A Guide to Regression Discontinuity Design in 2026
Regression Discontinuity Design (RDD) is a quasi-experimental method used to estimate the causal effect of an intervention by leveraging a sharp cutoff point. It compares individuals just above and below a specific threshold, creating a localized natural experiment to isolate the treatment's true impact, free from confounding variables like pre-existing ability. For analysts, it’s a powerful tool for causal inference when a randomized trial isn't feasible.
How do you measure the impact of a program when the people in it were already set up to succeed? It's a classic challenge, and this is where PlotStudio can make a difference. Our agentic analytics platform is built for the individual analyst and researcher, not as an enterprise BI dashboard bot. You upload a dataset and our AI data analyst autonomously plans the analysis, writes and runs real Python locally on your machine, and produces a saved, reproducible "Analysis Page" with narrative, charts, and auditable code.
Finding the Natural Experiment at the Cutoff
RDD's genius lies in focusing on the people right at the edge of the cutoff. Imagine a scholarship offered only to students who score 80% or higher on a test. A simple comparison between all recipients and non-recipients is useless—the kids who got the scholarship were already the top performers. But what about the students who scored 79% versus those who scored 81%?
What's the real difference between them? A lucky guess? A single misunderstood question? On average, these two groups are nearly identical in motivation, study habits, and background knowledge. The only meaningful difference is that one group got the scholarship, and the other didn't.
This is the core idea: that arbitrary cutoff creates a small, localized randomized experiment for us.
By comparing the future success (like college graduation rates or income) of the 81% group to the 79% group, we can isolate the true causal effect of the scholarship itself, stripping away the bias from pre-existing ability.
From Theory to Practice
This isn't just an academic exercise. Regression Discontinuity Design is a workhorse for policy evaluation and causal inference analysis in the real world. It's used to measure the effects of:
- Financial aid programs based on strict income thresholds.
- Remedial education classes assigned to students just below a proficiency score.
- Political outcomes by analyzing elections won by a razor-thin margin.
This is the kind of rigorous analysis that separates a simple summary from a defensible insight. It's also the kind of complex methodology that can be tricky to implement correctly.
This is where a tool like PlotStudio shines. Because it's built like a researcher, not a chatbot, its agentic analytics framework can autonomously plan and execute the entire workflow. It investigates across multiple steps, from checking RDD's strict assumptions to visualizing the discontinuity and narrating the findings. Instead of a one-shot answer, it produces a fully documented, reproducible Analysis Page with the narrative, charts, and underlying code, all on your local machine.
Table of Contents
- Finding the Natural Experiment at the Cutoff
- From Theory to Practice
- What is Regression Discontinuity Design? Understanding the Logic and Assumptions
- How to Run a Regression Discontinuity Analysis
- Validating Your RDD Results with Robustness Checks
- Common Pitfalls and Advanced RDD Extensions
- Agentic Analytics vs. Chatbots for Regression Discontinuity Design
- Frequently Asked Questions (FAQ)
What is Regression Discontinuity Design? Understanding the Logic and Assumptions
At its heart, a Regression Discontinuity Design (RDD) is a clever workaround. It’s what you use when you can’t run a real randomized trial but still need a credible, causal answer about a program's impact.
The entire method hinges on a single, powerful idea: comparing people who are just barely on either side of a cutoff. Think of a scholarship awarded to students with a GPA of 3.5 or higher. The intuition is that a student with a 3.49 GPA is, for all practical purposes, identical to a student with a 3.51 GPA. They have similar study habits, similar backgrounds, and similar motivation.
The only meaningful difference between them? One got the scholarship, and the other didn't. By focusing intensely on this narrow band around the cutoff, RDD creates a "local" randomized experiment.
But this clever setup only works if a few key assumptions hold true. If they don't, the whole analysis falls apart.
The Continuity Assumption
This is the most important rule of RDD. The continuity assumption states that if the program didn't exist, the relationship between the score (your "running variable") and the outcome you care about would be a smooth, continuous line.
Imagine plotting GPA against future income. Without any scholarship, you'd expect a smooth, upward-sloping trend. No sudden jumps. A student with a 3.51 GPA shouldn't magically earn way more than a student with a 3.49 just because they crossed some arbitrary line.
Any sharp, discontinuous "jump" that appears right at the 3.5 GPA cutoff must be the effect of the treatment itself—the scholarship.
Key Takeaway: The discontinuity or "jump" at the cutoff is the treatment effect. If the outcome would have jumped anyway for some other reason, your RDD is invalid.
