Applications of Finite Mixture Models

Presented by Paul Cislo, PhD., Statistical Research and Data Science Center, Pfizer, Inc.

Wednesday, March 1, 2023
4:00 PM ET
AUST 163
Webex Meeting Link

Standard parametric modelling approaches by applied researchers assume that a single distribution representative of population being sampled and modifications to that distribution are driven covariates in the regression model. In many cases, the implicit restrictions of the model structure (e.g., a linear function can describe the data) and underlying distribution assumption (e.g., the data has a normal distribution) does not impede reasonable analysis of the data at hand. However, when the data is difficult to fit, finite mixture models may provide useful alternative approach. An overview of various mixture distribution will be presented as well as a description of how EM-algorithm is used to estimate parameters for finite mixture. The local maxima problem associated with finite mixture models and few possible solutions will be presented. An example of a how a finite mixture model was used to estimate risk of exposure to ticks infected with Borrelia burgdorferi (the etiologic agent of Lyme Disease). The example will illustrate how the flexible nature of the mixture model allows the relationship between ecological values and risk of exposure can change through space and time. A second example of using mixture models to fit difficult to fit survival data will be presented. The example will illustrate how loosening the assumption of a single underlying distribution can allow for important improvements to fit of the data.

Speaker Bio:

Paul Cislo received a BS in Biology from UCONN (Class 1992), an MS in Marine Science from University of South Carolina (Class of 1994) and a PhD in Biostatistics from Yale University (Class 2011). His dissertation work focused on development of spatial mixture models and applying those models to estimate risk of exposure to Borrelia burgdorferi (the etiologic agent of Lyme Disease) in the lower 48 states. After leaving an academic setting, Paul has primarily focused on health economics and outcomes research within the pharmaceutical industry gaining experience at Bayer, BMS, and Pfizer. Paul is a member of the Health Economics & Outcomes Research Statistics Group at Pfizer, which part of its Statistical Research & Data Science Center.