Abstract
Two statistical inference problems in nonlinear mixed models (NLMM) are considered under only moment conditions on random effects and random errors. First, higher-order moment estimates of random effects and random errors in NLMM are proposed and they turn out to be strongly consistent. Second, a difference-type test \(T_{mDs}\) is developed to test whether some sub-vector of random effects exists or not, which is easy to implement without requiring the Monte Carlo method. Its theoretical properties including the power properties are obtained. Moreover, in the special case of testing the existence of random effects, two kinds of tests are also constructed: the global difference-type test \(T_{mDG}\), which is a special case of \(T_{mDs}\), and the modified score-type test \(ST_{n0}\), which is motivated by \(ST_{nru}\) in Russo et al. (TEST 21:519–545, 2012). The simulation study indicates that \(T_{mDs}\) is the most powerful. A real data analysis is also conducted to investigate the applicability of the procedures.
Similar content being viewed by others
References
Fang Y, Wu P, Zhu LX (2009) Statistical analysis for nonlinear mixed models with random design. Chin J Appl Prob Stat 25(3):327–336
Giampaoli V, Singer JM (2009) Likelihood ratio tests for variance components in linear mixed models. J Stat Plan Inference 139:1435–1448
Jamison B, Orey S, Pruitt W (1965) Convergence of weighted averages of independent random variables. Z Wahrsch Verw Gebiete 4:40–44
Jennrich RI (1969) Asymptotic properties of nonlinear least square estimators. Ann Math Stat 25:357–381
Li ZX (2011) Estimation in linear mixed models for longitudinal data under linear restricted conditions. J Stat Plan Inference 141:869–876
Li ZX, Chen F, Zhu LX (2014) Variance components testing in ANOVA-type mixed models. Scand J Stat 41:482–496
Li ZX, Zhu LX (2010) On variance components in semiparametric mixed models for longitudinal data. Scand J Stat 37:442–457
Nobre JS, Singer JM, Sen PK (2013) U-tests for variance components in linear mixed models. TEST 22:580–605
Petrov VV (1975) Sums of independent random variables. Springer, New York
Russo CM, Aoki R, Paula GA (2012) Assessment of variance components in nonlinear mixed-effect elliptical models. TEST 21:519–545
Sen PK, Pedroso-de-Lima AC (2011) Contiguity and irreconcilable nonstandard asymptotics of statistical tests. Braz J Prob Stat 25:444–470
Shapiro A (1988) Towards a unified theory of inequality constrained testing in multivariate analysis. Int Stat Rev 56:49–62
Silvapulle MJ, Silvapulle P (1995) A score test against one-sided alternatives. J Am Stat Assoc 90:342–349
Vonesh EF (2012) Generalized linear and nonlinear models for correlated data: theory and applications using SAS. SAS Institute Inc, Cary
Vonesh EF, Carter RL (1992) Mixed-effects nonlinear regression for unbalanced repeated measures. Biometrics 48:1–17
Vu HT, Zhou S (1997) Generalization of likelihood ratio tests under nonstandard conditions. Ann Stat 25:897–916
Wu CF (1981) Asymptotic theory of nonlinear least squares estimation. Ann Stat 9:501–513
Acknowledgements
The author is very grateful to the Editor, the Associate Editor and the referees for their constructive comments and suggestions that substantially improved an earlier manuscript. The author thanks Professor Vonesh, EF for sharing the hemodialysis data with me and also thanks Professor Pan, GM in Nanyang Technological University for helping me refine the English throughly. This paper is supported by the National Natural Science Foundation of China (No. 11671398, 11001267), State Key Laboratory of Coal Resources and Safe Mining in China (No. SKLCRSM16KFB03), the Fundamental Research Funds for the Central Universities in China (No. 2009QS02) and Beijing Higher Education Young Elite Teacher Project (No. YETP0945).
Author information
Authors and Affiliations
Corresponding author
Electronic supplementary material
Below is the link to the electronic supplementary material.
