Covariate selection and propensity score specification in causal inference