Comparing principal stratification and selection models in parametric causal inference with nonignorable missingness