Difference-in-differences with as few as two cross-sectional units -- A new perspective to the democracy-growth debate
Pith reviewed 2026-05-23 21:35 UTC · model grok-4.3
The pith
A temporal difference-in-differences estimator estimates unit-specific treatment effects with only two cross-sectional units and finds democratization increased Benin's average growth.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
The T-DiD estimator leverages temporal variation in the data to estimate unit-specific average treatment effects on the treated (ATT) with as few as two cross-sectional units. Under asymptotic parallel trends, limited anticipation, and temporal dependence conditions, the proposed DiD estimator is shown to be asymptotically normal. Provided at least two control units are available, the method is further complemented with an identification test that, unlike pre-trends tests, is more powerful and can detect violations of parallel trends in post-treatment periods. Empirical results using the DiD estimator suggest Benin's economy would have been 6.4% smaller on average over the 1993-2018 period.
What carries the argument
The Temporal Difference-in-Differences (T-DiD) estimator that uses temporal variation to identify unit-specific ATT.
If this is right
- Unit-specific effects can be estimated without large panels of cross-sectional units.
- The estimator is asymptotically normal, enabling standard inference.
- An identification test can detect parallel trends violations after treatment.
- The application shows a positive effect of democratization on economic growth in Benin.
Where Pith is reading between the lines
- If the method holds, many single-country policy changes can be evaluated with minimal data requirements.
- The test for identification could be used in other settings with few units to validate assumptions.
- Results for Benin suggest re-evaluating democracy-growth links with country-specific methods rather than pooled panels.
Load-bearing premise
The asymptotic parallel trends condition holds for the temporal variation in the data.
What would settle it
Finding that the parallel trends assumption is violated in the post-treatment period for the Benin democratization study would invalidate the 6.4 percent estimate.
read the original abstract
Pooled panel analyses often mask heterogeneity in unit-specific treatment effects. This challenge, for example, crops up in studies of the impact of democracy on economic growth, where findings vary substantially due to differences in country composition. To address this challenge, this paper introduces the Temporal Difference-in-Differences (T-DiD) estimator that leverages temporal variation in the data to estimate unit-specific average treatment effects on the treated (ATT) with as few as two cross-sectional units. Under asymptotic parallel trends, limited anticipation, and temporal dependence conditions, the proposed DiD estimator is shown to be asymptotically normal. Provided at least two control units are available, the method is further complemented with an identification test that, unlike pre-trends tests, is more powerful and can detect violations of parallel trends in post-treatment periods. Empirical results using the DiD estimator suggest Benin's economy would have been 6.4% smaller on average over the 1993-2018 period had she not democratised.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper introduces the Temporal Difference-in-Differences (T-DiD) estimator to recover unit-specific average treatment effects on the treated (ATT) from temporal variation alone, applicable when the number of cross-sectional units N is as small as 2 and T is large. It states that the estimator is asymptotically normal under asymptotic parallel trends, limited anticipation, and temporal dependence conditions. The manuscript also proposes an identification test for parallel-trends violations that can be applied in post-treatment periods. In the empirical application, the T-DiD estimator is used to conclude that Benin’s economy would have been 6.4 percent smaller on average over 1993–2018 in the absence of democratization.
Significance. If the asymptotic results and the identification test are valid, the approach would allow unit-specific causal estimates in the small-N large-T environments typical of cross-country growth studies, addressing the heterogeneity problem that arises in pooled panel regressions of democracy on growth. The post-treatment identification test could strengthen the credibility of DiD designs where pre-trends tests are uninformative. The practical value, however, depends on whether the asymptotic parallel trends condition can be maintained at the rate required for √T-consistency when N is fixed at 2.
major comments (3)
- [Abstract] Abstract and theoretical section: the asymptotic normality result is derived under the joint conditions of asymptotic parallel trends, limited anticipation, and temporal dependence, yet with N fixed at 2 any fixed (non-vanishing) violation of parallel trends produces a non-vanishing bias term that cannot be averaged away; the manuscript does not supply the rate condition on the violation or Monte Carlo evidence showing robustness at plausible violation rates.
