Abstract
This paper is concerned with the semiparametric regression model \(y_i=x_i\beta +g(t_i)+\sigma _ie_i,~~i=1,2,\ldots ,n,\) where \(\sigma _i^2=f(u_i)\), \((x_i,t_i,u_i)\) are known fixed design points, \(\beta \) is an unknown parameter to be estimated, \(g(\cdot )\) and \(f(\cdot )\) are unknown functions, random errors \(e_i\) are widely orthant dependent random variables. The p-th (\(p>0\)) mean consistency and strong consistency for least squares estimators and weighted least squares estimators of \(\beta \) and g under some more mild conditions are investigated. A simulation study is also undertaken to assess the finite sample performance of the results that we established. The results obtained in the paper generalize and improve some corresponding ones of negatively associated random variables.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
As we known that semiparametric regression models (or partially linear models) rely on a dimension reduction assumption, while being still flexible enough due to the presence of a nonparametric term. In recent years, semiparametric regression models have attracted a growing number of statisticians to study it. Based on the effect of weather on electricity demand, Engle et al. (1986) studied the following semiparametric regression model,
Hong (1991) studied the model (1.1), and gave the estimators \({\hat{\beta }}_n\) and \(g_n^{*}\) of \(\beta \) and g respectively by the methods of least squares and the nearest neighbor weight functions. In addition, he also obtained the asymptotic normality for \({\hat{\beta }}_n\) and the strong consistency for \(g_n^{*}\). Based on the independent and identically distributed (i.i.d.) samples, Gao (1992) proposed the kernel estimator \({\hat{g}}_n (\cdot )\) of \(g(\cdot )\) and the least squares estimator \({\hat{\beta }}_n\) of \(\beta \) in the semiparametric regression model as follows,
and obtained some strong and weak consistencies and convergence rates for estimators of \(\beta \) and \(g(\cdot )\). Hu (1999) defined the least squares estimators \({\tilde{\beta }}_{\tau }\) of \(\beta \) and the estimator \({\tilde{g}}_{\tau }(t)\) of g, respectively. In the case of the independent random errors, he established some asymptotic properties for the estimators \({\tilde{\beta }}_{\tau }\) and \({\tilde{g}}_{\tau }(t)\), including the strong consistency, uniform strong consistency, r-th (\(r>2\)) mean consistency and r-th (\(r>2\)) mean uniform consistency. Pan et al. (2003) discussed the semiparametric model (1.2) with \(L^q\) mixingale errors, and obtained r-th (\(r>2\)) mean consistency and complete consistency for estimators of \(\beta \) and g. Hu (2006) studied the model (1.2) with linear time series errors and obtained the r-th (\(r>2\)) mean consistency and complete consistency for the estimators \({\hat{\beta }}_n\) and \({\hat{g}}_n(t)\) of \(\beta \) and g, respectively. Based on model (1.2), Gao et al. (1994) proposed a more general semiparametric regression model,
where \(\sigma _i^2=f(u_i)\), \((x_i,t_i,u_i)\) are known fixed design points, \(\beta \) is an unknown parameter to be estimated, \(g(\cdot )\) and \(f(\cdot )\) are unknown functions defined on compact set \(A\subset \mathbb {R}\), \(e_i\) are random errors. Additionally, Gao et al. (1994) gave the least squares estimators (LSE) and the weighted least squares estimators (WLSE) of \(\beta \) and g and the estimator of f, and further proved the asymptotic normality for the two estimators of \(\beta \) under i.i.d. random errors. Chen et al. (1998) investigated the strong consistency for the two estimators of \(\beta \) if \(\{e_i,i\ge 1\}\) is i.i.d.. Based on negatively associated random errors, Baek and Liang (2006) studied the strong consistency of estimators of \(\beta \), g and f, and the asymptotic normality of \(\beta \); Zhou and Hu (2010) obtained p-th (\(p>2\)) mean consistency for LSE and WLSE of \(\beta \) and g. For more details about the asymptotic properties of the estimators in semiparametric regression models, one can refer to Chen (1988), Speckman (1988), Hamilton and Truong (1997), Mammen and Van de Geer (1997), Aneiros and Quintela (2001), Zhou and Lin (2013) among others.
