In observational biomedical research, model misspecification, non-randomized treatment assignment and missingness in the outcome measures contribute to bias the treatment effect estimate. The aim of this study is to compare different statistical approaches to limit the size of such biases. We compared covariates adjustment, propensity score adjustment, matching, Inverse-Probability-of-Treatment Weighting, and more complex approaches that combine ensemble of machine learning algorithms with double robust semiparametric maximum likelihood estimation like the targeted maximum likelihood estimator (TMLE) in terms of bias, standard error and coverage probability. Two scenarios were simulated: (i) with complete data set; (ii) with 20% missingness mechanism on the outcome. Model misspecification and near-positivity violations were addressed. Finally, comparisons were applied to the Italian multicenter retrospective study DApagliflozin Real Word evidence in Type 2 Diabetes (DARWIN-T2D). Preliminary results suggest that in both scenarios TMLE has the lowest bias, even if a slightly larger standard error, in the treatment estimate compared to the classical approaches.