The Impact of Confounder Selection in Propensity Scores for Rare Events Data - with Applications to Birth Defects