Dynamic Regression Discontinuity: An Event-Study Approach

arXiv:2307.14203 · econ.EM · Submitted 2023-07-26 · 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: "Dynamic Regression Discontinuity: An Event-Study Approach".

Warren: Here's the gist of the paper. The paper proposes a novel argument for identifying economically interpretable intertemporal treatment effects in dynamic regression discontinuity designs (RDD) by developing a dynamic potential outcomes model and reformulating two key assumptions from the difference-in-differences literature—no anticipation and common trends—to achieve point identification of cutoff-specific impulse responses.

Ruth: The estimand of each target parameter is expressed as the sum of two static RDD contrasts, which permits nonparametric estimation using standard local polynomial methods.

Warren: Furthermore, a nonparametric approach is introduced to aggregate treatment effects across calendar time and treatment paths by leveraging a limited path independence restriction to reduce the dimensionality of the parameter space.

Ruth: The identification argument relies on Assumptions 1 (Continuity at the Cutoff), Assumption 2 (No Anticipation), and Assumption 3 (Common Trends).

Paper summary: Ruth: So we're looking at this paper today called "Dynamic Regression Discontinuity: An Event-Study Approach" by Francesco Ruggieri and others. What they are doing is proposing a new way to figure out those intertemporal treatment effects in dynamic regression discontinuity designs, or RDDs.

Warren: They're developing a dynamic potential outcomes model and reformulating two key assumptions from difference-in-differences literature—no anticipation and common trends—to get point identification for cutoff specific impulse responses. That sounds like they’re trying to nail down exactly what the treatment does at that precise threshold.

Iris: So, the core idea is that they take those two old assumptions and rewrite them in a way that lets you pinpoint the effect right at the cutoff, which is usually hard to do in these sequential designs. But how do you actually measure those effects when everything depends on what happened before?

Ruth: Well, they express each target parameter as a sum of two static RDD contrasts, and that lets them use standard local polynomial methods for nonparametric estimation. It's an elegant way to handle the complexity without needing a perfect functional form for everything.

Warren: And they go further by introducing a nonparametric approach to aggregate treatment effects across calendar time and treatment paths, using a limited path independence restriction to shrink the parameter space. That’s smart if you’re dealing with many time periods in your model.

Iris: So, the identification argument relies on four assumptions: continuity at the cutoff, no anticipation, common trends in some sense, and then they add two more for longer-term effects. They propose that under these conditions, the instantaneous average direct treatment effect is point identified as a standard regression discontinuity estimand.

Ruth: That instantaneous average direct treatment effect is expressed as a standard regression discontinuity estimand called RD one zero c. It’s defined as the limit of the expected outcome when you approach the cutoff from below minus the limit when you approach it from above, which they call RD one c.

Warren: Then for those long-term effects, they introduce assumption six, which is common trends. This leads to proposition six, where the cumulative average direct treatment effect is expressed as a sum of two standard sharp regression discontinuity estimands over time.

Iris: The real strength here seems to be how they generalize this to an arbitrary number of time periods using those assumptions four and five, leading to proposition five. It basically says that the instantaneous average direct treatment effect can be identified as RD one c, even when you look across different time points t.

Ruth: And for those long-term effects, proposition six gives you the cumulative average direct treatment effect as a sum of two sharp regression discontinuity estimands over time. It’s a pretty detailed way to capture the sequence of effects.

Paper summary: Warren: The estimation itself uses local polynomial methods to build bias corrected nonparametric confidence intervals around these estimates. That’s how they handle the noise that comes with estimating parameters near the cutoff.

Iris: One limitation they mention is that when you have a lot of time periods, the number of estimands grows exponentially, so they had to introduce that limited path independence assumption to reduce the dimensionality of what you're trying to estimate. That’s a necessary simplification for computation.

Ruth: But what this means practically is that even with these simplifications, you get economically interpretable results where you can actually see how the treatment shifts things over time rather than just looking at a single snapshot.

Warren: They apply this method to estimate the dynamic effects of school district expenditure authorizations on housing prices in Wisconsin. They found that referendum approval leads to a cumulative increase in housing prices of approximately twelve percent within five years.

Iris: That's an interesting finding because the paper also noted that outcomes measured in negative relative time periods remain unaffected by the approval of referenda at time t equals g. That’s a specific result about how things behave when you look backward from the event.

Ruth: It suggests that all these estimated differences, even those over time, are not statistically distinguishable from zero at the zero point five significance level. So, maybe that local common trends assumption they relied on is plausible in this context.

Warren: If you only listen to this part, it tells you that while they found a twelve percent cumulative increase in housing prices over five years following referendum approval, the statistical evidence suggests it might just be noise at the five point zero level.

Iris: So what we have here is a paper that takes a complex dynamic design, breaks it down into manageable static parts, and then puts them back together nonparametrically to get causal effects over time. It’s about making sense of those sequential decisions.

Ruth: The title "Dynamic Regression Discontinuity: An Event-Study Approach" points directly at that—it's not just a one-off effect; it’s looking at how the treatment assignment changes things across time periods.

Warren: What this means for us is that if you’re working on policy evaluations where decisions are sequential, like government spending or regulatory changes, this method gives you a framework to get results that aren't just a single point in time.

Iris: The real implication is that it makes it more feasible to test causal questions about the path of treatment effects rather than just the initial jump at the cutoff.

Ruth: We need to keep an eye on how these dynamic models interact with other methods, because if you can get this kind of identification for long-term effects, it opens up new avenues for analyzing complex public finance data.

Conclusion: Ruth: So, we've been looking at this paper called "Dynamic Regression Discontinuity: An Event-Study Approach." It’s about how to actually measure treatment effects over time in these sequential designs.

