3
Introduction
The study of selection bias in political science has almost exclusively focused on
situations where the quantity of interest is either continuous or binary. Yet studies
increasingly use empirical models that fall into neither of these categories. Because the
effects of selectivity are not well understood in a broad array of empirical models, its
consequences for studies relying on these methods are unknown. This lack of
understanding has resulted in less attention to the potential for selection bias in these
models than in cases using OLS or probit models where its consequences are better
understood.
In this paper we focus on the effect of self-selectivity in an increasingly popular
class of models: continuous time duration models. Common versions include the
exponential, Weibull and Cox models. As the use of these models has increased over the
last decade or so, increasing attention has been paid to a variety of related statistical
issues. Since Box-Steffensmeier and Jones (1997) provided an introduction to duration
models for political science and raised many of these issues, researchers have focused
primarily on two main areas: estimating duration dependence (Beck 1999; Beck, Katz
and Tucker 1998; Bennett 1999; Zorn 2000) and, more recently, how to account for
repeated failures (Box-Steffensmeier and Zorn 2002; Branford and Jones 2003). Yet the
potential consequences of ignoring selection bias have not been explored in a systematic
fashion.
We argue in this paper that the consequences are just as serious for duration
models as they are for standard regression models. Further, as the use of duration models
has increased, we believe that there are many situations where duration models are
Convention