Appendices
Appendix 1: Notation
Denote \({\mathbf{S}}_X({\varvec{\beta }}) \equiv \frac{\partial {\mathbf{f}}(\mathbf{X}, {\varvec{\beta }})}{\partial {\varvec{\beta }}^\top }\), \({\mathbf{S}}_X^{(1)}({\varvec{\beta }})=\frac{\partial \left[ \text{ vec }\{{\mathbf{S}}_X({\varvec{\beta }})\}\right] }{\partial {\varvec{\beta }}^\top }\), \({\mathbf{Y}}=\left( {\mathbf{Y}}_1^\top ,\ldots ,\mathbf{Y}_m^\top \right) ^\top \),
where \(\tilde{\mathbf{S}}^{(1)}(\mathbf{X},{\varvec{\beta }})\) and \(\tilde{\mathbf{S}}^{(1)}_{(1)}(\mathbf{X},{\varvec{\beta }})\) are defined similar to \({\mathbf{S}}_X^{(1)}({\varvec{\beta }})\).
In particular, define the matrix functions of \({\varvec{\beta }}\) as follows:
Appendix 2: Regularity conditions
In the paper, \(n_i\ge 1\) with \(n_i's\) bounded and then \(n\rightarrow \infty \) means \(m\rightarrow \infty \).
C.1. Assume \(f(x_{ij},{\varvec{\beta }})\) is a twice continuous differentiable function of \({\varvec{\beta }} \in {\mathscr {B}}\) for each \(x_{ij}\) and a measurable function of \(x_{ij}\) for each \(\beta \) where \({\mathscr {B}}\) is a compact subset of a Euclidean space.
C.2. \(Q_{n0}({\varvec{\beta }},{\varvec{\beta }}')\equiv \frac{1}{n} \left\{ \mathbf{f}(\mathbf{X},{\varvec{\beta }})-\mathbf{f}(\mathbf{X},{\varvec{\beta }}')\right\} ^\top \left\{ \mathbf{f}(\mathbf{X},{\varvec{\beta }})-\mathbf{f}(\mathbf{X},{\varvec{\beta }}')\right\} \) converges uniformly to a continuous function \(Q_0({\varvec{\beta }},{\varvec{\beta }}')\) and \(Q_0({\varvec{\beta }},{\varvec{\beta }}')=0\) if and only if \({\varvec{\beta }}'={\varvec{\beta }}\).
C.3. Assume \(\frac{1}{n} \mathbf{G}({\varvec{\beta }})^\top \mathbf{G}({\varvec{\beta }})\) with \(\mathbf{G}({\varvec{\beta }})=\left( {\mathbf{f}}({\mathbf{X}},{\varvec{\beta }}),{\mathbf{S}}_{X}({\varvec{\beta }}),{\mathbf{S}}^{(1)}({\mathbf{X}},{\varvec{\beta }})\right) \) and the following items all converge uniformly in \({\varvec{\beta }} \in {\mathscr {B}}\):
-
(1)
\( \lim _{n \rightarrow \infty } \frac{1}{n} {\mathbf{S}}_{X}({\varvec{\beta }})^\top {\mathbf{S}}_{X}({\varvec{\beta }})\equiv {\varvec{\Sigma }}_{S_X}({\varvec{\beta }})\) with \({\varvec{\Sigma }}_{S_X}({\varvec{\beta }})\) being a nonsingular matrix;
-
(2)
\( \lim _{n \rightarrow \infty } \frac{1}{n} {\mathbf{S}}_{X}({\varvec{\beta }})^\top {\mathbf{P}}_{Z^\bot } {\mathbf{S}}_{X}({\varvec{\beta }})\equiv {\varvec{\Sigma }}_{\widetilde{S}_X}(\beta )\) with \({\varvec{\Sigma }}_{\tilde{S}_X}({\varvec{\beta }})\) being a nonsingular matrix;
-
(3)
\( \lim _{n \rightarrow \infty } \frac{1}{n} {\mathbf{S}}_{X}({\varvec{\beta }})^\top {\mathbf{P}}_{Z^{(1)\bot }} {\mathbf{S}}_{X}({\varvec{\beta }})\equiv {\varvec{\Sigma }}_{\widetilde{\mathbf{S}}_X(1)}({\varvec{\beta }})\) with \({\varvec{\Sigma }}_{\tilde{S}_X(1)}({\varvec{\beta }})\) being a nonsingular matrix.