This is the bedrock of a valid Regression Discontinuity Design. You can explore more about this family of methods in our complete guide to causal inference analysis.
The No Manipulation Assumption
Here's the second critical rule: people can’t be able to perfectly game the system. If individuals can precisely manipulate their score to land just on the "winning" side of the cutoff, your quasi-experiment is broken.
For example, if students who scored a 3.48 can easily convince a professor to bump them up to a 3.50, the group just above the cutoff is no longer a random-like sample. It's now filled with the most persuasive, resourceful, or well-connected students. They are systematically different from the group just below the line, and not just because of the scholarship.
Analysts have to test for this. You look for suspicious "bunching" in the data right at the cutoff—a telltale sign that the score was manipulated.
Sharp vs. Fuzzy RDD
Finally, you need to know whether your RDD is "sharp" or "fuzzy." The distinction is all about whether the cutoff rule is applied with perfect, absolute strictness. This choice directly impacts how you'll estimate the effect.
Here’s a quick breakdown of the two main types of Regression Discontinuity Design.
Sharp RDD vs. Fuzzy RDD Comparison
| Characteristic | Sharp RDD | Fuzzy RDD |
|---|---|---|
| Treatment Assignment | Deterministic. Everyone above the cutoff gets the treatment; everyone below does not. No exceptions. | Probabilistic. Crossing the cutoff increases the probability of treatment but doesn't guarantee it. |
| Example | A law providing benefits to anyone turning 65. Age is precise, and the assignment is certain. | A program recommending tutoring for students below a certain test score, but some eligible students opt-out and some ineligible students manage to opt-in. |
| Estimation | Simpler. A local regression comparing outcomes just above and below the cutoff is usually enough. | More complex. Requires an instrumental variable (IV) approach to account for the "fuzzy" non-compliance on both sides. |
In a Sharp RDD, the cutoff is a steel door—you're either in or you're out. In a Fuzzy RDD, it's more like a strong recommendation—most people follow it, but not all.
Understanding these foundational rules—continuity, no manipulation, and the sharp/fuzzy distinction—is the first step toward conducting a sound Regression Discontinuity Design and producing a result you can actually defend.
How to Run a Regression Discontinuity Analysis
Running a Regression Discontinuity Analysis isn't just about plugging numbers into a model. It's a clear, sequential process that moves from visual intuition to statistical estimation and, finally, to rigorous validation.
The first thing you should always do is plot your data. Seriously. Before you run a single regression.
A simple scatter plot of your outcome variable against the running variable is the best diagnostic you have. This one chart should immediately show you whether a "jump" or discontinuity actually exists at the cutoff. If you can't see a clear break with your own eyes, your hypothesis might be dead on arrival.
Choosing Your Estimation Strategy
Once you have some visual evidence, it's time for formal estimation. It can be tempting to just fit a high-order polynomial across the entire dataset to capture the trend, but modern practice screams against this. Global polynomials are notoriously badly behaved, especially at the edges of the data—which is precisely where your cutoff is.
The preferred method today is local linear regression. This approach is much smarter. It fits separate, simple regression lines on each side of the cutoff, but it only uses data within a specific window, or bandwidth. This focuses the analysis only on the observations that are truly comparable, giving you a much cleaner estimate of the jump right at the threshold.
The infographic below breaks down the core logic that makes this whole process work.
This visual roadmap shows how the cutoff creates that "natural experiment," which hangs entirely on the continuity and no-manipulation assumptions to give you a causal estimate you can trust.
Selecting the Right Bandwidth
Here’s where the art meets the science. The single most critical decision you'll make in a local linear RDD is bandwidth selection.
- A narrow bandwidth uses data very close to the cutoff. This is great for reducing bias because the individuals on either side are almost identical. The downside? You have fewer data points, which increases the variance of your estimate.
- A wide bandwidth uses more data, which shrinks your variance. But now you risk introducing bias, as you start including people who are less and less comparable to those right at the cutoff.
Standard statistical packages have algorithms to calculate an "optimal" bandwidth, but this is a starting point, not a final answer. A good analyst always tests how sensitive the results are to different bandwidth choices.
Key Takeaway: The treatment effect you estimate is only valid for the population right at the cutoff. Your choice of bandwidth defines how "local" that estimate is. A solid analysis requires checking how the results change as you vary that window.
The power of RDD has been proven in major policy evaluations. Researchers used it to show a California health insurance mandate led to an 18.3% increase in coverage at a specific income cutoff. The method was also central to evaluating the impact of the Affordable Care Act's Medicaid expansion.