Inspired by the above literatures, we will study the p-th (\(p>0\)) mean consistency and strong consistency for LSE and WLSE of \(\beta \) and g in model (1.3) under the random errors \(\{e_i,i\ge 1\}\) being zero mean widely orthant dependent random variables. We will give the details in Sect. 2. Now let us recall the definition of widely orthant dependence structure.
Definition 1.1
For the random variables \(\{X_n,n\ge 1\}\), if there exists a finite real sequence \(\{h_{U}(n),n\ge 1\}\) satisfying for each \(n\ge 1\) and for all \(x_i\in (-\infty ,+\infty )\), \(1\le i\le n\),
then we say that the \(\{X_n,n\ge 1\}\) are widely upper orthant dependent (WUOD, in short); if there exists a finite real sequence \(\{h_{L}(n),n\ge 1\}\) satisfying for each \(n\ge 1\) and for all \(x_i\in (-\infty ,+\infty )\), \(1\le i\le n\),
then we say that the \(\{X_n,n\ge 1\}\) are widely lower orthant dependent (WLOD, in short); if they are both WUOD and WLOD, then we say that the \(\{X_n,n\ge 1\}\) are widely orthant dependent, and \(h_{U}(n),h_{L}(n),n\ge 1\) are called dominating coefficients.
The concept of WOD random variables was firstly introduced by Wang et al. (2013). And they therein gave some examples to show that the class of WOD random variables includes some common negatively dependent random variables, some positively dependent random variables and some others. Subsequently, various properties and applications were obtained. For instance, Liu et al. (2012) gave the asymptotically equivalent formula for the finite-time ruin probability under a dependent risk model with constant interest rate, Shen (2013a) established the Bernstein type inequality for WOD random variables and gave some applications, He et al. (2013) provided the asymptotic lower bounds of precise large deviations with nonnegative and dependent random variables, Wang et al. (2014) further established the complete convergence for arrays of row-wise WOD random variables and gave its applications in nonparametric regression models, Wang and Hu (2015a) studied the consistency of the nearest neighbor estimator of the density function based on WOD samples, Shen et al. (2016) provided some exponential probability inequalities for WOD sequence and gave applications in complete convergence and complete moment convergence, Chen et al. (2016) established a more accurate inequality of WOD random variables, and obtained some limit theorems including the strong law of large numbers, the complete convergence, the almost sure elementary renewal theorem and the weighted elementary renewal theorem, and so on.
Obviously, \(h_U(n)\ge 1\), \(h_L(n)\ge 1\), \(n\ge 1\). If \(h_U(n)=h_L(n)=M\) for any \(n\ge 1\), it is easily seen that the random variables \(\{X_n,n\ge 1\}\) are extended negatively dependent (END, in short), where M is a positive constant. More particularly, if \(M=1\), then the random variables \(\{X_n,n\ge 1\}\) are called negatively orthant dependent (NOD, in short). In other words, NOD is a special case of END. For details about NOD and END sequence, one can refer to Volodin (2002), Asadian et al. (2006), Liu (2009), Wang and Wang (2013), Shen (2013b), Shen et al. (2015), Wang et al. (2015b), and so forth. Furthermore, Joag-Dev and Proschan (1983) pointed out that negatively associated (NA, in short) random variables are NOD. Meanwhile, Hu (2000) introduced the concept of negatively superadditive dependence (NSD, in short) and pointed out that NSD implies NOD [see Property 2 of Hu (2000)]. By the above description, the class of WOD random variables contains END random variables, NOD random variables, NSD random variables, NA random variables and independent random variables as special cases. Thus, it is of practical significance to study the mean consistency and the strong consistency of estimators in the semiparametric model (1.3) with WOD random errors.
The organization of the paper is as follows. In Sect. 2, we first present the LSE and the WLSE of \(\beta \) and \(g(\cdot )\), and some basic assumptions; and then we will establish the main results, including the mean consistency and strong consistency for the LSE and the WLSE of \(\beta \) and \(g(\cdot )\); a numerical simulation to study the consistency of LSE for \(\beta \) and \(g(\cdot )\) is also carried out; finally, some important lemmas to prove the main results are provided. In Sect. 3, we mainly give the proofs of the main results. In “Appendix”, we present the proofs of Lemmas 2.4 and 2.5.
