Continuous difference-in-differences with double/debiased machine learning

arXiv:2408.10509 · econ.EM · Submitted 2024-08-20 · 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: I'm Ruth.

Warren: And I'm Warren, here with Iris.

Ruth: We're looking at the paper: "Continuous difference-in-differences with double/debiased machine learning".

Warren: Here's the gist of the paper. Continuous difference-in-differences with double/debiased machine learning extends difference-in-differences to settings with continuous treatments by proposing estimators under a conditional parallel trends assumption and establishing their asymptotic normality using the double/debiased machine learning framework <ref:2408.10509#pg2>.

Ruth: Identification Framework The paper extends difference-in-differences to settings with continuous treatments by proposing estimators under a conditional parallel trends assumption and establishing their asymptotic normality using the double/debiased machine learning framework <ref:2408.10509#pg4>.

Warren: The identification results are based on a conditional parallel trends assumption, allowing researchers to account for covariates non-parametrically <ref:2408.10509#pg7>.

Ruth: The target parameter is the average treatment effect on the treated (ATT for short) at any given continuous treatment intensity d, defined as AT T(d):= E

Yt(d) − Yt(0)D = d: <ref:2408.10509#pg3>.

Paper summary: Ruth: So to recap the main thrust of this paper, they are extending difference-in-differences to settings with continuous treatments by proposing estimators under a conditional parallel trends assumption and establishing their asymptotic normality using the double/debiased machine learning framework <ref:2408.10509#pg4>.

Warren: The core claim is that they can now identify the average treatment effect on the treated at any level of treatment intensity, AT T(d), even when you have continuous treatments <ref:2408.10509#pg3>.

Iris: They address the problem by first estimating infinite-dimensional nuisance parameters, specifically the conditional density of that continuous treatment, which they know can cause substantial bias <ref:2408.10509#pg4>.

Ruth: To counter that, they propose estimators for causal parameters using the double/debiased machine learning framework and prove their asymptotic normality <ref:2408.10509#pg4>. They introduce an approximate causal parameter AT Th using a kernel function to get around the direct application of those ML methods <ref:2408.10509#pg4>.

Warren: The motivation is clear: continuous treatments are common, and you can investigate how outcomes vary across different treatment intensities within the treated group <ref:2408.10509#pg2>.

Iris: And they use a conditional parallel trends assumption to make the identification robust, which lets them condition on covariates that influence outcome dynamics <ref:2408.10509#pg3>.

Ruth: The implication is that instead of just looking at an aggregate effect, you can look at how the effect changes as the treatment dose increases <ref:2408.10509#pg2>.

Warren: It’s a way to investigate those varying levels of exposure without having to rely on simpler models that might miss important nuances <ref:2408.10509#pg8>.

Conclusion: Ruth: So looking at "Continuous difference-in-differences with double/debiased machine learning," the authors are basically giving us a way to handle continuous treatment variables in difference-in-differences settings <ref:2408.10509#pg1>.

Warren: They’ve shown that by using double/debiased machine learning, you can get estimators for the causal parameters and prove they have asymptotic normality <ref:2408.10509#pg4>.

Iris: The real implication is that this gives researchers new tools for examining the impacts of continuous treatment variables in empirical research <ref:2408.10509#pg8>.

Ruth: It means we can move past just binary treatments and start looking at how outcomes vary across different treatment intensities with more rigor and precision #page two of that work reads: "there is growing interest in extending DiD to continuous treatments."

Warren: So, for the listener, it’s about having a new way to analyze policy or exposure effects that aren't just on or off, but are spread out across a continuum <ref:2408.10509#pg1>.

Iris: It’s about giving researchers more tools to look at these complex exposures with better identification and inference than what was previously available in this area <ref:2408.10509#pg8>.

