Abstract
Covariate measurement error problems have been recently studied for current status failure time data but not yet for multivariate current status data. Motivated by the three-hypers dataset from a health survey study, where the failure times for three-hypers (hyperglycemia, hypertension, hyperlipidemia) are subject to current status censoring and the covariate self-reported body mass index may be subject to measurement error, we propose a functional inference method under the proportional odds model for multivariate current status data with mismeasured covariates. The new proposal utilizes the working independence strategy to handle correlated current status observations from the same subject, as well as the conditional score approach to handle mismeasured covariate without specifying the covariate distribution. The asymptotic theory, together with a stable computation procedure combining the Newton–Raphson and self-consistency algorithms, is established for the proposed estimation method. We evaluate the method through simulation studies and illustrate it with three-hypers data.
Similar content being viewed by others
References
Carroll RJ, Ruppert D, Stefanski LA, Crainiceanu CM (2006) Measurement error in nonlinear models: a modern perspective, 2nd edn. Chapman & Hall, Boca Raton
Chang IS, Wen CC, Wu YJ (2007) A profile likelihood theory for the correlated gamma-frailty model with current status family data. Stat Sinica 17:1023–1046
Chen MH, Tong X, Sun J (2007) The proportional odds model for multivariate interval-censored failure time data. Stat Med 26:5147–5161
Chen MH, Tong X, Sun J (2009) A frailty model approach for regression analysis of multivariate current status data. Stat Med 28:3424–3436
Clayton DG (1978) A model for association in bivariate life tables and its application in epidemiological studies of familial tendency in chronic disease incidence. Biometrika 65:141–151
Dunson DB, Dinse GE (2002) Bayesian models for multivariate current status data with informative censoring. Biometrics 58:79–88
Goggins WB, Finkelstein DM (2000) A proportional hazards model for multivariate interval-censored failure time data. Biometrics 56:940–943
Gumbel EJ (1960) Bivariate exponential distributions. J Am Stat Assoc 55:698–707
Heyde CC, Morton R (1998) Multiple roots in general estimating equations. Biometrika 85:954–959
Hosmer DW, Lemeshow S (2000) Applied logistic regression. Wiley, New York
Huang J (1996) Efficient estimation for the Cox model with interval censoring. Ann Stat 24:540–568
Korosok MR (2008) Introduction to empirical processes and semiparametric inference. Springer, New York
Ma S (2009) Cure model with current status data. Stat Sinica 19:233–249
Murphy SA, van der Vaart AW (1999) Observed information in semiparametric models. Bernoulli 5:381–412
Rudin W (1973) Functional analysis. McGraw-Hill, New York
Stefanski LA, Carroll RJ (1987) Conditional scores and optimal scores in generalized linear measurement error models. Biometrika 69:331–342
Su PF, Chi Y (2014) Marginal regression approach for additive hazards models with clustered current status data. Stat Med 33:46–58
Tsiatis AA, Davidian M (2001) A semiparametric estimator for the proportional hazards model with longitudinal covariates measured with error. Biometrika 88:447–458
van der Vaart AW (1998) Asymptotic statistics. Cambridge University Press, Cambridge
van der Vaart AW, Wellner JA (1996) Weak convergence and empirical processes. Springer, New York
Wei LJ, Lin DY, Weissfeld L (1989) Regression analysis of multivariate incomplete failure time data by modeling marginal distributions. J Am Stat Assoc 84:1065–1073
Wen CC, Chen YH (2011) Nonparametric maximum likelihood analysis of clustered current status data with the gamma frailty Cox model. Comput Stat Data Anal 55:1053–1060
Wen CC, Chen YH (2012) Conditional score approach to errors-in-variable current status data under the proportional odds model. Scand J Stat 39:635–644
Wen CC, Huang SYH, Chen YH (2011) Cox regression for current status data with mismeasured covariates. Can J Stat 39:73–88
Acknowledgments
We wish thank the Associate Editor and two anonymous referees for their many helpful comments and suggestions that greatly improved the paper. This research was supported by the National Science of Council of Taiwan grants.
