The M-Value: A Simple Sensitivity Analysis for Bias Due to Missing Data in Treatment Effect Estimates

Area of Science:

• Biostatistics
• Epidemiology
• Clinical Research Methodology

Background:

• Complete-case analyses are frequently used but can yield biased treatment effect estimates when data are not missing completely at random.
• Bias can arise from differences between retained and non-retained participants or from induced non-causal pathways between treatment and outcome.

Purpose of the Study:

• To propose simple sensitivity analysis methods for complete-case estimates of treatment effects.
• To provide tools for assessing the robustness of findings to potential bias from missing data without requiring complex modeling.

Main Methods:

• Develop sensitivity analyses that bound the overall treatment effect by specifying the unobserved treatment effect among non-retained participants.
• Quantify the strength of unobserved associations with exposure and outcome using a novel "M-value," analogous to the E-value.
• These methods utilize only simple summary data, avoiding distributional assumptions or specifying precise missingness mechanisms.

Main Results:

• The proposed methods allow bounding treatment effects by considering hypothetical scenarios of missing data.
• The M-value quantifies the strength of induced confounding associations needed to nullify the observed treatment effect.
• These approaches subsume existing worst-case imputation methods while allowing for less conservative assumptions.

Conclusions:

• The proposed sensitivity analyses offer a straightforward way to evaluate the potential impact of missing data on treatment effect estimates.
• The M-value provides a valuable metric for assessing the robustness of complete-case analyses to unobserved confounding.
• These methods enhance the interpretability and reliability of findings from studies with missing data.