Testing for Cointegrating Relationships with Near-Integrated Data

Political Analysis - Tập 8 Số 1 - Trang 99-117 - 1999
Suzanna De Boef1, Jim Granato2
1Pennsylvania State University and Harvard University
2Michigan State University

Tóm tắt

Testing theories about political change requires analysts to make assumptions about the memory of their time series. Applied analyses are often based on inferences that time series are integrated and cointegrated. Typically analyses rest on Dickey—Fuller pretests for unit roots and a test for cointegration based on the Engle—Granger two-step method. We argue that this approach is not a good one and use Monte Carlo analysis to show that these tests can lead analysts to conclude falsely that the data are cointegrated (or nearly cointegrated) when the data are near-integrated and not cointegrating. Further, analysts are likely to conclude falsely that the relationship is not cointegrated when it is. We show how inferences are highly sensitive to sample size and the signal-to-noise ratio in the data. We suggest three things. First, analysts should use the single equation error correction test for cointegrating relationships; second, caution is in order in all cases where near-integration is a reasonable alternative to unit roots; and third, analysts should drop the language of cointegration in many cases and adopt single-equation error correction models when the theory of error correction is relevant.

Từ khóa


Tài liệu tham khảo

10.1017/S0007123400007420

10.1093/pan/4.1.249

10.1093/biomet/74.3.535

10.1093/pan/4.1.127

MacKinnon, 1990, Critical Values for Cointegration Tests

Hendry David , Neale Adrian , and Ericsson Neil . 1990. “PC-NAIVE: An Interactive Program for Monte Carlo Experimentation in Econometrics,” Version 6.01.

10.1093/0198283164.001.0001

10.1111/j.1468-0343.1994.tb00099.x

10.2307/1911990

10.2307/1913236

Elliott Graham , and Stock James . 1992. “Inference in Time Series Regression when the Order of Integration of a Regressor is Unknown.” National Bureau of Economic Research Technical Working Paper No. 122.

Hamilton, 1994, Time Series Analysis, 10.1515/9780691218632

10.2307/2938963

De Boef, 1999, Testing for Cointegrating Relationships when Data are Near-Integrated

10.2307/2231972

10.1111/j.1475-6765.1996.tb00648.x

10.2307/2991774

10.1002/jae.3950070306

10.1111/j.1468-0084.1986.mp48003005.x

O refers to the order of probability. A sequence of random variables is said to be Op (T–½) if for every v > 0 there exists an M > 0 such that P{|Xt | > (M/)} < v for all T (Hamilton 1994).

s may be thought of as a “nuisance parameter.”

All integrals range from 0 to 1.

Assumptions about exogeneity become particularly important in the single-equation setup. For much of the literature in political science that relies on error correction representations, the necessary assumption of weak exogeneity is reasonable.

The ECM is estimated with an extra lagged independent variable to break homogeneity.

It may be that the absence of cointegration will be caught in the second step, but this is not known and we can provide no evidence in this regard.

For q to equal 0, β must equal 1. (In all other cases β is set to 0.5.) In this case, we chose unit variances so that the variance ratio, s, equals 1. Analytical results indicate that the statistics are invariant to s in this case. See the Appendix.

The results below should apply equally for larger samples and values of ρ closer to 1. For ρ = 0.95, we would falsely accept the null 50% of the time in samples as large as 130.

The derivation of s is given by De Boef and Granato (1999, Appendix, Part C).

Under the alternative hypothesis of cointegration, or near-cointegration, wt is a stationary process so standard asymptotic results apply. The distribution of both tests under the null hypothesis have been derived elsewhere and relevant results are presented in the Appendix.

To see the common factor restriction, express yt in terms of wt as yt = xt + wt . Nowgiven Eq. (11), rewrite wt as Wt = (1 + b)wt –1 + ψ. Substituting into the equation for yt , we have yt = xt + (1 + b)wt –1 + ψ. Finally, substitute for wt – lagged one period and collect terms: yt [1 – (1 + b)L] = [1 – (1 + b)L]xt + ψ, where L is a lag operator. Both yt and xt share a common factor in this representation, where the common factor is given by [1 – (1 + b)L] (Hendry and Mizon 1978). This restriction is not made in the ECM t test. In other words, the ECM t test does not impose a common factor restriction.

We assume that α = 0 without loss of generality.

The pretest analyses proceed in the same manner as follows below for η.

10.2307/2232053

Rejection frequencies are unaffected by the variance of the error term.

We might prefer to think of tests for cointegration in the context of near-integrated processes as tests to distinguish significant long-run relationships from spurious relationships rather than as tests for cointegration, but given our definition of near-cointegration, they are the same in finite samples.

The ECM representation may be applicable when the data are stationary as well. Hendry (1995) elaborates on the versatility of ECM representations of stationary data.

The ADL(1,1) model can be written as yt = α + ϕyt– 1 + π0 xt + π 1 xt –1 + εt , where λ = φ – 1, β = π0, and π0 + π1 + φ – 1 = α from the ECM representation. The DGP can thus be thought of as a process whose levels are caused both by its own past values and the current and past values of other variables. Alternatively, we can think of the DGP as a process whose changes are caused by past changes and the distance the series is from its long-run relationship (i.e., whether approval is too high or too low for current levels of economic conditions.).

