Propensity score matching (PSM) is shown to often increase imbalance, inefficiency, model dependence, and bias in causal inference. The paper argues that PSM, which aims to approximate a completely randomized experiment, is less effective than other matching methods that approximate a fully blocked randomized experiment. PSM is blind to important sources of information in observational studies and fails to reduce imbalance, model dependence, and bias, especially in data that is already balanced. The authors define the PSM paradox as the phenomenon where PSM increases imbalance and model dependence when applied to balanced data. They argue that other matching methods, such as coarsened exact matching (CEM) and Mahalanobis distance matching (MDM), are more effective at reducing model dependence and bias. The paper also highlights the dangers of random matching, which can increase imbalance and model dependence. The authors conclude that while PSM has other useful applications, it is not recommended for causal inference due to its limitations. The paper provides simulations and real-world examples to support these claims.Propensity score matching (PSM) is shown to often increase imbalance, inefficiency, model dependence, and bias in causal inference. The paper argues that PSM, which aims to approximate a completely randomized experiment, is less effective than other matching methods that approximate a fully blocked randomized experiment. PSM is blind to important sources of information in observational studies and fails to reduce imbalance, model dependence, and bias, especially in data that is already balanced. The authors define the PSM paradox as the phenomenon where PSM increases imbalance and model dependence when applied to balanced data. They argue that other matching methods, such as coarsened exact matching (CEM) and Mahalanobis distance matching (MDM), are more effective at reducing model dependence and bias. The paper also highlights the dangers of random matching, which can increase imbalance and model dependence. The authors conclude that while PSM has other useful applications, it is not recommended for causal inference due to its limitations. The paper provides simulations and real-world examples to support these claims.