Biometrika (1984), 71, 3, pp. 599-607 599 Printed in Great Britain Testing for unit roots in autoregressive-moving average models of unknown order By SAID E. SAID Department of Mathematics, Kast Carolina University, Greenville, North Carolina, U.S.A. anp DAVID A. DICKEY Department of Statistics, North Carolina State University, Raleigh, North Carolina, U.S.A. Summary Recently, methods for detecting unit roots in autoregressive and autoregressive- moving average time series have been proposed. The presence of a unit root indicates that the time series is not stationary but that differencing will reduce it to stationarity. The tests proposed to date require specification of the number of autoregressive and moving average coefficients in the model. In this paper we develop a test for unit roots which is based on an approximation of an autoregressive-moving average model by an autoregression. The test statistic is standard output from most regression programs and has a limit distribution whose percentiles have been tabulated. An example is provided. Some key words: Mixed model; Nonstationary; Time series; Unit root. 1. Ivrropverron Box & Jenkins (1970) discuss a class of models known as arma (p,d,q) models. The term Arima (p,d,q) refers to an autoregressive-moving average model of the form By ty Ly Oy Dy p = +B, Gato + Byer ge where Z, is the order d difference of the data and ¢, is a white noise sequence, These have become quite popular for modelling time series data. Methods are available for testing the p autoregressive and q moving average coefiicients for significance. The value of d indicates the amount of differencing necessary to reduce the series to stationarity. The number d also equals the number of unit roots in the characteristic equation for the time series. While much has been written about estimation when d = 0, relatively little has been written about cases with d > 0. D. A. Dickey, in his Iowa State University Ph.D. thesis, Dickey & Fuller (1979) and Hasza & Fuller (1979) discuss cases with q = 0. Dickey & Said (1981) discuss a method for testing the hypothesis Ho: d= 1 when p and q are known. Unfortunately, p and q are usually unknown and it is necessary to determine d before estimating p and q if we use the approach described by Box & Jenkins (1970). In the present paper we show that it is possible to approximate an ariMa (p, 1,q) by an autoregression whose order is a function of the number of observations n. Using least squares to estimate coefficients in this autoregressive approximation produces statistics whose limit distributions are the same as those tabulated by Dickey and listed by Fuller (1976, p.373). Thus it is possible to test the null hypothesis Ho: d = 1 without knowing p org.600 Sarp E. Satp anp Dayip A. DickEy The development is illustrated for p = and normal errors then extended to the general p,q case with independent identically distributed errors, 2. Estimation PROCEDURE We first consider the model described by Y= pY-1+%, aL,» +e + Be» (b= (2:1 wh 1) where it is assumed that || <1, |/| <1, Yo =0 and {;} is a sequence of independent identically distributed random variables which, for the moment, we will assume to be normal. If |p| <1 then, except for transitory start up effects, Y, is a stationary ARMA (2, 1) series in the terminology of Box & Jenkins (1970). If p = 1 the series (21) is an anima (I, 1,1) process. We consider p = 1 as a null hypothesis to be tested. Notice that —BY(Z,j- 92,51) and it follows that Ye= Vann = (P= VW) (Yes) + (0+ B) (Ze 1 Ba + BP Zp 3.0) 4 (22) Under the null hypothesis that p = 1, we see that Z, = Y,— Y,-;. This motivates us to estimate the coefficients in (22) by regressing the first difference ¥,= Y,—Y,-, on Yy-4, Yy-1,--., Y;-4 where & is a suitably chosen integer, To get consistent estimates of the coefficients in (2-2) it is necessary to let & be a function of n. We shall assume that n-'?