Primary Submission Category: Machine Learning and Causal Inference
Economic Burden of Breast Cancer in Denmark: Estimation and Evaluation of Casual Methods
Authors: Emily Johnson, Angela Chang, Liza Sopina,
Presenting Author: Emily Johnson*
The economic burden of breast cancer has long been a topic of interest within the literature, and in Nordic countries the existence of the extensive administrative data systems allows for causal analysis of this topic. Prior studies have applied matching to administrative data estimate the effects of disease burden on income. However, none of these studies have empirically evaluated the dependence of these results on the methods used.
This paper applies machine learning to evaluate the validity of different causal models in measuring the cost of breast cancer in Denmark. Data are sourced from the national administrative registries which capture the Danish population from 2000-2018. Matching methods evaluated include propensity score matching, coarsened exact matching, inverse probability weighting, and Bayesian additive regression trees, and matching parameters tested include age, socioeconomic status, education, family size, and baseline household income. Models are compared using synthetic validation, which can approximate cross validation for predictive estimation. These metrics provide more information about model fit than evaluations of matching performance, which can assess the quality of a match but not the bias of a measured causal effect. By applying synthetic validation to a topic of interest in health economics literature, this paper demonstrates a mechanism to evaluate the validity of causal inference results where empirical assessment of bias is often lacking.