Abstract
The adequacy check of regression models is a fundamental approach to avoid model misspecifications. Three types of tests: the weighted integrated squared distance test, the U-statistic test and the empirical process based test, are very popular due to attractive theoretical merits such as consistency and satisfactory performances in practice. In this paper, we apply these three tests to check the adequacy of a mean parametric regression model with measurement error. By rigorously investigating the asymptotic properties of three testing methods under the null, local and global alternative hypotheses, we make detailed comparisons for the three tests. To the best of our knowledge, the results of these theoretical comparisons are novel. We conduct simulation studies and a real data analysis to compare the finite sample behaviors of the proposed methods.
Similar content being viewed by others
References
Andrews DF, Herzberg AM (2012) Data: a collection of problems from many fields for the student and research worker. Springer, New York
Cai Z, Naik PA, Tsai CL (2000) Denoised least squares estimators: an application to estimating advertising effectiveness. Stat Sin 10:1231–1242
Cui H, He X, Zhu L (2002) On regression estimators with de-noised variables. Stat Sin 12:1191–1206
Escanciano JC (2006) A consistent diagnostic test for regression models using projections. Econom Theory 22:1030–1051
González-Manteiga W, Crujeiras RM (2013) An updated review of goodness-of-fit tests for regression models. Test 22(3):361–411
González-Manteiga W, Peréz-González A (2006) Goodness-of-fit tests for linear regression models with missing response data. Can J Stat 34:149–170
Guo X, Zhu L (2017) A review on dimension-reduction based tests for regressions. In: From statistics to mathematical finance. Springer, Cham, pp 105–125
Guo X, Wang T, Zhu L (2015) Model checking for parametric single-index models: a dimension reduction model-adaptive approach. J R Stat Soc B 61:1–23
Hall P (1984) Central limit theorem for integrated square error of multivariate nonparametric density estimators. J Multivar Anal 14:1–16
Härdle W, Mammen E (1993) Testing parametric versus nonparametric regression. Ann Stat 21:1926–1947
Koul HL, Song W et al (2009) Minimum distance regression model checking with Berkson measurement errors. Ann Stat 37(1):132–156
Li L, Greene T (2008) Varying coefficients model with measurement error. Biometrics 64:519–526
Li Q, Wang S (1998) A simple consistent bootstrap test for a parametric regression function. J Econom 87:145–165
Liu Z, Liu C, Sun Z (2019) Consistent model check of errors-in-variables varying-coefficient model with auxiliary variable. J Stat Plan Inference 198:13–28
Masry E (1996) Multivariate local polynomial regression for time series: uniform strong consistency and rates. J Time Ser Anal 17:571–599
Meintanis SG, Allison J, Santana L (2016) Goodness-of-fit tests for semiparametric and parametric hypotheses based on the probability weighted empirical characteristic function. Stat Pap 57(4):957–976
Pollard D (2012) Convergence of stochastic processes. Springer, New York
Song W (2009) Lack-of-fit testing in errors-in-variables regression model with validation data. Stat Probab Lett 79(6):765–773
Stute W (1997) Nonparametric model checks for regression. Ann Stat 25:613–641
Sun Z, Wang Q, Dai P (2009) Model checking for partially linear models with missing responses at random. J Multivar Anal 100:636–651
van der Vaart AW, Wellner JA (1996) Weak convergence and empirical processes. Springer series in statistics. Springer, New York
Wang M, Liu C, Xie T, Sun Z (2020) Data-driven model checking for errors-in-variables varying-coefficient models with replicate measurements. Comput Stat Data Anal 141(1):12–27
Xie C, Zhu L (2019) A goodness-of-fit test for variable-adjusted models. Comput Stat Data Anal 138:27–48
Xu W, Zhu L (2015) Nonparametric check for partial linear errors-in-covariables models with validation data. Ann Inst Stat Math 67(4):793–815
Zhang C, Dette H (2004) A power comparison between nonparametric regression tests. Stat Probab Lett 66(3):289–301
Zhang J, Li G, Feng Z (2015) Checking the adequacy for a distortion errors-in-variables parametric regression model. Comput Stat Data Anal 83:52–64
Zhao P, Xue L (2010) Variable selection for semiparametric varying coefficient partially linear errors-in-variables models. J Multivar Anal 101:1872–1883
Zheng JX (1996) A consistent test of functional form via nonparametric estimation techniques. J Econom 75:263–289
Zhu L, Cui H (2005) Testing the adequacy for a general linear errors-in-variables model. Stat Sin 15:1049–1068
Acknowledgements
This research was supported by National Natural Science Foundation of China (11571340, 11971404, 11971045), MOE (Ministry of Education in China) Project of Humanities and Social Sciences (19YJC910010), and the Open Project of Key Laboratory of Big Data Mining and Knowledge Management, Chinese Academy of Sciences.
