All models are wrong, but which are useful? Comparing parametric and nonparametric estimation of causal effects in finite samples