& — 0 and that there exist ¢ > 0,r > 0 such that ck > n'", 3. Derinrrions AND Norarion Let do = p—1 and d; = (7+ )(—f)'~! (i > 0) be the coefficients in (22). Let X= Zpntyee Bran), Ur = (Yr X)) (do,dy,...,d,). Now consider a truncated version of (2:2) Y= (BN) Feat (tt B) {Ze 1 Boa + B ya (BY Za} +e. (31) Notice that ey is not a white noise series. In fact ¢, = Y,—U;d. Define a = (8, 4, to be the vector of estimated coefficients in the regression of ¥, on ¥,1, Yy-ts 05 Yenes hereafter referred to as regression (3:1). Further, define a sum of squares and cross- products matrix Ro and a normalizing matrix D, by and d’ E¥2, U¥X aa EY1X, EX,X; | a D, = diag {(n—k)“4, (n—h)“4, ..., (n=) 4}. shall develop the limit distribution of Dy —d) = (Dy Ro Dy)" Dy Y Uren. (33)Unit roots in autoregressive-moving average models 601 ‘The first step will be to show that D, Ro D,—R converges to zero, where R is a block diagonal matrix. Let y,(i—j) = H(Z,-;Z,—3) be the autocovariance function of the stationary series Z,, I = y.(i—j), where Tis a kx k matrix, and y! = {7,(1),....7(k)}. It is well known (Fuller, 1976, p.239), that (n—k)7! X,X,>T, (nk)! Y XZ, 49, wf ehh where the convergence is in probability. Finally, let W, = e; +... +¢,and define the block diagonal matrix R by R= diag {(1—a) (1+ f)?(n—k) "7 E W?_,, PF}. (34) The ideas presented in the next section follow the lines of Berk (1974). He uses the standard Euclidean norm, || «|| = (#z)$, of a column vector « to define a matrix norm | B|| and defines | BY =sup {| Bel: le] <1} Notice that || B |? is bounded by the sum of squares of the elements of 2 and that || B || is bounded by the largest modulus of the eigenvalues of B. Using this norm we investigate the convergence of Ro. 4, AsyMPToric BQUIVALENCE OF D, Ro D, AND R In this section, we let @ = D, Ry D,— Rand show that || Q || converges in probability to 0. Let the (i, j)th element of Q be denoted by gy. Luvma 41, Under the assumptions in §3, if p=1 and nk +0 then KI QI) converges in probability to 0. Proof. First, Gir = (nk) 7{E YP —(1—a) (1. +B)? ZW}. For p =1 we have ¥,-1 =Zy+...+Z,-; 80 that en Farina = St Bea) 8ly and thus. (1-a) ¥,-y+a%,-, = (148) % e+ Bleo— es) +aZ0 (41) or Yo. = (1a) 1+ B) Wy. + O,(1). Dickey & Fuller (1979) show that EW}, =0,(n). Their arguments show that (n—k) E(qi;) < Cy, where C, is some constant. Berk (1974) shows that for some C;, (mk) Binh) EZ, Z,_ jy DY? SCz (GF = Vy rh) For i > 1, j > 1, we note that 9, = (n—k)"' 2 Z,_;Z,_;—y.(i—j). We now consider q for i=1, j>1. The vector of these qj from (32) is (n—k)-9?E ¥,_, X;, Now for some C3, (m= k) E(qi) = (nb)? EE, BY Bp—7 Vu-1 Zy-j) < Or.602 Sarp E. Sarp AnD Davip A. DickEy This follows from the facts that 1E(¥,2,)| 3 nb-a]< X_ lrld)1 = Ca < 20, IEYY St YX ly(d)| = tCy. To finish the proof, let C = max (C,, C2, C3), so that (n—k) E(q}) < C for all i, j; note that C,, C; and C do not depend on » or k. Since @ has dimension (k-+ 1) x (+1) we see B(\ QI?) < C(k+1)?/(n—k) and if k?/n + 0 then ||Q|| converges in probability to 0. Under our assumption that k3/n + 0 we have k# | Q|| converging to 0 also. Luma 42. Under the assumptions of model (2:1) with p =1 fa-a-ta+ prem s wasp =0,(1). : fet Proof. Now y W2,=e'Ae, A where e' = (¢,,...,¢,1) and the (i, j)th element of A is n—max (i, j). Thus net CAe= Y 4,23, a where {Z,} is an independent N(0,02) sequence, and (Dickey & Fuller, 1979) Yin = 47 * see? {(2n—1)"'(n—i) mn} (i= 1,...,n). Also, for any fixed i,n“,, > 4{(2i—1)n}~? > 0 as n + 00, so that, given ¢ > 0, we can find M, such that D 1 orf Wa) > a} < pr(n? yj, Zz? > M,) = pr(n-? y;,! 23 < Mz) 0. Here we have used the fact that pr(l| Ql] > 1) ean be made arbitrarily small by choice of n since p ||@| > 0. 5. CONSISTENCY OF LEAST SQUARES ESTIMATORS In §2 we propose the use of least squares autoregression to estimate p in model (2'1). We estimate p—1 by the first element of a vector @ given by expression (3:3). In §4 we looked at D, Ro D,, so it remains to consider D,¥ U,e,. We show that this vector converges to 0 and thus prove consistency. Lemma 5:1. Asn > co Proof. We have » 2 | Dy Y Usleu-e) oh 1 (nk) {E, Ysa e)}? + (nh) * x {E,Z,-Aeu—e)}. (51) é Since {Z,} is a stationary invertible RMA process, we can use (3-1) and (3-2) to see that [net x B2,1eu-2))| 0. ‘The order of the first term in (5+1) is established by expressing