Leveraging lavaan for iterated conditional expectations: A streamlined parametric g-formula

Area of Science:

• Causal Inference
• Longitudinal Data Analysis
• Psychological Research Methods

Background:

• Longitudinal designs are crucial for understanding time-dependent causal effects.
• Time-varying confounding presents significant challenges in causal effect estimation.
• Existing methods like parametric g-formula can be complex and prone to misspecification.

Purpose of the Study:

• To introduce a streamlined variant of the parametric g-formula called iterated conditional expectation (ICE).
• To provide an accessible implementation of ICE for psychological researchers using R.
• To demonstrate the application of ICE for estimating causal effects in longitudinal data.

Main Methods:

• Iterated conditional expectation (ICE) approach to parametric g-formula.
• Requires a sequence of regression models for mean outcomes at consecutive time points.
• Implementation provided using the R package lavaan for user-friendliness.

Main Results:

• ICE simplifies causal inference by avoiding simulation of joint distributions of confounders and outcomes.
• ICE is less susceptible to model misspecification and handles various confounder types (continuous, binary, categorical).
• A doubly robust version of ICE enhances protection against outcome model misspecification.

Conclusions:

• Parametric g-formula using ICE is an elegant, computationally efficient, and practical method for causal inference.
• ICE is well-suited for longitudinal data with repeatedly measured outcomes, common in psychological research.
• The method facilitates clearer understanding of causal assumptions and effect estimation.