SEnsembles – uma abordagem para melhorar a qualidade das correspondências de instâncias disjuntas em estudos observacionais explorando características idênticas e ensembles de regressores

Abstract

Introduction. The datasets used in observational studies have instances belonging to two distinct groups (i.e. treatment group and control group), which are compared in order to estimate the effect of the treatment over the results. For such, in one of the approaches, called Propensity Score Matching (PSM), the propensity score for the instances of both groups is estimated and, subsequently, the correspondence of these instances is performed based on the values for the propensity score. The propensity score is the probability of attribution of a treatment based on the observed characteristics (e.g. income, sex and age). In this context, the logistic regression is widely used to estimate the propensity score and there is an great variety of instance correspondence methods. Objective. This doctor´s thesis has as its main objective to investigate computational alternatives in order to improve the quality of the instance correspondence in datasets that are manipulated in observational studies. Methodology. Techniques that estimate the propensity score and methods to perform the instance correspondence in observational studies were investigated. Thus, it was possible to investigate how the identical characteristics of the instances could be exploited in a new process to perform correspondence and, how ensembles could substitute the logistic regression by estimating the propensity scores of the instances, in the context of the PSM process. Proposal. This thesis proposes a new approach in the context of the PSM process, called “SEnsembles”, which aims to improve the quality of instance correspondence based on two main processes, which use techniques that separately consider the identical characteristics of the instances and the ensembles of regressors, more precisely, bagging, random forest and boosting. Results. The proposed approach “SEnsembles” improves the quality of the instance correspondence for the majority of calipers used (i.e. zero, 0.05, 0.10, 0.15, 0.20, 0.25 and 0.30) when compared to the baseline Nearest Neighbor Matching (NNM). Based on the experiments, when there was an improvement over the baseline, the technique that separates the identical characteristics of the instances presented improvements of up to 53.8% in the quality of correspondence, with an average of gains of 12.1%; and only 2.7% of average in the reduction of the number of pairs of instances matched. The technique which substituted the logistic regression for ensembles of regressors, in turn, presented the best correspondence with the caliper zero and with the values 0.20, 0.25 and 0.30, with improvements of up to 36.3% and an average of gains of 12.7%; and a slightly reduction of 7.6% in the number of pairs of instances matched.