Choice of parametric models in survival analysis: applications to monotherapy for epilepsy and cerebral palsy

Summary. In the analysis of medical survival data, semiparametric proportional hazards models are widely used. When the proportional hazards assumption is not tenable, these models will not be suitable. Other models for covariate effects can be useful. In particular, we consider accelerated life models, in which the effect of covariates is to scale the quantiles of the base-line distribution. Solomon and Hutton have suggested that there is some robustness to misspecification of survival regression models. They showed that the relative importance of covariates is preserved under misspecification with assumptions of small coefficients and orthogonal transformation of covariates. We elucidate these results by applications to data from five trials which compare two common anti-epileptic drugs (carbamazepine versus sodium valporate monotherapy for epilepsy) and to survival of a cohort of people with cerebral palsy. Results on the robustness against model misspecification depend on the assumptions of small coefficients and on the underlying distribution of the data. These results hold in cerebral palsy but do not hold in epilepsy data which have early high hazard rates. The orthogonality of coefficients is not important. However, the choice of model is important for an estimation of the magnitude of effects, particularly if the base-line shape parameter indicates high initial hazard rates.     

Choice of Parametric Accelerated Life and Proportional Hazards Models for Survival Data: Asymptotic Results

We discuss the impact of misspecifying fully parametric proportional hazards and accelerated life models. For the uncensored case, misspecified accelerated life models give asymptotically unbiased estimates of covariate effect, but the shape and scale parameters depend on the misspecification. The covariate, shape and scale parameters differ in the censored case. Parametric proportional hazards models do not have a sound justification for general use: estimates from misspecified models can be very biased, and misleading results for the shape of the hazard function can arise. Misspecified survival functions are more biased at the extremes than the centre. Asymptotic and first order results are compared. If a model is misspecified, the size of Wald tests will be underestimated. Use of the sandwich estimator of standard error gives tests of the correct size, but misspecification leads to a loss of power. Accelerated life models are more robust to misspecification because of their log-linear form. In preliminary data analysis, practitioners should investigate proportional hazards and accelerated life models; software is readily available for several such models.     

Extended Hazard Regression Model for Reliability and Survival Analysis

We propose an extended hazard regression model which allows the spread parameter to be dependent on covariates. This allows a broad class of models which includes the most common hazard models, such as the proportional hazards model, the accelerated failure time model and a proportional hazards/accelerated failure time hybrid model with constant spread parameter. Simulations based on sub-classes of this model suggest that maximum likelihood performs well even when only small or moderate-size data sets are available and the censoring pattern is heavy. The methodology provides a broad framework for analysis of reliability and survival data. Two numerical examples illustrate the results.     

Covariate effect on the life expectancy and percentile residual life functions under the proportional hazards and the accelerated life models

The present paper is concerned with statistical models for the dependence of survival time or time to occurrence of an event, such as time to tumor, on a vector X of covariates or prognostic variables such as age, sex, blood pressure, length of exposure to a toxic material, etc., measured on a group of individuals in biomedical investigations. It is assumed that the covariates influence the distribution of time to tumor only through a linear predictor μ =β^′X.

The object of our paper is to investigate the effect due to the covariates on the Life Expectancy and the Percentile Residual Life (PRL) function of a family of organisms under the proportional hazards and the accelerated life models. The key result is that the families of survival distributions under these models have the 'setting the clock back to zero' property if the family of baseline survival distributions does. This property is a generalization of the lack of memory property of the exponential distribution. Simple examples of the members of this family are the linear hazard exponential, Pareto and Gompertz life distributions.

As a simple application of the main results obtained in the present paper, we have considered a stochastic survival model recently proposed by Chiang and Conforti (1989) for the time-to-tumor distribution in the context of a large-scale serial sacrifice experiment by the National Center of Toxicological Research (NCTR). This involves some mice that were fed 2-AAF from infancy and those that developed bladder and/or liver neoplasms, see Farmer et al (1980). It is shown that their stochastic model for tumor incidence intensity at time t leads to a family of survival models that has the setting the clock back to zero property. The survival functions and the effect of the vector X of covariates on the PRL and the tumor-free life expectancies are evaluated for the proportional hazards and accelerated life models.     

