Learning Causally Predictable Outcomes from Psychiatric Longitudinal Data.

Area of Science:

• Biomedical data analysis
• Psychiatric research
• Causal inference

Background:

• Causal inference in longitudinal biomedical data is challenging, particularly in psychiatry.
• Symptom heterogeneity and latent confounding complicate classical estimation methods.
• Existing approaches often assume a fixed outcome and rely on covariate adjustment, which may not hold in practice.

Purpose of the Study:

• To address limitations in causal inference for longitudinal psychiatric data.
• To develop a method that optimizes outcome definition for improved causal identifiability.
• To minimize observed and latent confounding in treatment effect estimation.

Main Methods:

• Introduced the DEBIAS (Durable Effects with Backdoor-Invariant Aggregated Symptoms) algorithm.
• DEBIAS learns clinically interpretable weights for outcome aggregation.
• Leverages time-limited direct effects of prior treatments to minimize confounding.

Main Results:

• DEBIAS maximizes durable treatment effects and minimizes confounding.
• The algorithm provides an empirically verifiable test for outcome unconfoundedness.
• DEBIAS outperforms state-of-the-art methods in recovering causal effects for composite outcomes in depression and schizophrenia.

Conclusions:

• DEBIAS offers a novel approach to causal inference in complex longitudinal psychiatric data.
• The method enhances the identifiability and reliability of treatment effect estimation.
• This algorithm improves the understanding of treatment efficacy in mental health research.