Inferring Support From Endorsement Experiments

Manuscript

• Manuscript (pdf and .tex)

The Problem: Endorsement Experiments Are Underidentified

Blair et al. (2013) conducted endorsement experiments in Pakistan and concluded:

"We find that Pakistanis in general are weakly negative toward Islamist militant organizations."

"The coefficients are negative and statistically significant...suggesting that Pakistanis hold militant groups in low regard."

This interpretation is incomplete.

The sign of the average treatment effect (ATE) tells us the balance of switchers, not the level of support. With 80% baseline policy support and a -0.011 treatment effect, implied support for militant groups ranges from 39% to 99% depending on assumptions about individual switching behavior:

Model	Assumption	Implied Support
Symmetric effects	Both groups shift ±0.05	39%
Binary switching	All supporters → 1, opponents → 0	79%
Extreme reaction	Opponents react maximally	99%

No model supports Blair et al.'s characterization of "low regard." But the precise level of support is sensitive to assumptions the data cannot adjudicate.

See the manuscript for the full analysis.

The Identification Problem

The ATE decomposes as:

β = p × δ⁺ + (1-p) × δ⁻

Three unknowns (p, δ⁺, δ⁻), one equation. The system is underidentified without restrictions on switching behavior. Different assumptions yield very different estimates:

Without Baseline	With Baseline (α=0.8)	Model
Max 49.5%	39%	Symmetric (d=0.05)
Max 49.5%	79%	Binary switching
Max 49.5%	99%	Extreme reaction

Key insight: Baseline information is crucial and routinely ignored.

Guidance for Practitioners

1. Report treatment group averages, not just ATEs. Under binary switching, the treatment average directly estimates the supporter proportion.

2. Do not equate the ATE sign with "sentiment." The sign tells you which switching force dominates, not whether the group is popular.

3. Discuss identifying assumptions explicitly. Present sensitivity analyses across plausible behavioral models.

4. Test the binary switching assumption. Compare treatment variance to p(1-p). Large discrepancies indicate binary switching is inappropriate.

Reproducing the Paper

python3 scripts/generate_tables_and_figures.py
cd ms && pdflatex endorsement.tex

File Structure

├── ms/endorsement.tex                      # Manuscript
├── scripts/generate_tables_and_figures.py  # Generates tabs/ and figs/
├── tabs/                                   # LaTeX tables for manuscript
├── figs/                                   # Figures for manuscript
└── README.md

Citation

Sood, Gaurav. (2025). Inferring Support from Endorsement Experiments.

Key Takeaways

1. Endorsement experiments are underidentified for the proportion of supporters without behavioral assumptions
2. Baseline support information is crucial and changes estimates dramatically
3. The ATE sign does not measure "sentiment" - it measures the balance of switchers
4. Report ranges, not point estimates unless you can defend a specific behavioral model
5. Binary switching is testable via the variance implication - and often rejected