On model selection and model misspecification in causal inference