Bias and Its Remedy in Interval-Censored Time-to-Event Applications - Interval-Censored-Time-to-Event Data: Methods and Applications - page 311

11.1 Introduction

In time-to-event data analysis, the most studied data type is right-censored even

though there are left-censored as well as interval-censored data. In fact, interval-
censored data commonly arise in many studies, especially in oncology recurrence

studies and public health applications. In these studies after undergoing an inter-
vention (such as surgery for solid tumors) that leaves the patient without measurable
disease, patients are followed periodically for disease recurrence. The patient might

be disease-free when checked at time t 1 , but have the disease when checked at a later
time t 2 . In this situation, the exact time (i.e., t 1 and t 2 ) of recurrence of the disease
is unknown. Interval-censored data are also common in HIV/AID epidemiological
studies where the HIV infection time is not known exactly but usually only known

to be between times of administering surveys. This type of time-to-recurrence data
is called interval-censored data and is extensively discussed in Sun (2006).

Because there are no existing, easy-accessible statistical analysis methods for
this type of data, analysts usually define the event time to be the time at which the
event was rst observed at t 2 or utilize the midpoint imputation of the interval;
then proceed as though the data are right-censored and use methods developed for
analyzing right-censored data, such as the Cox proportional hazards model. This

practice may lead to bias and erroneous statistical inferences. In analyzing cancer
progression-free survival, Panageas et al. (2007) investigated the bias and the con-

sequences from this imputation. In this chapter, through extensive computational
simulations, we investigate the bias associated with statistical inference arising from
treating the event as occurring at the midpoint or the first observed endpoint of the

observation interval followed using common right-censored statistical methods. We
also introduce the \IntCox" method available in R to remedy this application bias.

This chapter is organized as follows. In Section 11.2, we briefly review the well-
known Cox proportional hazards regression model for analyzing right-censored time-

to-event data and introduce the background of \IntCox" methods to analyze interval-
censored time-to-event data. We then set up simulation studies in Section 11.3 to

illustrate the bias generated from direct use of Cox regression and the improvement