From Dependence to Policy Decisions: Copula-based Causal Inference (Beginner; Added Fee)

Non-experimental observational data are invaluable for informing real-world policy evaluation and decision-making. However, such data are often subject to important threats to causal identification, particularly confounding and selection bias. For example, treatment exposure may be correlated with unobserved confounders, violating the assumption that regressors are independent of the error term. This regressor–error dependence can lead to biased estimates of policy effectiveness when standard models assume no unmeasured confounding. Classical approaches to addressing these challenges typically rely on instrumental variables, which must satisfy strong and often difficult-to-verify conditions such as the exclusion restriction and are infeasible in many studies.   In recent research, instrument-free copula methods have been increasingly used to model regressor-error dependence.  These approaches re-estimate parameters of interest under models that explicitly account for such dependence using robust copula specifications. Copula correction can be implemented in a straightforward manner through a two-stage control function approach. We have developed a suite of statistical methods for this purpose, now available in an R package named Rcope.   In this short course, we will describe these methods and train users to apply them in practice to correct for potential confounding and selection bias in observational data.     Outline:   1.    Overview of copula-based regressor-error dependence modeling. 2.    Copula methods to correct for confounding and selection bias for both continuous and discrete treatment variables. 3.    Copula correction in common statistical models: a.    Independent outcomes — the generalized linear model. b.    Longitudinal/clustered outcomes — ·         Fixed-effect model. ·         Linear mixed model.  ·         Generalized linear mixed model. 4.    Description of the R package Rcope 5.    Applications: a.    Prenatal smoking exposure on birthweight data b.    Glucocorticoid dose-response administrative data c.    Food consumption DataTarget audience: Biostatisticians, methodologists, and practitioners in observational studies.