Lucas Zheng Zhang

University of California, Los Angeles · Bates White Economic Consulting

econ.EM

Submitted: 2024-08-20

Updated: 2026-11-08

Journal ref: The Econometrics Journal, 2025

DOI: 10.1093/ectj/utaf024

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

Importance score: 87/100

The gist: Continuous difference-in-differences with double/debiased machine learning extends difference-in-differences to settings with continuous treatments by proposing estimators under a conditional…

Key concepts

Difference-in-Differences (DiD)
A statistical method used to estimate the causal effect of a policy or intervention by comparing changes in outcomes between a group that received the treatment and a control group that did not. It relies on the assumption that, without the treatment, both groups would have followed similar trends over time.
Continuous Treatment Intensity (d)
This refers to situations where the treatment is not binary (yes/no) but rather a continuous variable, such as a percentage of patients receiving a new medical procedure. The goal is to estimate how the outcome changes for every possible level of this continuous treatment intensity.
Double/Debiased Machine Learning (DML)
A powerful machine learning framework used to estimate causal parameters in high-dimensional settings. It helps reduce bias by using a combination of nuisance parameter estimation and machine learning techniques, ensuring that the final causal estimate is robust and asymptotically normal.

Terminology

Summary

Continuous difference-in-differences with double/debiased machine learning extends difference-in-differences to settings with continuous treatments by proposing estimators under a conditional parallel trends assumption and establishing their asymptotic normality using the double/debiased machine learning framework <ref:2408. p. 010509#pg2>.

Identification Framework

The paper extends difference-in-differences to settings with continuous treatments by proposing estimators under a conditional parallel trends assumption and establishing their asymptotic normality using the double/debiased machine learning framework <ref:2408.10509#pg4>. The identification results are based on a conditional parallel trends assumption, allowing researchers to account for covariates non-parametrically <ref:2408. p. 110509#pg7>. The target parameter is the average treatment effect on the treated (ATT for short) at any given continuous treatment intensity d, defined as AT T(d):= E[Yt(d) − Yt(0)D = d] <ref:2408. p. 510509#pg3>.

Estimation Methodology

Estimating the ATT in this framework requires first estimating infinite-dimensional nuisance parameters, particularly the conditional density of the continuous treatment, which can introduce substantial bias <ref:2408.10509#pg4>. To address this challenge, the authors propose estimators for causal parameters under the double/debiased machine learning (DML) framework and establish their asymptotic normality <ref:2408. p. 010509#pg4>. They introduce an approximate causal parameter AT Th using a kernel function to circumvent the direct application of DML methods <ref:2408. p. 210509#pg4>.

Asymptotic Properties and Inference

The paper establishes the asymptotic normality of these estimators, showing that AT T [(d) − AT T(d) σN (d)/√N → d N(0, 1) <ref:2408.10509#pg5>. The asymptotic variance has two components, both depending inversely on the kernel bandwidth h: one arising from the kernels in the orthogonal score ψh, and the other from the linear expansion of the estimator with respect to the kernel density estimator <ref:2408.10509#pg5>. They construct a cross-fitted variance estimator and provide uniform confidence bands using a multiplier bootstrap procedure <ref:2408. p. 210509#pg5>.

Empirical Application

The effectiveness of the approach is demonstrated by revisiting a previous study on the 1983 Medicare Prospective Payment System reform, reframing it as a DiD with continuous treatment and non-parametrically estimating its effects <ref:2408.10509#pg7>. The results suggest significant heterogeneity in the impact of the PPS reform across hospitals with different shares of Medicare inpatients <ref:2408. p. 310509#pg7>.

Simulation Results

Monte Carlo simulations demonstrate that our estimators perform well despite the highly non-linear relationship between the continuous treatment and the high-dimensional covariates <ref:2408.10509#pg8>. The simulation estimates appear to follow a normal distribution in each case, which is consistent with our asymptotic theory <ref:2408. p. 210509#pg8>. The coverage of the estimated confidence intervals is close to 95 percent <ref:2408.10509#pg8>.

