Abstract
Function-based hypothesis testing in two-sample location-scale models has been addressed for uncensored data using the empirical characteristic function. A test of adequacy in censored two-sample location-scale models is lacking, however. A plug-in empirical likelihood approach is used to introduce a test statistic, which, asymptotically, is not distribution free. Hence for practical situations bootstrap is necessary for performing the test. A multiplier bootstrap and a model appropriate resampling procedure are given to approximate critical values from the null asymptotic distribution. Although minimum distance estimators of the location and scale are deployed for the plug-in, any consistent estimators can be used. Numerical studies are carried out that validate the proposed testing method, and real example illustrations are given.
Similar content being viewed by others
References
Ahmed N, Subramanian S (2016) Semiparametric simultaneous confidence bands for the difference of survival functions. Lifetime Data Anal 22:504–530
Andersen PK, Borgan O, Gill RD, Keiding N (1993) Statistical models based on counting processes. Springer, New York
Bhattacharya R, Subramanian SS (2014) Two-sample location-scale estimation from semiparametric random censorship models. J Multivar Anal 132:25–38
Bickel P, Götze F, van Zwet W (1997) Resampling fewer than n observations: gains, losses and remedies for losses. Stat Sin 7:1–31
Bickel PJ, Sakov A (2008) On the choice of \(m\) in the \(m\) out of \(n\) bootstrap and confidence bounds for extrema. Stat Sin 18:967–985
Chen S, Dahl DB, Khan S (2005) Nonparametric identifiaction and estimation of a censored location-scale regression model. J Am Stat Assoc 100:212–221
Dette H, Heuchenne C (2012) Scale checks in censored regression. Scand J Stat 39:323–339
Doksum K (1974) Empirical probability plots and statistical inference for nonlinear models in the two-sample case. Ann Stat 2:267–277
Doksum K, Sievers GL (1976) Plotting with confidence: graphical comparisons of two populations. Biometrika 63:421–434
Epps TW, Singleton KJ (1986) An omnibus test for the two-sample problem using the empirical characteristic function. J Stat Comput Simul 26:177–2003
Fleming TR, Harrington DP (1991) Counting processes and survival analysis. Wiley, New York
Gill RD, Johansen S (1990) A survey of product-integration with a view toward application in survival analysis. Ann Stat 15:1501–1555
Hall P, Lombard F, Potgieter CJ (2013) A new approach to function based hypothesis testing in location-scale families. Technometrics 55:215–223
Heuchenne C, Van Keilegom I (2010) Estimation in nonparametric location-scale regression models with censored data. Ann Inst Stat Math 62:439–463
Hjort NL, McKeague IW, Van Keilegom I (2009) Extending the scope of empirical likelihood. Ann Stat 37:1079–1111
Hsieh F (1996) Empirical process approach in a two-sample location-scale model with censored data. Ann Stat 24:2705–2719
Kalbfleisch JD, Prentice RL (2002) The statistical analysis of failure time data. Wiley, New Jersey
Klein JP, Moeschberger ML (2003) Survival analysis. Techniques for censored and truncated data. Springer, New York
Kyle R, Therneau T, Rajkumar V, Offord J, Larson D, Plevak M, Melton LJ III (2002) A long-term study of prognosis in monoclonal gammopathy of undetermined significance. New Engl J Med 346:564–569
Li G (1995) On nonparametric likelihood ratio estimation of survival probabilities for censored data. Stat Probab Lett 25:95–104
Li G (2003) Nonparametric likelihood ratio goodness-of-fit tests for survival data. J Multivar Anal 86:166–182
Li G, Wang Q-H (2003) Empirical likelihood regression analysis for right censored data. Stat Sin 13:51–68
Lv X, Zhang G, Ren G (2017) Gini index estimation for lifetime data. Lifetime Data Anal 23:275–304
Lin DY, Wei LJ, Ying Z (1993) Checking the Cox model with cumulative sums of martingale-based residuals. Biometrika 80:557–572
Lu HHS, Wells MT, Tiwari RC (1994) Inference for shift functions in the two-sample problem with right-censored data: with applications. J Am Stat Assoc 89:1017–1026
McKeague IW, Zhao Y (2002) Simultaneous confidence bands for ratios of survival functions via empirical likelihood. Stat Probab Lett 60:405–415
McKeague IW, Zhao Y (2005) Comparing distribution functions via empirical likelihood. Int J Biostat. https://doi.org/10.2202/1557-4679.1007