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
1.1 Estimating procedure
Set \(\xi =\mathrm{E}({\tilde{\xi }}|V):=\xi (V).\) Let \(\omega _v=\text{ diag }\{k((V_1-v)/h_n),\ldots , k((V_n-v)/h_n)\}\) for \(v\in {\mathbb {R}}^q,\) where \(k(\cdot )\) is a kernel function and \(h_n\) is a bandwidth. By applying the local linear smoothing method, we can obtain the estimate \({\hat{\xi }}_n(v)\) of \(\xi (v)\):
where \(\tilde{\mathbb {{X}}}= ({\tilde{\xi }}_{1}, \ldots , {\tilde{\xi }}_{n})^\top , {\mathbb {V}}_v= (V_{1v}, \ldots , V_{nv})^\top \) with \(V_{iv} = \left( 1, (V_i-v)^\top /h_n\right) ^\top .\) Then we can define an estimator \({\hat{\theta }}_n,\) which solves the estimating equation
where \({\dot{g}}_{\theta }(\xi ,\theta )\) is the partial derivative of \(g(\xi ,\theta )\) with respect to \(\theta .\)
1.2 Conditions
We list some conditions needed to prove Theorems 1–6.
-
(C1)
With probability 1, both \(\xi \) and V lie in compact sets. Assume that \(\mathrm{E}(Y^2|\xi )<\infty ,\) and \({\mathrm{E}}({\tilde{\xi }}^2|V)<\infty .\)
-
(C2)
(i) The function \(g(\xi ,\theta )\) is three times continuously differentiable with finite first, second and third order partial derivatives respect to both \(\xi \) and \(\theta ;\) (ii) the function \(\xi (V) = \mathrm{E}({\tilde{\xi }}|V)\) is twice continuously differentiable with finite partial derivatives.
-
(C3)
\(\varSigma _0=\mathrm{E}[{\dot{g}}_{\theta }(\xi ,\theta _0){\dot{g}}_{{\theta }}(\xi ,\theta _0)^\top ]\) is positive definite.
-
(C4)
The densities of \(\xi \) and V are bounded away from zero and infinity.
-
(C5)
The functions \(k(\cdot )\) and \(\lambda (\cdot )\) are bounded kernel functions of order 2.
-
(C6)
(i) \(\ln n/(nh^q_n)\rightarrow 0\) and \(nh_n^{2q}\rightarrow 0;\) (ii) \(nb_n^q\rightarrow \infty \) and \(nb_n^{2q}\rightarrow 0;\) (iii) \(h_nb_n^{-q/2}(\ln n)^{-1}\) \(\rightarrow 0\) and \(b_n^{2q}h_n^{-q/2}(\ln n)^{-1}\) \(\rightarrow 0.\)
Remark 1
Conditions (C1)–(C3) are necessary for the asymptotic normality of the estimator defined in (A.2). Condition (C4) aims for avoiding tedious proofs of the theorems. Conditions (C5)–(C6) are common assumptions for the convergence of the local smoothing method.
1.3 Some useful lemmas
Let \(V_{i,j}=V_i-V_j\) for \(i,j=1,\ldots ,n.\) We first list two lemmas needed for the proofs of main theorems.
