🔍 What Ecological Inference Is and What Was Changed
Ecological inference (EI) aims to learn about individual behavior from aggregate data. This study relaxes a common restriction by allowing for linear contextual effects—assumptions that prior work treated as plausible but avoided because they produce nonidentification.
🔎 How Nonidentification Is Handled
Instead of forcing point estimates under nonidentification, the approach derives bounds. This partial-identification framework provides a conceptual method to improve on the classic Duncan–Davis bound (first derived more than 65 years ago) while acknowledging inherent limits to identification.
📊 Large-Scale Test Using Known Ground Truth
- Collected and analyzed 8,430 2×2 EI datasets with known ground truth from several sources.
- This collection brings considerably more data to bear than the roughly dozen datasets previously used to evaluate EI estimators.
📈 Key Findings
- 88% of the real datasets in the collection satisfy a proposed rule for applying the new bounds.
- For those datasets, the method reduces the width of the Duncan–Davis bound by about 44% on average.
- The tightened bounds still capture the true district-level parameter about 99% of the time.
- The remaining 12% of datasets revert to the original Duncan–Davis bound.
⚖️ Why This Matters
Relaxing to linear contextual effects and using bounds rather than forced point estimates yields substantially tighter uncertainty intervals for most real cases while preserving near-perfect coverage. This provides researchers and practitioners a more informative and honest alternative to the long-standing Duncan–Davis bound when analyzing aggregate data.
