Causal inference under over-simplified longitudinal causal models

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Causal models in epidemiology frequently involve longitudinal data for exposures, confounders, and mediators.
• Practical limitations often restrict the use of repeated measurements, leading to simplified causal models.
• This simplification can overlook the time-varying nature of exposures, potentially biasing results.

Purpose of the Study:

• To evaluate the relationship between causal effects estimated using simplified models and true longitudinal causal effects.
• To determine conditions under which simplified model estimates approximate true causal effects.
• To assess the implications of using over-simplified causal models in epidemiological research.

Main Methods:

• Derivation of sufficient conditions for simplified causal estimates to represent weighted averages of longitudinal causal effects.
• Theoretical analysis of the relationship between estimates from misspecified and correctly specified causal models.
• Simulation studies to quantify the bias between estimated and true longitudinal causal effects.

Main Results:

• Sufficient conditions for simplified estimates to approximate longitudinal causal effects are highly restrictive.
• In general, quantities estimated under simplified models do not correspond to the true longitudinal causal effects of interest.
• Simulations demonstrate substantial bias between estimated quantities and weighted averages of longitudinal causal effects.

Conclusions:

• Estimates from over-simplified causal models in epidemiology should be interpreted with caution.
• The use of repeated measurements is crucial for accurate causal effect analysis.
• Sensitivity analyses are recommended when repeated measurements are unavailable.