Lemma 1
Under Conditions (C1)–(C5), (C6)(i) and the alternative hypothetical models \( {\mathcal {H}}_{1n} \) (3.1) with \(0\le \rho \le q/2,\) we have
Proof
We only prove the result with \(\rho =0.\) The proof of the results with \(0<\rho \le q/2\) is similar and we omit the details. Note that
Then \(A_n\) can be divided into
By some simple computations, we can decompose \(A_{n1}\) into
When the alternative hypothetical models \(Y=g(\xi ,\theta _0)+n^{-1/2}b_n^{-\rho }S(\xi )+\varepsilon \) with \(\rho =0\) hold, we have \(Y_i-g(\xi _i,\theta _0)=n^{-1/2}S(\xi _i)+\varepsilon _i\) for \(i=1,\ldots , n.\) By Taylor expansion and the law of large numbers, we have
Thus it yields
with \(\varSigma _0\) defined in Condition (C3). For \(A_{n2},\) by Taylor expansion and Condition (C2), we can validate that
Let \({\dot{g}}_{\theta }(\xi _i, \theta ){\dot{g}}_{\xi }(\xi _i,\theta )^\top =:\vartheta (\xi _i, \theta ).\) Following the result of Theorem 6 in Masry (1996), we have \(\sup _v|{\hat{\xi }}_n(v) -\xi (v)|=O_P(\{\ln n /(nh^q_n)\}^{1/2} + h_n^{2}).\) Then by Taylor expansion and Conditions (C1)–(C2), \(n^{-1/2}\sum _{i=1}^n\{\vartheta (\xi _i, \theta _0)-\vartheta (\xi _i, {\hat{\theta }}_{n})\}\{{\hat{\xi }}_n(V_i)-\xi _i\}=O_P(\{\ln n /(nh^q_n)\}^{1/2} + h_n^{2})=o_p(1).\) Thus we have
Note that under Conditions C4, C5 and C6(i),
By the definition of \({\hat{\xi }}_n(v),\) we have
By Conditions C2, C5 and C6(i), we can validate that
Lemma 3 which is listed below yields the result:
Since \(n^{1/2}A_{n}=0,\) from (A.4), (A.6), (A.7) and (A.8), we can obtain that
Thus Lemma 1 for \(\rho =0\) is proved. \(\square \)
Lemma 2
Under Conditions (C1)–(C6)(i) and \({\mathcal {H}}_0,\) we have
Remark 2
Lemma 2 can be proved by letting \(S(\xi )=0\) in the proof of Lemma 1. We omit the details.
Lemma 3
Under the same conditions of Lemma 1, we have
Proof
We notice that \(n^{-1/2}\sum _{i=1}^n\{{\dot{g}}_{\theta }({\hat{\xi }}_n(V_i),{\hat{\theta }}_{n})-{\dot{g}}_{\theta }(\xi _i,{\hat{\theta }}_{n})\}\{Y_i-g({\hat{\xi }}_n(V_i),{\hat{\theta }}_{n})\}\) is just \(A_3\) in (A.4). It can be decomposed into two parts:
Let \(\ddot{g}_{\theta ,\xi }(\xi ,\theta )\) be the twice partial derivatives related to \(\theta \) and \(\xi .\) For the first term \(A_{n31},\) we have
By the similar method to prove (A.7), we can prove that \(n^{-1}\sum _{i=1}^n\ddot{g}_{\theta ,\xi }(\xi _i,\theta )^\top \{{\hat{\xi }}_n(V_i)-\xi _i\}S(\xi _i)=O_p(n^{-1/2})=o_p(1).\) So further by the definition of \({\hat{\xi }}_n(v)\) and some simple computations, it yields