C.4. (1) \({\mathbf{b}}_i's\) are independently and identically distributed (i.i.d.), which are also independent of the i.i.d. \(\varepsilon _{ij}'s\); (2) Random effects \({\mathbf{b}}_i's\) and random errors \(\varepsilon _{ij}'s\) have the finite fourth-order moments.
C.5. (1) \(n_i \ge q_1\) and there exists infinite \(n_i\) such that \(n_i> q_1\) as \(m\rightarrow \infty \);
(2) \((n_i-q_i)/{\sum _{l=1}^i(n_l-q_l)}=O({1}/{i})\); (3) \(\mathrm{rank}\{{\mathbf{S}}_X({\varvec{\beta }})\}\) is bounded and \(rz-rz_1\rightarrow \infty \).
C.6. (1) \(n^{-1}\sum _{i=1}^m \text{ tr }\left\{ \text{ diag }\left( {\mathbf{P}}_{Z_i^\bot }\right) \right\} ^2\) are bounded away from zero;
(2) \(\sum _{i=1}^\infty \frac{\text{ tr }\{\text{ diag }^2 {\mathbf{P}}_{Z_i^\bot }\}}{\sum _{l=1}^i \text{ tr }\{\text{ diag }^2 {\mathbf{P}}_{Z_l^\bot }\}}< \infty \).
C.7. (1) \(m^{-1}\sum _{i=1}^m ({\mathbf{Z}}_i^\top {\mathbf{Z}}_i)^{-1}=O(1)\);
(2) As \(m\rightarrow \infty \), \(\text{ tr }\left( {\mathbf{Z}}^\top {\mathbf{Z}} \right) ^{-2}-\left[ \text{ tr }\left\{ ({\mathbf{Z}}^\top {\mathbf{Z}})^{-1}\right\} \right] ^2 /n \) converges to infinity;
(3) The maximal eigenvalues of \({\mathbf{Z}}_i^\top {\mathbf{Z}}_i\) are bounded.
Remark 2
Conditions C.1–C.3 are the requirements for the nonlinear function, which are similar to those in Vonesh and Carter (1992) and Wu (1981). Condition C.4 is about the independence and the moments of random effects and random errors. Distributional assumptions are not needed for model errors except for the moment conditions in C.4 since our procedures are distribution-free. Conditions C.5–C.7 are conditions of covariates \(\mathbf{X}_i,\mathbf{Z}_i\) and \(n_i\). Condition C.7 is the requirement for \({\mathbf{Z}}_i\), which is mild. In fact, for the special case with \({\mathbf{Z}}_i=1_{n_0}\), \(\mathrm{tr}\left( {\mathbf{Z}}^\top {\mathbf{Z}} \right) ^{-2}-\left[ \mathrm{tr}\left\{ ({\mathbf{Z}}^\top {\mathbf{Z}})^{-1}\right\} \right] ^2 /n= (n_0-1)m/n_0^3 \rightarrow \infty \) as \(m\rightarrow \infty \).
Appendix 3: Proofs
Before proving the theorems, we will introduce some lemmas first.
Lemma 1
Let \(\{\xi _i\}\) be a sequence of independent random variables with zero means. If \(\{a_i\}_{i=1}^k\) is a sequence of real numbers satisfying \(a_i\uparrow \infty \) and \(\sum _{i=1}^\infty {\mathrm{E}|\xi _i|^\delta }/{a_i^\delta } < \infty \) for some \(\delta \in [1,2]\), then \(\sum _{i=1}^m \xi _i /a_m \mathop {\rightarrow }\limits ^{a.s.}0\).