Author information
Authors and Affiliations
Corresponding author
Appendix
Appendix
We use the notation \({\mathbb P}_n, P_0,\) and \(P\) for the expectations taken under the empirical distribution, the true underlying distribution, and a given model, respectively. For simplicity the assumptions and proofs for theories are presented under the simpler setting where the distribution of \((C_1,\ldots ,C_K)\) is independent of \((X,Z_1,\ldots ,Z_K),\) though the proposed method can allow the dependence case. We consider \(H_1,\ldots ,H_K\) are functions in \(\mathcal{H},\) the set of right-continuous non-decreasing functions that are uniformly bounded on the study period \([0,\tau ^*].\) Let \(\ell ({\varvec{\theta }})=\log L({\varvec{\theta }}).\) The asymptotic theories are based on the following regularity assumptions, which have been similarly made in the studies of univariate current status data (e.g. Huang 1996; van der Vaart 1998; Ma 2009; Wen and Chen 2012).
-
(C1)
The examination times \(C_1,\ldots ,C_K\) possess a common continuous density whose support is an interval \([\tau _1,\tau _2]\) with \(0<\tau _1<\tau _2<\tau ^*.\)
-
(C2)
The true parameter \(({\varvec{\beta }}_0,\sigma ^2_0)\) lies in the interior of a compact subset \(\mathcal B \times \mathcal{Q}\) of \({\mathbb R}^d \times (0,\infty )\); all \(H_{10},\ldots , H_{K0} \) are continuously differentiable on \([\tau _1,\tau _2]\) and satisfy \(-M<H_{k0}(\tau _1)<H_{k0}(\tau _2)<M\) for \(k=1,\ldots ,K.\)
-
(C3)
The distribution of \((X,\mathbf{Z}_1,\ldots ,\mathbf{Z}_K)\) is not concentrated on any proper subspace of \({\mathbb R}^{1+\sum _k d_k} \) and has a bounded support, where \(d_k=\dim (\mathbf{Z}_k)\).
-
(C4)
The functions \(\mathbf{g}_1^*,\ldots ,\mathbf{g}_K^*\) given in (6) are differentiable with bounded derivatives on \([\tau _1,\tau _2]\).
-
(C5)
The information matrix \(\mathcal{I}\) defined in (7) is invertible.
Proof of Theorem 1
Let \((\widehat{H}_{1,({\varvec{\beta }},\sigma ^2)}, \ldots , \widehat{H}_{K,({\varvec{\beta }},\sigma ^2)} )\) denote the maximizer of \(L_n\) with \(({\varvec{\beta }}, \sigma ^2)\) fixed. We first apply Theorem 5.7 of van der Vaart (1998) to establish the consistency of \((\widehat{H}_{1,({\varvec{\beta }}_0,\widehat{\sigma }^2)}, \ldots , \widehat{H}_{K,({\varvec{\beta }}_0,\widehat{\sigma }^2)} )\). Since the class of monotone and uniformly bounded functions is a Donsker class, by Theorem 2.10.6 of van der Vaart and Wellner (1996) and conditions (C1)-(C3), we know that the class \(\{ \ell ({\varvec{\beta }}_0,H_1,\ldots ,H_K,\) \( \sigma ^2_0) \ | \ H_1,\ldots ,H_K \in \mathcal{H} \}\) is Donsker and hence Glivenko-Cantelli. Further, by Jensen’s inequality, we have
and the equality holds only if \(H_k=H_{k0}\) on \((\tau _1,\tau _2)\) for all \(k=1,\ldots ,K.\) This indicates that
Furthermore, note that
where the inequality follows from the definition of \((\widehat{H}_{1,({\varvec{\beta }},\sigma ^2)}, \ldots , \widehat{H}_{K,({\varvec{\beta }},\sigma ^2)} )\), and two equalities are obtained by the mean value theorem and the consistency of \(\widehat{\sigma }^2.\) Therefore, by Theorem 5.7 of citev98, we have \(\sum _{k=1}^K \Vert \widehat{H}_{k,({\varvec{\beta }}_0,\widehat{\sigma }^2)}-H_{k0}\Vert _2 {\rightarrow } \ 0,\) and so \(\widehat{H}_{k,({\varvec{\beta }}_0,\widehat{\sigma }^2)}(t) \mathop {\rightarrow }\limits ^{P} H_{k0}(t),\) \(k=1,\ldots , K,\) for every \(t\) in \([\tau _1,\tau _2]\).
