Instrumental Variable Analyses and Selection Bias.

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Instrumental variables (IV) methods are crucial for estimating causal effects of exposure (E) on outcome (Y) when unmeasured confounding is present.
• Standard IV assumptions include IV's effect on E, IV's sole effect on Y through E, and no common cause between IV and Y.
• However, even with met IV assumptions, selection bias can introduce non-causal pathways, leading to biased effect estimates.

Purpose of the Study:

• To demonstrate how selection bias can affect instrumental variables (IV) analyses.
• To evaluate the effectiveness of inverse probability of selection weights in mitigating selection bias within IV frameworks.

Main Methods:

• Utilized a simulated dataset to model selection into the sample as a collider on a non-causal exposure-outcome path.
• Applied instrumental variables (IV) analysis to the simulated data.
• Implemented inverse probability of selection weighting to adjust for observed selection bias.

Main Results:

• The instrumental variables (IV) analysis on the simulated data exhibited bias due to selection.
• The application of inverse probability of selection weights successfully eliminated the observed selection bias.
• Demonstrated that IV methods, while robust to unmeasured confounding, are susceptible to selection bias.

Conclusions:

• Selection bias poses a threat to the validity of instrumental variables (IV) analyses, irrespective of IV assumptions.
• Inverse probability of selection weights offer a viable strategy to minimize or eliminate selection bias in conjunction with IV approaches.
• Careful consideration of potential selection mechanisms is essential for reliable causal inference using IV.