Difference-in-Differences Estimators for Treatments Continuously Distributed at Every Period

arXiv:2201.06898 · econ.EM · Submitted 2022-01-18 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: We're looking at the paper: "Difference-in-Differences Estimators for Treatments Continuously Distributed at Every Period".

Paper summary: Ruth: We started by looking at the paper, "Difference-in-Differences Estimators for Treatments Continuously Distributed at Every Period." The authors are tackling the problem of estimating effects of taxes or prices in panel data where the treatment is not just a simple switch.

Warren: They develop difference-in-differences estimators specifically for these continuous treatments, which they partition units into switchers, who change their treatment between periods, and stayers who have constant treatment.

Iris: The main thesis is that under a parallel-trends assumption—meaning no treatment means everyone follows the same path—the slopes of those switchers’ potential outcomes with respect to the treatment can be identified.

Ruth: They claim this identification happens by comparing switchers and stayers who share the same baseline treatment level, and they emphasize that conditioning on that baseline treatment is key for making the parallel-trends assumption work with time-varying effects.

Warren: Then, they go on to study two weighted averages of those slopes, calling them the Average of Slopes and the Weighted Average of Slopes.

Iris: The identification results show that the AS, or Average of Slopes, is identified by an estimand comparing outcome evolutions between switchers and stayers with the same period one treatment, which doesn't impose any restrictions on how much treatment effect varies across units.

Ruth: But then they have the WAS, or Weighted Average of Slopes, which they identify using a method that combines regression adjustment with propensity score reweighting.

Warren: The key distinction is that the WAS estimator can be estimated at the parametric rate even if some units can have an arbitrarily small treatment change.

Iris: The AS estimator converges at the parametric rate provided switchers don't experience arbitrarily small treatment changes, while the WAS estimator always converges at that parametric rate and under some conditions achieves a strictly lower asymptotic variance than the AS estimator.

Ruth: They also extend this framework to instrumental variable settings where the parallel-trends assumption is made with respect to an instrument instead of the treatment itself, which restricts treatment effect heterogeneity.

Warren: Finally, they propose extensions for more than two periods by comparing consecutive pairs of time periods and pre-trend estimators that look at switchers and stayers from period t minus two to t minus one.

Conclusion: Ruth: So looking back at the paper "Difference-in-Differences Estimators for Treatments Continuously Distributed at Every Period," it’s about developing a way to handle continuous treatments that aren't simple binary ones.

Warren: The authors propose the AS and WAS as different targets, with WAS offering advantages in estimation rate and asymptotic variance compared to the AS.

Iris: In simple terms, this means they have two tools for measuring how treatment affects things: one for the average slope, and one that handles more subtle variations in how much treatment varies between units.

Ruth: It’s about moving past older methods that assumed treatments were fixed or binary and showing you can still get solid estimates when the treatment keeps changing.

Warren: The implication is that we might have a more robust way to estimate these effects, especially in fields like economics where taxes are often continuously distributed across time.

Sciences Po Paris · CREST-ENSAE · Princeton University · University of California, Santa Barbara

econ.EM

Submitted: 2022-01-18

Updated: 2026-11-17

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

The gist: Difference-in-differences estimators for treatments continuously distributed at every period are developed to provide robust and efficient methods for estimating treatment effects in settings where…

Key concepts

Switchers
These are units whose treatment changes between consecutive time periods. They are central to the analysis because their changing treatments allow researchers to identify how the slope of the outcome response relates to the change in treatment over time.
Stayers
These are units whose treatment remains constant across all observed time periods. Stayers serve as a crucial comparison group, helping researchers isolate the effect of treatment changes from baseline differences between units.
Average of Slopes (AS)
This parameter represents the average slope of potential outcomes for switchers relative to their period-one and period-two treatment levels. It is identified using comparisons between switchers and stayers sharing the same initial treatment level, requiring no restrictions on how much the treatment changes.
Weighted Average of Slopes (WAS)
This parameter weights switchers based on the absolute magnitude of their treatment change. It is estimated using a combination of regression adjustment and propensity-score reweighting, offering advantages in estimation rate and asymptotic variance compared to the AS estimator.

