We use a large, rich Canadian micro-level dataset to examine the channels through which family socio-economic status and unobservable characteristics aff ect children's decisions to drop out of high school. First, we document the strength of observable socio-economic factors: our data suggest that teenage boys with two parents who are themselves high school dropouts have a 16% chance of dropping out, compared to a dropout rate of less than 1% for boys whose parents both have a university degree. We examine the channels through which this socio-economic gradient arises using an extended version of the factor model set out in Carneiro, Hansen, and Heckman (2003). Specifically, we consider the impact of cognitive and non-cognitive ability and the value that parents place on education. Our results support three main conclusions. First, cognitive ability at age 15 has a substantial impact on dropping out. The highest ability individuals are predicted never to drop out regardless of parental education or parental valuation of education. In contrast, the lowest ability teenagers have a probability of dropping out of approximately .36 if their parents have a low valuation of education. Second, parental valuation of education has a substantial impact on medium and low ability teenagers. A low ability teenager has a probability of dropping out of approximately .03 if his parents place a high value on education but .36 if their educational valuation is low. These eff ects are estimated while conditioning on ability at age 15. Thus, under some assumptions, they reflect parental influences during the upper teenage years and are in addition to any impact they might have in the early childhood years leading up to age 15. Third, parental education has no direct e ffect on dropping out once we control for ability and parental valuation of education. Overall, our results point to the importance of whatever determines ability at age 15 (including, potentially, early childhood interventions) and of parental valuation of education during the teenage years. Our work also provides a small methodological contribution by extending the standard factor based estimator to allow a more non-linear relationship between the factors and a co-variate of interest. We show that allowing for non-linearities has a substantial impact on estimated effects.