Regression Discontinuity Design is a quasi-experimental method used to estimate cause and effect by comparing data points immediately surrounding a predetermined cutoff or threshold where an intervention is applied.
Researchers, economists, and policy analysts studying causal inference read this when they need to determine the impact of an intervention in situations where random assignment of participants is not possible.
External context
This technique allows for the estimation of average treatment effects by comparing individuals who fall just above versus those who fall just below a specific cutoff point. While it provides a powerful way to model causality without randomization, users must remember that RDD cannot account for all potential confounding variables, meaning true causal inference remains theoretically impossible.
Regression discontinuity design Wikipedia contributors, “Regression discontinuity design”, en.wikipedia.orgLicence01How the Mechanism Works
The core idea behind RDD relies on a continuous variable (the 'forcing' or 'running' variable) that determines whether an intervention occurs. There must be a clear, arbitrary cutoff point—for example, scoring 75% on a quiz or having a conversion rate of $10 per month. The model estimates the average outcome for those just above this threshold and compares it to those just below it. By analyzing the local trend around the discontinuity, you isolate the effect of crossing that specific boundary. This method is powerful because it bypasses many standard assumptions required by traditional A/B testing, making it useful when randomized control trials are impossible or unethical.
Simply put, RDD helps you measure if something caused a change in performance when that change only happens after crossing a specific line or score. You look at people who scored 49 and compare them directly to people who scored 51—the difference between those two groups is your estimate of impact.
The comparison focuses only on the immediate vicinity of the cutoff point, not the entire distribution.
02Concrete Actions for Marketers
When reviewing your brand performance data this week, look for natural thresholds that dictate visibility or feature access. For instance, if Google Search Central documentation suggests a specific minimum page speed score (e.g., 60/100) is required to rank in certain results, the score itself becomes your cutoff variable. Your action should be to segment your data into groups just below and just above that critical score. You are not testing the whole spectrum; you are pinpointing the impact of crossing that specific line. This requires deep segmentation using tools capable of handling local polynomial regression.
Identify a variable in your product or search ranking data that acts as a clear pass/fail boundary.
03What to Look For When Measuring RDD
To notice the effect, you must plot the outcome variable (e.g., click-through rate or organic impressions) against your forcing variable (the score). If a significant causal effect exists due to crossing the cutoff, the graph will show a noticeable, sharp jump or drop precisely at that threshold point. This vertical gap represents the estimated treatment effect. You are looking for a discontinuity in the trend line—a break that cannot be explained by gradual changes in the surrounding data points. A smooth transition suggests the variable is not acting as a true cutoff.
The magnitude of the jump right at the threshold estimates the immediate impact of crossing the boundary.
How the record puts it
Regression discontinuity designs (RDD) are a quasi-experimental pretest–posttest design that attempts to determine the causal effects of interventions by assigning a cutoff or threshold above or below which an intervention is assigned.
04Common Mistakes to Avoid
Applying RDD requires strict adherence to statistical assumptions. Failing to meet these can lead to wildly inaccurate conclusions about your brand's performance.
- warn: Ignoring 'smearing': If the variable determining treatment (e.g., score) is correlated with other factors that also affect visibility, the results will be biased.
- warn: Using too wide a bandwidth: The analysis must focus locally. Including data points far away from the cutoff dilutes the effect and violates the core assumption of local comparison.
- check: Ensuring continuity in non-observed variables: You must assume that all factors affecting ranking or visibility change smoothly across the cutoff, except for the one factor you are testing.
05When RDD Does Not Apply
RDD is not a universal solution. It fails when the cutoff point is arbitrary or non-existent, or if there are major external factors that affect performance regardless of the score. For example, if brand visibility drops due to a global algorithm update rather than crossing a specific internal threshold, RDD cannot measure that decline because it assumes local stability around the discontinuity. Furthermore, if the variable determining treatment is not truly continuous (e.g., only integer values are possible), the method loses its statistical power and applicability.
The entry above is written by GetLoopLoop. What follows is what independent catalogues hold about the same term — none of it is the source of this page.
- Also called
- RDD, Regression discontinuity design, RDD
- Kind of thing
- scholarly method
The same term on Wikipedia
Catalogued in 10 languagesFrequently asked questions
How does using Regression Discontinuity Design differ from a standard A/B test when measuring brand visibility?
It differs because RDD is used when you cannot randomly assign users to treatment groups, but rather when an intervention naturally occurs based on a continuous score. Instead of randomizing the population, it compares outcomes for individuals falling just above and just below a known cutoff threshold, which simulates randomization locally.
When should I use RDD instead of simple correlation analysis when reviewing my brand performance data?
You should consider RDD when you suspect that the variable determining visibility (the forcing variable) is directly causing a change in outcome, and there is a clear, measurable threshold. Simple correlation only shows association; RDD attempts to establish localized causality by controlling for all factors except whether or not the cutoff was crossed.
If my 'forcing' variable—the score that determines feature access—is highly noisy, how does that affect the reliability of the results?
The noise in the forcing variable weakens the assumption of local comparability, making it harder to assume that individuals just above and below the cutoff are truly comparable. High variability can obscure a true treatment effect, requiring more robust statistical methods or clearer data segmentation.
What is the fundamental assumption required for RDD to provide reliable estimates of cause and effect?
The core assumption is that individuals immediately above the cutoff are comparable to those immediately below it, meaning there should be no systemic difference in unobserved factors between the two groups. If there are underlying factors related to the score that also influence visibility, the results will be biased.
If I fail to account for trends or time-varying confounders near the cutoff threshold, what kind of bias might my brand performance metrics show?
You risk incorporating a general trend into your estimate, leading to an overestimation or underestimation of the true local effect. Failing to model these external factors means you are not isolating the impact solely due to crossing the threshold.
Wikimedia Commons
Related visuals with source and licence credit
Asked out loud
spoken, not typedThe same term in the words somebody uses speaking to an assistant rather than typing into a box — written from the situation, which is why each one carries the situation it came from.
It depends on whether the population distribution of scores is continuous around that threshold. If you can assume that people just above 70% are comparable to those just below it, then a localized comparison is statistically appropriate for estimating cause and effect.
You can estimate this localized causal effect, but you cannot definitively 'prove' it without meeting strict statistical assumptions. The analysis compares outcomes for those just above and below the cutoff to isolate the impact of crossing that specific boundary.
Yes, it can be a very powerful technique for measuring localized impact. This method estimates cause and effect by focusing only on data points immediately surrounding the cutoff threshold, which is ideal when random testing isn't possible.