Abstract
In the literature, most works of the specification tests focus on the problem with one-dimensional response or fixed multidimensional responses. In this paper, we develop a new specification test for the parametric models with non-stationary regressor under multidimensional setup, where the dimension of responses may tend to infinity, which fills a gap in the literature. The theoretical results about the asymptotic properties of the proposed test are studied and the optimal rate of the local departure under the alternative hypothesis is also given which ensures the models underpinning by the null and alternative hypotheses can be differentiated. Some simulation studies are done to evaluate the performance of the proposed test with the finite sample. Besides, a real data example based on the US aggregate consumers’ consumption data is employed to illustrate the performance. The results of simulation studies and real data analysis both demonstrate the efficiency of our proposed method.
Similar content being viewed by others
References
Chen J, Gao J, Li D, Lin Z (2013) Nonparametric specification testing in nonlinear and nonstationary time series models: theory and practice. SSRN Electron J. https://doi.org/10.2139/ssrn.2235356
Cho H, Kim S (2017) Model specification test in a semiparametric regression model for longitudinal data. J Multivar Anal 160:105–116
Dette H (1999) A consistent test for the functional form of a regression based on a difference of variance estimators. Ann Stat 27(3):1012–1040
Dong C, Gao J (2018) Specification testing driven by orthogonal series for nonlinear cointegration with endogeneity. Econom Theory 34(4):754–789
Dong C, Gao J, Tjøstheim D, Yin J (2017) Specification testing for nonlinear multivariate cointegrating regressions. J Econom 200(1):104–117
Eubank RL, Hart JD (1992) Testing goodness-of-fit in regression via order selection criteria. Ann Stat 20(3):1412–1425
Fan Y, Li Q (1996) Consistent model specification tests: omitted variables and semiparametric functional forms. Econometrica 64(4):865–890
Fan Y, Li Q (2002) A consistent model specification test based on kernel sum of squares residuals. Econom Rev 21(3):337–352
Gao J (2012) Model specification between parametric and nonparametric cointegration. SSRN Electron J. https://doi.org/10.2139/ssrn.2140996
Gao J, Casas I (2008) Specification testing in discretized diffusion models: theory and practice. J Econom 147(1):131–140
Gao J, Dong C (2012) Specification testing driven by orthogonal series in nonlinear and nonstationary time series models. SSRN Electron J. https://doi.org/10.2139/ssrn.2175290
Gao J, King M, Lu Z, Tjøstheim D (2009a) Nonparametric specification testing for nonlinear time series with nonstationarity. Econom Theory 25:1869–1892
Gao J, King M, Lu Z, Tjøstheim D (2009b) Specification testing in nonlinear and nonstationary time series autoregression. Ann Stat 37(6B):3893–3928
Gao J, Wang Q, Yin J (2011) Specification testing in nonlinear time series with long-range dependence. Econom Theory 27(2):260–284
Gao J, Chen J, Li D, Lin Z (2015) Specification testing in nonstationary time series models. Econom J 18(1):117–136
Geman D, Horowitz J (1980) Occupation densities. Ann Probab 8(1):1–67
Guerre E, Lavergne P (2005) Data-driven rate-optimal specification testing in regression models. Ann Stat 33(2):840–870
Hardle W, Mammen E (1993) Comparing nonparametric versus parametric regression fits. Ann Stat 21(4):1926–1947
Hart J, Wehrly TE (1992) Kernel regression when the boundary region is large, with an application to testing the adequacy of polynomial models. J Am Stat Assoc 87(420):1018–1024
Hong Y, Lee YJ (2013) A loss function approach to model specification testing and its relative efficiency. Ann Stat 41(3):1166–1203
Koul H, Stute W (1998) Lack of fit tests in regression with non-random design. Appl Stat Sci III 25:53–69
Koul HL, Song W, Liu S (2014) Model checking in tobit regression via nonparametric smoothing. J Multivar Anal 125:36–49