Throughout the paper, denote \(h(n)=\max \{h_U(n),h_L(n)\}\). \(a_n=O(b_n)\) denotes that there exists a positive constant C such that \(a_n\le Cb_n\). Let c, \(c_1\), \(c_2\), C, \(C_1\), \(C_2\), \(\ldots \) denote the positive constants whose values may vary at each occurrence.
2 Main results and lemmas
2.1 Estimators and basic assumptions
The LSE and the WLSE of \(\beta \) and \(g(\cdot )\) given in Gao et al. (1994) are as follows:
where \(W_{ni}(\cdot )\) are weight functions only depending on the designed points \(t_i~(i=1,2,\ldots ,n)\), \({\tilde{x}}_i=x_i-\sum \nolimits _{j=1}^n W_{nj}(t_i)x_j\), \({\tilde{y}}_i=y_i-\sum \nolimits _{j=1}^n W_{nj}(t_i)y_j\), \(S_n^2=\sum \nolimits _{i=1}^n {\tilde{x}}_i^2\), \(a_i=\frac{1}{f(u_i)}\), \(T_n^2=\sum \nolimits _{i=1}^na_i{\tilde{x}}_i^2\).
In this paper, we will consider the following assumptions:
-
\(\mathbf {H_1}\) \(~(i)~\lim \limits _{n\rightarrow \infty }\frac{1}{n(h(n))^{2r}}\sum \nolimits _{i=1}^n \tilde{x}_i^2=\Gamma \) (\(0<\Gamma <\infty \)), \(\exists ~r\ge 1\);
\((i)^{'}\) \(~\lim \limits _{n\rightarrow \infty }\frac{1}{n(h(n))^{r}}\sum \nolimits _{i=1}^n \tilde{x}_i^2=\Gamma \) (\(0<\Gamma <\infty \)), \(\exists ~r>0\);
(ii) \(0<m_0\le \min \limits _{1\le i\le n}f(u_i)\le \max \limits _{1\le i\le n}f(u_i)\le M_0<\infty \);
(iii) \(g(\cdot )\) and \(f(\cdot )\) are continuous on compact set A.
-
\(\mathbf {H_2}\) \(\max \limits _{1\le j\le n}\left| \sum \nolimits _{i=1}^nW_{ni}(t_j)-1\right| =o(1)\);
-
\(\mathbf {H_3}\) \(\max \limits _{1\le j\le n}\sum \nolimits _{i=1}^n|W_{ni}(t_j)|I(|t_i-t_j|>a)=o(1),~\forall ~a>0\);
-
\(\mathbf {H_4}\) \(\max \limits _{1\le j\le n}\sum \nolimits _{i=1}^n|W_{ni}(t_j)|=O(1)\);
-
\(\mathbf {H_5}\) \(\max \limits _{1\le i,j\le n}|W_{ni}(t_j)|=O(n^{-s}(h(n))^{-r}),~\exists ~s>0,~r\ge 1\);
-
\(\mathbf {H_5^{'}}\) \(\max \limits _{1\le i,j\le n}|W_{ni}(t_j)|=O(n^{-s}(h(n))^{-r}),~\exists ~s>0,~r>0\).
Remark 2.1
(\(H_1\))(i) (with \(h(n)=1\)) (ii) are some regular conditions, which are assumed in Gao et al. (1994), Chen et al. (1998), Baek and Liang (2006) and so on. Moreover, it can be deduced from (\(H_1\))(i) (or \((i)^{'})\) (ii) here that
Remark 2.2
Remark 2.3 in Baek and Liang (2006) mentioned that the following two weight functions satisfy assumptions \((H_2)\)–\((H_5)\) with \(h(n)=\log n\), \(s=\frac{1}{2}\) and \(r=1\):
where \(s_i=(t_i+t_{i+1})/2,~i=1,2,\ldots ,n-1,~s_0=0,~s_n=1\), \(0\le t_1\le t_2\le \cdots \le t_n\le 1\), \(K(\cdot )\) is the Parzen-Rosenblatt kernel function, and \(h_n\) is a bandwidth parameter.
2.2 Consistency
Let \(\{e_i,i\ge 1\}\) be a sequence of mean zero WOD random errors with dominating coefficient h(n), which is stochastically dominated by a random variable e, that is
for all \(x\ge 0\), \(n\ge 1\) and some \(C>0\).
