Pre-processing data to reduce biases: full matching incorporating an instrumental variable in population-based...

Area of Science:

• Epidemiology
• Biostatistics
• Public Health

Background:

• Epidemiological studies face challenges with unobserved biases causing confounding.
• Assessing risks in humanitarian emergencies, such as sexual violence, is complex due to potential biases.

Purpose of the Study:

• Introduce a novel design: Full-IV Matching (full matching incorporating an instrumental variable).
• Demonstrate its utility in reducing observed and unobserved biases to enhance inference accuracy.
• Apply the method to analyze sexual violence risk differences based on displacement settings in humanitarian emergencies.

Main Methods:

• Conducted 1000 Monte Carlo simulations based on a post-earthquake Haitian population survey.
• Utilized earthquake damage severity as an instrumental variable (IV) and 'social capital' as an unmeasured variable.
• Compared standardized mean differences (SMDs) and mean risk differences (RDs) across various matching designs.

Main Results:

• Naive and pair matching analyses showed similar, overstated risk differences for sexual violence.
• Full matching reduced covariate imbalances but not those related to unobserved 'social capital'.
• Full-IV Matching successfully reduced imbalances across observed covariates and 'social capital', yielding results closest to the true effect.

Conclusions:

• Full-IV Matching is a novel and promising approach for improving inference accuracy.
• This method is particularly valuable when unmeasured confounding is suspected.
• The approach enhances the reliability of findings in complex population-based studies.