Implementation of Instrumental Variable Bounds for Data Missing Not at Random.

Area of Science:

• Epidemiology
• Biostatistics
• Econometrics

Background:

• Instrumental variables (IV) are standard for causal inference with unmeasured confounding.
• IV methods for nonignorable missing data are less common but exist.
• Existing IV methods for missing data often require untestable parametric assumptions.

Purpose of the Study:

• To present an alternative instrumental variable approach for selection bias.
• To provide informative inferences on selection bias without third untestable assumptions.
• To introduce a tool for empirical evidence of selection bias and bounds calculation.

Main Methods:

• Leveraging a valid instrumental variable.
• Developing an approach for selection bias inference.
• Creating an Excel spreadsheet tool for bounds and Bayesian credible intervals.
• Applying the tool to HIV prevalence data from the 2007 Zambia Demographic and Health Survey.

Main Results:

• The proposed method offers insights into selection bias presence, direction, and magnitude.
• The Excel tool facilitates calculation of bounds and Bayesian 95% credible intervals.
• Empirical analysis of HIV data demonstrated the tool's utility.

Conclusions:

• Instrumental variables can be effectively utilized for selection bias analysis in epidemiology.
• The presented approach avoids restrictive parametric assumptions.
• The developed tool provides a practical method for assessing selection bias and its impact.