Universal Difference-in-Differences for Causal Inference in Epidemiology

Area of Science:

• Epidemiology
• Biostatistics
• Econometrics

Background:

• Difference-in-differences (DiD) is a prevalent method for causal inference in observational studies.
• The standard DiD relies on a parallel trends assumption, which may be violated with binary, count, or polytomous outcomes, or non-additive confounder effects.
• Violations of the parallel trends assumption limit the credibility of standard DiD in many real-world scenarios.

Purpose of the Study:

• To introduce a novel causal inference method, Universal Difference-in-Differences (UDD).
• To replace the restrictive parallel trends assumption with a more flexible odds ratio equi-confounding assumption.
• To enable robust causal effect estimation in settings where standard DiD assumptions are untenable.

Main Methods:

• The proposed Universal Difference-in-Differences (UDD) method utilizes an odds ratio equi-confounding assumption.
• It employs a generalized linear model linking pre-exposure outcomes and exposure to identify causal effects.
• Both fully parametric and semiparametric UDD estimators are developed and presented.

Main Results:

• The UDD method successfully estimates causal effects, including nonlinear ones like quantile treatment effects.
• The approach is demonstrated through a real-world application assessing the Zika virus outbreak's impact on birth rates in Brazil.
• The study illustrates the practical application and robustness of the developed UDD estimators.

Conclusions:

• Universal Difference-in-Differences provides a powerful alternative to standard DiD when parallel trends are not met.
• The method enhances causal inference capabilities for complex data structures and non-linear relationships.
• UDD offers a more universally applicable framework for evaluating interventions in observational research.