Abstract
This paper examines the linearity tests against smooth transition autoregressive models under the null hypothesis of a random walk with drift. The results show that the limiting distribution of the Wald-type statistic \(W\) proposed by Teräsvirta (1994) and the\(W_T\) statistic proposed by Harvey and Leybourne (2007) follow the standard \({\chi }^2\) distribution, whereas the robust \(W_\lambda \) statistic proposed by Harvey et al. (2008) diverges at a rate of \(T\). The finite sample simulations show \(W\) is oversized, and the robust \({W_T^{*}} \) test proposed by Harvey and Leybourne (2007) has better power performance despite being slightly undersized. Therefore, the robust \({W_T^{*}} \) statistic is recommended to be used as a conservative test in practical applications.
Similar content being viewed by others
Notes
For brevity, we introduce these two tests briefly. Interested readers are referred to the corresponding papers.
For the high order STAR models, the results are generalized readily.
Although we impose a strict assumption of i.i.d errors, we can extend our results to cases where the data-generating process has serial correlation in a very similar manner to the ADF test, and this extension should not affect the asymptotic distributions of the three tests. Theorem 1 rules out the presence of heteroscedasticity in the conditional second moments of errors, and we leave this possible extension to future research.
References
Harris D, McCabe BP, Leybourne SJ (2003) Some limit theory for autocovariances whose order depends on sample size. Econ Theory 19:829–864
Harvey DI, Leybourne SJ (2007) Testing for time series linearity. Econ J 10:149–165
Harvey DI, Leybourne SJ, Xiao B (2008) A powerful test for linearity when the order of integration is unknown. Study Nonlinear Dyn Econ 12:1–23
Kiliç R (2004) Linearity tests and stationarity. Econ J 7:55–62
Kruse R (2011) A new unit root test against ESTAR based on a class of modified statistics on a class of modified statistics. Stat Papers 52:71–85
Luukkonen R, Saikkonen P, Teräsvirta T (1988) Testing linearity against smooth transition autoregressive models. Biometrika 75:491–499
Teräsvirta T (1994) Specification, estimation, and evaluation of smooth transition autoregressive models. J Am Stat Assoc 89:208–218
White H (1984) Asymptotoic theory for econometricians. Academic press, Orlando
Acknowledgments
The research was funded by Grants of the National Social Science Fund of China (No.13CJY011).
Author information
Authors and Affiliations
Corresponding author
Appendix
Appendix
1.1 Proof of Theorem 1
(a) According to Eq. (2) and the assumption of Theorem 1, when the true DGP is a random walk with drift, \(y_t =a_0 +y_{t-1} +\varepsilon _t \). This equation implies that \(y_t =a_0 t+y_0 +\xi _t ,\xi _t =\sum {\varepsilon _t } \). Thus, we obtain
Let \(\Upsilon _1 ,\tilde{\Upsilon }_1 \) be the scaling matrix,\(\Upsilon _1 =diag\,(T^{1 / 2},T^{3 / 2},T^{5 /2},T^{7 /2},T^{9 / 2})\), \(\tilde{\Upsilon }_1 =diag(T^{5 /2},T^{7 /2},T^{9 /2})\), and \(\tilde{\Upsilon }_1 \mathbf{R}_1 =\mathbf{R}_1 \Upsilon _1 \), \({\mathbf{R}}_1 =\left[ {\begin{array}{llllll} 0&{} 0 &{}1&{} 0 &{}0 \\ 0&{} 0&{} 0 &{}1 &{}0 \\ 0&{} 0&{} 0&{} 0&{} 1 \\ \end{array}} \right] \). Let X be the matrix of independent variables in Eq. (2), \({\mathbf {\beta }}\) the coefficient vector, \(\mathbf{b}_T \) the OLS estimate of \({\mathbf {\beta }}\), and \({\mathbf {\varepsilon }}\) the error vector. Thus, we obtain
All h \(_{1}\) elements follow a Gaussian distribution, i.e.,
Here, only the proof of \(T^{-3/2}\sum {y_{t-1} } \varepsilon _t \Rightarrow N\left( 0,\frac{a_0^2 }{3}\sigma ^2\right) \) is given, and others can be readily obtained.
Clearly, \(t /T\varepsilon _t \) is a martingale difference sequence with a variance of \(\sigma _t^2 =E(t/ T\varepsilon _t )^2={t^2} /T^2\sigma ^2\). Some conditions are satisfied, i.e., \(\frac{1}{T}\sum {\sigma _t^2 } =\frac{1}{T}\sum {{t^2} /T^2\sigma ^2} \rightarrow \frac{1}{3}\sigma ^2\), \(\frac{1}{T}\sum {(t/T\varepsilon _t )^2} \buildrel p \over \longrightarrow \frac{1}{3}\sigma ^2\), and \(E(t/T\varepsilon _t )^r<\infty \) for some \(r>\)2 and all t. Thus, according to White (1984, corollary 5.25), we obtain \(T^{-3 /2}\sum {y_{t-1} } \varepsilon _t \Rightarrow N(0,\frac{a^2}{3}\sigma ^2)\).