Proof
It is Theorem 12 in Chap. 9 in Petrov (1975). \(\square \)
Lemma 2
Suppose that \(\{\xi _i\}\) is a sequence of i.i.d. random variables with \(\mathrm{E}|\xi _1|<\infty \), \(\{a_i\}_{i=1}^k\) is a sequence of positive real numbers. Denote \(A_k=\sum _{i=1}^k a_i\) and \(B_k=\#\{i:A_i/a_i \le k\}/k\) with \(\#\{i:A_i/a_i \le k\}\) being the number of i satisfying \(A_i/a_i \le k\), and then \(\sum _{i=1}^k a_i \xi _i/A_k \mathop {\rightarrow }\limits ^{a.s.} \mathrm{E} \xi _1\) if \(A_k \rightarrow \infty \); in particular, if \(c_1\le a_i \le c_2\) for some positive constants \(c_1,c_2\), \(\sum _{i=1}^k a_i \xi _i/A_k \mathop {\rightarrow }\limits ^{a.s.} \mathrm{E}\xi _1\).
Proof
It is the theorem in Jamison et al. (1965). \(\square \)
Lemma 3
(1) Suppose conditions C.1, C.2, C.3, C.4(1) hold. Then \(n^{-1} {\mathbf{g}}^{(1)}({\varvec{\beta }})+{\varvec{\Sigma }}_{S_X}({\varvec{\beta }}) \mathop {\rightarrow }\limits ^{a.s.} 0 \) under \(H_{0G}\); \(n^{-1} \tilde{\mathbf{g}}_{(1)}^{(1)}({\varvec{\beta }})+{\varvec{\Sigma }}_{\tilde{S}_X(1)}({\varvec{\beta }})\mathop {\rightarrow }\limits ^{a.s.} 0\) under \(H_{0}\), and \(n^{-1} \tilde{\mathbf{g}}^{(1)}({\varvec{\beta }})+{\varvec{\Sigma }}_{\tilde{S}_X}({\varvec{\beta }})\mathop {\rightarrow }\limits ^{a.s.} 0 \) under \(H_{A}\). All converge uniformly in \({\varvec{\beta }} \in {\mathscr {B}}\);
(2) Under conditions C.1, C.2, C.4(1),
Proof
(1) Recall the notation of \({\mathbf{g}}^{(1)}({\varvec{\beta }})\), denote \({\mathbf{II}}_1^{g(1)} ={\mathbf{I}}_p\otimes \{{\mathbf{Y}}-{\mathbf{f}}({\mathbf{X}},{\varvec{\beta }})\}^\top {\mathbf{S}}_X^{(1)}({\varvec{\beta }})\), let \({\mathbf{II}}_{1k}^{g(1)}\) be the kth row of \({\mathbf{II}}_1^{g(1)}\) for \(k=1,2,\ldots ,p\), and then it is sufficient to verify \(n^{-1} {\mathbf{II}}_{1k}^{g(1)} \mathop {\rightarrow }\limits ^{a.s.} 0\) under \(H_{0G}\). In fact, conditions C.3, C.4(1) indicate that \(n^{-1}{\mathbf{II}}_{1k}^{g(1)}= n^{-1} \sum _{i=1}^m \sum _{j=1}^{n_i} {\mathbf{S}}_{x_{ij}}({\varvec{\beta }}) \varepsilon _{ij} \mathop {\rightarrow }\limits ^{a.s.} 0\). Similarly, the conclusions about \(n^{-1} \tilde{\mathbf{g}}_{(1)}^{(1)}({\varvec{\beta }})\) under \(H_0\) and \(n^{-1} \tilde{\mathbf{g}}^{(1)}({\varvec{\beta }})\) under \(H_A\) in Lemma 3 (1) can be derived, respectively.
