Individual Submission Summary
Share...

Direct link:

Download

Inference with Extremes: An Extreme Value and Zero Inflated Regression Model (Pre-Recorded)

Sat, October 2, 8:00 to 9:30am PDT (8:00 to 9:30am PDT), TBA

Abstract

Modeling data with extreme values, such as data on different types of political violence, often pose substantial difficulties for the researcher, as different methods for handling these extreme values may be either arbitrary, cause the results to shift, or be difficult to implement. To solve this problem, we propose an Extreme Value and Zero Inflated Negative Binomial (EVZINB) regression model which can model data with both excess zeroes and extreme values. The EVZINB model can be thought of as a latent state regression model where the latent states correspond to different data generating processes, which may be modelled separately. Multiple latent states allows for a more nuanced analysis of how covariates affect the outcomes of interest, allowing for not only an analysis of the mean effects of covariates in the well-behaved count process but also the likelihood of observing extreme values and the potential magnitude of these extreme values. Disaggregating the analysis of the effects into different components gives researchers and policy makers more tools of assessing the effects risks involved in processes which sometimes, but not always, generate extreme values. We show that the EVZINB model can be motivated from both a statistical and a theoretical perspective, and that the EVZINB model properly retrieves the parameter estimates in a simulation study. Lastly, we show the empirical usefulness of the model by replicating a recently published study on one-sided violence against civilians in which the EVZINB model outperforms the negative binomial and the zero-inflated negative binomial models on a number of key performance metrics, and where the EVZINB model allows for more nuanced inference about the relationship between the covariates and outcome.

Authors