Evaluation of Methods Adjusting for Unmeasured Confounding Using Large Healthcare Databases: An Empirical Study...

Area of Science:

• Pharmacoepidemiology
• Causal Inference
• Health Data Science

Background:

• Large healthcare databases offer opportunities for pharmacoepidemiologic research.
• Mitigating unmeasured confounding is crucial for valid causal inference in these datasets.
• Comparing causal inference methods in real-world, high-dimensional data is essential.

Purpose of the Study:

• To compare the performance of G-computation (GC), Targeted Maximum Likelihood estimation (TMLE), and Propensity Score methods in reducing measured and unmeasured confounding.
• To evaluate these methods using a machine learning LASSO algorithm on a high-dimensional, real-world healthcare database.
• To assess the ability of these methods to identify true and false positive drug-prematurity associations.

Main Methods:

• Utilized the French National Healthcare Claims Database (SNDS) with 2,172,702 pregnancies (≥22 weeks gestation, 2011-2014).
• Employed G-computation (GC), Targeted Maximum Likelihood estimation (TMLE), and Propensity Score methods adapted for high-dimensional data.
• Assessed 42 negative and 13 positive reference drugs for prematurity risk, estimating odds ratios and confidence intervals.

Main Results:

• All causal inference methods outperformed a crude model in reducing false positives.
• Targeted Maximum Likelihood estimation (TMLE) yielded the lowest false positive rate (45.2%).
• G-computation (GC) achieved the highest true positive rate (92.3%) and demonstrated good performance and ease of implementation.

Conclusions:

• Causal inference methods are valuable for leveraging large healthcare databases.
• G-computation (GC) shows particular promise for pharmacoepidemiologic studies due to its performance and implementation simplicity.
• These findings support the use of advanced causal inference techniques to address confounding in real-world health data research.