Terminology

Summary

Difference-in-differences estimators for treatments continuously distributed at every period are developed to provide robust and efficient methods for estimating treatment effects in settings where treatments change over time continuously When studying the effects of taxes, tariffs, or prices using panel data, the treatment is often continuously distributed in every period The authors propose a parallel-trends assumption on the outcome evolution of switchers and stayers with the same baseline treatment to identify two target parameters: the Average of Slopes (AS) and the Weighted Average of Slopes (WAS)

Core Estimation Framework

The paper partitions units into switchers, whose treatment changes between consecutive periods, and stayers, whose treatment remains constant Under a parallel-trends assumption, the slopes of switchers’ potential outcomes with respect to the treatment are nonparametrically identified by difference-in-differences comparisons between switchers and stayers sharing the same baseline treatment level The authors propose two target parameters: the Average of Slopes (AS), which is the average slope of switchers’ period-two potential outcome function between their period-one and period-two treatment levels, and the Weighted Average of Slopes (WAS), which weights switchers in proportion to the absolute magnitude of their treatment change

Identification Results for AS and WAS

The AS is identified by an estimand comparing the outcome evolutions of switchers and stayers with the same period-one treatment, which requires no restriction on treatment-effect heterogeneity The WAS is identified by an estimand that combines regression adjustment with propensity-score reweighting, showing it can be estimated at the parametric rate even if units can experience an arbitrarily small treatment change The AS estimator converges at the parametric rate provided switchers cannot experience arbitrarily small treatment changes, while the WAS estimator always converges at the parametric rate and under some conditions achieves a strictly lower asymptotic variance than the AS estimator

Extensions to Instrumental Variables and Time Periods

The framework is extended to instrumental variables (IV) settings where a parallel-trends assumption is made with respect to an instrument, showing that this restricts treatment-effect heterogeneity The authors propose extensions for more than two periods by comparing outcome evolutions of consecutive pairs of time periods and proposing pre-trend estimators that compare the outcome evolutions of switchers and stayers from period t−2 to t−1 Furthermore, the framework is generalized to incorporate covariates, allowing for identification under an alternative parallel trends assumption conditional on the first-period covariates

Empirical Application and Findings

The results are applied to a yearly 1966–2008 US state-level panel of gasoline taxes on consumption and prices, where the WAS estimates yield significantly negative effects on consumption and significantly positive effects on prices, while the AS estimates are close in magnitude but substantially less precise The placebo analysis shows that before switchers change their gasoline taxes, switchers’ and stayers’ consumption and prices do not follow detectably different evolutions

Asymptotic Properties and Efficiency

The asymptotic normality of the estimators is established under specific assumptions, with the asymptotic variance derived through a cross-fitting procedure The efficiency of the influence function is shown by demonstrating that it belongs to the tangent space T, implying that it is pathwise differentiable and efficient The authors also show that under certain conditions, the asymptotic variance of the WAS estimator is lower than that of the AS estimator, and this ratio goes to infinity when a specific condition on switchers' treatment change is relaxed

Robustness Checks

The paper includes several robustness checks, such as placebo analyses comparing outcome evolutions restricted to stayers to test the parallel-trends assumption The authors also propose estimators robust to dynamic treatment effects up to a pre-specified treatment lag by restricting the estimation sample, which is shown to be consistent for average effects of the current treatment on the outcome under certain assumptions Additionally, they discuss extensions for more than two time periods and how placebo tests can calibrate bias in these generalized estimators The final results show that IV-WAS estimates are negative but not significant when the instrument’s first stage is not very strong

Conclusion

The paper proposes new doubly-robust, semiparametrically efficient difference-in-difference (DID) estimators for treatments continuously distributed at every time period The AS and WAS serve different purposes, with the WAS offering advantages in estimation rate and asymptotic variance compared to the AS The authors also consider the instrumental-variable case, proposing widely applicable IV-DID estimators