Consider the joint distribution of the h \(_{1 }\)elements. Any linear combination of these five elements takes the following form:
Moreover,\(\left( \lambda _1 +\lambda _2 \frac{at}{T}+\lambda _3 \frac{a^2t^2}{T^2}+\lambda _4 \frac{a^3t^3}{T^3}+\lambda _5 \frac{a^4t^4}{T^4}\right) \varepsilon _t \) is a martingale difference sequence with a positive variance given by
and \(\frac{1}{T}\sum {\sigma _t^2 } \rightarrow \sigma ^2 \mathbf{{\lambda }^{\prime }Q}_1 {\mathbf {\lambda }}\). Furthermore,
where \({\mathbf {\lambda }}=\left[ {\lambda _1 \lambda _2 \lambda _3 \lambda _4 \lambda _5 } \right] ^\prime \). Thus, any linear combination of the five h \(_{1}\) elements is asymptotically Gaussian, which implies a joint Gaussian distribution of h \(_{1 }\)according to the Cramer-Wold theorem. Thus, \({\mathbf{h}}_1 \Rightarrow N( {{\mathbf{0}},\sigma ^2{\mathbf{Q}}_1 })\) and
The limiting distribution of statistic \(W\) can be derived by
where \(s_T ^2\) is the sample estimate of \(\sigma ^2\), and \({\mathbf{R}}_1 \Upsilon _1 ({\mathbf{b}}_T\! -\!{{\mathbf {\beta }}})\equiv {\mathbf{z}}\Rightarrow N({\mathbf{0}},\sigma ^2{\mathbf{R}}_1 {\mathbf{Q}}_1^{-1} {\mathbf{{R}^{\prime }}}_1 )\). Therefore,\(W\Rightarrow \chi ^2(3)\).
(b) According to the proof of (a), the limiting distribution of the Wald statistic \(W_{0}\) can be readily derived as \(\chi ^2(2)\).
Under the assumption of Theorem 1, the unrestricted model and restricted model of Equation (4) are misspecified because neither nests the true DGP. This result in the population value \(\gamma _2 ,\gamma _3 \) is non-zero, and thus, the OLS estimator \(\hat{\gamma }_2 ,\hat{\gamma }_3 \) converges in probability to the non-zero constants. As a result, \({RSS_1^r } /{RSS_1^u }=O_p (1)\), but it does not converge in probability to 1. Therefore,
The robust Wald statistic is \(W_\lambda =(1-\lambda )W_0 +\lambda W_1 \), where \(\lambda (U,S)=\exp ( {-g( {\frac{U}{S}})^2})\), \(U\) and \(S\) denote the standard ADF unit root statistic and the nonparametric stationarity statistic proposed by Harris et al. (2003), respectively, and \(g\) is a finite positive constant derived through simulation.
According to the Lemma 1 developed by Harvey et al. (2008),
Under the null hypothesis of \(H_{0,1} :\gamma _2 =\gamma _3 =0\), \(y_{t}\) is a linear unit root, and then\(\left| U \right| =O_p (1)\) because it converges to the Dickey-Fuller limit distribution; under the alternative \(H_{1,1} :\gamma _2 \ne \text{ and }/\text{ or } \gamma _3 \ne 0\), \(y_{t}\) is nonlinear and a unit root, and then \(\left| U \right| =O_p (1)\) according to Harvey et al. (2008). Similarly, according to Harris et al. (2003), under both \(H_{0,1}\) and \( H_{1,1}\), \(\left| S \right| \)diverge to \(+\infty \), and thus, under both \(H_{0,1}\) and\( H_{1,1}^{, }\lambda \buildrel p \over \longrightarrow 1\) and \(W_\lambda =O_p (T)\).
(c) Harvey and Leybourne (2007) construct an auxiliary regression in terms of the observed \(y_{t}\)
In terms of 21, the null hypothesis of linearity test can be stated as
They define a Wald statistic \(W_{T}\) under the null of \(H_0 \) . Under the assumption of Theorem 1, the limiting distribution of \(W_T \) can be derived directly as \(\chi ^2(4)\) using the proof of (a) and standard asymptotic results of Wald tests.
Rights and permissions
About this article
Cite this article
Yan, R., Zhang, L. Linearity tests under the null hypothesis of a random walk with drift. Stat Papers 57, 407–418 (2016). https://doi.org/10.1007/s00362-015-0659-1
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00362-015-0659-1