We can prove that the second moment of each component of the vector \(A^{[1]}_{n31,1}\) is \(O(n^{-1}h_n^{-q}).\) So \(A^{[1]}_{n31,1}=o_p(1)\) by Condition (C6). Note that \((nh_n^{q})^{-1/2}\sum _{i=1}^n\) \(\varepsilon _i\) \(\ddot{g}_{\theta ,\xi }(\xi _i,\theta )^\top ({\tilde{\xi }}_i -\xi _i)\) \(k(V_{i,j}/h_n)/f_v(V_i)\) \(=O_p(1).\) Then it yields \(A^{[1]}_{n31,2}=O_p(n^{-1}h_n^{-q/2})=o_p(1)\) by the condition of the bandwidth. Recall that \(\xi _j-\xi _i=\xi (V_j)-\xi (V_i).\) Via Taylor expansion of \(\xi (V_j)-\xi (V_i),\) we can obtain that \(A^{[1]}_{n31,2}=O_p(h_n^2)=o_p(1).\) Thus we have the result that \(A^{[1]}_{n31}=o_p(1).\) For \(A^{[2]}_{n31},\) we have \(A^{[2]}_{n31}=({\hat{\theta }}_{n}-\theta )n^{-1/2}\sum _{i=1}^n \breve{g}_{\theta ,\xi ,\theta }(\xi _i,\theta )\}^\top \{{\hat{\xi }}_n(V_i)-\xi _i\}\{n^{-1/2}S(\xi _i)+\varepsilon _i\}\) with \(\breve{g}_{\theta ,\xi ,\theta }(\xi _i,\theta )\) the three times partial derivative related to \(\theta ,\xi ,\theta .\) By similar method to prove \(A^{[1]}_{n31}=o_p(1),\) we can prove that \(n^{-1/2}\sum _{i=1}^n \breve{g}_{\theta ,\xi ,\theta }(\xi _i,\theta )\}^\top \{{\hat{\xi }}_n(V_i)-\xi _i\}\{n^{-1/2}S(\xi _i)+\varepsilon _i\}=o_p(1).\) Further we can obtain \(A^{[2]}_{n31}=o_p(1)\) and then \(A_{n31}=o_p(1).\) For \(A_{n32},\) by denoting \(\tau (\xi ,\theta )={\dot{g}} _{\xi }(\xi _i,\theta )\ddot{g}_{\theta ,\xi }(\xi ,\theta )^\top ,\) we have
By the existing result \(\sup _v|{\hat{\xi }}_n(v) -\xi (v)|=O_P(\{\ln n /(nh^q_n)\}^{1/2} + h_n^{2}),\) we can prove that both terms in (A.11) are \(o_p(1).\) Recalling that we have proved \(A_{n31}=o_p(1),\) therefore we obtain that \(A_{n3}=o_p(1)\) from (A.10). \(\square \)
Recall that \(\varDelta _{\theta }(\xi ,{\tilde{\xi }}) =({\tilde{\xi }}-\xi )^\top {\dot{g}}_{\xi }(\xi ,\theta ).\) Then we prove a lemma needed in Proof of Theorem 5.
Lemma 4
Let \({\hat{\xi }}_{i}= {\hat{\xi }}_n(V_i)\) and \(\lambda _{b_n}(V_{i,j})= \lambda ((V_i-V_j)/b_n)/b_n^q.\) Assume that Conditions (C1)–(C6) hold, then we have
Proof
Let \(U_n =n^{-1}b_n^{q/2}\sum _{i\ne j}(\xi _i-{\hat{\xi }}_{i})^\top {\dot{g}}_{\xi }(\xi _i,\theta _0){\dot{g}}_{\xi }(\xi _j,\theta _0)^\top (\xi _j-{\hat{\xi }}_{j})\lambda _{b_n}(V_{i,j}).\) Then
We first focus on \(U_{n2}.\) Let \(k_{h_n}(V_{j,s})=k((V_j-V_s)/h_n)/h_n^q.\) Note that \(\sup _j \Vert {\hat{\xi }}_j -n^{-1}\sum _{s=1}^n{\tilde{\xi }}_s k_{h_n}(V_{j,s})/f_v(V_j) \Vert _2 = o_p(1).\) Then
By law of large numbers, we can prove that \(\frac{1}{n}\sum _{j=1}^n{\dot{g}}_{\xi }(\xi _j,\theta _0)^\top k_{h_n}(V_{j,s})\lambda _{b_n}(V_{i,j})/f_v(V_j)=(1+o_p(1)){\dot{g}}_{\xi }(\xi _s,\theta _0)^\top \lambda _{b_n}(V_{i,s}).\) For \(U_{n2}^{[1]},\) we have
Similarly,
Thus
Similar to prove (A.14), we can obtain that
Recall that \(\varDelta _{\theta }(\xi ,{\tilde{\xi }}) =({\tilde{\xi }}-\xi )^\top {\dot{g}}_{\xi }(\xi ,\theta ).\) Combining (A.13)–(A.15), we have
Since \(n^{-1}b_n^{q/2}\sum _{i=1}^n\sum _{j\ne i}^n \varDelta _{\theta _0}(\xi _i,{\tilde{\xi }}_i) \varDelta _{\theta _0}(\xi _j,{\tilde{\xi }}_j) \lambda _{b_n}(V_{i,j}) = O_p(1)\) by verifying \(\mathrm{E} (T_n) = 0\) and \(\mathrm{E} (T_n^2) = O(1),\) the lemma is proved. \(\square \)
1.4 Proofs of the Theorems
Theorems 2.1–2.3 are special cases of Theorems 4–6 with \(S(\xi )=0.\) As follows, we give the detailed proofs for Theorems 4–6.
