Gold-standard causal method yields spurious effects in large administrative datasets

The widespread availability of large administrative datasets and the adoption of quasi-experimental methods have fueled a "credibility revolution" in evidence-based policymaking across the sciences. These methods, such as the Regression Discontinuity Design (RDD), are prized for their potential to yield credible causal inferences. However, their validity rests on strong, often untestable, assumptions about the data-generating process. Here we show that, without rigorous falsification, these methods can produce a powerful illusion of causality. Using a sharp RDD to analyze administrative data for over three million households in a nationwide anti-poverty program, we first identify a statistically significant and seemingly robust 4.0 percentage point reduction in poverty. We then demonstrate this finding to be entirely spurious. A pre-specified falsification framework, including placebo tests at arbitrary cutoffs, reveals illusory effects of nearly identical magnitude. This failure indicates the model is systematically detecting structural artifacts in the data, likely from unobserved policy confounding, rather than a true program effect. Our findings provide a stark, empirical warning that statistical robustness is insufficient to establish causality and that aggressive falsification must be a non-negotiable standard for credible science and policy.