Ladd GW (1965) Experiments with autoregressive error estimation. Res Bull (Iowa Agriculture and Home Economics Experiment Station) 35(533):1
Lin Z, Li Q, Sun Y (2014) A consistent nonparametric test of parametric regression functional form in fixed effects panel data models. J Econom 178(1):167–179
Revuz D, Yor M (1999) Continuous martingales and Brownian motion. Springer, Berlin
Stute W (1997) Nonparametric model checks for regression. Ann Stat 25(2):613–641
Stute W, Zhu LX (2002) Model checks for generalized linear models. Scand J Stat 29(3):535–545
Stute W, Manteiga WG, Quindimil MP (1998a) Bootstrap approximations in model checks for regression. J Am Stat Assoc 93(441):141–149
Stute W, Thies S, Zhu LX (1998b) Model checks for regression: an innovation process approach. Ann Stat 26(5):1916–1934
Sun Z, Wang Q, Dai P (2009) Model checking for partially linear models with missing responses at random. J Multivar Anal 100(4):636–651
Sun Z, Chen F, Zhou X, Zhang Q (2017) Improved model checking methods for parametric models with responses missing at random. J Multivar Anal 154:147–161
Toda AA (2011) Income dynamics with a stationary double pareto distribution. Phys Rev E Stat Nonlinear Soft Matter Phys 83(4):046122
Vetter M, Dette H (2012) Model checks for the volatility under microstructure noise. Bernoulli 18(4):1421–1447
Wang Q, Phillips PC (2009) Asymptotic theory for local time density estimation and nonparametric cointegrating regression. Econom Theory 25(3):710–738
Wang Q, Phillips PCB (2012) A specification test for nonlinear nonstationary models. Ann Stat 40(2):727–758
Wang Q, Wu D, Zhu K (2018) Model checks for nonlinear cointegrating regression. J Econom 207(2):261–284
Acknowledgements
This research work is supported by NSFC No. 11801034 and NSAF No. U1430125. The authors are grateful to Professor Ping-Shou Zhong for his kindly help.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix
Appendix
Here, the proofs of Theorems 1–3 are presented. In order to prove these theoretical results, some lemmas are needed to be shown firstly.
Lemma 1
Under Assumptions A-D, we have
-
(1)
\(\sum \nolimits _{t=1}^n\dfrac{1}{d_t}\sim \dfrac{n^{\frac{\mu }{2}}}{p}\), \(n\rightarrow \infty .\)
-
(2)
\(\sum \nolimits _{t=2}^n\sum \nolimits _{s=1}^{t-1}\dfrac{1}{d_{ts}}\dfrac{1}{d_s}<Cn^{\mu }\), \(n\rightarrow \infty \), C is a constant.
-
(3)
\(\sum \nolimits _{k=3}^n\sum \nolimits _{t=2}^{k-1}\sum \nolimits _{s=1}^{t-1}\dfrac{1}{d_{kt}}\dfrac{1}{d_{ts}}\dfrac{1}{d_s}<Cn^{\frac{3}{2}\mu }\), C is a constant.
Proof
-
(1)
$$\begin{aligned} \lim _{n\rightarrow \infty }\dfrac{\sum \nolimits _{t=1}^n\dfrac{1}{d_t}}{n^{\mu /2}/p}=\lim _{n\rightarrow \infty }\dfrac{\sum \nolimits _{t=1}^n\dfrac{1}{t^{1-\mu /2}log(t)}}{n^{\mu /2}/log(n)}=\lim _{n\rightarrow \infty }\dfrac{log(n)}{\frac{\mu }{2}log(n)-1}=\dfrac{2}{\mu }. \end{aligned}$$
Since \(\dfrac{2}{\mu }>0\) is a constant, we have \(\sum \nolimits _{t=1}^n\dfrac{1}{d_t}\sim \dfrac{n^{\frac{\mu }{2}}}{p}\).
-
(2)
$$\begin{aligned}&\sum \limits _{t=2}^n\sum \limits _{s=1}^{t-1}\dfrac{1}{d_{ts}}\dfrac{1}{d_s}=C_4 \sum \limits _{t=2}^n\sum \limits _{s=1}^{t-1}\dfrac{1}{log(t)(t-s)^{1-\mu /2}}\dfrac{1}{log(s) s^{1-\mu /2}} \\&\quad<C_4 \sum \limits _{t=2}^n\sum \limits _{s=1}^{t-1}\dfrac{1}{(t-s)^{1-\mu /2}}\dfrac{1}{s^{1-\mu /2}}\sim C_4 \sum \limits _{s=1}^{n-1}\dfrac{1}{s^{1-\mu /2}}\displaystyle {\int _{s+1}^n (t-s)^{\mu /2-1}\mathrm {d}t} \\&\quad <\dfrac{2}{\mu }C_4 \sum \limits _{s=1}^{n-1} s^{\mu /2-1}n^{\mu /2}\sim Cn^{\mu }. \end{aligned}$$
-
(3)
$$\begin{aligned}&\sum \limits _{k=3}^n\sum \limits _{t=2}^{k-1}\sum \limits _{s=1}^{t-1}\dfrac{1}{d_{kt}}\dfrac{1}{d_{ts}}\dfrac{1}{d_s}\le C_5\sum \limits _{k=3}^n\sum \limits _{t=2}^{k-1}\sum \limits _{s=1}^{t-1}\dfrac{1}{(k-t)^{1-\mu /2}}\dfrac{1}{(t-s)^{1-\mu /2}}\dfrac{1}{s^{1-\mu /2}} \\&\quad<C_6n^{\mu /2}\sum \limits _{s=1}^{n-2}\dfrac{1}{s^{1-\mu /2}}\sum \limits _{t=s+1}^{n-1}\dfrac{1}{(t-s)^{1-\mu /2}}<Cn^{\frac{3}{2}\mu }. \end{aligned}$$
\(\square \)
The following Lemma 2 shows some properties of the Hermite functions which which is proved in Dong and Gao (2018).