Theorem 2.1
(mean consistency) Let \(p>0\). Suppose that conditions \((H_1)\)(i, ii, iii) and \((H_2)\)–\((H_5)\) hold. If \(Ee^2<\infty \) for \(0<p\le 2\) or \(E|e|^p<\infty \) for \(p>2\), then
In addition, if \(\max _{1\le j\le n}\left| \sum _{i=1}^n W_{ni}(t_j)x_i\right| =O(1)\), then
In particular, if h(n) is a constant function, we can get the following corollary by Theorem 2.1.
Corollary 2.1
Let \(p>0\). Assume that \(\{e_i,i\ge 1\}\) be a sequence of END random errors with mean zero. Suppose that conditions \((H_1)\)–\((H_5)\) hold with \(h(n)=1\). If \(\sup _i Ee_i^2<\infty \) for \(0<p\le 2\) or \(\sup _iE|e_i|^p<\infty \) for \(p>2\), then (2.4) and (2.5) hold. In addition, if \(\max \nolimits _{1\le j\le n}\left| \sum \nolimits _{i=1}^n W_{ni}(t_j)x_i\right| =O(1)\), then (2.6) and (2.7) hold.
Remark 2.3
Under the NA sequence and conditions \((H_1)\)–\((H_5)\) with \(h(n)=1\) and \(s=1/2\), Zhou and Hu (2010) obtained results (2.4)–(2.7) under \(\sup _{i}E|e_i|^p<\infty \) for some \(p>2\), while Corollary 2.1 in our paper gives these results for some \(p>0\). Since NA sequence is END, we extend Theorem 2.1 of Zhou and Hu (2010) to the case of END sequence. Furthermore, \(s>0\) in condition \((H_5)\) of Corollary 2.1 is more general than \(s=1/2\) in Theorem 2.1 of Zhou and Hu (2010).
The next theorem gives the strong consistency of estimators under some analogous conditions.
Theorem 2.2
(strong consistency) Suppose that conditions \((H_1)\) (\(i^{'}\), \(\textit{ii}\), \(\textit{iii}\)), \((H_2)\)–\((H_4)\) and \((H_5^{'})\) hold. If \(Ee^2<\infty \) and \(\sum \nolimits _{i=1}^{n}i^{-s}(h(i))^{-r}=O(n^{s})\) for some \(s>0\) and \(r>0\), then
In addition, if \(\max _{1\le j\le n}\left| \sum _{i=1}^n W_{ni}(t_j)x_i\right| =O(1)\), then
Remark 2.4
Since \(h(i)\ge 1\), \(\sum \nolimits _{i=1}^{n}i^{-s}(h(i))^{-r}=O(n^{s})\) in Theorem 2.2 always holds as long as \(s\ge 1/2\) and \(r>0\).
Particularly, if h(n) is a constant function and \(s=1/2\) in Theorem 2.2, we can get the following corollary.
Corollary 2.2
Assume that \(\{e_i,i\ge 1\}\) be a sequence of mean zero END random errors, which is stochastically dominated by a random variable e. Suppose that conditions \((H_1)\)–\((H_5)\) hold with \(s=1/2\), \(h(n)=1\). If \(Ee^2<\infty \), then (2.8) and (2.9) hold. In addition, if \(\max _{1\le j\le n}\left| \sum \nolimits _{i=1}^n W_{ni}(t_j)x_i\right| =O(1)\), then (2.10) and (2.11) hold.
Remark 2.5
Under mean zero NA random errors, Theorem 2.1 of Baek and Liang (2006) gave the results (2.8)–(2.11) under the conditions \((H_1)\)–\((H_5)\) with \(s=1/2\), \(h(n)=1\) and \(\sup _iE|e_i|^p<\infty \) for some \(p>2\). Compared with it, Corollary 2.2 (i) extends the case of NA random variables to END random variables; (ii) lowers the order of the moment from \(p>2\) to 2.