- [Empirical application] Empirical application (Benin results): the headline 6.4 percent figure is reported without standard errors, confidence intervals, or any indication of the underlying data sources, variable definitions, or sample construction, preventing assessment of whether the asymptotic parallel trends condition is plausible in the 1993–2018 period.
- [Identification test section] Identification test: the claim that the proposed post-treatment test is “more powerful” than conventional pre-trends tests is stated without a formal power analysis, local-alternative derivation, or simulation study comparing size and power under the maintained temporal-dependence conditions.
minor comments (2)
- [Abstract] The abstract introduces the acronym T-DiD without spelling out “Temporal Difference-in-Differences” on first use.
- [Notation and definitions] Notation for the unit-specific ATT and the precise definition of the “asymptotic parallel trends” condition should be stated explicitly in the main text before the asymptotic normality theorem is invoked.
Simulated Author's Rebuttal
We thank the referee for the constructive comments. We address each major point below and indicate planned revisions.
read point-by-point responses
-
Referee: [Abstract] Abstract and theoretical section: the asymptotic normality result is derived under the joint conditions of asymptotic parallel trends, limited anticipation, and temporal dependence, yet with N fixed at 2 any fixed (non-vanishing) violation of parallel trends produces a non-vanishing bias term that cannot be averaged away; the manuscript does not supply the rate condition on the violation or Monte Carlo evidence showing robustness at plausible violation rates.
Authors: We agree that with N fixed at 2, asymptotic normality requires the parallel-trends violation to vanish at a rate that makes the bias term o_p(T^{-1/2}). The manuscript states the asymptotic parallel trends assumption but does not derive the explicit rate or supply Monte Carlo evidence. We will add the required rate condition to the theoretical section and include Monte Carlo simulations demonstrating performance under plausible violation rates. revision: yes
-
Referee: [Empirical application] Empirical application (Benin results): the headline 6.4 percent figure is reported without standard errors, confidence intervals, or any indication of the underlying data sources, variable definitions, or sample construction, preventing assessment of whether the asymptotic parallel trends condition is plausible in the 1993–2018 period.
Authors: The referee is correct that the empirical section omits standard errors, confidence intervals, and full data documentation. We will revise the application to report standard errors and confidence intervals for the 6.4 percent estimate and to include complete details on data sources, variable definitions, and sample construction. revision: yes
-
Referee: [Identification test section] Identification test: the claim that the proposed post-treatment test is “more powerful” than conventional pre-trends tests is stated without a formal power analysis, local-alternative derivation, or simulation study comparing size and power under the maintained temporal-dependence conditions.
Authors: The manuscript claims greater power on the basis of post-treatment data availability, but we agree that no formal power analysis or simulation study is provided. We will either add a simulation study under the maintained temporal-dependence conditions or qualify the claim accordingly in the revision. revision: partial
Circularity Check
No circularity: T-DiD asymptotic normality derived under explicit assumptions
full rationale
The paper claims asymptotic normality of the unit-specific ATT estimator under the joint conditions of asymptotic parallel trends, limited anticipation, and temporal dependence (with N fixed, T→∞). This is a standard derivation from stated regularity conditions rather than a reduction by construction to fitted parameters, self-citations, or renamed inputs. The identification test is presented as an additional tool that can detect post-treatment violations but does not relax or replace the core assumptions. No self-definitional steps, fitted-input predictions, or load-bearing self-citations appear in the derivation chain. The Benin 6.4% figure is an application of the estimator, not a circular output.
Axiom & Free-Parameter Ledger
axioms (3)
- domain assumption Asymptotic parallel trends condition
- domain assumption Limited anticipation
- domain assumption Temporal dependence conditions
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.