Lemma 2
Let \({\mathscr {T}}_k(x)=\dfrac{1}{k}\sum \nolimits _{i=0}^{k-1}{\mathscr {H}}_i^2(x)\) and \({\mathscr {T}}(x)=\dfrac{1}{2\pi }\sqrt{4-x^2}I\{|x|\le 2\}\), then \({\mathscr {T}}_k(x)\) converges to \({\mathscr {T}}(x)\) for any \(x\in R\) as \(k\rightarrow \infty \). Moreover, \(\int _{-\infty }^{\infty }|{\mathscr {T}}_k(x)-{\mathscr {T}}(x)|\,\mathrm{d}x\rightarrow 0\) as \(k\rightarrow \infty \).
Proof of Theorem 1
Under \(H_0\), \({\hat{e}}_{t}^{(i)}=y_{t}^{(i)}-g(x_{t};{\hat{\theta }}^{(i)})=y_{t}^{(i)}-g(x_{t};\theta _{0}^{(i)})+g(x_{t};\theta _{0}^{(i)})-g(x_{t};{\hat{\theta }}^{(i)})\), denote \({\hat{g}}^{(i)}(x)=g(x;\theta _0^{(i)})-g(x;{\hat{\theta }}^{(i)})\), then \({\hat{e}}_{t}^{(i)}=e_t^{(i)}+{\hat{g}}^{(i)}(x_{t})\). The test statistic \(T_{np}\) can be rewritten as
We first consider the term \(T_{np1}\),
Using the notations in Lemma 2, we have
Let \(h(n)=p^2\sim log^2(n)\), so h(n) is a slowly varying function. Then according to Corollary 2.1 in Wang and Phillips (2009) and \(\int {\mathscr {T}}\,\mathrm{d}x=1\), we have
Noting that the standard normal density function is a bounded function, by Lemmas 1(1) and 2, we have
Hence, we have \(\dfrac{1}{n^{\mu /2}K\sigma _0}T_{np1}'\xrightarrow {D}L_{W_{1-\mu /2}}(1,0)\).
For \(T_{np1}''\), noting that \(\Vert Z(x)\Vert ^2\le O(1)K\) uniformly for any x and \(\int \Vert Z(x)\Vert ^2\,\mathrm{d}x=K\) by orthogonality and by Assumption B(b), we have
Now we move on to \(T_{np1}'''\),
Let \({\mathscr {F}}_s\) be the information flow of \(\{u_1,\ldots ,u_s\}\). According to \({\int [Z(x)'Z(x_s)]^2\mathrm {d}x}=\Vert Z(x_s)\Vert ^2\), we have
Thus, by Lemma 1(2), we have
and together with the last several results we yield
For \(T_{np3}\), we will show that \(\dfrac{1}{n^{\mu /2}K\sigma _0}T_{np3}=o_p(1)\). Set \(\xi _n\sim n^{\lambda _2}\), \(n\rightarrow \infty \) where \(\lambda _2\) satisfies \(\lambda _0<\lambda _2<-\dfrac{\mu }{4}\). For any \(\delta \), \(\epsilon >0\),
For the first term of the right-hand side of the inequality, apply the Boole’s inequality,
Then for the second term, according to the virtue of Markov’s inequality, we have
where \(I(\cdot )\) stands for the indicator function.