Y,_ and ey —e, in terms of {Z;} and the proof is thereby completed. ‘At this point we will establish the probability order of which, in combination with Lemma 5:1 and Theorem 41, will establish consistency. DS Ue aka604 Sarp E. Sarp anp Davip A. DICKEY Lemma 52. Asn > 00, |>. oie =0,(k#). Proof. Now ko fnck 2 BD, &, 0,6? = (n—k)-? HE Y,_ 1)? +(n— ky a(S 4-54) , Ne Since Z, is a stationary invertible autoregressive-moving average process (nb! x (S fey ‘) kys0) 0, Here we have used the independence of Z,-; and ¢, for j > 0. This same fact shows that BOE, Yous)? = 08D, BYE) = O(n’). Combining Lemmas 5:1 and 5-2, we see that 1,2, U,eu | = Og(k*) 62) and we thus are in #. position to prove our main result. TneoreM 5:1. Under the assumptions of model (2:1) with p =1, |4—d || converges in probability to 0. Proof. We write Dy'(a—d) = (RY R-) DE, eq + Ro" Dy E, Ope; + R7* DE, Ue —e,). (53) ‘Taking norms of the three terms above we obtain, respectively, op(1), O,(k#) and O,(n~") so that || D, (@—d) | = O,(k#). Since || D,*|| = O(n—k) and k= O(n") the proof is complete. 6. Limrr DISTRIBUTION OF UNIT ROOT TEST STATISTICS Having shown that D,'(@—d) =0,(1) we now develop the limit distribution of D,,\(@—d) which, by (5:3), is the same as that of R~! D,E,U,e,. The first element of D, (a—d) is (nk) (6-1) = (nk) PE VE nk) EY By Lemma 41 (n—h)7E ¥?_, = (1—a)-7(1 +B) EW? +0, (074), (6-1) By the arguments of Lemma 5:1, (= WE Yee = (WWE Vy 1+ 0,(0-) (62) and, if we use result (41), (=k) EY, G = (n—h) (1a) (14+ BYE Wye, +O,(n- 4). Dickey & Fuller (1979) define random variables (I”, ¢) and show that (07? n-7EW2 4, 07?n EW, 6)Unit roots in autoregressive-moving average models 605 converges in law to (I, €). If we use (61) and (6-2), the distribution of P~* € is the same as, the limit distribution of (1a) *(1-+B) (mk) (8 —do) = (1a) "(1 +B) (mn) (= 1), If we use the results of Hasza & Fuller (1979) the above limiting distribution is unchanged if the ¢, sequence is taken to be independently and identically distributed with zero mean and variance o?. Further, since R is block diagonal, the results of Berk (1974) apply to the coefficient vector (2, —dy, ..., 44 —d,). That is, the limit distribution of this part of a—d is the same whether or not we include Y,_, on the right-hand side in regression (3:1). Thus, for large n, significance tests for these coefficients are not affected by including ¥,_, in the regression Since the distribution of n(p—1) involves the unknown parameters a and f we would be unable to use p as a test for unit roots at the ideritification stage of analysis. We now show that the limit distribution of the studentized ¢ statistic associated with p does not depend on any unknown parameters, Turorem 6:1. Under the assumptions of Theorem 4-1, define t = (C;, 67)” *(p—1), where Oy, is the upper left-hand corner element of Rg, p—1 = &o—do, and 6? is the error mean square from regression (3:1). Then the limit distribution of « is the same as the distribution of rte. Proof. By our previous consistency results, 6? converges in probability to a”. Using the arguments of Lemma 4:1 we see that t—{n? (E71) 1} Fm(p—1) > 0. If we use (61) and (6-2) it follows that 1-6 (LW2 4) FEM 20 and thus t > T7#é. Percentiles of the distribution of t are given by Fuller (1976, p.373) and can be used to test the null hypothesis of a unit root. An illustration is given in §8. 7. EXTENSIONS ‘The general anita (p, 1,q) model is defined by Y= pY-1t+% (b= 1,2...) (71) 2+ § ot, et S bien (72) with Yq = 0, ¢, a normal pure noise sequence and p = 1, We assume that Z, is stationary and invertible. Thus there exists a sequence of real numbers {dj}, a number 0 <4 <1 and a number M such that ¢ = ZdjZ,-; and |d,| <.Mi (Puller, 1976, Theorem 27-2) Rearranging, we can write, in analogy to (31), Y= (PW Kp Za Za On (73) Notice that the consistency proofs in the anima (I,1,1) case depend only on the existence of exponentially decreasing bounds on the d; and so they generalize im- mediately to the higher order case,606 Samp E. Sarp anp Davip A. DickEY If p = 1, Z,-;= Y,-j— ¥,-j-1 = ¥;-; and in regression (7-3) we will refer to the least. squares coefficient on Y,_, as p—1. Summation of (7-2) on both sides as goes from 1 to n shows that, in analogy to (4-1) with p = 1, (L404... ay) Y= (148i +...