Baixe o app para aproveitar ainda mais
Prévia do material em texto
Economics Letters 21 (1986) 21-25 North-Holland 21 ON USING DURBIN’S h-TEST TO VALIDATE THE PARTIAL-ADJUSTMENT MODEL H.E. DORAN University of New England, Armdale, N.S. W. 2351, Australia Received 20 November 1985 Final version received 18 December 1985 A Lagrange multiplier test for diagnostically checking the partial-adjustment model is presented. This test is more powerful than the commonly used Durbin h-test because a non-linear restriction on the parameters is taken into account. 1. Introduction The partial-adjustment (PA) model, originally proposed by Nerlove (1958) has had long and continued usage in empirical research. Its attraction lies in its simplicity - it involves few parameters and can be estimated by ordinary least squares (OLS). It still has an important role in situations where inadequate data make it necessary to invest in a simple model. However, if simple models are to be used to represent complex reality, it is necessary to have valid diagnostic procedures available to detect when the models are inappropriate. The PA model may be formulated as follows: y:=a+px,-,, Y,- Y-1 = (1 - Y)(y: - y,-,> + u,> (1)7(2) where Y,* is the desired or optimal level of the variable Y, (often production or investment), u, is i.i.d. N(0, a*) and 0 5 y < 1. When Y,* is eliminated from (1) and (2), the estimating equation r,=a(l-y)+P(l-Y)X,-,+vrc-1+u, (3) is obtained. A common method of diagnostically validating the PA model is to regress Y on X,_ , and Y_ , and examine the residuals for autocorrelation using Durbin’s (1970) h-statistic. There is theoretical justification for this procedure. Suppose the PA model is nested in the more general partial-adjustment-adaptive-expectations (PAAE) model, in which X,_ 1 is replaced by X:, given by x;c - x:_, = (1 -6)(X,-, - X7-i). ’ (4) ‘Eq. (4) imphes X,! = (l- s)c~$‘X,_,_~ 0165-1765/86/$3.50 0 1986, Elsevier Science Publishers B.V. (North-Holland) 22 H. E. Dorm / Validating the partial-adjustment model The estimating model then becomes y=b,+b,X,-, + b,Y,_, +&I’_, + u,, where b, = a(1 - y)(l - S), b, = P(1 - Y)(I - 61, b,=y+6, 6,= -ys, (5) and u,=u,--u,_,, O$S<l. (6) The PA model may be validated by testing the hypothesis S = 0 in (5) and (6). These equations represent a dynamic model with a first-order moving average disturbance and Breusch (1978) has shown that the h-test is asymptotically equivalent to a Lagrange multiplier (LM) test of the hypothesis 13 = 0. However, an important feature of the PAAE model is that there is a non-linear restriction on the parameters. From the definition of 6, and b, above, b, = -S(b, -S). (7) Use of Durbin’s h-test ignores this restriction and therefore results in a loss of power. The purpose of this note is to derive an LM test which does incorporate the restriction (7) and to demonstrate its relationship to the h-test. 2. A Lagrange multiplier test When x,_, is replaced by X;C = (1 - S)~~OgS’X,_,--l in (3), the PAAE model becomes v,=&+P,( .F,SX_,_‘) +P*r,-, +u,> where PO = 41- Y>, P, = P(l - Y)(l - a)? P*=u, and the symbols have been chosen so that under the null 6 = 0, p, = b, (i = 0, 1, 2). The restriction (7) is implicit in this formulation. If we define 8 = (&, &, &)‘, then the likelihood function L(6, 8, a*) is given by L= -+ ln(2s)-t Ino’---$u’U, where N is the sample size and u = (v,, v2, . . . , uN)‘. The Lagrange multiplier statistic LM has the form where [see, for example, +* = 0, and ‘*’ indicates asymptotically distributed the form H. E. Dorm / Validating the partial-adjumnent model 23 Breusch and Pagan (1980)] 3,, = -plim(N-‘a*L/&#+Q~) with +, = S, evaluation under the null hypothesis. When the null is true, LM is as x:. Doran (1985) has shown that when testing the PA model, LM takes plim( N- 'RSS ), (10) where RSS is the residual sum of squares which is obtained when aC/&V is regressed on afi/M. From (8) ac - --#= -p,x_*, ai ^ 1 a6 P, a; -zz as -,O’,=,im,, a= -[jt x-1, y-113 j being a vector of ones. Thus, in finite samples LM = ( NC?