Using Taylor expansion for \(g(x;\theta )\) with respect to \(\theta \) in a neighborhood of \(\theta _0^{(i)}\), we have
where \({\bar{\theta }}^{(i)}\) is on the line segment joining \(\theta _0^{(i)}\) and \({\hat{\theta }}_0^{(i)}\). Therefore,
Thus,
Since \(\lambda _2<-\dfrac{\mu }{4}\) and \(p\sim \log (n)\), \(\xi _n^2 n^{\mu /2}p\sim n^{2\lambda _2+\mu /2}p=o(1)\). For \(T_{np3}'\), due to the arbitrariness of \(\epsilon \), by results (1) and (2) in Lemma 1, we have
Similarly, for \(T_{np3}'''\), we have
Apply Cauchy–Schwartz inequality, we have
Thus,
Apply Cauchy–Schwartz inequality again, we have
Finally, we get
Then, the proof of Theorem 1 is finished. \(\square \)
Proof of Theorem 2
Under \(H_1\) and Assumption D, we have \(m^{(i)}(x_t)=g(x_t;\theta _0^{(i)})+\Delta _n^{(i)}(x_t)=g(x_t;\theta _0^{(i)})+\delta _n \Delta ^{(i)}(x_t)\). Thus
The test statistic \(T_{np}\) can be rewritten as
It is already showed in the proof of Theorem 1 that \(\dfrac{1}{n^{\mu /2}k\sigma _0}T_{np3}=o_p(1)\), and by Cauchy–Schwarz inequality, \(|T_{np2}|\le 2\sqrt{T_{np1}T_{np3}}\). So we only need to show \(\dfrac{1}{n^{\mu /2}k\sigma _0}T_{np1}\xrightarrow {p}\infty \) to finish the proof. Consider the term \(T_{np1}\) firstly, we have
From the proof of Theorem 1, we know that \(\dfrac{1}{n^{\mu /2}K\sigma _0}T_{np1}'\xrightarrow {D}L_{W_{1-\mu /2}}(1,0).\)
For \(T_{np1}'''\),
where \(l_i\) is the first integer such that \(\int {\mathscr {H}}_{l_i}(x)\Delta ^{(i)}(x)\mathrm {d}x\ne 0\) under dimension i. Such \(l_i\) must exist, otherwise, for any l, \(\int {\mathscr {H}}_{l}(x)\Delta ^{(i)}(x)\mathrm {d}x=0\), which means \(\Delta ^{(i)}(x)\) is orthogonal with every \({\mathscr {H}}_l(x)\), i.e., \(\Delta ^{(i)}(x)\) is a zero function, that contradicts the assumption.
Apply Corollary 2.1 in Wang and Phillips (2009) again, as \(n\rightarrow \infty \), we have
By continuous mapping theorem, we get
as \(n\rightarrow \infty \).
According to Assumption D, \(\dfrac{\delta _n^2 n^{\mu /2}}{p^2 K\sigma _0}\rightarrow \infty \), thus
It follows that \( \dfrac{1}{n^{\mu /2}K\sigma _0}T_{np1}'''=\varOmega _P\left( \dfrac{\delta _n^2 n^{\mu /2}}{p^2 K}\right) \rightarrow \infty .\)
For \(T_{np1}''\), it can be rewritten as
We first consider the term \(T_{np11}''\),
Then, we move on to \(T_{np12}''\),
For \(T_{np121}''\), by Lemma 1(2), \(\dfrac{\delta _n^2}{n^{\mu }K^2\sigma _0^2}T_{np121}''\le \dfrac{O(1)\delta _n^2 p^2}{K}\rightarrow 0.\)
For \(T_{np122}\), by Lemma 1(3),
Thus, \(\dfrac{1}{n^{\mu /2}K\sigma _0}T_{np12}''=O_p(\delta _n p n^{\mu /4} K^{-1/2}).\)
Similarly, we can get
Then
As we just have proved earlier,
while
so the divergence rate of \(\dfrac{1}{n^{\mu /2}K\sigma _0}T_{np1}'''\) dominates the divergence rate of \(\dfrac{1}{n^{\mu /2}K\sigma _0}T_{np1}''\). Thus,
\(\square \)
Proof of Theorem 3
As \({\hat{\sigma }}\) is a consistent estimator for \(\sigma _0\), this theorem is obvious due to Theorems 1 and 2. \(\square \)
Rights and permissions
About this article
Cite this article
Wang, J., Wang, D. & Tian, Y. Multidimensional specification test based on non-stationary time series. TEST 31, 348–372 (2022). https://doi.org/10.1007/s11749-021-00780-0
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11749-021-00780-0