Estimating causal effects of time-dependent exposures on a binary endpoint in a high-dimensional setting

Area of Science:

• Causal inference
• High-dimensional data analysis
• Biostatistics

Background:

• The intervention calculus when the DAG is absent (IDA) method estimates causal effects from observational data but was limited to non-time-varying exposures.
• Clinical settings often involve repeated biomarker measurements over time, necessitating an extension of existing causal inference methods.
• Existing methods do not adequately account for the temporal nature of biomarker data in causal effect estimation.

Purpose of the Study:

• To extend the Peter Clarks (PC)-algorithm, a component of IDA, to handle time-dependent exposures.
• To adapt causal inference methods for binary outcomes with repeated, time-stamped biomarker measurements.
• To develop a chronologically aware algorithm for causal discovery in longitudinal observational studies.

Main Methods:

• Generalised the PC-algorithm to incorporate the chronological order of exposure measurements, creating the chronologically ordered PC-algorithm (COPC-algorithm).
• Incorporated Firth's correction into the COPC-algorithm for improved statistical stability.
• Validated the COPC-algorithm through a simulation study and applied it to estimate causal effects of time-dependent immunological biomarkers on patient outcomes.

Main Results:

• The COPC-algorithm generated completed partially directed acyclic graphs (CPDAGs) that were structurally closer to the true causal graph than those from the standard PC-algorithm.
• Causal effects estimated using COPC-algorithm-derived CPDAGs were more accurate.
• COPC-algorithm effectively removed non-chronological arrows and reduced bidirected edges, indicating improved causal structure identification and less variability in effect estimates.

Conclusions:

• The COPC-algorithm successfully preserves the chronological structure of time-dependent biomarker data.
• This method enables more accurate estimation of lower bounds for causal effects of time-dependent biomarkers on clinical endpoints like toxicity, death, and progression.
• The COPC-algorithm offers a robust approach for causal inference in high-dimensional, longitudinal observational health data.