Proof of Theorem 4
First define \( B_n(v)= n^{-1/2}b_n^{-q/2}\sum _{i=1}^{n}\{Y_i-g({\hat{\xi }}_n(V_i),{\hat{\theta }}_{n})\} \lambda ((V_i-v)/b_n). \) Then \(T_n^{\mathtt {WISD}}=\int B^2_n(v)dv.\) For \(B_n(v),\) it can be decomposed into
For the alternative hypothetical models (3.1) when \(\rho =q/2,\) we can prove that
For the second term \(B_{n2}(v),\) we can obtain that
By the definition of \({\hat{\xi }}_n(v),\) it follows that
Let \(\psi _v(V_i)=\lambda ((V_i-v)/b_n){\dot{g}}_{\xi }(\xi _i,\theta _0)^\top /f_v(V_i). \) For \(B^{[1]}_{n2}(v),\) we can get
Then combining (A.18) and (A.19), we have
For \(B_{n3}(v),\) via Taylor expansion and a simple decomposition, we have
Note that \(1/(nb^q_n)\sum _{i=1}^{n}{\dot{g}}_{\theta }(\xi _i,\theta _0)\lambda ((V_i-v)/b_n)=f_v(v){\dot{g}}_{\theta }(\xi (v),\theta )+o_p(1).\) Therefore by the result of Lemma 1, it yields
By the result of Lemma 1 with \(\rho =q/2,\) we have \(\Vert \theta _0-{\hat{\theta }}_{n}\Vert _2=O_{p}(n^{-1/2}b_n^{-q/2}).\) Then with Condition C2, we have
where c is a finite positive number. Note that \(1/(nb^q_n)\sum _{i=1}^{n}\lambda ((V_i-v)/b_n)=f_v(v)+o_p(1)\) and \(\sup _i \Vert {\hat{\xi }}_n(V_i) - \xi (V_i)\Vert _2 = o_p(1).\) Note that \(f_v(v)\) is bound away from zero and infinity in Conditions C4. Then it yields \(B_{n3}^{[1]}(v)=o_{p}(1).\) Combining the above discussions and (A.21), we have that
Thus from (A.16)–(A.17), (A.20) and (A.22), it yields
Note that the kernel function \(\lambda (\cdot )\) is a real function, then the class \(\{\{\varepsilon +(X-{\tilde{\xi }})^\top {\dot{g}}_{\xi }(\xi ,\theta _0)\}\lambda ((V-v)/b_n): v\in {\mathbb {R}}^q\}\) is a V–C class by the Lemma 28 of Pollard (2012) and Lemma 2.6.18 of van der Vaart and Wellner (1996). Thus we can prove that \(B_n(v)\) converges to a Gauss process \(\varUpsilon (v).\) Further by Theorem 1.11.1 of van der Vaart and Wellner (1996), we prove the result of \(T_n^{\mathtt {WISD}}.\) \(\square \)
Proof of Theorem 5
Recall that \(T_n^{\mathtt {U}} = n^{-1}b_n^{q/2}\sum _{i\ne j}(Y_i-g({\hat{\xi }}_n(V_i),{\hat{\theta }}_{n}))(Y_j-g({\hat{\xi }}_n(V_j),{\hat{\theta }}_{n})) \lambda (V_{i,j}/b_n)/b_n^q.\) Let \(\lambda _{b_n}(\cdot )= \lambda (\cdot /b_n)/b_n^q\) and \({\hat{\xi }}_i = {\hat{\xi }}_n(V_i).\) Then
where
For \(D_{n1},\) under the local alternative models (3.1) with \(\rho =q/4,\) it is shown that
Following the similar argument as Lemma 3.3b in Zheng (1996), we have
Then with Condition C6(ii),
