A More Efficient Causal Mediator Model Without the No-Unmeasured-Confounder Assumption

Area of Science:

• Causal inference
• Statistical modeling
• Epidemiology

Background:

• Traditional mediator models often rely on the unverifiable assumption of no unmeasured confounding.
• Violations of this assumption can lead to spurious identification of mediator variables.
• The rank preserving model (RPM) offers an alternative using a no-effect-modifier assumption, but can be inefficient.

Purpose of the Study:

• To introduce a semi-parametric nonlinear extension of the rank preserving model (nRPM).
• To improve the efficiency and robustness of mediator effect estimation in the presence of potential unmeasured confounding.
• To address the limitations of existing mediator models in real-world scenarios.

Main Methods:

• Development of the nonlinear rank preserving model (nRPM) using thin plate regression splines.
• A simulation study to compare the performance of nRPM against the standard RPM.
• Application of the nRPM to a dataset on CD4 cell counts in HIV patients.

Main Results:

• The nRPM demonstrates robust estimates even with violations of the no-effect-modifier assumption.
• nRPM provides substantively more efficient estimates compared to the traditional RPM.
• The simulation study confirmed the superior performance of nRPM.

Conclusions:

• The nRPM is a valuable advancement for estimating mediator effects, offering improved efficiency and robustness.
• This model is particularly useful in situations where the no-unmeasured-confounding assumption is questionable.
• The nRPM provides a more reliable tool for causal inference in complex datasets, as demonstrated in the HIV context.