The paper focuses on proximal causal inference, a framework that allows researchers to estimate causal effects even when some important confounders are unmeasured, by using observed proxy variables. Traditionally, this approach has been applied to settings with fixed treatments or interventions based on baseline characteristics.
This work extends the framework to more complex situations with continuous treatments, specifically considering scenarios where each individual’s observed treatment is systematically modified (known as modified treatment regimes).
The main contributions are:
– proposing a flexible method that does not require all confounders to be measured;
– using modern debiased machine learning techniques to avoid strong parametric assumptions.
The methodology is motivated by studies on COVID-19 vaccines, where key factors like an individual’s immune capacity are not directly observed. The approach is validated using real data and simulation studies.