Lecture 6 - Testing Restrictions on the Disturbance Process (References – Sections 2.7 and 2.10, Hayashi) We have developed sufficient conditions for the consistency and asymptotic normality of the OLS estimator that allow for conditionally heteroskedastic and serially correlated regressors conditionally heteroskedastic, serially uncorrelated disturbances. In this lecture we will consider testing the disturbances for conditional heterokedasticity and serial correlation. It is clear why we would want to test for serial correlation in the distrubances: one of the maintained assumptions of the model is that there is no serial correlation. The case for testing conditional heteroskedasticity is less compelling since our theory provides procedures that do not depend on whether the disturbances are conditionally homoskedastic or conditionally heteroskedastic. But, such tests still might be useful. Under assumptions A.1-A.5, if the disturbances are conditionally homoskedastic we showed last time that the inference procedures that are appropriate in the classical linear regression model with strictly exogenous regressors and spherical disturbances can be applied for testing and interval construction and these procedures will be asymptotically valid. If the disturbances are conditionally homskedastic, procedures that take this into account may perform better in finite samples. (Monte Carlo studies might be helpful in this regard.) If the disturbances are conditionally heteroskedastic, the OLS estimator is not asymptotically efficient. If the disturbances are conditionally heteroskedastic the FGLS estimator will be asymptotically efficient. (To apply the FGLS estimator we need to be able to model the conditional heteroskedasticity and have a consistent estimator of the parameters of that model.) The most popular nonparametric tests for serial correlation (Box-Pierce and Ljung-Box tests) rely on the assumption that the disturbances are conditionally homoskedastic. There are two ways to approach these testing problems: parametric tests and nonparametric tests. A parametric test specifies a model for the serial correlation or the conditional heteroskedasticity and then tests a set of restrictions on the parameters of that model. For instance, we might think that if the disturbances are conditionally heteroskedastic it is because the disturbances follow the first-order ARCH process: t vt 0 1 t 1 In this case, conditional homoskedasticity is the restriction that 1 = 0 and so a natural way to test for conditional homoskedasticity would be to construct a test of the null hypothesis, H0: 1=0. A nonparametric test does not require you to formulate a specific model of the conditional heteroskedasticity or the serial correlation under the alternative hypothesis. Testing for Conditional Heteroskedasticity Most tests for conditional heteroskedasticity in time series regressions are parametric tests that formulate the possible heteroskedasticity in terms of an ARCHtype model. These models will be discussed in more detail in Econ 674. Note Engle’s TR2 test here Section 2.7 of Hayashi’s textbook outlines a nonparametric test of heteroskedasticity developed by White (1980). This test relies on the assumption of i.i.d regressors and so it is not particularly useful in time series settings. The idea is that when regression disturbances are conditionally homoskedastic, there are (at least) two consistent estimators of the matrix S that appears in the asymptotic variance formula for the OLS 2 ' E ( x x estimator, where S = t t t ): 1 T 2 ' ˆ S ˆt xt xt T 1 and 1 T 2 2 xt xt' s Sxx = s T , where s2 = SSR/T. 1 So, if the disturbances are homoskedastic, the difference between these two estimators should be getting “small” as sample size increases. In fact, the difference will converge in probability to zero if the disturbances are homoskedastic. White constructed a statistic based on this difference that converges in distribution to a χ2 random variable if the disturbances are conditionally homoskedastic. Testing for Serial Correlation Both parametric and nonparametric tests are widely used in time series to test for serial correlation in the regression disturbances. The next major section of the course will be