Zoom Meeting ID: 950 0967 8295
Speaker: Erin Austin, PhD
Assistant Professor, Mathematical and Statistical Sciences, UCDenver
Using Data Consistent Inversion to describe and quantify the impact of sources of uncertainty on predictions of lung function in the COPDGene cohort
Data Consistent Inversion (DCI) is a new iterative methodology within the field of uncertainty quantification to identify distributional forms for a model’s parameters that lead to predictions consistent with observed outcomes. That is, DCI is a new way to quantify if a hypothesized model can be used to generate predictions that are distributionally similar to that of the observed outcomes. The method has traditionally been limited to areas where hard science has credibly determined the functional form (model) of a behavior, and uncertainty quantification is often used to study small changes to this known model. In our work, we show that statistical reasoning can substitute for the role of hard science and allow us to credibly hypothesize models for our phenomenon of interest. We can then adapt DCI tools to provide researchers a new metric to assess whether our hypothesized models give data consistent predictions. Further, the flexibility of DCI permits us to easily consider modifications to linear models; for example, embed unobserved measurement error in observed variables. The process results in a model form, including distributions for all its parameters, that has been optimally adapted for consistency with our data; a model that can then be used to better quantify the effect of the different model components on prediction.
Specifically, we use DCI to study different hypothesized model forms for predicting 10-year measurements of lung health from baseline and 5-year measurements for participants in the COPDGene project. Using DCI we demonstrate how a linear model that accounts for bias in previous measurements is sufficient for predictions. Further, we show that classic regression approaches underestimate the true range in 10-years outcomes as they fail to adequately capture the true uncertainty in the data.