Causal Inference with Unobserved Confounding: Leveraging Negative Control Outcomes Using Lavaan.

Area of Science:

• Epidemiology
• Causal Inference
• Biostatistics

Background:

• Causal conclusions from non-randomized studies rely on the untestable assumption of no unobserved confounding.
• Unobserved confounding is a pervasive threat in real-world observational data.
• Estimating unbiased causal effects in the presence of unobserved confounding remains a significant challenge.

Purpose of the Study:

• Introduce negative control outcomes as a method to address unobserved confounding.
• Explain the mechanism by which negative control outcomes counteract bias.
• Demonstrate the practical implementation and utility of the Control Outcome Calibration Approach (COCA).

Main Methods:

• Utilize negative control outcomes, a concept from causal inference and epidemiology.
• Employ the Control Outcome Calibration Approach (COCA) for estimation.
• Implement COCA in R using the lavaan package for statistical modeling.

Main Results:

• Demonstrated the application of COCA using two real-world datasets.
• Showcased COCA as a practical and straightforward method for causal effect estimation.
• Provided evidence that COCA can achieve unbiased causal effect estimation under specific assumptions, even with unobserved confounding.

Conclusions:

• Negative control outcomes provide a viable strategy to mitigate bias from unobserved confounding.
• The Control Outcome Calibration Approach (COCA) is an accessible and effective tool for implementing this strategy.
• COCA facilitates more reliable causal inference in observational studies where unobserved confounding is a concern.