(2) According to Theorem 1 (1), under conditions C.1, C.2, C.4(1), \({\tilde{{\varvec{\beta }}}}\mathop {\rightarrow }\limits ^{a.s.} \beta \) where \({\tilde{{\varvec{\beta }}}}\) is the estimate of \({{\varvec{\beta }}}\), which may be \({\hat{{\varvec{\beta }}}}_f\), \({\hat{{\varvec{\beta }}}}_0\) or \(\hat{{\varvec{\beta }}}_{0G}\) defined in (8). Then by the Taylor expansions, \(\mathbf{f}(\mathbf{X},{\tilde{{\varvec{\beta }}}})=\mathbf{f}(\mathbf{X},{{\varvec{\beta }}})+\mathbf{S}_X({{\varvec{\beta }}}) ({\tilde{{\varvec{\beta }}}}-{\varvec{\beta }})+o({\tilde{{\varvec{\beta }}}}-{\varvec{\beta }})\) a.s., and
which yield the conclusions of Lemma 3(2) by recalling the notations of \(\tilde{\mathbf{g}}({\varvec{\beta }})\), \(\tilde{\mathbf{g}}_{(1)}({\varvec{\beta }})\) and \({\mathbf{g}}({\varvec{\beta }})\), respectively. \(\square \)
Lemma 4
For the error \({\varepsilon }\) in model (1) in the text, which has independent components, we have the following conclusions for any \(n\times n\) symmetric matrix \(\mathbf A _n\),
-
(i)
\(\text{ var }({{\varvec{\varepsilon }}}^\top \mathbf A _n {{\varvec{\varepsilon }}})= (\kappa -3 \sigma ^4 ) \cdot \text{ tr }\{\text{ diag }^2(\mathbf A _n )\}+ 2 \sigma ^4 \cdot \text{ tr }(\mathbf A _n^2 )\);
-
(ii)
\({\nu _{max}(\mathbf{A}_n^2)}/{\mathrm{tr}(\mathbf{A}_n^2)}\rightarrow 0\), with \(\nu _{max}(\mathbf{A}_n^2)\) being the maximal eigenvalue of the matrix \(\mathbf{A}_n^2\), implies
$$\begin{aligned} \frac{{{\varvec{\varepsilon }}}^\top \mathbf A _n {{\varvec{\varepsilon }}}- \sigma ^2 \cdot \text{ tr } (\mathbf A _n)}{\sqrt{ (\kappa -3 \sigma ^4 ) \cdot \mathrm{tr}\{\text{ diag }^2(\mathbf A _n)\}+2 \sigma ^4 \cdot \mathrm{tr}(\mathbf A _n^2)}} \mathop {\rightarrow }\limits ^{d} \text{ N }(0, 1). \end{aligned}$$
Proof
This lemma is Lemma 3 in the supplement of Li et al. (2014). \(\square \)
Lemma 5
Assume that \({\varvec{\varsigma }}=({\varsigma }_1,{\varsigma }_2,\ldots ,{\varsigma }_k)^\top \) is a \(k\times 1\) random vector with its elements having the fourth-order moments, and then for any \(k\times k\) matrix \({ \mathbf M} \), \(\mathrm{E}({\varvec{\varsigma }} ^\top \mathbf{M}{\varvec{\varsigma }})^2= \mathrm{vec}^\top (\mathbf{M}^\top \otimes \mathbf{M}) \cdot {\varvec{\gamma }}_{4\zeta }\) with \( \gamma _{4\zeta }\equiv \mathrm{E}\{({\varvec{\varsigma }} \otimes {\varvec{\varsigma }} )^{\otimes 2}\} \).
