Dynamic latent trait models: An application to Latin American banking crises q Guillermo Rosas * Department of Political Science, Washington University in St. Louis, One Brookings Drive, Campus Box 1063, St. Louis, MO 63130, United States Keywords: Dynamic factor analysis Hierarchical models Time dependence Bank crises Latin America abstract Dynamic latent trait models combine information from a variety of manifest variables, possibly measured on different scales, that are presumed to be indicators of an unobserved latent phenomenon, while allowing appropriate consideration of the longitudinal char- acter of time series. I use a Bayesian dynamic latent trait model of banking sector financial accounts measured at the country/quarter level to build an indicator of banking system robustness in Latin America. As a methodological innovation, I extend dynamic latent trait models to take into account country-specific effects of bank regulatory regimes through hierarchical modeling of factor loadings. I suggest how these models can be applied to other types of phenomenadfor example to combine existing political regime indicators to build a more informative measure of democracy. Ó 2009 Elsevier Ltd. All rights reserved. Across all subfields of political science it is very common to encounter problems that require estimation of latent traits. Whether interest is in capital openness (Chinn and Ito, 2002), political risk (Quinn, 2004), democracy (Pem- stein et al., 2008; Treier and Jackman, 2008), or ideological positions of parties or legislators (Clinton et al., 2004; Huber and Gabel, 2000), to suggest but a few examples, scholars combine information from a variety of observed indicators to construct latent scores through scaling tech- niques such as factor analysis or item-response theory. Several extensions to normal factor analysis permit esti- mation of principled measurement models that accom- modate different assumptions regarding the distribution of manifest indicators. These models further our ability to analyze cross-sectional data without a time component. In many instances, however, manifest indicators are structured temporally, and the analyst may desire to exploit this information to build measurement models that account for serial dependence in repeated observations within subjects. We may care for example about changes in a polity’s propensity towards adoption of a fully democratic regime, or we might want to model explicitly the temporal shifts in an economy’s degree of capital openness, or understand if ideological changes in a party manifesto are more or less rapid after a party loses a crucial election. Even if we are not intrinsically interested in the dynamic component of a particular phenomenon, accounting for time dependence in latent trait models can lead to improved estimates of parameters of interest. In this paper, I use a Bayesian dynamic latent trait model to build an index of financial distress in national banking systems. The growing political economy and policy litera- ture on the subject has been aided by the existence of dichotomous expert scores that code whether country i at time t suffered from widespread insolvency in its banking sector (see, for example, Keefer, 2007; Montinola, 2003; Rosas, 2006, 2009). These scores provide a great resource to explore testable implications of different theories, but are q This paper was prepared during a sabbatical stay at the Institut Barcelona d’Estudis Internacionals (IBEI). Early versions were presented at Nuffield College and IBEI. I am grateful to Dave Armstrong, Ryan Bak- ker, Mark Pickup, and Anton Westveld for their advice. None of these individuals are responsible for any remaining errors. * Tel.: þ1 (314) 935 7456; fax: þ1 (314) 935 5856. E-mail address: grosas@wustl.edu Contents lists available at ScienceDirect Electoral Studies journal homepage: www.elsevier.com/locate/electstud 0261-3794/$ – see front matter Ó 2009 Elsevier Ltd. All rights reserved. doi:10.1016/j.electstud.2009.05.013 Electoral Studies 28 (2009) 375–387