2.3 Simulation
In this section, we will carry out a numerical simulation to study the consistency of LSE for \(\beta \) and \(g(\cdot )\). The data is generated from the model (1.3). Choose \(\sigma _i=1\), \(x_i=(-1)^{i}\frac{i}{n}\), \(i=1,2,\ldots ,n\). We take random error vector \((e_1,e_2,\ldots ,e_n)^{'}\sim N (\mathbf {0},\mathbf {\Sigma })\), where \(\mathbf {0}\) is a zero column vector, and
It is obvious that \(e_1,e_2,\ldots ,e_n\) generated as the above method are NA by Joag-Dev and Proschan (1983), which is a special case of WOD (\(h(n)=1\)). Especially, we choose the nearest neighbor weights to be weight functions \(W_{ni}(\cdot )\). Without loss of generality, let \(A = [0,1]\) and \(t_i=\frac{i}{n}\), \(i=1,2,\ldots ,n\). For any \(t\in A\), we rewrite \(|t_1-t|,|t_2-t|,\ldots ,|t_n-t|\) as follows
if \(|t_i-t|=|t_j-t|\), \(|t_i-t|\) is permuted before \(|t_j-t|\) if \(i<j\). Let \(k_n=\lfloor n^{0.6} \rfloor \) and define the nearest neighbor weight functions as follows
For any \(t=t_i\), \(i=1,2,\ldots ,n\), it is easily checked that
which imply that the assumptions in our results are satisfied. Next, we compute \({\hat{\beta }}_n-\beta \) and \({\hat{g}}_n(t)-g(t)\) for 1000 times and obtain the corresponding boxplots by taking \(t=\frac{1}{n},\frac{100}{n}\) and the sample sizes n as 200, 400, 800, 1400, 3400 respectively when \(\beta \) and g(t) are in two different forms.
Case 1 \(\beta =2\), \(g(t)=t^2\).
Case 2 \(\beta =3\), \(g(t)=\sin t\).
In Figs. 1, 2, 3, 4, 5, 6, 7 and 8, \({\hat{\beta }}_n-\beta \) and \({\hat{g}}_n(t)-g(t)\), regardless of the values of t, fluctuate to zero and the variation ranges decrease markedly as the sample size n increases. These verify the validity of our results.
3 Proof of main results
It is easy to see that
where \({\tilde{g}}(t_i)=g(t_i)-\sum \nolimits _{j=1}^nW_{nj}(t_i)g(t_j)\).
Proof of Theorem 2.1
We only prove (2.5) and (2.7), as the proofs of (2.4) and (2.6) are respectively analogous. Denote
From (3.2) and \(C_{r}\) inequality, we have
Note that \(H_{1n}=\sum \nolimits _{i=1}^n(T_n^{-2}a_i\sigma _i{\tilde{x}}_i)e_i\doteq \sum \nolimits _{i=1}^nb_{ni}e_i\), and
by \(H_1(i, ii)\) and (2.3). Hence, we obtain by Lemma A.1 that
Observe that \(H_{2n}=\sum \nolimits _{j=1}^n\left( \sum \nolimits _{i=1}^n T_n^{-2}a_i{\tilde{x}}_i W_{nj}(t_i)\sigma _j\right) e_j\doteq \sum \nolimits _{j=1}^nd_{nj}e_j,\) and
by \(H_1(ii)\), \(H_5\) and (2.3). Therefore, we have by Lemma A.1 that
We now discuss \(H_{3n}\). It follows from \(H_1(iii)\), \(H_2\), \(H_3\) and \(H_4\) that
and
So, we can obtain by (3.12) and (2.3) that
which, together with (3.5), (3.8) and (3.11), yields (2.5).