Chi-Squared Goodness-of-Fit Tests for Parametric Accelerated Failure Time Models

We give chi-squared goodness-of fit tests for parametric regression models such as accelerated failure time, proportional hazards, generalized proportional hazards, frailty models, transformation models, and models with cross-effects of survival functions. Random right censored data are used. Choice of random grouping intervals as data functions is considered.     

Empirical Likelihood for Censored Linear Regression and Variable Selection

The linear regression model for right censored data, also known as the accelerated failure time model using the logarithm of survival time as the response variable, is a useful alternative to the Cox proportional hazards model. Empirical likelihood as a non‐parametric approach has been demonstrated to have many desirable merits thanks to its robustness against model misspecification. However, the linear regression model with right censored data cannot directly benefit from the empirical likelihood for inferences mainly because of dependent elements in estimating equations of the conventional approach. In this paper, we propose an empirical likelihood approach with a new estimating equation for linear regression with right censored data. A nested coordinate algorithm with majorization is used for solving the optimization problems with non‐differentiable objective function. We show that the Wilks' theorem holds for the new empirical likelihood. We also consider the variable selection problem with empirical likelihood when the number of predictors can be large. Because the new estimating equation is non‐differentiable, a quadratic approximation is applied to study the asymptotic properties of penalized empirical likelihood. We prove the oracle properties and evaluate the properties with simulated data. We apply our method to a Surveillance, Epidemiology, and End Results small intestine cancer dataset.     

Bayesian analysis of generalized odds-rate hazards models for survival data

In the analysis of censored survival data Cox proportional hazards model (1972) is extremely popular among the practitioners. However, in many real-life situations the proportionality of the hazard ratios does not seem to be an appropriate assumption. To overcome such a problem, we consider a class of nonproportional hazards models known as generalized odds-rate class of regression models. The class is general enough to include several commonly used models, such as proportional hazards model, proportional odds model, and accelerated life time model. The theoretical and computational properties of these models have been re-examined. The propriety of the posterior has been established under some mild conditions. A simulation study is conducted and a detailed analysis of the data from a prostate cancer study is presented to further illustrate the proposed methodology.     

Fraction of design space plots for generalized linear models

The use of graphical methods for comparing the quality of prediction throughout the design space of an experiment has been explored extensively for responses modeled with standard linear models. In this paper, fraction of design space (FDS) plots are adapted to evaluate designs for generalized linear models (GLMs). Since the quality of designs for GLMs depends on the model parameters, initial parameter estimates need to be provided by the experimenter. Consequently, an important question to consider is the design's robustness to user misspecification of the initial parameter estimates. FDS plots provide a graphical way of assessing the relative merits of different designs under a variety of types of parameter misspecification. Examples using logistic and Poisson regression models with their canonical links are used to demonstrate the benefits of the FDS plots.     

Performance and Prediction for Varying Survival Time Scales

The Cox proportional hazards model is widely used for analyzing associations between risk factors and occurrences of events. One of the essential requirements of defining Cox proportional hazards model is the choice of a unique and well-defined time scale. Two time scales are generally used in epidemiological studies: time-on-study and chronological age. The former is the most frequently used time scale, both in clinical studies and longitudinal observation studies. However, there is no general consensus on which time scale is the most appropriate for a given question or study. In this article, we address the question of robustness of the results using one time scale when the other is actually the correct one. We use three criteria to measure the performances of these models through simulations: magnitude of the bias of the regression coefficients, mean square errors, and the measure of overall predictive discrimination of the models. We conclude that the time-on-study models are more robust to misspecification of the underlying time scale.     

Smooth Semi‐nonparametric Analysis for Mixture Cure Models and Its Application to Breast Cancer

Mixture cure models are widely used when a proportion of patients are cured. The proportional hazards mixture cure model and the accelerated failure time mixture cure model are the most popular models in practice. Usually the expectation–maximisation (EM) algorithm is applied to both models for parameter estimation. Bootstrap methods are used for variance estimation. In this paper we propose a smooth semi‐nonparametric (SNP) approach in which maximum likelihood is applied directly to mixture cure models for parameter estimation. The variance can be estimated by the inverse of the second derivative of the SNP likelihood. A comprehensive simulation study indicates good performance of the proposed method. We investigate stage effects in breast cancer by applying the proposed method to breast cancer data from the South Carolina Cancer Registry.     

