When Controlling Backfires: How Fixed Effects Amplify and Unmask Bias

🔍 Problem and Approach

Conditioning on observed covariates is a standard strategy for reducing confounding in non-experimental causal inference, but some covariates can increase rather than decrease bias. This article develops a clearer way to think about when and why that happens by decomposing omitted-variable bias into three distinct parts and studying the mechanisms that produce bias increases when conditioning.

🧩 A New Way to Break Down Omitted-Variable Bias

The omitted-variable bias is decomposed into three constituent components:

• bias due to an unobserved confounder
• bias due to excluding observed covariates
• bias due to amplification (the change in bias that results from conditioning)

This decomposition clarifies how adding controls can alter bias through multiple channels, not just by blocking confounding paths.

⚠️ Main Surprises

• Prior work focused mostly on instruments as a source of bias amplification; this analysis shows that adding group fixed effects can also amplify bias. Many practitioners treat fixed effects as a catch-all, harmless control for group-level confounding; the results show that fixed effects can instead increase bias.
• The paper introduces the concept of "bias unmasking," where conditioning can reveal or enlarge bias that had been offset by other factors. In some cases, bias unmasking can be more harmful than conventional bias amplification because conditioning removes offsetting influences and leaves a larger net bias.

🧪 How the Effects Are Demonstrated

• Analytical derivations develop the new decomposition and identify conditions under which amplification and unmasking occur.
• Constructed observational placebo studies using real data illustrate both bias amplification and bias unmasking in practice, showing concrete examples of how added controls can worsen estimates.

📈 Practical Recommendation

A proposal is made to augment graphical sensitivity displays with bias-decomposition information so researchers can visualize the potential for amplification and unmasking. These enhanced displays aim to help practitioners diagnose when commonly used adjustments—especially fixed effects—might backfire and to guide more cautious sensitivity analysis.

Why it matters: Conditioning is not always benign. The decomposition and the concepts of bias amplification and bias unmasking provide actionable diagnostics for applied researchers worried about misleading causal estimates in observational studies.