Empirical Example

The empirical example focuses on the Medicare Prospective Payment System (PPS) reform, where the share of Medicare inpatients can be interpreted as a continuous treatment variable <ref:2408.10509#pg7>. The analysis complements Callaway et al. p. 9 (2024), offering an alternative perspective on the effects of this policy reform <ref:2408. p. 310509#pg7>.

Conclusion

The extension of difference-in-differences models to the continuous treatment setting has important implications for empirical research <ref:2408.10509#pg8>. Our methods provide researchers with new tools for examining the impacts of continuous treatment variables <ref:2408. p. 610509#pg8>. The paper concludes that our estimators perform well despite the highly non-linear relationship between the continuous treatment and the high-dimensional covariates <ref:2408. p. 210509#pg8>.

References

Abadie, A. (2005). Semiparametric difference-in-differences estimators. p. 0 Review of Economic Studies 72, 1–19 <ref:2408. p. 010509#pg3>.

Acemoglu, D. p. 0 and Finkelstein, A. p. 0 (2008). Input and technology choices in regulated industries: evidence from the health care sector. Journal of Political Economy 116, 837–880 <ref:2408. p. 010509#pg7>.

Ananat, E., Glasner, B., Hamilton, C., and Parolin, Z. (2022). Effects of the expanded child tax credit on employment outcomes: evidence from real-world data from April to December 2021. p. 1 Technical Working Paper 29823, National Bureau of Economic Research <ref:2408. p. 010509#pg8>.

Athey, S. and Imbens, G. p. 1 p. 1 W. (2006). Identification and inference in nonlinear difference-indifferences models. p. 0 Econometrica 74, 431–497 <ref:2408. p. 010509#pg3>.

Belloni, A., Chernozhukov, V. p. 2 p. 2 p. 2 p. 2, Fern´andez-Val, I., and Hansen, C. (2017). Program evaluation and causal inference with high-dimensional data. p. 2 Econometrica 85, 233–298 <ref:2408. p. 010509#pg8>.

Bibaut, A. F. and van der Laan, M. p. 3 J. (2017). Data-adaptive smoothing for optimal-rate estimation of possibly non-regular parameters, arXiv:1706. p. 107408 <ref:2408. p. 010509#pg3>.

Callaway, B. p. 1 p. 1 p. 1 p. 1 and Sant’Anna, P. p. 1 p. 1 p. 1 H. (2018). Difference-in-differences with multiple time periods and an application on the minimum wage and employment, arXiv:1803. p. 009015v2 <ref:2408. p. 010509#pg3>.

Callaway, B. and Sant’Anna, P. H. (2021). Difference-in-differences with multiple time periods Journal of Econometrics 225, 200–230 <ref:2408. p. 010509#pg7>.

Callaway, B. (2023). Difference-in-differences for policy evaluation Handbook of Labor, Human Resources and Population Economics, 1–61 <ref:2408. p. 010509#pg7>.

Callaway, B., Goodman-Bacon, A., and Sant’Anna, P. p. 1 H. (2024). Difference-in-differences with a continuous treatment Technical Working Paper 32117, National Bureau of Economic Research <ref:2408. p. 010509#pg7>.

Cattaneo, M. D. and Jansson, M. (2021). Average density estimators: efficiency and bootstrap consistency. p. 9 Econometric Theory 38, 1140–1174 <ref:2408. p. 010509#pg8>.

Chang, N. C. (2020). Double/debiased machine learning for difference-in-differences models Econometrics Journal 23, 177–191 <ref:2408. p. 010509#pg7>.

Chernozhukov, V., Chetverikov, D. p. 0 p. 0 p. 0, and Kato, K. (2014). Gaussian approximation of suprema of empirical processes. p. 1 Annals of Statistics 42, 1564–1597 <ref:2408. p. 010509#pg8>.