By the consistency shown above and the mean value theorem, we have
Since the right hand side of the above display set to zero is an unbiased estimating equation, a consistent sequence of solutions of \({\varvec{\beta }}\) to it exists, indicating the existence of a consistent solution of \({\varvec{\beta }}\) to the CS estimating equation \({\mathbb P}_n {\varvec{\ell }}_0({\varvec{\beta }},\widehat{H}_{1,({\varvec{\beta }},\widehat{\sigma }^2)}, \ldots , \widehat{H}_{K,({\varvec{\beta }},\widehat{\sigma }^2)},\widehat{\sigma }^2)=0.\)
Now we shall prove
by verifying the conditions (3.5) and (3.6) in Theorem 3.2 of Murphy and van der Vaart (1999). Here \(\Vert \cdot \Vert \) is the Euclidean norm, \(\Vert H\Vert _2^2=\int H^2(u)dQ(u),\) and \(Q\) denotes the marginal distribution of \(C_k.\) The rates of convergence and consistencies of \(\widehat{H}_k, k=1,\ldots ,K,\) can then be obtained by the consistency of \((\widehat{\varvec{\beta }}, \widehat{\sigma }^2).\)
Given two functions \(l\) and \(u\), the bracket \([l,u]\) is the set of all functions \(f\) with \(l\le f \le u\). An \(\varepsilon \)-bracket in \(L_2(P)=\{f: Pf^2<\infty \}\) is a bracket \([l,u]\) with \(P(u-l)^2<\varepsilon ^2.\) For a subclass \(\mathcal{C}\) of \(L^2(P),\) the bracketing number \(N_{[ \ ]}(\varepsilon ,\mathcal{C},L_2(P))\) is the minimum number of \(\varepsilon \)-bracket need to cover \(\mathcal C\) (see van der Vaart 1998). Let \(\varPsi =\{ \ell ({\varvec{\theta }}) : {\varvec{\theta }}\in \mathcal{B}\times \mathcal{H}^K \times \mathcal{Q}\}.\) Note that each element in \(\varPsi \) is uniformly bounded and satisfies \(P_0\{ \ell ({\varvec{\theta }}) -\ell ({\varvec{\beta }},H_{10},\ldots ,H_{K0},\sigma ^2) \}^2 \preceq \sum _{k=1}^K \Vert H_k-H_{k0}\Vert _2^2+\Vert {\varvec{\beta }}-{\varvec{\beta }}_0\Vert ^2+\Vert \sigma ^2-\sigma ^2_0\Vert ^2.\) The notation \(\preceq \) means smaller than, up to a constant. Lemma 1 below gives the bracketing integral \(J(\varDelta ,{\varPsi },L_2(P))\), defined as \(\int _0^{\varDelta } \{1+\log N_{[ \ ]}(\varepsilon ,\varPsi ,L_2(P))\}^{1/2}d\varepsilon ,\) is \(O(\varDelta ^{1/2})\). It then follows from Lemma 3.3 of Murphy and van der Vaart (1999) that their condition (3.6) is satisfied for \(\phi _n(\varDelta )=\varDelta ^{1/2}.\)
A similar Taylor series argument used in Lemma 2 below gives \(P_0\{\ell ({\varvec{\theta }}_0)-\ell ({\varvec{\beta }},H_{10},\ldots ,H_{K0},\sigma ^2))\}\preceq \Vert {\varvec{\beta }}-{\varvec{\beta }}_0\Vert ^2+\Vert \sigma ^2-\sigma ^2_0\Vert ^2.\) This and Lemma 2 imply
which is condition (3.5) of Murphy and van der Vaart (1999). This completes the proof.\(\square \)
Lemma 1
\(\log N_{[ \ ]}(\varepsilon ,\varPsi ,L_2(P_0)) =O( {1/ \varepsilon })\).