Of course, running the regression is just one part of the job; correctly interpreting the output is another beast entirely. If you want to go deeper, you can review our guide on how to interpret regression results. For an RDD analysis, this means focusing squarely on the coefficient for the treatment indicator right at the cutoff.
This entire workflow—visualization, local linear regression with optimal bandwidth, and narrative interpretation—is exactly what an agentic analytics tool like PlotStudio automates. Its domain-aware logic can apply field-specific methods like this, delivering a reproducible Analysis Page without any manual coding.
Validating Your RDD Results with Robustness Checks
So you found a statistically significant jump at the cutoff. That's a great start, but it's not the end of the story—not by a long shot. A credible RDD analysis depends on a series of rigorous checks to prove that your jump is a real treatment effect, not a ghost in the machine or a flaw in your data.
Think of it as defending your findings from the toughest possible critic: yourself.
The most critical threat to any RDD is manipulation. If people can game the system to get just over the cutoff line, your whole assumption of "as-if-random" assignment goes out the window. Your treated and control groups are no longer comparable.
Your first job is to check for this.
Testing for Data Manipulation and Placebo Effects
The most intuitive way to spot manipulation is to just look. Plot a histogram or a density chart of your running variable, zoomed in on the cutoff. Do you see a suspicious, unnatural pile-up of observations just on one side of the line? That’s a massive red flag.
This visual check is often formalized with a McCrary Test, which is designed specifically to detect a discontinuity in the density of the running variable itself. A non-significant p-value here gives you some confidence that people weren't strategically sorting themselves around the threshold.
Next up are placebo tests. This is where you try to find an effect where you know there shouldn't be one. If you find one, it means your model is probably broken.
Placebo Cutoffs: If your real cutoff is a test score of 80, run the exact same analysis with a fake cutoff at 70. Then try another one at 90. You should find nothing. If you find a "significant" jump at these arbitrary points, it suggests there are other things going on in your data that have nothing to do with the treatment.
Placebo Outcomes: You can also run your RDD on outcomes the treatment couldn't have possibly affected. For instance, did pre-treatment demographics like age or gender also "jump" at the cutoff? They absolutely should not have. Finding a jump in a placebo outcome is a sign of a deep problem in your research design. You can learn more about related validation methods in our article on nonparametric tests.
Finally, you need to check how sensitive your results are to the bandwidth you chose. Does your effect hold up if you use a slightly narrower or wider window around the cutoff?
If your amazing result only appears at one, hyper-specific bandwidth and vanishes everywhere else, it’s probably not a real finding. A robust effect should be stable.
This whole suite of tests—manipulation checks, placebos, and bandwidth sensitivity—is what separates a flimsy analysis from a defensible one. It’s what an agentic analytics platform like PlotStudio is built to do automatically. Because it’s designed to think like a researcher, not just run code, it autonomously performs these diagnostics as part of its analytical plan. The result isn't just an answer; it's a piece of auditable, reproducible intelligence you can actually stand behind.
Common Pitfalls and Advanced RDD Extensions
Even with all its statistical power, Regression Discontinuity Design is surprisingly easy to get wrong. Several common pitfalls can turn a sharp analysis into a misleading artifact, and no automated tool can save you if your core judgment is off.
The most frequent mistake is fitting global high-order polynomials across the entire dataset. These models are just too flexible. They will bend and twist to fit the noise, creating the illusion of a discontinuity where none actually exists. It's a classic case of overfitting.
This is often tied to the second major error: improper bandwidth selection. We’ve already touched on this, but it’s impossible to overstate its importance. A window that’s too wide pulls in observations that are fundamentally different from the units near the cutoff, biasing your results. Get the bandwidth wrong, and the whole analysis is compromised.
But the most subtle and dangerous error is misinterpreting what you've found. RDD doesn't give you a universal truth; it gives you a Local Average Treatment Effect (LATE).
Key Takeaway: The effect you measure only applies to the specific group of individuals right at the cutoff. It is a profound error to generalize that finding and claim the program would have the same impact on people far from the threshold.
Navigating Complexity with Advanced Extensions
Once you master the basics, you'll inevitably run into scenarios where a standard RDD doesn't quite fit. The good news is the methodology has a whole family of extensions designed for these more complex, real-world problems.
- RDD with Multiple Cutoffs: Some programs don't have just one eligibility threshold, but several. This design lets you estimate treatment effects at each point, giving you a way to test if the effect is consistent across different groups.
