DEALING WITH INFORMATIVE CENSORING IN SURVIVAL ANALYSIS

Childhood cancer is a worldwide public health concern being a leading cause of death in children. While, high-income countries have improved the probability of surviving after cancer diagnosed in pediatric age (it now reaches 80%), children and adolescents with cancer living in low and middle-income countries (LIC/LMIC) have a dismal outcome. One important initiative is the AHOPCA (Asociación de Hemato-Oncología Pediátrica de Centro América) network, which is a group of hospital units specialized in childhood cancer treatment from Central America and the Caribe. These countries faces common difficulties such as widespread poverty (25 to 60% of their populations live below the poverty line), malnutrition, illiteracy, poor infrastructure, difficult access to health services, inconsistent drug availability, lack of supportive care and low priority of cancer treatment if compared to the priority of other health issues (mostly infectious diseases). This network developed from twinning programs between Italy, Switzerland, United States institutions and the countries of Central America joined formally into the AHOPCA collaborative group. Survival analysis which is the methodology typically used to describe the outcome in cancer clinical trials and is also used as an indicator of their efficacy in disease management and care, studying the time elapsed between some initial event defined as a starting point (such as date of diagnosis or start of treatment) and the time of occurrence of some event (failure) of interest (such as disease relapse or death), deals with censoring. A typical complexity of observed survival data is the presence of right censoring on the survival time, which occurs generally when the survival time is shorter than the failure time. Censoring is due to a limitation on the observability of the failure/survival time itself (for this reason it is called administrative censoring) and has to be accounted for in the analysis. Statistical methods in survival analysis were developed mostly to address for the presence of censoring and for the non-symmetric shape of the distribution of survival time. In the classical survival analysis theory, the censoring distribution is reasonably assumed to be independent from the survival time distribution, i.e. censoring is non-informative on the “true” survival time. This assumption implies that the velocity of occurrence of failure can be estimated by considering the survival experience of the non-censored times. Treatment abandonment is a relevant problem in LIC/LMIC and, according to the experience of these countries, some of these children who abandon treatment are seen later alive and in complete remission, others return to the clinic with relapse or progressive disease or die, most of them are not retraced and their status is unknown. Given these considerations, it is clear that abandonment is not the standard administrative censoring and is not independent from the survival experience. Considering abandonment of therapy as an event (failure) likely leads to underestimate the protocol effect but considering it as administrative censoring can lead to overestimate the effect. The current approach perform the estimation of EFS (event-free survival) in two ways: by treating abandonment as a failure censoring. This project aims at estimating the survival outcome of childhood cancer in LIC/LMIC countries where treatment abandonment is a relevant issue with approaches that can deal with the informative nature of the related censored information. The project will develop the following two points: 1. Handling informative censoring on survival time due to abandonment of treatment, using the non-standard statistical method of Marginal Structural Model. 2. Comparing the classic with the non-standard statistical methods in evaluating the effects of treatment protocols in children with of acute lymphoblastic leukemia treated in LIC/LMIC.