Mondal S, Subramanian S (2016) Simultaneous confidence bands for Cox regression from semiparametric random censorship. Lifetime Data Anal 22:122–144
O’Brien PC (1988) Comparing two samples: extensions of the t, rank sum, and log-rank tests. J Am Stat Assoc 83:52–61
Patil PN (1993) On the least squares cross-validation bandwidth in hazard rate estimation. Ann Stat 21:1792–1810
Politis D, Romano J (1994) Large sample confidence regions on subsamples under minimal assumptions. Ann Stat 22:2031–2050
Potgieter CJ, Lombard F (2016) Nonparametric two-sample estimation of location and scale parameters from empirical characteristic functions. J Stat Comput Simul 86:3225–3242
Qin G, Tsao M (2003) Empirical likelihood inference for median regression models for censored survival data. J Multivar Anal 85:416–430
Satten G, Datta S (2001) The Kaplan–Meier estimator as an inverse-probability-of-censoring weighted average. Am Stat 55:207–210
Thomas DR, Grunkemeier GL (1975) Confidence interval estimation of survival probabilities for censored data. J Am Stat Assoc 70:865–871
Ying Z, Jung S-H, Wei LJ (1995) Survival analysis with median regression models. J Am Stat Assoc 90:178–184
Zhang Z, Li G (1996) A simple quantile approach to the two-sample location-scale problem with random censorship. J Nonparametric Stat 6:323–335
Zhang Z, Yu Q (2002) A minimum distance estimation approach to the two-sample location-scale problem. Lifetime Data Anal 8:289–305
Acknowledgements
The author expresses his sincere thanks to an Associate Editor and two reviewers whose comments improved the overall quality of the manuscript.
Author information
Authors and Affiliations
Corresponding author
Appendix
Appendix
1.1 Review of the MDE of \(\varvec{\theta }\)
See also Bhattacharya and Subramanian (2014), where semiparametric estimation in two-sample location-scale models is investigated. Let \(\mathcal{Q}_{\varvec{\theta }}(s) = {\hat{Q}}_2(s)-\varphi _{\varvec{\theta }}({\hat{Q}}_1(s))\), where \({\hat{Q}}_i(s), i=1,2\), are the Kaplan–Meier quantiles. Let G(s) denote a positive measure on (0, 1). To obtain the MDE \({\hat{\varvec{\theta }}}\), one minimizes \(\mathcal{S}(\varvec{\theta })\), a Cramér-von Mises type discrepancy (Zhang and Yu 2002), defined as
Define \({\varvec{d}}=(d_1,d_2)^T\) where \(d_l=\langle {\hat{Q}}_1^{l-1}{\hat{Q}}_2, G\rangle \), \(l=1,2\), and let \({\varvec{A}}=(A_{il})_{2\times 2}\) where \(A_{il}=\langle {\hat{Q}}_1^{i-1}{\hat{Q}}_1^{l-1}, G\rangle \). Minimizing \(\mathcal{S}(\varvec{\theta })\) given by Eq. (A.1) yields \({\varvec{A}}\varvec{\theta }= {\varvec{d}}\). When \({\hat{Q}}_1\) is not a constant, A is non-singular (Zhang and Yu 2002). Then \({\hat{\varvec{\theta }}}={\varvec{A}}^{-1}{\varvec{d}}\). More specifically, let
Then \({\hat{\varvec{\theta }}}\equiv {\varvec{g}}({\hat{\varvec{U}}})=(g_1({\hat{\varvec{U}}}),g_2({\hat{\varvec{U}}}))^T\), where \(g_1\) and \(g_2\) are functions of \({\varvec{u}}=(u_1, u_2,u_3,u_4)^T\):
where \(c=\langle 1, G\rangle \). Zhang and Yu (2002) derived a large-sample representation for \({\hat{\varvec{\theta }}}\). To develop an equivalent martingale representation for \({\hat{\varvec{\theta }}}\), let \({\varvec{D}}=[D_{il}]_{2\times 4}=\frac{d{\varvec{g}}}{d{\varvec{u}}}\vert _{{\varvec{u}}={\varvec{U}}}\), where
Then \({\hat{\varvec{\theta }}}-\varvec{\theta }_0 = {\varvec{D}}({\hat{\varvec{U}}}-{\varvec{U}}) + o_{{\mathbb {P}}}(n^{-1/2})\). Let \(\tau _{i,l}\) and \(\tau _{i,u}\) denote the infimum and supremum of the support set for \(G\circ F_i(\cdot )\equiv G(F_i(\cdot ))\). To deal with \({\hat{\varvec{U}}}-{\varvec{U}}\), for \(s \in (\tau _{i,l}, \tau _{i,u})\), define
For the first component of \({\hat{\varvec{U}}}-{\varvec{U}}\), say, apply Eq. (2.4), then apply Theorem II.6.2 of Andersen et al. (1993) for \({\hat{F}}_1-F_1\), and finally interchange the order of integration to obtain
The other components of \({\hat{\varvec{U}}}-{\varvec{U}}\) can be treated likewise. Let \(\varvec{W}(s)\equiv [W_{kl}(s)]_{4\times 2}\), where \(W_{11}(s)=e_1(s), W_{21}(s)=2{\tilde{e}}_{11}(s)\), \(W_{32}(s)=e_2(s)\), \(W_{41}(s)={\tilde{e}}_{12}(s)\), \(W_{42}(s)={\tilde{e}}_{21}(s)\), and \(W_{12}(s)=W_{22}(s)=W_{31}(s)=0\). It follows that \({\hat{\varvec{\theta }}}- \varvec{\theta }_0\) has the large-sample representation
Note that the last step above is obtained by applying Eq. (2.1). We will exploit Eq. (A.3) when we derive the large-sample null distribution of the plug-in EL.