+8,) Wet Vor (74) where J, is a linear combination of a finite number of Z, and ¢, and W, = ¢+...+e as before. If we use (7-4) and the arguments of §6 it follows that, for the general anima (p, 1,4) case, then in law (Lt ay +... +p) (L+ By +... + Bq) (P-L) > ToS, (75) as does the studentized statistic from the least squares regression (7'3) Finally, if the series mean ¥ is subtracted from each ¥, prior to analysis, we replace (74) by (Lt +... tay) (Ye— P) = (14, +... +B) (We ) + Ve (76) where W = n~* W,. We note that Z, = (¥,—)—(¥,-;—Y) = ¥, when p = 1. Using the results of Dickey & Fuller (1979) we get distributions similar to (7-5) except that the limit distribution percentiles now correspond to the p, and 2, tables of Fuller (1976, p. 371-3). This follows from the fact that for fixed j, 3 = 0,(n), so that the order results following (4-1) are unchanged if the series mean 7’ is subtracted prior to analysis, ‘The limit matrix R is still block diagonal. The only difference from (3°3) is that the upper left-hand element of R is now, if we use (6'1), (m—k)-7(L 40, +... +p) 2(1 + Bi +... + By)? (W,— W?), and a similar modification in (62) yields results like (7'5) 8, EXaMPLe Box & Jenkins (1970, p. 525) list 197 concentration readings from a chemical process. ‘The authors conclude that an ana (I, 1) or an ARIMA (0,1,1) model should be fitted to the data. We use our testing procedure to test the hypothesis that the model form is arma (p, 1,9). Using sas (Barr, Goodnight & Sall, 1979) we fit the model —0-1601(¥,_;— ¥)—0-4941 Y,_, -0-2919Y,_,—0-2640¥,_5 —0-2477 ¥,_ 4-0-2682 Y,_ 5-0-1888 Y,_, where the error mean square is 0-0938 and Y, = Y,— ¥,_. The coefficient standard errors are 0-0785, 0-0963, 0-0985, 0-0947, 0-0903, 0-0858 and 0-0726. The studentized statistic 2, = (00785) *(—0-1601) = —2-04 is compared to the 2, tables of Fuller (1976, p.373). We do not reject the unit root hypothesis using a one sided 10% significance level test. The limit theory, of course, does not specify the value of k for any given n. To select k, we initially fit regression (3:1) for k = 6, 7, 8, 9 and 10, Assuming that an autoregressive order 10 model gives a sufficient approximation to the data we used the standard regression F' test that the coefficients on Y,_, to Y,_y9 are simultaneously zero. The justification for this as an approximate test is the fact from §6 that Berk’s results hold for the coefficients on Y;_; in regression (3:1). The sequential sums of squares for Y,_, to Y,_19 are 0:0279, 0-0019, 0-0013 and 01341. The error sum of squares isUnit roots in autoregressive-moving average models 607 166423 with 174 degrees of freedom and we compute F't,4 = 0-43. None of the individual L tests for these 4 lags was significant, even at the 20% level. Furthermore, when regression (3'1) is fitted with k = 7, 8, 9 and 10, the t, statisties are —1-931, — 1-830, — 1-796 and —2-013 so that our unit root test is not affected over this range of k. One will now want to fit a model to the differenced data. Once this has been done, unit root tests based on known p and q are available (Dickey & Said, 1981). Reverences Bane, A. J, Goooionr, JH. & Sat, J.P. (1979). SAS User's Guide. Cary, North Carolina: sas Institute Bunk, K_N. (1974). Consistent autoregressive spectral estimates. Ann. Siatst. 2, 489-502 Box, G. E. P. & dunxins, G. M. (190). Time Series Analysis Forecasting and Control. San Francisco: Holden-Day. Dickey, D. A. & Futter, W. A. (1979). Distribution of the estimators for autoregressive time series with a unit root. J. Am. Statist. Assoc. 74, 427-31 Diokzy, D. A. & Sam, 8. B. (1981). Testing anata (p,1,q) versus anata (p+1,q). Proc. Bus. Heon. Statist. Sect, Am. Statist. Assoe., 318-22 Funune, W. A, (1976). Introduction to Statistical Time Series. New York: Wiley: Hasa, D. P.& Pour, W. A. (1979). Estimation for autoregressive processes with unit roots. Ann. Statist. 7, 1106-20, [Received November 1983. Revised February 1984]