Proof
For fixed \(({\varvec{\beta }}, \sigma ^2)\) and \(1\le k \le K,\) the functions in \(\varPsi \) depend on \(H_k\) monotonically for \(\varDelta _{k}=1\) and \(\varDelta _{k}=0\) separately. Thus, given a \(\varepsilon \)-bracket \(H^{L}_k\le H_k \le H^{U}_k,\) it follows from monotonicity of \(\mathcal{E}_{k} \) in \(H_k\) that we can get a bracket \((\ell ^L,\ell ^U)\) for \(\ell ({\varvec{\theta }})\) where
Further, by the mean value theorem, we have \(|\ell ^L-\ell ^U|^2 \preceq \sum _{k=1}^K (H_k^{U}-H_k^{L})^2 (C_{k}).\) Thus brackets for \(H_k\) of \(\Vert \cdot \Vert _2\)-size \(\varepsilon \) can translate into brackets for \(\ell ({\varvec{\theta }})\) of \(L_2(P_0)\)-size proportional to \(\varepsilon .\) By Example 19.11 of van der Vaart (1998), we can cover the set of all \(H_k\) by \(\exp (C/\varepsilon )\) brackets of size \(\varepsilon \) for some constant \(C.\) Next we allow \(\zeta =({\varvec{\beta }}',\sigma ^2)'\) to vary freely as well. Because \(\mathcal{B}\times \mathcal{Q}\) is finite-dimensional and \((\partial /\partial \zeta ) \ell ({\varvec{\theta }})(O)\) is uniformly bounded in \(({\varvec{\theta }},O),\) this increases the entropy only slightly. Lemma 1 is thus proved.\(\square \)
Lemma 2
For \({\varvec{\theta }}\) near \({\varvec{\theta }}_0,\) \(P_0\{ \ell ({\varvec{\theta }})-\ell ({\varvec{\theta }}_0) \} \preceq -\{\sum _{k=1}^K\Vert H_k-H_{k0}\Vert _2^2+\Vert {\varvec{\beta }}-{\varvec{\beta }}_0\Vert ^2+\Vert \sigma ^2-\sigma ^2_0\Vert ^2\}.\)
Proof
Let \({\varvec{\theta }}_1=({\varvec{\beta }},H_{1},\ldots ,H_{K},\sigma ^2_0).\) It suffices to show
Let \(V=\{(C_k,S_k(\beta _{10},\sigma ^2_0),Z_k), k=1,\ldots ,K \}.\) By the Kullback-Leibler inequality, \(E_{0}\{ \ell ({\varvec{\theta }}_1)(O)|V \}\) is maximized at \(({\varvec{\beta }}_0,H_{10}(C_1),\ldots ,H_{K0}(C_K)).\) So its first derivative is equal to 0 there (this can also be verified directly by the fact that \(E_0\{\varDelta _k|V\}=\mathcal{E}_k({\varvec{\theta }}_0)(O)\)). Since \((C_k,S_k,Z_k)\)’s have bounded support and the parameter spaces are compact, a Taylor’s expansion gives \(E_0\{ \ell ({\varvec{\theta }}_1)(O)-\ell ({\varvec{\theta }}_0)(O) |V \} \preceq -\{\sum _{k} (H_k(C_k)-H_{k0}(C_k))^2+\Vert {\varvec{\beta }}-{\varvec{\beta }}_0\Vert ^2\},\) which implies (9).
We now prove (10). Denote \(p_{{\varvec{\theta }}}\) the density of observed data \(O\) given the model \({\varvec{\theta }}.\) Because the densities \(p_{{\varvec{\theta }}_0}/p_{{\varvec{\theta }}_1}\) are uniformly bounded above and below by a positive constant, (10) is equivalent to
Let \(V_1=\{(C_k,S_k(\beta _{1},\sigma ^2_0),Z_k), k=1,\ldots ,K \}.\) By the fact that \(E_{{\varvec{\theta }}_1}\{\varDelta _k|V_1\}=\mathcal{E}_k({\varvec{\theta }}_1)(O),\) we can show directly that the first derivative of \(E_{{\varvec{\theta }}_1}\{ \ell ({\varvec{\theta }})(O)|V_1 \}\) with respect to \(\sigma ^2\) at \(\sigma ^2=\sigma ^2_0\) is equal to 0. Note that the second derivative of \(E_{{\varvec{\theta }}_1}\{ \ell ({\varvec{\theta }})(O)|V_1 \}\) with respect to \(\sigma ^2\) is given by \(-E_{{\varvec{\theta }}_1}\{\beta _1^4 \sum _k \mathcal{V}_k({\varvec{\theta }})(O)/(4m^2)|V_1 \}.\) Consequently, a Taylor’s expansion around \(\sigma ^2=\sigma ^2_0\) can thus yield \(E_{{\varvec{\theta }}_1}\{ \ell ({\varvec{\theta }})(O)-\ell ({\varvec{\theta }}_1)(O) |V_1 \} \preceq - \Vert \sigma ^2-\sigma ^2_0\Vert ^2,\) which implies (11). This completes the proof.\(\square \)