References

Aklin, M. and A. Kern (2021). p. 1 The side effects of central bank independence. p. 0 American Journal of Political Science 65 (4), 971–987 Anderson, T. p. 1 W. and H. Rubin (1949). Estimation of the parameters of a single equation in a complete system of stochastic equations. p. 3 Annals of Mathematical Statistics 20 (1), 46–63 Andrews, D. W. (1991). Asymptotic optimality of generalized cl, cross-validation, and generalized cross-validation in regression with heteroskedastic errors. p. 7 Journal of Econometrics 47 (2-3), 359–377 Angrist, J. D., K. Graddy, and G. W. Imbens (2000). p. 7 The interpretation of instrumental variables estimators in simultaneous equations models with an application to the demand for fish. p. 2 The Review of Economic Studies 67 (3), 499–527 Bang, H. p. 0 and J. M. Robins (2005). Doubly robust estimation in missing data and causal inference models. p. 5 Biometrics 61 (4), 962–973 Bertrand, M. p. 5, E. Duflo, and S. Mullainathan (2004). How much should we trust differences-indifferences estimates? p. 3 The Quarterly Journal of Economics 119 (1), 249–275 Biau, G. p. 1 (2012). Analysis of a random forests model. p. 4 Journal of Machine Learning Research 13 (38), 1063–1095 Bojinov, I. p. 0, A. Rambachan, and N. Shephard (2021). p. 1 Panel experiments and dynamic causal effects: A finite population perspective. p. 0 Quantitative Economics 12 (4), 1171–1196 Borusyak, K. p. 1, X. Jaravel, and J. Spiess (2024). p. 1 Revisiting event-study designs: robust and efficient estimation. p. 0 Review of Economic Studies 91 (6), 3253—-3285 Callaway, B. p. 0, A. Goodman-Bacon, and P. p. 1 H. Sant’Anna (2021). p. 3 p. 3 Difference-in-differences with a continuous treatment. p. 0 arXiv preprint arXiv:2107. p. 002637 Callaway, B. p. 3 and P. H. Sant’Anna (2021). Difference-in-differences with multiple time periods. p. 0 Journal of Econometrics 225, 200–230 Chamberlain, G. p. 3 (1982). Multivariate regression models for panel data. p. 0 Journal of econometrics 18 (1), 5–46 Chen, X. and H. White (1999). Improved rates and asymptotic normality for nonparametric neural network estimators. p. 2 IEEE Transactions on Information Theory 45 (2), 682–691 Chernozhukov, V., D. Chetverikov, M. Demirer, E. Duflo, C. Hansen, W. Newey, and J. Robins (2018). Double/debiased machine learning for treatment and structural parameters Chernozhukov, V. p. 3, W. K. Newey, V. Quintas-Martinez, and V. Syrgkanis (2024). p. 1 Automatic debiased machine learning via Riesz regression. p. 0 arXiv preprint arXiv:2104. p. 014737 Chiu, A. p. 0, X. Lan, Z. Liu, and Y. p. 3 Xu (2026). Causal panel analysis under parallel trends: Lessons from a large reanalysis study. p. 1 American Political Science Review 120 (1), 245–266 de Chaisemartin, C. p. 0 (2010). A note on instrumented difference in differences. p. 0 Working Paper de Chaisemartin, C. p. 3, D. Ciccia, X. D’Haultfœuille, and F. p. 0 Knau (2025). p. 3 Difference-in-differences estimators when no unit remains untreated. p. 0 arXiv preprint arXiv:2405. p. 004465 de Chaisemartin, C. p. 0 and X. D’Haultfœuille (2020). p. 1 Two-way fixed effects estimators with heterogeneous treatment effects. p. 0 American Economic Review 110 (9), 2964–2996 de Chaisemartin, C. p. 0 and X. d’Haultfoeuille (2025). p. 3 Treatment-effect estimation in complex designs under a parallel-trends assumption. p. 3 arXiv preprint arXiv:2508. p. 007808 de Chaisemartin, C. p. 0, X. D’Haultfœuille, F. p. 0 Pasquier, and G. p. 0 Vazquez-Bare (2022). p. 0 Differencein-differences for continuous treatments and instruments with stayers. p. 0

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improve estimation for continuous treatments by using doubly robust, semiparametrically efficient estimators for both Average of Slopes (AS) and Weighted Average of Slopes (WAS). This allows systems to estimate effects even when units experience arbitrarily small treatment changes, and potentially achieves a strictly lower asymptotic variance than AS under some conditions.

  2. Enhance causal inference in instrumental variable settings by proposing widely applicable IV-DID estimators. These estimators are robust to heterogeneity restrictions by making a parallel trends assumption with respect to an instrument, allowing for the identification of effects like price elasticity of demand even when direct parallel trends assumptions fail.

  3. Extend the framework to multiple time periods by aggregating period specific estimators into AS and WAS generalizations for T greater than two. This allows systems to estimate effects across longer time horizons while providing a robustness check against dynamic treatment effects up to a specified lag using placebo tests on stayers.

  4. Generalize estimation methods to include covariates by redefining nuisance functions as functions of the covariate X1. This enables systems to estimate AS and WAS under parallel trends assumptions conditional on covariates, allowing for richer modeling of heterogeneity based on observed characteristics like initial treatment status.

  5. Develop machine learning based estimators for nuisance functions when covariate dimensionality is high or when classical nonparametric estimators fail to satisfy required convergence rates. This ensures that the asymptotic normality of the resulting DID estimators holds even in complex settings where sample sizes are limited or covariates are numerous.

Sources

Related papers