Chernozhukov, V., Chetverikov, D., and Kato, K. (2014). Anti-concentration and honest, adaptive confidence bands. p. 0 Annals of Statistics 42, 1787–1818 <ref:2408. p. 010509#pg8>.

Chernozhukov, V., Chetverikov, D., Demirer, M., Duflo, E., Hansen, C., Newey, W., and Robins, J. (2018). Double/debiased machine learning for treatment and structural parameters. p. 0 Econometrics Journal 21, C1–C68 <ref:2408. p. 010509#pg8>.

Colangelo, K. p. 3 and Lee, Y. p. 3 Y. (2025). Double debiased machine learning non-parametric inference with continuous treatments Journal of Business & Economic Statistics 1–26 <ref:2408. p. 010509#pg8>.

Cook, L. D., Jones, M. E., Logan, T. D., and Ros´e, D. (2023). The evolution of access to public accommodations in the United States The Quarterly Journal of Economics 138, 37–102 <ref:2408. p. 110509#pg7>.

de Chaisemartin, C. p. 1 p. 1, D’Haultfoeuille, X. p. 1 p. 1, Pasquier, F., and Vazquez-Bare, G. (2022). Difference-indifferences estimators for treatments continuously distributed at every period arXiv:2201. p. 406898 <ref:2408. p. 010509#pg7>.

de Chaisemartin, C. and D’Haultfoeuille, X. p. 1 (2023). Two-way fixed effects and differences-indifferences with heterogeneous treatment effects a survey Econometrics Journal 26, C1–C30 <ref:2408. p. 110509#pg7>.

D’Haultfoeuille, X., Hoderlein, S., and Sasaki, Y. (2023). non-parametric difference-in-differences in repeated cross-sections with continuous treatments Journal of Econometrics 234, 664–690 <ref:2408. p. 010509#pg8>.

Fan, J., Yao, Q., and Tong, H. (1996). Estimation of conditional densities and sensitivity measures in nonlinear dynamical systems Biometrika 83, 189–206 <ref:2408. p. 310509#pg8>.

Fan, J. and Yao, Q. (2003). Nonlinear Time Series: non-parametric and Parametric Methods (Vol. p. 1 20) New York: Springer <ref:2408. p. 010509#pg8>.

Fan, Q., Hsu, Y. C., Lieli, R. P., and Zhang, Y. p. 0 (2022). Estimation of conditional average treatment effects with high-dimensional data Journal of Business and Economic Statistics 40, 313–327 <ref:2408. p. 210509#pg8>.

Galvao, A. F. and Wang, L. p. 6 (2015). Uniformly semiparametric efficient estimation of treatment effects with a continuous treatment Journal of the American Statistical Association 110, 1528–1542 <ref:2408. p. 210509#pg8>.

Haddad, M. F., Huber, M., and Zhang, L. p. 0 Z. (2024). Difference-in-differences with time-varying continuous treatments using double/debiased machine learning arXiv:2410. p. 021105 <ref:2408. p. 010509#pg8>.

H¨ardle, W. (1990). Applied non-parametric regression (No. p. 119) United Kingdom: Cambridge University Press <ref:2408. p. 010509#pg8>.

Hettinger, G. p. 3, Lee, Y., and Mitra, N. (2025). Multiply robust difference-in-differences estimation of causal effect curves for continuous exposures Biometrics 81, ujaf015 <ref:2408. p. 010509#pg8>.

Heckman, J. p. 1 p. 1 J., Ichimura, H., and Todd, P. E. (1997). Matching as an econometric evaluation estimator: Evidence from evaluating a job training programme Review of Economic Studies 64, 605–654 <ref:2408. p. 010509#pg8>.

Heckman, J., Ichimura, H., Smith, J., and Todd, P. E. (1998). Characterizing selection bias using experimental data Econometrica 66, 1017–1098 <ref:2408. p. 110509#pg8>.

