In this paper, we consider what can be learned about causal effects when one uses a contaminated instrumental variable. In particular, we consider what inferences can be made about the causal effect of teenage childbearing on a teen mother's subsequent outcomes when we use the natural experiment of miscarriages to form an instrumental variable for teen births. Miscarriages might not meet all of the conditions required for an instrumental variable to identify such causal effects for all of the observations in our sample. However, it is an appropriate instrumental variable for some women, namely those pregnant women who experience a random miscarriage. Although information from typical data sources does not allow one to identify these women, we show that one can adapt results from Horowitz and Manski (1995) on identification with data from contaminated samples to construct informative bounds on the causal effect of teenage childbearing. We use these bounds to re-examine the effects of early chilbearing on the teen mother's subsequent educational and labour market attainment as considered in Hotz, McElroy and Sanders (1995a, 1995b). Consistent with their study, these bounds indicate that women who have births as teens have higher labour market earnings and hours worked compared to what they would have attained if their childbearing had been delayed.

This paper analyses a new estimator for the structural parameters of dynamic models of discrete choice. Based on an inversion theorem due to Hotz and Miller (1993), which establishes the existence of a one-to-one mapping between the conditional valuation functions for the dynamic problem and their associated conditional choice probabilities, we exploit simulation techniques to estimate models which do not possess terminal states. In this way our Conditional Choice Simulation (CCS) estimator complements the Conditional Choice Probability (CCP) estimator of Hotz and Miller (1993). Drawing on work in empirical process theory by Pakes and Pollard (1989), we establish its large sample properties, and then conduct a Monte Carlo study of Rust's (1987) model of bus engine replacement to compare its small sample properties with those of Maximum Likelihood (ML).