Parallel Trends as a Spectrum
Abstract
Difference-in-differences (DiD) designs are typically used to identify and estimate average causal effects, even though treatment can affect the entire outcome distribution. This has led to a growing interest in distributional DiD methods. However, identifying the full counterfactual distribution requires a parallel trends assumption on the entire distributional dynamics, which can be too strong in practice. We address this problem by introducing a spectrum of identifying restrictions that lie between the average and fully distributional DiD by imposing parallel trends on L-moments. This corresponds to a basis expansion of the distributional dynamics. Importantly, each L-moment is interpretable as familiar distributional components: location, dispersion, skewness, kurtosis, etc. Researchers can pick and choose which features they want to target. For example, the second L-moment identifies the Gini mean difference, but does not require imposing mean-level or higher-order parallel trends. Taking a collection of L-moments can approximate distributional dynamics under weaker conditions since it leaves untargeted components unrestricted. We develop conditions under which higher-order L-moments decay, efficient estimators, and simultaneous inference across growing collections of moments. We demonstrate the benefits of our framework by exploring distributional changes to healthcare spending after the rollout of Medicare Part D.
Related
- Parallel transport for Higgs bundles over p-adic curves
- Pluriclosed manifolds with parallel Bismut torsion
- Fedosov manifolds with parallel symplectic Weyl curvature
- The geometric convergence of a parallel domain decomposition method on manifolds
- Parallel Optimized Particle Swarm Optimization: An Efficient Parallel Approach for Neural Network Training
BibTeX
@misc{schindl2026parallel,
title = {{Parallel Trends as a Spectrum}},
author = {Kyle Schindl and Joseph Fry and Florian Gunsilius and Larry Wasserman},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.31978},
eprint = {2610.31978},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.31978},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Parallel Trends as a Spectrum
AU - Kyle Schindl
AU - Joseph Fry
AU - Florian Gunsilius
AU - Larry Wasserman
PY - 2026
DA - 2026/10/22
PB - arXiv
DO - 10.48550/arxiv.2610.31978
UR - https://arxiv.org/abs/2610.31978
AB - Difference-in-differences (DiD) designs are typically used to identify and estimate average causal effects, even though treatment can affect the entire outcome distribution. This has led to a growing interest in distributional DiD methods. However, identifying the full counterfactual distribution requires a parallel trends assumption on the entire distributional dynamics, which can be too strong in practice. We address this problem by introducing a spectrum of identifying restrictions that lie between the average and fully distributional DiD by imposing parallel trends on L-moments. This corresponds to a basis expansion of the distributional dynamics. Importantly, each L-moment is interpretable as familiar distributional components: location, dispersion, skewness, kurtosis, etc. Researchers can pick and choose which features they want to target. For example, the second L-moment identifies the Gini mean difference, but does not require imposing mean-level or higher-order parallel trends. Taking a collection of L-moments can approximate distributional dynamics under weaker conditions since it leaves untargeted components unrestricted. We develop conditions under which higher-order L-moments decay, efficient estimators, and simultaneous inference across growing collections of moments. We demonstrate the benefits of our framework by exploring distributional changes to healthcare spending after the rollout of Medicare Part D.
ER -