Remark A.1
For a purely location model, \(\theta _2=1\). Then \(\theta _1\equiv \theta \) is the only parameter and its MDE is \({\hat{\varvec{\theta }}}\equiv {\hat{\theta }} = \langle {\hat{Q}}_2 - {\hat{Q}}_1, G\rangle \). We can write \(\hat{\theta }- \theta _0={\varvec{D}}^T({\hat{\varvec{U}}}-{\varvec{U}})\), where \(D=(-1,1)^T\) and \({\hat{\varvec{U}}}-{\varvec{U}}=(\langle {\hat{Q}}_1 - Q_1, G\rangle , \langle {\hat{Q}}_2 - Q_2, G\rangle )^T\). Then \({\hat{\theta }} - \theta _0\) has the large-sample representation
which one can exploit when testing the adequacy of a location model, see Remark 7.
Recall that \({\tilde{\kappa }}_i(t)\) is the number of distinct uncensored lifetimes by time t for each sample. In the notation to handle ties, the two cumulative hazard estimators are given by
Let \(\Vert h\Vert _{\alpha _1}^{\alpha _2} = \sup _{t\in [\alpha _1,\alpha _2]}|h(t)|\). Lemma 1 gives the asymptotic equivalence between \({\hat{\Lambda }}_i\) and \({\hat{\zeta }}_i\).
Lemma 1
Assume that \(S_i(t), i=1,2\), are continuous. Then \(\Vert \hat{\Lambda }_i-{\hat{\zeta }}_i\Vert _{\alpha _1}^{\alpha _2} = o_{{\mathbb {P}}}(n^{-1/2})\).
Proof
Following McKeague and Zhao (2005), see their Eq. (A.6), we can show that
By continuity of \(S_i\), the first quantity on the right hand side (RHS) of inequality (A.6) is
From Eq. (A.5), the second quantity on the RHS of inequality (A.6) is bounded above by \(\hat{\Lambda }_i(\alpha _2)=O_{{\mathbb {P}}}(1)\). Combined with inequalities (A.6) and (A.7), this completes the proof. \(\square \)
Lemma 2
The Lagrange multiplier, \({\hat{\lambda }}_{\varvec{\theta }}(t)\), solving Eq. (2.10), satisfies
Proof
From Eq. (2.10), \({\hat{\lambda }}_{\varvec{\theta }}(t)\) satisfies the equation \(I_1(t)-I_2(t)=0\), where
Consider \({\hat{\lambda }}_{\varvec{\theta }}(t)<0\). Follow Li (1995), leading to his Eq. (2.12) on p. 101, to obtain
Applying some elementary calculations (cf. Ahmed and Subramanian, 2015), we can show that, for large enough \(n_2\), \(0<|{\hat{\lambda }}_{\varvec{\theta }}(t)|/n_2<1\) almost surely. Now, as in McKeague and Zhao (2005), use the fact that \(1/(1-x)\ge 1+x\) for \(0<x<1\), to obtain
Since \(-{\hat{\lambda }}_{\varvec{\theta }}(t)>0\), follow Li (1995) leading to his Eq. (2.13) on p. 101, and obtain
Combine inequalities (A.10) and (A.11) to obtain inequality (A.8) from
When \({\hat{\lambda }}_{\varvec{\theta }}(t)>0\), the above approach can be repeated by applying Li’s (1995) technique in reverse order. First obtain the inequality
Then obtain the inequality
Combine inequalities (A.12) and (A.13) to obtain inequality (A.9) from
\(\square \)
Remark A.3
Define \(\mu (\cdot )=-\log (\cdot )\), which is a continuously differentiable function. Then \({\hat{\zeta }}_2(t)=\mu ({\hat{S}}_2(t))\). It follows by Lemma 1 of Ying et al. (1995) that
Lemma 3
The estimate \({\hat{\lambda }}_{{\hat{\varvec{\theta }}}}(t)\), solving Eq. (2.10) with \(\varvec{\theta }={\hat{\varvec{\theta }}}\), satisfies \(\Vert {\hat{\lambda }}_{{\hat{\varvec{\theta }}}}\Vert _{\alpha _1}^{\alpha _2} = O_{{\mathbb {P}}}(n^{1/2})\).
