Exploring the dynamics of competing risk models for multiple discrete outcomes

Abstract

[ACCESS RESTRICTED TO THE UNIVERSITY OF MISSOURI AT AUTHOR'S REQUEST.] The objective of this research is to explore the predictive power of the correlated competing risks model in two different economic contexts. In both cases there are multiple discrete outcomes influenced by unobserved heterogeneity components. In the first essay, I analyzed the dynamics of firm's credit rating changes in a correlated competing risk framework. In each period, a firm's credit rating may be upgraded, downgraded or left unchanged. The probabilities of these events are negatively correlated and are linked to the observable features of the firm. The competing risk model is more flexible than the conventional modeling techniques like ordered (where the rating of each period is treated as static); hazard model (where only the dynamics of a non-terminal and terminal states are considered); and cohort approach(where the probability of rating transition is not linked to firm's features). An additional contribution of this paper is finding better predictors of credit rating changes within the framework of dependent competing risk model. Using the S&P Long Term Issuer's credit rating from 1986 - 2006, I found that the dependent competing risk model outperforms the ordered logit model and cohort approach in both in-sample and out-of-sample predictions. In the second essay, I model teachers' post-retirement behavior. At the beginning of every academic year a teacher can make any of the three choices, i.e., in the next year, he/she may continue to work without taking retirement, or take retirement and did not come back to teach again the same public school retirement system, or may take retirement but continue working, which is popularly known as "double-dipping". The probabilities of making these choices are negatively correlated and are related to teacher's characteristics and their financial status. In the existing literature, a number of conventional discrete choice models have been developed for empirical analysis of these problems. I find that the competing risks model outperform other conventional discrete choice models in prediction. The correlated competing risk model is more flexible than conventional modeling techniques such as multinomial logit model(where teacher's decision in each period is treated as static) and cox-proportional hazard model (where only the dynamics of terminal and non-terminal states are considered). Using Missouri Teachers data set from 2002 to 2008 and selecting all the teachers who are above age 50 in 2002, I find that the correlated competing risk model outperforms multinomial logit model and cox-proportional hazard model in determining the significance of the explanatory variables and also in out-of-sample forecasting.