Now we turn to prove (2.7). It can be seen by (3.4) that
We can obtain from \(H_1(ii)\), \(H_4\) and \(H_5\) that \(Q_{1n}\rightarrow 0,~n\rightarrow \infty \) by applying Lemma A.1. From (2.5) and the assumption \(\max _{1\le j\le n}\left| \sum \nolimits _{i=1}^n W_{ni}(t_j)x_i\right| =O(1)\), we can get \(Q_{2n}\rightarrow 0,~n\rightarrow \infty \). \(Q_{3n}\rightarrow 0,~n\rightarrow \infty \) follows from (3.12). Therefore, the desired result (2.7) follows from (3.13) and \(Q_{1n}\rightarrow 0\), \(Q_{2n}\rightarrow 0\), \(Q_{3n}\rightarrow 0\), \(n\rightarrow \infty \). This completes the proof of the theorem. \(\square \)
Proof of Theorem 2.2
Using the notations in the proof of Theorem 2.1, we know that
Applying Lemma A.2, we have by (3.6) and (3.7) that \(H_{1n}\rightarrow 0~a.s.,~n\rightarrow \infty .\) Likewise, by (3.9) and (3.10), \(H_{2n}\rightarrow 0~a.s.,~n\rightarrow \infty \). From (2.3) and (3.12), we can easily obtain that
So (2.9) is proved. It follows from (3.4) that
From \(H_1(ii)\), \(H_4\) and \(H_5\), we obtain that \(R_{1n}\rightarrow 0~a.s.,~n\rightarrow \infty \) by applying Lemma A.2. According to (2.9) and the assumption \(\max _{1\le j\le n}\left| \sum _{i=1}^n W_{ni}(t_j)x_i\right| =O(1)\), we can get \(R_{2n}\rightarrow 0~a.s.,~n\rightarrow \infty \). \(R_{3n}\rightarrow 0,~n\rightarrow \infty \) follows from (3.12). Therefore, the desired result (2.11) follows from (3.14) and \(R_{1n}\rightarrow 0~a.s.\), \(R_{2n}\rightarrow 0~a.s.\), \(R_{3n}\rightarrow 0\), \(n\rightarrow \infty \). The proof is completed. \(\square \)
References
Aneiros G, Quintela A (2001) Asymptotic properties in partial linear models under dependence. TEST 10:333–355
Asadian N, Fakoor V, Bozorgnia A (2006) Rosenthal’s type inequalities for negatively orthant dependent random variables. J Iran Stat Soc 5(1–2):66–75
Baek J, Liang H (2006) Asymptotics of estimators in semiparametric model under NA samples. J Stat Plan Inference 136:3362–3382
Chen H (1988) Convergence rates for parametric components in a partly linear model. Ann Stat 16:136–146
Chen MH, Ren Z, Hu SH (1998) Strong consistency of a class of estimators in partial linear model. Acta Math Sin 41(2):429–439
Chen W, Wang YB, Cheng DY (2016) An inequality of widely dependent random variables and its applications. Lith Math J 56(1):16–31
Engle RF, Granger CWJ, Weiss RJ (1986) Nonparametric estimates of the relation weather and electricity sales. J Am Stat Assoc 81(394):310–320
Gao JT (1992) Consistency of estimation in a semiparametric regression model (I). J Syst Sci Math Sci 12(3):269–272
Gao JT, Chen XR, Zhao LC (1994) Asymptotic normality of a class of estimators in partial linear models. Acta Math Sin 37(2):256–268
Hamilton SA, Truong YK (1997) Local linear estimation in partly linear models. J Multivar Anal 60:1–19
He W, Cheng DY, Wang YB (2013) Asymptotic lower bounds of precise large deviations with nonnegative and dependent random variables. Stat Probab Lett 83:331–338
Hong SY (1991) Estimate for a semiparametric regression model. Sci China Math 12A:1258–1272
Hu SH (1999) Estimate for a semiparametric regression model. Acta Math Sci 19A(5):541–549
Hu SH (2006) Fixed-design semiparametric regression for linear time series. Acta Math Sci 26B(1):74–82
Hu TZ (2000) Negatively superadditive dependence of random variables with applications. Chin J Appl Probab Stat 16:133–144
Joag-Dev K, Proschan F (1983) Negative association of random variables with applications. Ann Stat 11(1):286–295
Liu L (2009) Precise large deviations for dependent random variables with heavy tails. Stat Probab Lett 79:1290–1298