Proof
Note that
\(\square \)
Lemma 6
Suppose \({\varvec{\xi }}=({\varvec{\xi }}_1^\top ,\ldots ,{\varvec{\xi }}_m^\top )^\top \) is a \(q \times 1\) random vector where \({\varvec{\xi }}_i's\) are independent \(q_i\times 1\) sub-vector for different i with mean 0 and covariance matrix \({\varvec{\Sigma }}_i\). For any \(q\times q\) symmetric diagonal block matrix \(\mathbf{A}_q=\mathrm{diag}({\mathbf{A}}_{11},\ldots ,{\mathbf{A}}_{mm})\) with the ith block \({\mathbf{A}}_{ii}\) being \(q_i\times q_i\),
-
(i)
\( \text{ var }({\varvec{\xi }}^\top \mathbf{A }_q {{\varvec{\xi }}})= \sum _{i=1}^m \left[ \text{ vec }^\top ({\mathbf{A}}_{ii}^\top \otimes {\mathbf{A}}_{ii}) \cdot \mathrm{E}\{({\varvec{\xi }}_i \otimes {\varvec{\xi }}_i)^{\otimes 2}\}- \left\{ \text{ tr }({\mathbf{A}}_{ii} {\varvec{\Sigma }}_i)\right\} ^2 \right] \);
-
(ii)
Assume for different i, \(q_i \equiv q_1\) and \(\mathrm{E}\{({\varvec{\xi }}_i \otimes {\varvec{\xi }}_i)^{\otimes 2}\}\equiv {\varvec{\gamma }}_{4\xi _1}\). If all the elements of \({\varvec{\xi }}_i\) have the \((4+\delta )\)th-order moments for some \(\delta >0\), then
$$\begin{aligned} \frac{{\varvec{\xi }}^\top \mathbf{A }_q {\varvec{\xi }}- \sum _{i=1}^m \text{ tr } ({\mathbf{A}}_{ii} {\varvec{\Sigma }}_i )}{\sqrt{ \text{ vec }^\top \left\{ \sum _{i=1}^m \left( {\mathbf{A}}_{ii}^\top \otimes {\mathbf{A}}_{ii}\right) \right\} \cdot {\varvec{\gamma }}_{4\xi _1}-\sum _{i=1}^m\left\{ \text{ tr }({\mathbf{A}}_{ii}{\varvec{\Sigma }}_i)\right\} ^2}} \mathop {\rightarrow }\limits ^{d} \text{ N }(0, 1). \end{aligned}$$
Proof
(i) Note that by Lemma 5
(ii) Let \(\eta _i={\varvec{\xi }}_i^\top {\mathbf{A}}_{ii} {\varvec{\xi }}_i\), and then \(\eta _i's\) are independent. Denote \(B_m^2=\sum _{i=1}^m \text{ var }(\eta _i)\) which can be written as \(B_m^2=\text{ vec }^\top \left\{ \sum _{i=1}^m \left( {\mathbf{A}}_{ii}^\top \otimes {\mathbf{A}}_{ii}\right) \right\} \cdot {\varvec{\gamma }}_{4\xi _1}-\sum _{i=1}^m\left\{ \text{ tr }({\mathbf{A}}_{ii}{\varvec{\Sigma }}_i)\right\} ^2\). According to the conditions in the lemma and Lyapunov cental limit theory (CLT), the conclusion is verified. \(\square \)
Proof of Theorem 1
(Asymptotic properties of estimates for \({\varvec{\beta }}\) and \(\sigma ^2\)):
(1) Note that \({\mathbf{P}}_{Z^{(1)^\bot }}\) and \({\mathbf{P}}_{Z^\bot }\) are idempotent matrices and the functions \(\tilde{\mathbf{f}}_{(1)}(\cdot )\) and \(\tilde{\mathbf{f}}(\cdot )\) satisfy conditions C.1, C.2(2), C.4(1), and then the strong consistency of \(\hat{{\varvec{\beta }}}_0\) and \(\hat{{\varvec{\beta }}}_f\) is obtained by Jennrich (1969).
(2) According to conditions C.1–C.3, C.4(1) and the proof of the asymptotic normality of \(\hat{{\varvec{\beta }}}_{0G}\) in Jennrich (1969), the conclusions in Theorem 1(2) can be derived by the mean-value theorem and the convergence of the estimators \(\hat{{\varvec{\beta }}}_f\) and \(\hat{{\varvec{\beta }}}_0\) in Theorem 1(1). \(\square \)
Proof of Theorems 2–5 and Corollary 1:
See the Supplement. \(\square \)
Appendix 4: derivation of \({\varvec{\gamma }}_{4b}\)
See the Supplement. \(\square \)
Rights and permissions
About this article
Cite this article
Li, Z. Inference of nonlinear mixed models for clustered data under moment conditions. TEST 26, 759–781 (2017). https://doi.org/10.1007/s11749-017-0532-8
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11749-017-0532-8