Dynamic regression models for survival data. (Q2493435): Difference between revisions

The Cox proportional hazard model has been used widely in medical research to study the relationship between covariates and a response variable expressed as time to an event which is often called as survival time. The crucial assumption, i.e., the proportional hazard assumption, underlying this model may not hold, however, in many applications. This is illustrated in the first chapter of this book using datasets from studies which investigated the prognostic importance of various risk factors on mortality for patients with, respectively, acute myocardial infarction and primary biliary cirrhosis of the liver. These datasets also indicate that the effects of some covariates on survival tend to change over time, which gives background and motivation to the approach which extends the Cox model by allowing some of the regression coefficients in the model to be functions of the time. This class of new models is characterized as ``dynamic regression models'' since they are able to capture time-varying dynamics of covariate effects. Historically, time-varying coefficients appeared earlier in the Aalen's additive hazard model which models the hazard function of the survival time itself, instead of its logarithm as in the Cox model, as a linear function of the covariates and can be considered also as an alternative to the Cox model. Partly because of this reason, after two chapters on the probabilistic and statistical backgrounds and one chapter on the review of nonparametric inference for survival data without covariates, the authors introduce first in Chapter 5 this additive model, followed by the extended Cox model with time-varying coefficients in Chapter 6 and a mixture of Aalen and Cox models in Chapter 7. The inferences on the cumulative regression coefficients, defined as the integration of time-varying regression coefficients over time, and goodness-of-fit procedures are the focus of these chapters. The next three chapters consider regression models with constant regression coefficients in various settings. Chapter 8 introduces semi-parametric accelerated failure time and transformation models, which assume a direct relationship between a function of the survival time and covariates, instead of through the hazard function of the survival time. These models were also proposed as alternatives to the classical Cox models. The marginal regression and frailty models for the clustered failure time data, which have some underlying correlation structure, are covered in Chapter 9, while Chapter 10 considers models for competing risk data where subjects may experience more than one type of failures. The last chapter deals with a seemingly totally different topic, i.e., the modeling of longitudinal data where both responses and covariates are collected over time for independent subjects. The inclusion of this topic in this book is justified in the first chapter based on the reason that the similar technique based on point processes are used to develop statistical theory and inference procedures for the modeling of both longitudinal and survival data. The regression coefficients in both nonparametric and semi-parametric models introduced are also allowed to vary over time as those covered in the earlier part of the book. Overall, this book provides a timely summary of the results for topics which are important to practical applications. The readers who are interested in further research in these areas will find the detailed derivations of mathematical results helpful. The examples involving analysis of real datasets and, especially the R-codes included in the book, will increase the chance that the procedures introduced in this book are adopted in practice. The rich exercises at the end of each chapter make this book an excellent choice as a textbook for an advanced survival analysis course.