Liu XJ, Gao QW, Wang YB (2012) A note on a dependent risk model with constant interest rate. Stat Probab Lett 8(4):707–712
Mammen E, Van de Geer S (1997) Penalized quasi-likelihood estimation in partial linear models. Ann Stat 25:1014–1035
Pan GM, Hu SH, Fang LB, Cheng ZD (2003) Mean consistency for a semiparametric regression model. Acta Math Sci 23A(5):598–606
Shen AT (2013a) Bernstein-type inequality for widely dependent sequence and its application to nonparametric regression models. Abstr Appl Anal 2013:9 (Article ID 862602)
Shen AT (2013b) On the strong convergence rate for weighted sums of arrays of rowwise negatively orthant dependent random variables. RACSAM 107(2):257–271
Shen AT, Zhang Y, Volodin A (2015) Applications of the Rosenthal-type inequality for negatively superadditive dependent random variables. Metrika 78:295–311
Shen AT, Yao M, Wang WJ, Volodin A (2016) Exponential probability inequalities for WNOD random variables and their applications. RACSAM 110(1):251–268
Speckman P (1988) Kernel smoothing in partial linear models. J R Stat Soc Ser B 50:413–436
Volodin A (2002) On the Kolmogorov exponential inequality for negatively dependent random variables. Pak J Stat 18(2):249–253
Wang KY, Wang YB, Gao QW (2013) Uniform asymptotics for the finite-time ruin probability of a new dependent risk model with a constant interest rate. Methodol Comput Appl Probab 15(1):109–124
Wang SJ, Wang XJ (2013) Precise large deviations for random sums of END real-valued random variables with consistent variation. J Math Anal Appl 402:660–667
Wang XJ, Xu C, Hu TC, Volodin A, Hu SH (2014) On complete convergence for widely orthant-dependent random variables and its applications in nonparametrics regression models. TEST 23(3):607–629
Wang XJ, Hu SH (2015a) The consistency of the nearest neighbor estimator of the density function based on WOD samples. J Math Anal Appl 429(1):497–512
Wang XJ, Zheng LL, Xu C, Hu SH (2015b) Complete consistency for the estimator of nonparametric regression models based on extended negatively dependent errors. Stat J Theor Appl Stat 49(2):396–407
Zhou XC, Hu SH (2010) Moment consistency of estimators in semiparametric regression model under NA samples. Pure Appl Math 6(2):262–269
Zhou XC, Lin JG (2013) Asymptotic properties of wavelet estimators in semiparametric regression models under dependent errors. J Multivar Anal 122:251–270
Acknowledgements
The authors are grateful to the Referee for carefully reading the manuscript and for providing helpful comments and constructive criticism which enabled them to improve the paper.
Author information
Authors and Affiliations
Corresponding author
Additional information
Supported by the National Natural Science Foundation of China (11671012, 11501004, 11501005), the Natural Science Foundation of Anhui Province (1508085J06) and the Key Projects for Academic Talent of Anhui Province (gxbjZD2016005).
Appendix
Appendix
Lemma A.1
Let \(p>0\) and \(\{X_n,n\ge 1\}\) be a sequence of zero mean WOD random variables with dominating coefficient h(n), which is stochastically dominated by a random variable X. Assume that \(\{a_{ni}(\cdot ), 1\le i\le n, n\ge 1\}\) is a function array defined on compact set A satisfying
and
If \(EX^2<\infty \) for \(0<p\le 2\), then
If \(E|X|^p<\infty \) for \(p>2\), then (3.17) still holds.
Remark A.1
Lemma A.1 also holds when the moment condition \(EX^2<\infty \) is changed to \(\sup _{i}EX_i^2<\infty \), \(E|X|^p<\infty \) is changed to \(\sup _{i}E|X_i|^p<\infty \) and the condition of stochastic domination is deleted. Under the similar modification, Theorem 2.1 also holds true.
Proof of Lemma A.1
Without loss of generality, we can assume that \(a_{ni}(z_j)>0\).
