Implementation and analysis of Causal Inference techniques for treatment effects estimation applied to Covid-19 cases

Abstract

The objective of this Master’s Thesis is to demonstrate the importance of applying causal reasoning when using Machine Learning to estimate the effect of a treatment on a disease. To this end, a series of problems will be proposed in the context of COVID-19, a disease caused by the SARS-CoV-2 that has been a worldwide pandemic, emerged in 2020 and is still present in 2022. Finding sufficiently complete COVID-19 databases is a highly complicated task due to the confidentiality of personal medical data. For this reason, a synthetic database generated from differential equations has been chosen as a source of data, following a model that manages to represent the evolution of the virus, starting from individual initial conditions. The treatments applied modify the behavior of the differential equations. In this way, we can know the effect of all treatments on each patient, which is very useful when evaluating the performance of the estimation techniques. To evaluate the performance of causal inference, several treatment problems will be solved in different scenarios of observational data, where confounders, unobserved confounders and selection bias are present. The results obtained will be compared with conventional associative Machine Learning techniques, such as linear regression or Neural Networks. For this purpose, several Python libraries and frameworks will be used, including Pytorch for the development of neural networks. By analyzing subpopulations and considering cause-effect relationships, causal inference should find the optimal individualized treatment, for instance, the most effective drug for a specific patient, and should lead to a significant improvement in the performance of the associative ML. The causal inference techniques will use some Python libraries such as DoWhy and EconML, that are useful when the three assumptions of the Potential Outcomes framework are satisfied. When these assumptions are not met, the problem becomes more difficult to address. We propose the use of deconfounding techniques as a solution to the problem of unobserved confounders is proposed. The results obtained compare the errors in the estimation of treatment effects at the individualized level using regression methods and using causal inference, both in the scenario of observed and unobserved confounders. In the second case, the performance of the deconfounding techniques will also be evaluated. evaluated