Hines, O., Dukes, O., Diaz-Ordaz, K., and Vansteelandt, S. (2022). Demystifying statistical learning based on efficient influence functions The American Statistician 76, 292–304 <ref:2408. p. 810509#pg8>.

Hirano, K. and Imbens, G. W. (2004). The propensity score with continuous treatments Applied Bayesian Modeling and Causal Inference from Incomplete-Data Perspectives 226164, 73–84 <ref:2408. p. 210509#pg8>.

Hong, S. H. (2013). Measuring the effect of napster on recorded music sales: difference-in-differences estimates under compositional changes Journal of Applied Econometrics 28, 297–324 <ref:2408. p. 010509#pg8>.

Kallus, N. and Zhou, A. p. 3 (2018). Confounding-robust policy improvement Advances in Neural Information Processing Systems 31 <ref:2408.10509#pg8>.

Kennedy, E. p. 3 p. 3 H., Ma, Z., McHugh, M. D., and Small, D. S. (2017). Non-parametric methods for doubly robust estimation of continuous treatment effects Journal of the Royal Statistical Society, Series B (Statistical Methodology) 79, 1229–1245 <ref:2408. p. 310509#pg8>.

Kennedy, E. H. (2024). Semiparametric doubly robust targeted double machine learning a review Handbook of Statistical Methods for Precision Medicine, 207–236 <ref:2408. p. 310509#pg8>.

Li, Q. and Racine, J.S. (2007). non-parametric Econometrics: Theory and Practice Princeton, NJ: Princeton University Press <ref:2408.10509#pg8>.

Roth, J., Sant’Anna, P. p. 1 H., Bilinski, A., and Poe, J. (2023). What’s trending in difference-indifferences? p. 0 A synthesis of the recent econometrics literature Journal of Econometrics 235, 2218–2244 <ref:2408. p. 010509#pg8>.

Sant’Anna, P. H. and Xu, Q. (2025). Difference-in-Differences with compositional changes arXiv:2304. p. 413925v2 <ref:2408. p. 010509#pg8>.

Semenova, V. p. 3 and Chernozhukov, V. p. 2 (2021). Debiased machine learning of conditional average treatment effects and other causal functions Econometrics Journal 24, 264–289 <ref:2408. p. 010509#pg8>.

Silverman, B. W. (2018). Density Estimation for Statistics and Data Analysis Routledge <ref:2408. p. 910509#pg8>.

van der Vaart, A. p. 3W. and Wellner, J.A. (1996). Weak Convergence and Empirical Processes: With Applications to Statistics New York: Springer <ref:2408.10509#pg8>.

Zeng, H. S., Danaher, B., and Smith, M. D. (2022). Internet governance through site shutdowns the impact of shutting down two major commercial sex advertising sites Management Science 68, 8234–8248 <ref:2408. p. 110509#pg8>.

Zimmert, M. p. 4 (2020). Efficient difference-in-differences estimation with high-dimensional common trend confounding arXiv:1809. p. 201643v5 <ref:2408. p. 010509#pg8>.

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. Adaptive treatment effect estimation for continuous treatments: The system can estimate the average treatment effect on treated at any level of treatment intensity by using kernel smoothing to approximate infinite-dimensional nuisance parameters, providing a detailed curve rather than a single aggregate effect.

  2. Robust causal parameter estimation using double debiased machine learning: The system can estimate causal parameters under the double/debiased machine learning framework, which reduces bias from high-dimensional nuisance parameter estimation compared to standard methods like propensity scores.

  3. Uniform confidence band construction via multiplier bootstrap: The system can construct uniform confidence bands for treatment curves using a multiplier bootstrap procedure, offering valid inference results even in settings with continuous treatments and high-dimensional covariates.

  4. Nonparametric estimation of effects with covariate conditioning: The system can incorporate covariates nonparametrically into both the identification and estimation procedures by modifying the parallel trends assumption to condition on covariates, leading to a more robust framework than unconditional assumptions.

Sources

Related papers