Warren: Yeah, it tackles those tricky intertemporal effects you can't just see with a single snapshot. The authors are essentially building a way to get point identification for impulse responses.

Ruth: Exactly. They take the standard assumptions from difference-in-differences and rework them so they give you specific numbers instead of just vague ideas about common trends.

Iris: From an econometric side, it’s interesting how they break down the cumulative effect into a sum of two standard regression discontinuity contrasts. That’s what makes the math tractable.

Warren: And they use local polynomial methods for that estimation, which means they get nonparametric confidence intervals around these estimates. It keeps things flexible when you don't have a perfectly smooth model.

Ruth: The big implication is that this gives us a much more robust way to evaluate policy changes that happen in stages, not just at one fixed moment.

Iris: It also addresses the problem of dimensionality when you have many time periods by introducing a path independence restriction for aggregation. That keeps the parameter space manageable.

Warren: So, it’s about making complex sequential decisions testable with real data and getting economically interpretable results instead of just abstract statistics.

Ruth: It moves us toward testing not just if a treatment works, but how its impact evolves across different time points. But we have to remember those assumptions they made about anticipation and continuity are still critical for the whole thing to hold up.

University of Chicago

econ.EM

Submitted: 2023-07-26

Updated: 2026-11-09

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

Importance score: 83/100

The gist: The paper proposes a novel argument for identifying economically interpretable intertemporal treatment effects in dynamic regression discontinuity designs (RDD) by developing a dynamic potential…

Key concepts

Dynamic Potential Outcomes Model
This model is used to estimate treatment effects over time. It helps determine how a treatment received at one point influences outcomes in subsequent time periods, moving beyond simple single-period comparisons to capture the full dynamic impact.
Point Identification of Cutoff-Specific Impulse Responses
Instead of estimating broad average effects, this method aims to pinpoint the exact effect resulting from a treatment occurring precisely at a specific cutoff point. This allows researchers to isolate and measure the immediate, localized impact of the intervention.
Nonparametric Estimation using Local Polynomial Methods
This is the statistical technique used to estimate the parameters. It allows researchers to estimate complex relationships without assuming a specific functional form for the data, which is crucial when dealing with dynamic effects over many time periods.

Terminology

Summary

The paper proposes a novel argument for identifying economically interpretable intertemporal treatment effects in dynamic regression discontinuity designs (RDD) by developing a dynamic potential outcomes model and reformulating two key assumptions from the difference-in-differences literature—no anticipation and common trends—to achieve point identification of cutoff-specific impulse responses. p. 0 The estimand of each target parameter is expressed as the sum of two static RDD contrasts, which permits nonparametric estimation using standard local polynomial methods. p. 0 Furthermore, a nonparametric approach is introduced to aggregate treatment effects across calendar time and treatment paths by leveraging a limited path independence restriction to reduce the dimensionality of the parameter space. p. 0

The identification argument relies on Assumptions 1 (Continuity at the Cutoff), Assumption 2 (No Anticipation), and Assumption 3 (Common Trends). p. 0 Under these assumptions, the instantaneous Average Direct Treatment Effect (ADTE) is point identified as a standard regression discontinuity estimand: ADTE 1,0(c) = r c E [Y 1R 1 = r] - r c E [Y 1R 1 = r] RD 1(c). p. 8 The cumulative Average Direct Treatment Effect (ADTE) is point identified as the sum of two standard sharp Regression Discontinuity estimands: ADTE 1,1(c) = r c E [Y 1R 1 = r] - r c E [Y 1R 1 = r] + r c E [Y 2 - Y 1R 1 = r, D 2 = 0] - r c E [Y 2 - Y 1R 1 = r, D 2 = 0]. p. 8

The authors generalize these results to an arbitrary number of time periods using Assumption 4 (Continuity at the Cutoff) and Assumption 5 (No Anticipation), leading to Proposition 5: ADTE 1,0(c) = r c E [Y tR 1 = r] - r c E [Y 1R 1 = r] RD 1(c). p. 8 For long-term effects, Assumption 6 (Common Trends) is introduced, leading to Proposition 6: ADTE 1, tau(c) = c E [Y tR 1 = r] - c E [Y 1R 1 = r] + c E [Y t - Y 1R 1 = r, D 2 = 0,, D t = 0] - c E [Y t - Y 1R 1 = r, D 2 = 0,, D t = 0]. p. 6 The estimation utilizes local polynomial methods to construct bias-corrected nonparametric confidence intervals. p. 3

The approach is presented as advantageous because it is applicable in settings where external information may be unavailable or fail to satisfy prerequisites for validity (e. p. 2g., observed characteristics co-determined by the treatment of interest), relies on well-understood identification assumptions with testable suggestive implications, and is easy to implement since estimators are expressed as sums of standard regression discontinuity contrasts. p. 3 The methodology is applied to estimate the dynamic effects of school district expenditure authorizations on housing prices in Wisconsin, finding that referendum approval leads to a cumulative increase in housing prices of approximately 12 percent within five years. p. 0 The results show that outcomes measured in negative relative time periods remain unaffected by the approval of referenda in period t=g, and all estimated differences are not statistically distinguishable from zero at the 0.05 significance level, suggesting the local common trends assumption might be plausible. 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. Improved AI systems can perform point identification of intertemporal treatment effects in dynamic regression discontinuity designs by developing a dynamic potential outcomes model and reformulating assumptions from difference-in-differences literature to attain point identification of cutoff-specific impulse responses.

  2. The improved system can aggregate treatment effects across calendar time and treatment paths using a limited path independence restriction to reduce the dimensionality of the parameter space, allowing for nonparametric estimation with standard local polynomial methods.

  3. The AI system can estimate dynamic effects by expressing target parameters as sums of two static RD outcome contrasts, thereby allowing for nonparametric estimation and inference with standard local polynomial methods.

Related papers