Research to ROI: New Method - Flexible Covariate Adjustments in Regression Discontinuity Designs.

Your team just ran a Regression Discontinuity Design (RDD) to measure the impact of your new scholarship program. The result is... inconclusive. The effect is somewhere between "highly positive" and "a complete waste of money."

What do you do? Do you kill a program that might be changing lives? Do you double down on an investment that might be worthless?

This is a high-stakes form of "Doing Data Wrong." It's not that the analysis is biased; it's that it's too noisy to be useful. The signal is lost in the static.

Let's take an example from an Education Group.

The Wrong Way (The Noisy RDD): An education foundation gives a scholarship (Treatment) to every student with a GPA of 3.5 or higher (the Cutoff). They correctly use an RDD to measure the impact on college completion rates (Outcome). But the result has a massive confidence interval. Why? Because college completion is affected by many things other than the scholarship—parental income, high school quality, choice of major, etc. (Covariates). The standard RDD struggles to isolate the scholarship's effect from all this other life "noise."

The Right Way (Sharpening the Picture with Flexible Covariate Adjustment): The standard RDD is a blurry photograph of the truth. We can see a jump at the cutoff, but it's fuzzy. What if we could use all that other student data to digitally sharpen the image?

This is the power of a new class of methods for Flexible Covariate Adjustment in RDD, laid out in cutting-edge research by economists like Noack, Olma, and Rothe.

The approach is a brilliant two-step process:

1. Predict the Noise: First, you use a flexible machine learning model (like a random forest or boosted trees) to predict the outcome (college completion) using only the rich covariate data (parental income, high school, etc.). This model learns the "predictable" part of the outcome—the noise you want to remove.

2. Analyze the "Unexplained": Second, you subtract this predicted noise from the actual outcome, creating a "residual" outcome. Then, you run your RDD analysis on this new residual outcome.

   
   

By stripping out the predictable variance, the true causal jump at the cutoff becomes much clearer and more precise. Your confidence intervals shrink dramatically. You move from an inconclusive guess to a confident, decision-grade estimate.

My career has been a bridge between the academic frontier and the pragmatic needs of the C-suite. From my research at institutions like Berkeley, HEC Paris, and FGV to my executive roles at Itaú and Ambev, I've learned that precision is the foundation of confident decision-making. A noisy estimate is a failure of imagination.

This knowledge isn't mine to keep. It's for all of us to demand more from our data.

For Managers: Don't accept "it's not statistically significant" as a final answer. Ask your team: "Have we used all available covariate information to reduce the variance and increase the precision of this estimate?"

For Data Professionals: Your job isn't just to run the RDD. It's to deliver the most precise estimate the data will allow. This means embracing two-stage estimation. Master the art of using ML for covariate adjustment before you analyze the discontinuity.

Let's stop letting noise dictate our strategy.

If you’re ready to build sharper, more precise causal insights, join my movement. Subscribe to my email list for more no-nonsense, research-backed playbooks.

And if you're sitting on a noisy RDD result right now, book a 20-minute, no-nonsense consultation with me. Let's sharpen the picture together.

#RDD #CausalInference #EducationAnalytics #MachineLearning #CovariateAdjustment #PrecisionMatters #DataDrivenEd #MariaAliceMaia