- Regression Kink Design (RKD): Sometimes the policy doesn't create a jump in the outcome, but a change in its slope. RKD is built for this. It's perfect for analyzing policies that alter incentives, like a tax rate that suddenly gets steeper after a certain income level.
- Geographic RDD: This extension simply applies the RDD logic to spatial data. Think of comparing outcomes for households living just on either side of a school district line or a state border.
Despite being developed back in the 1960s, RDD remains a sharp tool that is often left in the box. A 2022 review, for instance, found a surprisingly low number of RDD studies in crime and justice policy, highlighting a huge gap between its potential and its actual use. You can read the full research on RDD's adoption in policy fields.
These advanced methods demand careful implementation. Even with an agentic analytics tool like PlotStudio automating the complex code, the analyst's role is irreplaceable. Its "Plan Mode" allows users to review and edit the proposed analysis before execution, ensuring human expertise guides the AI. Choosing the right framework and correctly interpreting its highly localized results is where the human expert earns their keep.
Agentic Analytics vs. Chatbots for Regression Discontinuity Design
Anyone who has run a full Regression Discontinuity Design by hand knows the drill. It's a long, painstaking process that demands serious expertise and a lot of careful coding. You have to visualize the data, pick an estimation strategy like local linear regression, hunt for the optimal bandwidth, and then slog through a whole suite of robustness checks—from McCrary tests to placebo cutoffs. Each step is a judgment call.
This is where the contrast between agentic analytics and "chat-with-your-data" tools becomes stark.
Chatbots or copilots translate one question into one query. They give you an answer—a data point—but not a complete, reproducible analysis. Asking one to run an RDD would be a frustrating exercise in single-step commands that miss the bigger methodological picture.
An agentic analytics platform like PlotStudio operates differently. It’s not about answering single questions but about conducting a complete, multi-step investigation. You upload your dataset, and the AI analyst plans and executes the entire sequence on its own, mimicking the workflow a human researcher would follow. The system runs that plan using a local Python engine, so your data never leaves your machine, ensuring privacy. Of course, automating complex analyses brings up important questions about oversight. For anyone looking deeper into this, a 2026 AI agent governance roadmap offers a solid framework.
The result isn't just a single number or a chart. It's a complete, reproducible Analysis Page that compounds as persistent knowledge. This page weaves the entire RDD workflow into a single, coherent story, complete with the discontinuity plot, the statistical outputs from all the robustness checks, and every line of the auditable Python code used to generate them.
As an independent review from The Effortless Academic noted, PlotStudio is a purpose-built tool for researchers, not a simple chatbot, and is "superior to chat tools like ChatGPT/Claude/Julius AI for dedicated data work" precisely because it automates the full analytical workflow.
This automated approach doesn't make the analyst's judgment obsolete—it supercharges it. By taking care of the mechanical, repetitive work of coding and validation, it frees you up to focus on what actually matters: interpreting the results and understanding the real causal story behind the numbers.
Frequently Asked Questions (FAQ)
What's the difference between Regression Discontinuity Design and an RCT?
An RCT (Randomized Controlled Trial) is the gold standard for causal inference, assigning participants to treatment or control groups completely at random. A Regression Discontinuity Design is used when randomization isn't possible but assignment is determined by a strict cutoff rule (e.g., a test score or income level). RDD cleverly mimics the randomness of an RCT, but only for the population clustered right around that specific cutoff, creating a "local" randomized experiment.
How big of a sample size do I need for a Regression Discontinuity Design?
You're going to need a lot of data. Seriously. RDD is a data-hungry method because its entire logic hinges on having a high density of observations packed tightly around the cutoff. This is necessary to get enough statistical power to detect a significant jump. There's no single magic number, but an RDD often requires a dataset far larger than a comparable experiment would to yield a credible, precise estimate.
Can I use RDD if the cutoff rule isn't followed perfectly?
Yes. This is a common real-world scenario, and it's handled by a Fuzzy RDD. In a Sharp RDD, the rule is absolute. In a Fuzzy RDD, crossing the cutoff only increases the probability of receiving treatment; it doesn't guarantee it. For example, some eligible students might not sign up for a program, or some ineligible students might find a way in. The analysis is more complex, typically requiring an instrumental variable (IV) approach, but it's a well-established method for getting a valid estimate when assignment isn't perfectly enforced.
If you're a researcher looking to implement rigorous methods like Regression Discontinuity Design without the manual coding overhead, we can help. 1,000 free credits for researchers are available to get you started on your next project.