concerned with “time series models” and one of the applications of time series will be to parameterize serial correlation in regression models: that will allow to test for serial correlation and, if it appears to be present, apply the FGLS estimator and test procedures. Our focus today will be on nonparametric tests. The advantage of these tests is that they do not require us to formulate a specific model of serial correlation. The disadvantages are that 1) they will rely on the assumption of conditionally homoskedastic disturbances and 2) if we find evidence of serial correlation, then what do we do? Note: nonparametric tests for serial correlation in the presence of conditional heteroskedasticity do exist and nonparamteric adjustment procedures that account for potential serial correlation in regression models do exist. (HAC robust testing; HAC robust standard errors.) Let’s begin by assuming that {zt} is a stationary and ergodic process with finite variance (i.e., it is covariance stationary,too). Then the j-th autocovariance of the process is: cov( zt , zt j ) E[( zt z )( zt j z )] j for all t and j, where z E ( z t ) The sample j-th autocovariance is defined by: 1 T ˆ j ( zt zT )( zt j zT ) T t j 1 where zT is the sample mean of z1,…,zT Corresponding to these are the j-th autocorrelations and the sample j-th autocorrelations: j corr ( z t , z t j ) cov( z t , z t j ) / var( zt ) j / 0 and ˆ j ˆ j / ˆ0 We have already claimed as a fact that under the assumption that zt is a stationary and ergodic process, the sample autocovariances and autocorrelations are consistent estimators of the population autocovariances and autocorrelations.So, if zt is a serially uncorrelated process, the sample autocorrelations will converge almost surely to zero for all j > 0. (Note: Since the autocovariance and autocorrelation functions are symmetric around 0, we only have to consider j > 0.) Assume that {zt – μz} is a conditionally homoskedastic m.d.s. That is: zt = μz + t where E(t │ t-1,…) = 0 and E(t2 │ t-1,…) = 2 for all t. Then, for any positive integer p, T ˆ N (0, I p ) , ˆ [ ˆ1 ...ˆ p ]' d and, since uncorrelated normally distributed random variables are independent random variables, { T ̂ j } is asymptotically an i.i.d. N(0,1) sequence. To test for first-order serial correlation (H0: ρ1=0, vs. HA: ρ1≠0) use ˆ1 (1 / T ) N (0,1) d To test for j-th order serial correlation (H0: ρj=0, vs. HA: ρj≠0) use ˆ j (1 / T ) N (0,1) d To test H0: 1=0,2=0,…,p=0 vs. HA: ρ1≠0 or ρ2≠0… p Q T ˆ 2j 2 ( p) 1 d which is the Box-Pierce Q statistic. (This result follows from the fact that the T ̂ j ’s are asymptotically i.i.d. N(0,1).) The Ljung-Box Q statistic, QLB, p QLB = T (T 2) 1 ˆ 2j Tj is asymptotically equivalent to the B-P Q statistic 2 Q ( p) ), but seems to (i.e., Q-QLB 0 and LB p d work a little better in practice. These tests can be applied to any stationary time series to test for serial correlation. We are interested in applying them to test for serial correlation in regression disturbances. These disturbances are unobservable! It would seem natural to consider whether we can apply these tests using the residuals ( ˆ' s) from the fitted regression in place of the unobservable regression disturbances (’s). Let 1 T ~ ~ ~ ~ j j / 0 , where j t t j , for j = 0,1,2,… T j 1 and ˆ j ˆ j / ˆ0 , 1 T ˆ j ˆt ˆt j for j = 0,1,2,… T j 1 Recall, we have already modified our assumptions A.1-A.5 to restrict the disturbances to be conditionally homoskedastic. If, in addition, the regressors are strictly exogenous, T ˆ N (0, I p ) , ˆ [ ˆ1 ...ˆ p ]' d i.e., the limiting distribution of the t and Q statistics do not depend on whether we construct the sample autocorrelations using the ~ ’s or the ̂ ’s. Suppose the regressors are predetermined but not strictly exogenous. In particular, suppose: E(t │t-1,t-2,…,xt,xt-1,…) =0 and E(t2 │t-1,t-2,…,xt,xt-1,…) =2 for all t. N (0, I p ) , ˆ [ ˆ1 ...ˆ p ]' In this case, T ˆ d where Φ is a pxp matrix whose ij-th element is: Φij = E(xtt-i)’E(xtxt’)-1E(xtt-j)/2 This leads to the modified Box-Pierce Q statistic ˆ ) 1 ˆ QMBP = T ˆ ' ( I p where ̂ is the consistent estimator of equation (2.10.19) in Hayashi. , given by Under the null hypothesis of no serial correlation, QMBP Χ2(p) d