Proof
The denominators on the RHS of inequalities (A.8) and (A.9) are each \(O_{{\mathbb {P}}}(1)\), uniformly for \(t\in [\alpha _1,\alpha _2]\). This follows from \(\Vert \hat{\Lambda }_i-\Lambda _i\Vert _0^{\tau _i}=o_{{\mathbb {P}}}(1)\), where \(\tau _1=\alpha _2\) and \(\tau _2=\varphi _{\varvec{\theta }_0}(\alpha _2)\), see Sect. 2.2. Furthermore, the denominators on the RHS of inequalities (A.8) and (A.9) are bounded away from 0 for \(t\in [\alpha _1, \alpha _2]\). It remains to show that the numerators on the RHS of inequalities (A.8) and (A.9) are each \(O_{{\mathbb {P}}}(n^{1/2})\) uniformly for \(t\in [\alpha _1,\alpha _2]\). Consider the numerator on the RHS of inequality (A.8). Note that under the null hypothesis, \(-\log (S_2(\varphi _{\varvec{\theta }_0}(t))=\Lambda _2(\varphi _{\varvec{\theta }_0}(t))=\Lambda _1(t)\). Therefore,
By Lemma 1, \(\Vert {\hat{\zeta }}_1-{\hat{\Lambda }}_1\Vert _{\alpha _1}^{\alpha _2}=o_{{\mathbb {P}}}(n^{-1/2})\). Also, it is standard that \(\Vert {\hat{\Lambda }}_1-\Lambda _1\Vert _{\alpha _1}^{\alpha _2}=O_{{\mathbb {P}}}(n^{-1/2})\). By Remark A.2, the third expression within square brackets is \(o_{{\mathbb {P}}}(n^{-1/2})\) uniformly for \(t \in [\alpha _1, \alpha _2]\). By Remark A.3, the fourth expression within square brackets is \(o_{{\mathbb {P}}}(n^{-1/2})\) uniformly for \(t \in [\alpha _1, \alpha _2]\). By a Taylor expansion around \(\varphi _{\varvec{\theta }_0}(t)\) and the \(n^{1/2}\) consistency of the MDE \({\hat{\varvec{\theta }}}\) of \(\varvec{\theta }\), the fifth expression within square brackets is seen to be \(O_{{\mathbb {P}}}(n^{-1/2})\) uniformly for \(t\in [\alpha _1, \alpha _2]\). By a Taylor expansion, the last expression, which equals \(-\log {\hat{S}}_2(\varphi _{\varvec{\theta }_0}(t)) + \log S_2(\varphi _{\varvec{\theta }_0}(t))\), is also \(O_p(n^{-1/2})\) uniformly for \(t\in [\alpha _1, \alpha _2]\). Likewise, it can be shown that the numerator on the RHS of inequality (A.9) is also \(O_{{\mathbb {P}}}(n^{-1/2})\) uniformly for \(t\in [\alpha _1, \alpha _2]\). \(\square \)
In the proof of Theorem 1, we will show that \(-2R_{{\hat{\varvec{\theta }}}}(t)\) is asymptotically equivalent to the square of \(n^{1/2}{\hat{{\mathbb {V}}}}(t)\) scaled by the reciprocal of \(\sigma _{\mathrm{c}}^2(t)\) defined by Eq. (2.14),
Lemma 4
Under the null hypothesis, \({\hat{{\mathbb {V}}}}(t)\) is asymptotically linear with influence function \(\sum _{i=1}^2 J_{i1}(t)/\sqrt{\rho _i}\), where \(J_{11}(t)\) and \(J_{21}(t)\) are given by Eqs. (2.15) and (2.16) respectively.