A dependent Dirichlet process model for survival data with competing risks

In this paper, we first propose a dependent Dirichlet process (DDP) model using a mixture of Weibull models with each mixture component resembling a Cox model for survival data. We then build a Dirichlet process mixture model for competing risks data without regression covariates. Next we extend this model to a DDP model for competing risks regression data by using a multiplicative covariate effect on subdistribution hazards in the mixture components. Though built on proportional hazards (or subdistribution hazards) models, the proposed nonparametric Bayesian regression models do not require the assumption of constant hazard (or subdistribution hazard) ratio. An external time-dependent covariate is also considered in the survival model. After describing the model, we discuss how both cause-specific and subdistribution hazard ratios can be estimated from the same nonparametric Bayesian model for competing risks regression. For use with the regression models proposed, we introduce an omnibus prior that is suitable when little external information is available about covariate effects. Finally we compare the models’ performance with existing methods through simulations. We also illustrate the proposed competing risks regression model with data from a breast cancer study. An R package “DPWeibull” implementing all of the proposed methods is available at CRAN.

     

Mixture cure rate models with accelerated failures and nonparametric form of covariate effects

Two-component mixture cure rate model is popular in cure rate data analysis with the proportional hazards and accelerated failure time (AFT) models being the major competitors for modelling the latency component. [Wang, L., Du, P., and Liang, H. (2012), ‘Two-Component Mixture Cure Rate Model with Spline Estimated Nonparametric Components’, Biometrics, 68, 726–735] first proposed a nonparametric mixture cure rate model where the latency component assumes proportional hazards with nonparametric covariate effects in the relative risk. Here we consider a mixture cure rate model where the latency component assumes AFTs with nonparametric covariate effects in the acceleration factor. Besides the more direct physical interpretation than the proportional hazards, our model has an additional scalar parameter which adds more complication to the computational algorithm as well as the asymptotic theory. We develop a penalised EM algorithm for estimation together with confidence intervals derived from the Louis formula. Asymptotic convergence rates of the parameter estimates are established. Simulations and the application to a melanoma study shows the advantages of our new method.     

On selecting parametric link transformation families in generalized linear models

The use of parametric link transformation families in generalized linear models (GLM) has been shown to improve substantially the fit of standard analyses using a fixed link in some data sets (see Czado, 1993, for example). When link and regression parameters are globally orthogonal (Cox and Reid, 1987), then the variance inflation of the regression parameter estimates due to the additional estimation of the link is asymptotically zero. Parameter orthogonality also induces numerical stability which is seen in the reduction of computation time required for the calculation of parameter estimates. This stability remains a desirable property even for inferences which are conditional on a fixed link value. Czado and Santner (1992b), for binomial error, and Czado (1992), for GLMs have shown that only local orthogonality can be achieved in general. This paper provides conditions on the link family to extend the notion of local orthogonality at a point to orthogonality in a neighborhood asymptotically and shows that the resulting links are location and scale invariant. General concepts for the construction of such links are given, and it is shown how they relate to link families proposed in the literature. The ideas are illustrated by two examples.     

Sliced inverse regression for survival data

We apply the univariate sliced inverse regression to survival data. Our approach is different from the other papers on this subject. The right-censored observations are taken into account during the slicing of the survival times by assigning each of them with equal weight to all of the slices with longer survival. We test this method with different distributions for the two main survival data models, the accelerated lifetime model and Cox’s proportional hazards model. In both cases and under different conditions of sparsity, sample size and dimension of parameters, this non-parametric approach finds the data structure and can be viewed as a variable selector.     

Causal Proportional Hazards Models and Time-constant Exposure in Randomized Clinical Trials

The last decade saw enormous progress in the development of causal inference tools to account for noncompliance in randomized clinical trials. With survival outcomes, structural accelerated failure time (SAFT) models enable causal estimation of effects of observed treatments without making direct assumptions on the compliance selection mechanism. The traditional proportional hazards model has however rarely been used for causal inference. The estimator proposed by Loeys and Goetghebeur (2003, Biometrics vol. 59 pp. 100–105) is limited to the setting of all or nothing exposure. In this paper, we propose an estimation procedure for more general causal proportional hazards models linking the distribution of potential treatment-free survival times to the distribution of observed survival times via observed (time-constant) exposures. Specifically, we first build models for observed exposure-specific survival times. Next, using the proposed causal proportional hazards model, the exposure-specific survival distributions are backtransformed to their treatment-free counterparts, to obtain – after proper mixing – the unconditional treatment-free survival distribution. Estimation of the parameter(s) in the causal model is then based on minimizing a test statistic for equality in backtransformed survival distributions between randomized arms.     