‘)-‘&( iYX_,)2/( N-‘& RSS*), where RSS* = D*‘O*,D* being the residual vector obtained when X_, is regressed on j, X_ , and Y -1’ That is, on writing NG* = 26, LM = N( D’Xp,)*,‘[ (;‘;)( fi*‘O*)] . Also, because D* can be written in the form D* = X_, - ( j, X_ t, Y_ , )? for some vector P, fi’fi* = $X_, - C’( j, X_,, Y_,)F = 2X_,, by the OLS properties of D, which ensure that C’j = D’X-, = D’Y-, = 0. Finally then, LM= Nr’, where r2 is the ‘R* ’ which is obtained when i3 is regressed on O*. The LM test of the PA model thus consists of the following simple procedure: (11) (i) regress Y on X_, and Y-, to obtain the residuals D, (ii) regress X_, on X_, and Y_, to obtain the residuals G*, (iii) regress G on 6* to obtain r2, and (iv) compare Nr* with Xf. It should be noted that the additional information that S 2 0 can be utilised by constructing a one-sided test in which N’/* r is compared with the standardised normal random variable, with the null being rejected when N112 r exceeds the positive critical value. The above procedure can be justified intuitively by recognising that the three regressions are equivalent to estimating the coefficient & in the multiple regression model yI=Po+P,X,-1 +P2Y-l+&xI-2+~,. 24 H. E. Doran / Valrdating the part&adjusrmenr model A ‘large’ r2 is directly related to a ‘large’ &. Now from (S), the PAAE model may be written as ~=~,+~,X,_,+~,~_,+~,SX,_,+ (higherordersina)+u,, and the coefficient of XI+, is seen to be proportional to S. 3. Connection between the LM and h tests In order to recognise the role that the non-linear restriction (7) plays in the LM statistic, it will be convenient to consider the PAAE model in the form (5) and (6). The presence of the restriction implies that au/a8 is non-zero. Expressing the moving-average process (6) in the form u=M(6)u, 2 where M,, = 1, M, ,_] = -6 and all other elements of M are zero, we have (12) au awl au -= -I___ as as u+M a6 _ _M-l$++M-‘~. Thus. where V, ,p1 = 1 and otherwise y, = 0. Substituting into (lo), (13) where RSS is obtained by regressing Vii + i3G/at3 on Z = (j, X_ ,, Y_,). This expression shows how ait, enters both the numerator and denominator of LM. Suppose now that the restriction is ignored, and consequently, au/a8 is taken to be zero. Then the LM statistic, call it LM*, is given by LM* = (;l’Vii)*/[ Ns’plim( N-'RSS*)] , where RSS is obtained by regressing Vii on Z. That is, RSS* = ?I”[ I - Z( Z’Z)-‘Z’] Vii. From the definition of V, N N-l B’Vfi = c ii_,& = Ne2;, plimN_‘( ii’V’V;i) = plimN_’ c n: = u2, t=2 I=1 plimN_‘( Z’Vii) = [0, 0, u’]. *It is clear M(g) = I, and hence ic = 8. Thus, plimN-‘RSS* = a’plim 1 [ For finite samples, we use H. E. Dorm / Validating the partial-adjustment model 25 (15) When the square-root of LM* is taken, enabling one-sided tests to be performed, we have precisely Durbin’s h-statistic. Preliminary Monte Carlo studies reported in Doran (1985) indicate that, at least in cases in which X, has strong autocorrelation, the loss of power incurred from using the h-statistic instead of the LM statistic (11) is very considerable. References Breusch, T.S., 1978, Testing for autocorrelation in dynamic linear models, Australian Economic Papers 17, 334-355. Breusch, T.S. and A.R. Pagan, 1980, The Lagrange multiplier test and its application to model specification in econometrics, Review of Economic Studies 47, 239-253. Doran, H.E., 1985, Diagnostic tests for the partial adjustment and adaptive expectations models, University of New England working papers in econometrics and applied statistics,no. 17 (University of New England, Armidale). Durbin, J., 1970, Testing for serial correlation in least-squares regression when some of the regressors are lagged dependent variables, Econometrica 38, 410-421. Nerlove, M., 1958, The dynamics of supply: Estimation of farmers’ response to price (The Johns Hopkins University, Baltimore, MD).
Compartilhar