Proof
First apply a Taylor’s expansion for \(\log {\hat{S}}_1(t)\) about \(S_1(t)\). Then apply the Duhamel equation (Andersen et al. 1993) and Eq. (2.1) to yield an asymptotic representation for the first term of \({\hat{{\mathbb {V}}}}(t)\) as shown below:
For the second term of \({\hat{{\mathbb {V}}}}(t)\), applying Eq. (A.14) of Remark A.3, it follows that
Applying the delta method to the first and second terms on the RHS of Eq. (A.16) yields
the last step that leads to Eq. (A.17) following from Eq. (A.3), where \(\lambda _2\) is the hazard function associated with \(F_2\) and \({\varvec{c}}_t=(1,t)^T\). The third term on the RHS of Eq. (A.16) can be treated exactly the way the first term of \({\hat{{\mathbb {V}}}}(t)\) yielded Eq. (A.15):
Since \(\Lambda _1(t)=\Lambda _2(\varphi _{\varvec{\theta }_0}(t))\), from Eqs. (2.13)–(A.18) we obtain
\(\square \)
1.2 Proof of Theorem 1
Let \(\breve{\kappa }_1(t)={\tilde{\kappa }}_1(t)\), \(\breve{\kappa }_2(t)={\tilde{\kappa }}_2(\varphi _{{\hat{\varvec{\theta }}}}(t))\). From Eq. (2.10), \(h_1(-{\hat{\lambda }}_{{\hat{\varvec{\theta }}}}(t))-h_2({\hat{\lambda }}_{{\hat{\varvec{\theta }}}}(t))=0\), where
Note that \(h_1(0) = - {\hat{\zeta }}_1(t)\) and \(h_2(0)=- {\hat{\zeta }}_2(\varphi _{{\hat{\varvec{\theta }}}}(t))\) [cf. Eq. (A.5)]. Note also that
Clearly, \(n_1h_1^{'}(0)={\hat{\sigma }}_1^2(t)\) and \(n_2h_2^{'}(0)={\hat{\sigma }}_2^2(\varphi _{{\hat{\varvec{\theta }}}}(t))\), see Eq. (2.2). Let \(|{\hat{\xi }}_i|\le |{\hat{\lambda }}_{{\hat{\varvec{\theta }}}}(t)|, i=1,2\). Taylor’s expansion about 0 yields
Applying the Glivenko–Cantelli lemma to \(r_{ij}\), we have \(\Vert h_i''({\hat{\xi }}_i)\Vert _{\alpha _1}^{\alpha _2}=O_{{\mathbb {P}}}(n_i^{-2})\). Therefore, by Lemma 3, it follows that \(\Vert h^{''}_i({\hat{\xi }}_i)({\hat{\lambda }}_{{\hat{\varvec{\theta }}}}(t))^2\Vert _{\alpha _1}^{\alpha _2}=O_{{\mathbb {P}}}(n_i^{-1})=O_{{\mathbb {P}}}(n^{-1})\). Recalling \({\hat{{\mathbb {V}}}}(t)\) defined by Eq. (2.13), it follows from Eqs. (A.20), (A.21), and (2.14) that
Solving for \({\hat{\lambda }}_{{\hat{\varvec{\theta }}}}(t)\), we obtain
To complete the proof of Theorem 1, consider Eq. (2.12). Using Taylor expansions of \(\log (1+x)\) and \(\log (1-x)\) about 0 and applying Eq. (A.22), the leading term of \(-2R_{{\hat{\varvec{\theta }}}}(t)\) equals
uniformly over \([\alpha _1,\alpha _2]\). The subsequent terms of \(-2R(t)\) are proportional to
each of which is \(o_{{\mathbb {P}}}(1)\), uniformly for \(t\in [\alpha _1,\alpha _2]\). For example, when \(l=3\), we have
which, uniformly for \(t\in [\alpha _1,\alpha _2]\), equals
From Eq. (A.23) and Lemma 4 the proof of Theorem 1 is completed. \(\square \)
Rights and permissions
About this article
Cite this article
Subramanian, S. Function-based hypothesis testing in censored two-sample location-scale models. Lifetime Data Anal 26, 183–213 (2020). https://doi.org/10.1007/s10985-018-09456-8
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10985-018-09456-8