For fixed \(1\le k \le K,\) consider a parametric path \(H_{k,\varepsilon }\) in \(\mathcal H\) through \(H_k,\) that is, \(H_{k,\varepsilon } \in \mathcal{H}\) and \(H_{k,\varepsilon }=H_k\) when \(\varepsilon =0.\) Let \(\dot{\mathcal{H}}=\{g_k:({\partial }/{\partial \varepsilon })|_{\varepsilon =0} H_{k,\varepsilon }=g_k,k=1,\ldots ,K\}.\) Then the score for \(H_k\) along the direction \(g_k,\) define by \( ({\partial }/{\partial \varepsilon })|_{\varepsilon =0} \ell ({\varvec{\beta }},H_1,\ldots ,H_{k-1},H_{k,\varepsilon },H_{k+1},\ldots ,H_K,\sigma ^2),\) has the form
Also define \(\ell _{0k}({\varvec{\theta }})[g_k]=({\partial }/{\partial \varepsilon }) |_{\varepsilon =0} \ell _0({\varvec{\beta }},H_1,\ldots ,H_{k-1},H_{k,\varepsilon },H_{k+1},\ldots ,H_K,\sigma ^2)\) and \(\ell _{k'k}({\varvec{\theta }})[\tilde{g}_{k'},g_k]=({\partial }/{\partial \varepsilon }) |_{\varepsilon =0} \ell _{k'}({\varvec{\beta }},H_1,\ldots ,H_{k-1},H_{k,\varepsilon },H_{k+1},\ldots ,H_K,\sigma ^2)[\tilde{g}_{k'}],\) where \(1\le k,k'\le K,\) \(g_k\) and \(\tilde{g}_k'\) are in \(\dot{\mathcal{H}}.\) They have forms
and \(\ell _{k'k}({\varvec{\theta }})[\tilde{g}_k', g_k]=0\) if \(k'\ne k.\) Following semiparametric M-estimator theories (e.g. Korosok 2008), the function \({\varvec{\ell }}^*\) given in Sect. 3 is defined as \({\varvec{\ell }}^*({\varvec{\theta }})= \ell _0({\varvec{\theta }})- \sum _{k=1}^K \ell _k({\varvec{\theta }})[\mathbf g_k^*],\) where \(\mathbf{g}_k^*\) is the \(d\)-dimensional \((d=\dim ({\varvec{\beta }}))\) vector-valued function satisfying
for all \(g_k\) in \(\dot{\mathcal{H}}.\) Fixing one \(k\) \((1\le k\le K)\) and setting \(g_{k'}=0\) for all \(k' \ne k\) in (12), we have
which implies that \(\mathbf{g}_k^*\) is given by (6). Below we establish the asymptotic theory of the CS estimator.
Proof of Theorem 2
We first verify
Apply a Taylor expansion to \({\varvec{\ell }}^*({\varvec{\beta }}_0, H_1,\ldots ,H_K,\sigma ^2_0)(O)\) at the point \((H_{10}(C_{1}),\ldots , \) \(H_{K0}(C_{K}))\) to get
Using (12), the fact that \(P_0 {\varvec{\ell }}^*({\varvec{\theta }}_0)=0,\) and applying the rates of convergence on \(\widehat{H}_k, k=1,\ldots , K,\) to (14), we get (13).
Applying Theorem 2.10.6 of van der Vaart and Wellner (1996), it can be verified with condition (C4) that \(\{ {\varvec{\ell }}^*({\varvec{\theta }}) | {\varvec{\theta }}\in \mathcal{B} \times \mathcal{H}^K \times \mathcal{Q}\}\) and \(\{ \varphi (\sigma ^2) | \sigma ^2 \in \mathcal{Q} \}\) are uniformly bounded Donsker classes; the proof of which is technical and hence omitted here. Combining this with the consistency of \(\widehat{\varvec{\theta }}\) leads to
Adding (13) to the first row of preceding display and using the facts that \(P_0 {\varvec{\ell }}^*({\varvec{\theta }}_0)=0\) and \({\mathbb P}_n {\varvec{\ell }}^*({\widehat{\varvec{\theta }}})={\mathbb P}_n\varphi (\widehat{\sigma }^2)=0,\) it is seen that
By the mean value theorem, there exists \((\tilde{{\varvec{\beta }}},\tilde{\sigma ^2})\) lying between \((\widehat{{\varvec{\beta }}},\widehat{\sigma }^2)\) and \(({\varvec{\beta }}_0,\sigma ^2_0)\) such that
By the consistency of \((\widehat{\varvec{\beta }},\widehat{\sigma }^2)\) and condition (C5), we have
This completes the proof.\(\square \)
Rights and permissions
About this article
Cite this article
Wen, CC., Huang, YH. & Wu, YJ. A functional inference for multivariate current status data with mismeasured covariate. Lifetime Data Anal 21, 379–396 (2015). https://doi.org/10.1007/s10985-014-9296-6
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10985-014-9296-6