Randomization is rarely available in observational or real-world data, which makes treatment groups fundamentally different at baseline. This tutorial walks through how to handle such confounding using multiple propensity score designs in R. Using the Lalonde dataset, we compare widely used approaches from MatchIt and WeightIt, including nearest neighbor matching, optimal matching, full matching, subclassification, entropy balancing, CBPS weighting, overlap weighting and more. Instead of promoting a single preferred method, the tutorial shows how each design alters covariate balance, sample structure and estimand, and demonstrates how outcome estimates change after applying these adjustments. The material is intended as a hands-on reference for analysts working in causal inference, health economics and outcomes research, and real-world evidence.

Conteúdo resumido para a prova 3 de probabilidade II