The assumption of weak exogeneity is critical. It is also a reasonable assumption in many cases in which this type of DGP has been proposed in political science. Beck (1992) notes that in many cases the “symmetric treatment” of political time series does not make sense and that assuming weak exogeneity is reasonable. However, one should always test for weak exogeneity. Weak exogeneity cannot be tested directly, but tests for parameter constancy are the usual approach (Engle et al. 1983).

Even longer time series may be needed to distinguish near-integrated prcesses from unit root processes when the series is estimated, as in the case of tests on the residuals from a cointegrating regression.

We use PC-NAIVE (Hendry et al. 1990) to conduct the Monte Carlo simulations.

The form of the test on the residuals from the cointegrating regression is the same as that used in the pretest phase; however, different critical values are used to reflect the fact that the residual series is estimated.

The cointegrating regression may contain multiple independent variables as well as trends or events. In cases where the relationship is symmetric, the choice of which variable to treat as (in)dependent is irrelevant.

This is a special case of cointegration and the one we consider.

Alternative representations of cointegrating relationships are often used in economics. These include the Johansen vector error correction model (1988) and the Engle and Yoo three-step estimator (1991), but these techniques have not been widely used in political science. But see, for example, Granato and West (1994) or Clarke and Stewart (1994, 1995) Testing for Cointegrating Relationships

For an exchange on this debate see Beck (1992), Williams (1992), and Smith (1992).

The results of Kremers et al. (1992), with T = 20, show a little less power—91.6 and 94.3% rejection frequencies for p = 0.05 for MacKinnon and t critical values, respectively.

10.1017/S0007123400007134

Krause, 1998, Rivalry, Reciprocity, and the Dynamics of Presidential-Congressional Institution Building, 54, 325

10.1016/0165-1889(88)90041-3

The true DGP imposes homogeneity, that is, yt = xt in equilibrium. In practice, this assumption may be unreasonable. However, the results generalize in the absence of homogeneity. Note that the estimates of λ and β are unaffected whether we estimate Δyt = α + βΔxt + λ(yt– 1 – xt– 1) + γxt –1 + εt or Δyt = α + βΔxt + λ(yt –1 – θxt–1) + εt . DGPs without homogeneity can thus be similarly estimated by including an additional xt –1 term on the right (Banerjee et al. 1993).

Banerjee et al. (1986) developed and Kremers et al. (1992) corrected the distributional results for the ECM t ratio.

10.2307/2082609

10.2307/2600607

10.2307/2111780

Macroideology is measured as the percentage of respondents claiming to be liberal from the total number of respondents claiming to be either liberal or conservative. It may be thought of as the balance of liberal, relative to conservative, ideological self-placement.

Dickey, 1979, Distribution of the Estimators for Autoregressive Time Series with a Unit Root, Journal of the American Statistical Association, 84, 427

10.1017/S000712349721015X

Engle, 1991, Cointegrated Economic Time Series: An Overview with New Results, 237

The mean and variance of aunit root process depend on time, so that the mean does not converge and the series’ variance tends toward infinity. All references to integrated series in this paper refer more specifically to unit root processes.

10.2307/2586310

Other tests for cointegration include tests in the context of the Engle-Yoo three-step procedure (1991) and the Johansen (1988) procedure. If weak exogeneity appears to be an unreasonable assumption, then these tests take simultaneity into account. See Granato and West (1994) and Krause (1998) for applications of the Johansen procedure and Clarke and Stewart (1994, 1995) for use of the Engle-Yoo three-step procedure.

We fix the variance of σε to 1 and vary σμ to simulate different values of q.

Adding additional lagged values of changes in the residuals to pick up any autocorrelation in the DF t test, using the augmented Dickey–Fuller test (ADF), does not provide any significant improvement on the inferences drawn from the DF t test. This follows from the fact that the omitted dynamics in the DF t test are from xt in the cointegrating regression itself, and not from lagged changes in the residuals. Thus the residuals from the cointegrating regression can be white noise even if the common factor restriction is not valid. As a result, adding lagged changes in the residuals—using the augmented Dickey–Fuller test—will not help to solve the problem. We did repeat the experiments using the ADF t test to verify this claim. In none of the six sets of experiments does the ADF perform significantly better than the DF t test. At most, the true null hypothesis is rejected with a 2% lower frequency. This occurs with the smallest ρ and the larger sample—as intuition would suggest. When ρ is near 1, the rejection frequencies (using either set of critical values) differ by an average of less than 0.20. It is also the case that the test is more conservative whether the null hypothesis is true or false so that inferences will be neither clearly better nor worse using the ADF t test. Clearly, however, if the DGP is more general than that proposed here, the analyst should check the residuals from the DF t test to see if additional lags need to be included in the test.

10.1017/S0266466600009993

10.1093/pan/4.1.229

10.1016/0304-4076(93)01562-Z

10.1093/pan/4.1.237

10.2307/1911357

De Boef, 1999, Modeling Near-Integrated Data

10.1093/0198288107.001.0001

10.1111/j.1468-0084.1996.mp58003007.x

10.2307/1910998

10.2307/2111734

10.2307/1911521

It is important to note that the critical values for these test statistics vary given the presence of a constant or a trend in the cointegrating regression and the sample size. Further, these values are distinct from the critical values used when testing an individual series for a unit root. There is considerable controversy on unit root pretesting (Stock 1994; Elliott and Stock 1992). The critical values reported by MacKinnon (1990) are considered superior to those reported by Dickey and Fuller (1979) and Engle and Granger (1987).