Rewriting \(n^{-2}\sum _{j\ne i}S(\xi _i)S(\xi _j)\lambda _{b_n}(V_{i,j})\) as \(n^{-1}\sum _{i= 1}^n S(\xi _i)\frac{n-1}{n}\{(n-1)^{-1}\sum _{j= 1,j\ne i}^n\) \(S(\xi _j)\lambda _{b_n}(V_{i,j})\},\) then it is equal to \(n^{-1}\sum _{i= 1}^nS(\xi _i) E\{S(\xi )|V=V_i\}f_v(V_i)+ o_p(1)=\mathrm{E}(S^2(\xi )f_v(V))+o_p(1).\) Thus the expression (A.24) for \(D_{n1}\) reduces to
Next we turn to the term \(D_{n2}\) in (A.23). Define
By Taylor expansion and Conditions C2, C4 and C5, we have \(D_{n2} = D_{n2}^*+o_p(1).\) We only need to focus on \(D_{n2}^*\) for \(D_{n2}.\) By the result of Lemma 1 under the alternative hypothesis (3.1) with \(\rho =q/4,\) we have \(n^{1/2}b_n^{q/4}({\hat{\theta }}_{n}-\theta _0)=\varSigma _0^{-1}\mathrm{E}[{\dot{g}}_{\theta }(\xi ,\theta _0)S(\xi )]+o_p(1).\) Therefore, it follows that
Recall that \(\varDelta _{\theta }(\xi ,{\tilde{\xi }}) =({\tilde{\xi }}-\xi )^\top {\dot{g}}_{\xi }(\xi ,\theta ).\) By Lemma 4, we have
Let \(n^{-3/2}b_n^{q/2}\sum _{i\ne j} {\dot{g}}_{\theta }(\xi _i,\theta _0){\dot{g}}_{\xi }(\xi _j,\theta _0)^\top (\xi _j-{\hat{\xi }}_{j}) \lambda _{b_n}(V_{i,j})=:M_n.\) Then it yields
By denoting \({\dot{g}}_{\theta }(\xi _j,\theta _0)f_v(V_j){\dot{g}}_{\xi }(\xi _j,\theta _0)^\top \) as a new function, for example \(V(\xi _j,\theta _0),\) similar to the proof of (A.7), it yields \(M_n= n^{-1/2}b_n^{q/2} \sum _{j=1}^n {\dot{g}}_{\theta }(\xi _j,\theta _0)f_v(V_j){\dot{g}}_{\xi }(\xi _j,\theta _0)^\top \) \((\xi _j-{\hat{\xi }}_{j})+o_p(1).\) Further recalling that \(n^{1/2}b_n^{q/4}(\theta _0-{\hat{\theta }}_{n})=O_p(1),\) we have
From (A.26), (A.27), (A.28), (A.29) and \(D_{n2} = D_{n2}^*+o_p(1),\) we have
Under (3.1) with \(\rho =q/4,\) noticing that\(n^{1/2}b_n^{q/4}(\theta _0-{\hat{\theta }}_{n})=O_p(1),\) we have
We can prove that \( n^{-3/2}\sum _{i=1}^n\sum _{j\ne i}\varepsilon _i {\dot{g}}_{\theta }(\xi _j,{\theta }_{0}) \lambda _{b_n}(V_{i,j})\) \( = n^{-1/2}\sum _{i=1}^n\varepsilon _i\left\{ n^{-1}\sum _{j=1,j\ne i}^n {\dot{g}}_{\theta }(\xi _j,{\theta }_{0}) \lambda _{b_n}(V_{i,j})\right\} = n^{-1/2}\sum _{i=1}^n\varepsilon _i {\dot{g}}_{\theta }(\xi _i,{\theta }_{0})f_v(V_i)+o_p(1).\) Further by the fact that \(n^{1/2}b_n^{q/4}(\theta -{\hat{\theta }}_{n})=O_p(1),\) we have that
For \(D_{n3},\) we first consider the part \(n^{-1}b_n^{q/2}\sum _{i=1}^n\sum _{j=1}^n\varepsilon _i{\dot{g}}_{\xi }(\xi _j,{\theta }_{0})^\top {\hat{\xi }}_j \lambda _{b_n}(V_{i,j})\):
Thus it is easy to obtain that
By applying law of large numbers twice, it can be validated that
By Lemma 1 with \(\rho =q/4,\) \(n^{1/2}b_n^{q/4}({\hat{\theta }}_{n}-\theta _0)=\varSigma _0^{-1}\mathrm{E}[{\dot{g}}_{\theta }(\xi ,\theta _0)S(\xi )]+o_p(1).\) Combining the above two results, it yields
For \(D_{n3}^{[4]},\) we can validate that
