Data di Pubblicazione:
2018
Citazione:
Pseudo-values regression models in the presence of semi-competing risks / A. Orenti, E. Biganzoli, F. Ambrogi, P. Boracchi. ((Intervento presentato al convegno Networking biostatistics tenutosi a Milano nel 2018.
Abstract:
The course of a disease is frequently characterized by a sequence of non fatal events related to disease progression and death. The interest is frequently in estimating survival free from disease progression (disease free survival), however some patients die without previously experiencing events related to disease progression. This is a typical situation of “semi-competing risks”, as the occurrence of fatal event like death prevents the subsequent observation of non fatal event, but not vice versa.
To estimate disease free survival, methods based on censoring times to fatal event can be applied only if times to occurrence of fatal and non fatal events are independent; otherwise specific statistical methods assuming bivariate distribution of times, as for example Archimedean Copulas, must be considered.
In semi-competing risks, two regression methods have been proposed: one is based on method of moments (Fine and Peng) and one is based on conditional likelihood (Hsieh and Huang). They require ad hoc implementation in statistical software. As an alternative, we propose a regression model based on pseudo-values. Pseudo-values regression models have the advantage of enabling flexible modelling of the effect of time and the evaluation of time dependent covariate effects.
The model is implemented in three steps: 1) estimate semi-parametrically disease free survival under a defined copula, 2) compute disease free survival pseudo-values for every individual at predefined time points, 3) fit a GEE using pseudo-values as a response variable.
To evaluate the performance of the regression model based on pseudo-values, we performed a Monte Carlo simulation. As regards the simulation results, the method based on pseudo-observation gives almost unbiased estimates of the covariate coefficient. However, with small sample size it has biases and empirical standard deviations slight higher than those obtained by the method proposed by Hsieh and Huang, adopted as reference method in the simulation.
The pseudo-value method was applied to data of a randomized clinical trial on breast cancer to evaluate the treatment effect (QUART vs QUAD) on relapse free survival. Clayton copula was used after goodness of fit assessment though a graphical model checking. The results are comparable to those obtained by Hsieh and Huang method and indicate a significant smaller hazard of relapse for women treated with QUART. Hazard ratio of relapse at 10 years is 0.683 (95% CI: 0.490 - 0.885).
Tipologia IRIS:
14 - Intervento a convegno non pubblicato
Elenco autori:
A. Orenti, E. Biganzoli, F. Ambrogi, P. Boracchi
Link alla scheda completa: