Estimating Average and Individual Treatment Effects in the Presence of Time-Dependent Covariates

The causal inference in survival analysis framework provides a comprehensive evaluation of survival probabilities, considering the influence of time-dependent covariates. Incorporating potential outcomes and propensity scores within the causal inference in survival analysis framework establishes a robust method for estimating treatment effects in the presence of time-dependent covariates. We examine treatment effects on both a population level and an individual level, leading to a more nuanced comprehension of treatment outcomes. The accuracy of the treatment effect estimators within the framework is assessed, focusing on the DeepSurv, DeepHit, and the multitask learning deep neural network (MTL-DNN) models. Notably, the DeepHit model applied to both real and simulated datasets outperforms with an average concordance statistic (C-statistic) of 0.9997. This surpasses the C-statistic of DeepSurv and MTL-DNN, which are 0.8928 and 0.9996, respectively. These results show that the deep learning models have exceptional discrimination and agreement between observed and predicted survival probabilities. In addition, the bias values for DeepSurv (0.0174), DeepHit (0.0114), and MTL-DNN (0.0108) are notably small and comparable, indicating that these models provide accurate and unbiased estimates of treatment effects. Consequently, this research underscores the superior predictive accuracy of these deep learning models, suggesting their potential to enhance decision-making and deepen our understanding of treatment outcomes in survival analysis.