Estimating causal log-odds ratio using the case-control sample and its application in the pharmaco-epidemiology...

Area of Science:

• Pharmaco-epidemiology
• Causal inference
• Observational data analysis

Background:

• Pharmaco-epidemiology aims to establish causal links between drug exposure and clinical outcomes, but faces challenges like confounding from co-medications in observational data.
• Existing pharmaco-epidemiology datasets, often sampled using matched case-control designs from large medical records, may not represent the original patient population.
• Large sample sizes in these datasets exceed the capacity of traditional statistical analysis packages.

Purpose of the Study:

• To develop and validate a novel methodological and computational approach for causal inference in pharmaco-epidemiology.
• To address challenges of confounding, data representativeness, and large sample sizes in observational drug safety studies.
• To accurately estimate causal effects of drug exposures on binary adverse drug events.

Main Methods:

• Proposed a conditional causal log-odds ratio (OR) definition to characterize causal effects, adjusting for individual-level confounders.
• Developed a propensity score estimation method using only case samples within a case-control design, providing conditions for consistent causal log-odds ratio estimation.
• Implemented principal component analysis for dimensionality reduction of high-dimensional confounders and applied the method to a large-scale dataset (Indiana Network for Patient Care).

Main Results:

• Extensive simulation studies demonstrated the superior performance of the proposed method compared to existing approaches.
• Analysis of drug-induced myopathy data identified 70 out of 72 drugs with known myopathy side effects (p < 0.05).
• The identified drugs included three statins, confirming known associations with myopathy.

Conclusions:

• The proposed method provides a robust framework for causal inference in pharmaco-epidemiology, effectively handling confounding and large datasets.
• The approach successfully identified numerous drug-induced myopathy cases, highlighting its utility in real-world drug safety surveillance.
• This work advances the ability to understand drug-induced adverse events from observational data.