If \(0<p\le 2\), by Jensen’s inequality, Marcinkiewicz-Zygmund-type inequality (one can refer to Wang et al. (2014) for instance), (3.15), (3.16) and \(EX^2<\infty \), we have
If \(p>2\), we denote
thus, we only need to prove
For any \(t>0\), denote
For fixed \(t>0\) and \(1 \le j\le n\), we can see that \(\{Y_{ni}^j,1\le i\le n, n\ge 1\}\) and \(\{Z_{ni}^j,1\le i\le n, n\ge 1\}\) are both arrays of rowwise WOD random variables. Noting that \(X_{ni}^j=Y_{ni}^j-EY_{ni}^j+Z_{ni}^j-EZ_{ni}^j\), we have
First, we prove \(I_2\rightarrow 0,~n\rightarrow \infty \). Note that
Hence, for any \(t>n\varepsilon \) and all n large enough, we have \(\max _{1\le j\le n}\left| \sum \nolimits _{i=1}^nEZ_{ni}^j\right| \le t^{1/p}/4\), which implies that for all n large enough,
Next, we will show that \(I_1\rightarrow 0,~n\rightarrow \infty \). Taking \(q>p\), we have by Markov’s inequality and Rosenthal-type inequality (one can refer to Wang et al. (2014) for instance) that
According to the definition of \(Y_{ni}^j\), we have
In view of the proof of \(I_2\), we can get that \(I_{112}\rightarrow 0,~n\rightarrow \infty \). Next, we estimate the limit of \(I_{111}\) as \(n\rightarrow \infty \). It is easy to check that
Similar to the proof of (3.19), we have
and
which imply that \(I_{111}\rightarrow 0,~n\rightarrow \infty \). Noting that \(p>2\), \(\beta \ge 1\) and \(EX^2<\infty \), we have
The proof is completed. \(\square \)
Lemma A.2
Let \(\{X_n,n\ge 1\}\) be a sequence of zero mean WOD random variables with dominating coefficient h(n), which is stochastically dominated by a random variable X. Assume that \(\{a_{ni}(\cdot ), 1\le i\le n, n\ge 1\}\) is a function array defined on compact set A satisfying
and
If \(EX^2<\infty \) and \(\sum \nolimits _{i=1}^{n} i^{-\alpha }(h(i))^{-\beta }=O(n^{\alpha })\) for some \(\alpha >0\) and \(\beta >0\), then
Proof
Without loss of generality, we can assume that \(a_{ni}(z_j)>0\).
For any \(\varepsilon >0\), choose \(0<\delta <\alpha /2\) and large \(N\ge 1\), which will be specialized later. Denote \(X_{ni}(j)=a_{ni}(z_j)X_i\), and
Then
To prove (3.28), it suffices to show \(J_i\rightarrow 0~a.s.,~n\rightarrow \infty ,~i=1,2,3,4\). We first prove \(J_1\rightarrow 0~a.s.\), \(n\rightarrow \infty \). For each j, we know that \(\{Y_{ni}^{(1)}(j),1\le i\le n,n\ge 1\}\) is still an array of rowwise WOD random variables. In view of \(EX_i=0\), (3.26), (3.27) and \(EX^2<\infty \), we get
Hence, for all n large enough, \(\max \limits _{1\le j\le n}\left| \sum \nolimits _{i=1}^n E Y_{ni}^{(1)}(j)\right| <\frac{\varepsilon }{2}\). Applying Markov’s inequality and Rosenthal-type inequality, and taking
we have
Note that
and
We can see that \(J_1\rightarrow 0~a.s.\), \(n\rightarrow \infty \) by (3.30)–(3.32) and the Borel–Cantelli Lemma.
Next we turn to estimate \(J_2\). It follows from (3.27) that
Hence, to prove \(J_2\rightarrow 0~a.s.,~n\rightarrow \infty \), we only need to show
It can be checked by \(\sum \nolimits _{i=1}^{n} i^{-\alpha }(h(i))^{-\beta }=O(n^{\alpha })\) and \(EX^2<\infty \) that
which implies that (3.34) holds. Consequently, according to (3.33), (3.34) and Kronecker’s lemma, \(J_2\rightarrow 0~a.s.\), \(n\rightarrow \infty \).
From the definition of \(Y_{ni}^{(3)}(j)\), we know that
Therefore, by taking \(N>\max \left\{ \frac{2}{\alpha -2\delta },\frac{2}{\beta }\right\} \), we have
Hence, from the Borel–Cantelli lemma, we can obtain \(J_3\rightarrow 0~a.s.\) \(n\rightarrow \infty \). Note that
Similar to the proof of \(J_3\), we have \(J_4\rightarrow 0~a.s.\) \(n\rightarrow \infty \). This completes the proof of lemma. \(\square \)
Rights and permissions
About this article
Cite this article
Wang, X., Deng, X. & Hu, S. On consistency of the weighted least squares estimators in a semiparametric regression model. Metrika 81, 797–820 (2018). https://doi.org/10.1007/s00184-018-0659-y
Received:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00184-018-0659-y
Keywords
- Semiparametric regression model
- Widely orthant dependent random error
- Least squares estimator
- Consistency