Use of a mixture model for the analysis of contraceptive-use duration among long-term users

This paper introduces a mixture model that combines proportional hazards regression with logistic regression for the analysis of survival data, and describes its parameter estimation via an expectation maximization algorithm. The mixture model is then applied to analyze the determinants of the timing of intrauterine device (IUD) discontinuation and long-term IUD use, utilizing 14 639 instances of IUD use by Chinese women. The results show that socio-economic and demographic characteristics of women have different influences on the acceleration or deceleration of the timing of stopping IUD use and on the likelihood of long-term IUD use.     

Predictive comparison of joint longitudinal-survival modeling: a case study illustrating competing approaches

The joint modeling of longitudinal and survival data has received extraordinary attention in the statistics literature recently, with models and methods becoming increasingly more complex. Most of these approaches pair a proportional hazards survival with longitudinal trajectory modeling through parametric or nonparametric specifications. In this paper we closely examine one data set previously analyzed using a two parameter parametric model for Mediterranean fruit fly (medfly) egg-laying trajectories paired with accelerated failure time and proportional hazards survival models. We consider parametric and nonparametric versions of these two models, as well as a proportional odds rate model paired with a wide variety of longitudinal trajectory assumptions reflecting the types of analyses seen in the literature. In addition to developing novel nonparametric Bayesian methods for joint models, we emphasize the importance of model selection from among joint and non joint models. The default in the literature is to omit at the outset non joint models from consideration. For the medfly data, a predictive diagnostic criterion suggests that both the choice of survival model and longitudinal assumptions can grossly affect model adequacy and prediction. Specifically for these data, the simple joint model used in by Tseng et al. (Biometrika 92:587–603, 2005) and models with much more flexibility in their longitudinal components are predictively outperformed by simpler analyses. This case study underscores the need for data analysts to compare on the basis of predictive performance different joint models and to include non joint models in the pool of candidates under consideration.     

Orthogonalizing parametric link transformation families in binary regression analysis

This paper studies the application of the orthogonalization technique of Cox and Reid (1987) to parametric families of link functions used in binary regression analysis. The explicit form of Cox and Reid's condition (4), for orthogonality at a point, is derived for arbitrary link families. This condition is used to determine a transform of a family introduced by Burr (1942) and Prentice (1975, 1976) which is locally orthogonal when the regression parameter is zero. Thus the benefits of having orthogonal parameters are limited to “small” regression effects. The extent to which approximate orthogonality holds for nonzero regression coefficients is investigated for two data sets from the literature. Two specific issues considered are: (1) the ability of orthogonal reparametrization to reduce the variability of the regression parameters caused by estimation of the link parameter and (2) the improved numerical stability (and hence interpretability) of regression estimates corresponding to different link parameters.     

Conditional maximum likelihood estimation in semiparametric transformation model with LTRC data

Left-truncated data often arise in epidemiology and individual follow-up studies due to a biased sampling plan since subjects with shorter survival times tend to be excluded from the sample. Moreover, the survival time of recruited subjects are often subject to right censoring. In this article, a general class of semiparametric transformation models that include proportional hazards model and proportional odds model as special cases is studied for the analysis of left-truncated and right-censored data. We propose a conditional likelihood approach and develop the conditional maximum likelihood estimators (cMLE) for the regression parameters and cumulative hazard function of these models. The derived score equations for regression parameter and infinite-dimensional function suggest an iterative algorithm for cMLE. The cMLE is shown to be consistent and asymptotically normal. The limiting variances for the estimators can be consistently estimated using the inverse of negative Hessian matrix. Intensive simulation studies are conducted to investigate the performance of the cMLE. An application to the Channing House data is given to illustrate the methodology.     

An introduction to survival models: in honor of Ross Prentice

I review some key ideas and models in survival analysis with emphasis on modeling the effects of covariates on survival times. I focus on the proportional hazards model of Cox (J R Stat Soc B 34:187–220, 1972), its extensions and alternatives, including the accelerated life model. I briefly describe some models for competing risks data, multiple and repeated event-time data and multivariate survival data.