Therefore, with (A.31)–(A.36), we have
From (A.23), (A.24), (A.30) and (A.37), we can validate that
where \(\mu (\xi ) = S(\xi )-{\dot{g}}_{\theta }(\xi ,\theta _0)^\top \varSigma _0^{-1}\mathrm{E}[{\dot{g}}_{\theta }(\xi ,\theta _0)S(\xi )].\) Since the first term of (A.38) is a degenerated U-statistic, by similar method of Li and Wang (1998), we can validate that the conditions of Theorem 1 of Hall (1984) hold for \(T_n^{\mathtt {U}}.\) Thus we have proved the asymptotic normality of \(T_n^{\mathtt {U}}.\)
Proof of Theorem 6
Recall that \(\varphi _n(v)=n^{-1/2}\sum \nolimits _{i=1}^n\{ Y_i-g({\hat{\xi }}_n(V_i),{\hat{\theta }}_{n})\}{\mathbf {1}}(V_i\le v)\) and the EP based test \( T^{\mathtt {EP}}_n=\int \{\varphi _n(v)\}^2dF_n(v).\) Then we have
Under the alternative hypothetical model with \(\rho =0\) in (3.1), that is \(Y=g(\xi ,\theta _0)+n^{-1/2}S(\xi )+\varepsilon ,\) we have
For \(G_{n2},\) by Taylor expansion and Condition C2, we have
Lemma 1 shows that \( n^{1/2}({\hat{\theta }}_{n}-\theta _0)=n^{-1/2}\varSigma _0^{-1} \sum _{i=1}^n{\dot{g}}_{\theta }(\xi _i,\theta _0)\left\{ \varepsilon _i- \varDelta _{\theta _0}(\xi _i,{\tilde{\xi }}_i)\right\} +\varSigma _0^{-1}\mathrm{E}\{{\dot{g}}_{\theta }(\xi ,\theta _0)S(\xi )\}+o_p(1) \) under (3.1) with \(\rho =0.\) Thus we have \(n^{1/2}({\hat{\theta }}_{n}-\theta _0)=O_p(1).\) Further by law of large numbers, it yields
For \(G_{n2}^{[2]},\) similar to prove (A.7), we can prove
From the results (A.41)–(A.43), we have
From the definition of \(\mathtt {IF}_{v}^{\mathtt {EP}}(\varepsilon , \xi , {\tilde{\xi }},V),\) (A.39), (A.40) and (A.44), we have
Note that the function class \(\{{\mathbf {1}}(V\le v): v \in R^q \}\) is a V–C class since \({\mathbf {1}}(V\le v)\) is the product of q univariate indicator functions. Further by Lemma 2.6.18 of van der Vaart and Wellner (1996), we can obtain that the function class \(\{\mathtt {IF}_{v}^{\mathtt {EP}}(\varepsilon , \xi , {\tilde{\xi }}, V_i): v \in R^q \}\) is a V–C class. By Theorem 2.6.8 of van der Vaart and Wellner (1996), we can prove the weak convergence of \(\mathtt {IF}_{v}^{\mathtt {EP}}(\varepsilon , \xi , {\tilde{\xi }},V).\) The result of \(T^{\mathtt {EP}}_n \) can be proved from Theorem 1.11.1 of van der Vaart and Wellner (1996). \(\square \)
Rights and permissions
About this article
Cite this article
Sun, Z., Luo, D., Zhou, X. et al. Comparative studies on the adequacy check of parametric measurement error models with auxiliary variable. Stat Papers 62, 1723–1751 (2021). https://doi.org/10.1007/s00362-019-01154-3
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00362-019-01154-3