Optimal anchor percentiles for maximizing ABCD first-stage effects

Determine whether, in Anchoring-Based Causal Design (ABCD) experiments that treat beliefs via numerical anchors, choosing anchor values at approximately the 5th and 95th percentiles of the baseline distribution of the targeted belief maximizes the first-stage anchoring effect (i.e., the difference in mean posttreatment beliefs between the low-anchor and high-anchor groups) used for instrumental variable estimation.

Background

Anchoring-Based Causal Design (ABCD) treats beliefs by prompting respondents to compare a concept to a randomly assigned numeric anchor and then eliciting their posttreatment estimate of the belief. The instrument’s strength in the first stage is crucial for valid and precise IV estimation, and depends critically on the choice of anchor values.

Using multivalued anchor experiments on recession expectations and estimated mean household donations, the authors observed non-linear sensitivity to anchors, with strongest effects when anchors are near the baseline belief range. Based on these results, they propose a conjecture that anchor values at the 5th and 95th percentiles of the baseline belief distribution would approximately maximize the anchoring effect, thereby improving instrument strength and reducing variance in the IV estimator.