Estimating and Using Propensity Scores with Partially Missing Data