--- NO where ˆ 1 X ) 1 ] se( ˆ F GLS , i ) ˆ 2 [( X ' ii , ˆ 2 SSR /( n k ) and if Rβ = r, where R is qxk and r is qx1 then F= {(R ̂ -r)’[R(X’ F GLS ̂ -1 X)-1R’]-1(R ̂ -r)/q}/ ˆ 2 F GLS ~ F(q,n-k) --- NO That is, the “FGLS” t-statistic is not drawn from a t-distribution and the “FGLS” F-statistic is not drawn from an F-distribution. So, it appears that if the errors are heteroskedastic or serial correlated and we have incomplete information about the form of the heteroskedasticity or serial correlation (i.e., Ω is unknown), we are in a very bad situation. • The OLS estimator is unbiased, the FGLS is biased; neither has any optimal properties • Neither the OLS nor FGLS estimators can be used as the basis for constructing valid hypothesis tests or confidence intervals! Now, suppose we relax A.2 (strict exogeneity). This assumption is rarely plausible in time series regressions because the i-th observation of the dependent variable y will frequently be correlated with observations i+1,i+2,… of the explanatory variables, X: lagged dependent variables (yi = β1 + β2xi + β3yi-1 + εi) feedback effects between x and y (xi is determined by yi-1) Instead, the best we might hope for is that the x’s are predetermined: E(εi │x11,…,xk1,…, x1i,…,xki) = 0, all i In this case, the OLS estimator of β, • • • • is unbiased ---- NO exists and is unique ---- YES has variance σ2(X’X)-1--- NO is normally distributed --- YES, but … • is the BUE of β ---- NO • is the MLE of β ---- NO In addition, ti = (ˆi i ) / se(ˆi ) ~ t(n-k) ---- NO where se(ˆi ) ˆ 2 [( X ' X ) 1 ]ii , ˆ 2 SSR /( n k ) if Rβ = r, where R is qxk and r is qx1 then 2 -1 -1 ˆ ˆ ˆ F = {(R -r)’[R(X’X) R’] (R -r)/q}/ ~ F(q,n-k) --- NO So, the OLS estimator is not unbiased, does not retain any optimality properties, and inference based on the OLS t and F statistics are not longer valid. This is an even worse situation than when A.4 fails since, in that case, at least OLS is unbiased. Summary – In time series regressions A.2 and /or A.4 are often implausible assumptions. When these assumptions fail, the OLS estimator may or may not be unbiased (depending upon whether A.2 is or is not correct). The OLS t and F statistics will not be valid starting points for testing or interval construction. If A.4 fails but A.2 holds, the GLS procedure provides a simple solution to this problem, but relies on our knowing more about the variance of the error vector than is reasonable in most applications. That is, GLS will generally not be a feasible approach. The FGLS estimator is a biased estimator and the FGLS t and F statistics will not be valid starting points for hypothesis testing or interval construction. This would seem to leave us in a very bad position! However, there is a way around these problems - Formulate an alternative set of assumptions that will accommodate weakening of assumptions A.2 and A.4 to allow for predetermined (rather than strictly exogenous) regressors and for non-spherical disturbances allow us to apply asymptotic distribution theory to construct estimation and test procedures that will have desirable properties in applications with sufficiently large sample sizes. Note: In Econ 671 you were introduced to the application of asymptotic distribution theory to study large sample properties of OLS and FGLS under a particular set of assumptions. However those assumptions are not adequate for our purposes. We will formulate a set of assumptions that turn out to be more useful in time series regressions.