How Medical Cutoffs Can Help Researchers Identify Treatment Effects

A small difference with analytical value

Why did one patient receive an intervention while another did not? In observational research, the answer often involves differences in health that can also influence outcomes. Adjusting for recorded characteristics may leave unmeasured differences unresolved.

In their arXiv preprint, Matias D. Cattaneo and Rocío Titiunik explain how regression discontinuity, or RD, can address this problem when a known threshold changes treatment eligibility, recommendations or receipt. The manuscript also states that it was published online in the Journal of the American Medical Association on November 2, 2026. It is a methodological introduction, not a new clinical study.

The central idea is to compare patients whose measurements fall just above and below a cutoff. If those patients are otherwise comparable, an abrupt change in outcomes at the threshold can reveal the effect of the treatment rule.

What makes the comparison credible?

The measurement governing the rule is called the score or running variable. The authors describe two frameworks for interpreting comparisons around its cutoff.

In the continuity framework, the average outcomes patients would experience under either intervention are assumed not to jump at the threshold. An observed jump can then be attributed to the treatment change. Researchers commonly fit separate local linear regressions on either side and compare their estimated outcomes at the cutoff.

In the local randomization framework, falling above or below the threshold is assumed to behave like random assignment within a small surrounding window. Researchers can then use methods resembling those for randomized trials, commonly comparing average outcomes within that window.

Neither framework makes a threshold comparison automatically causal. Both depend on assumptions about comparability near the cutoff.

The rule and the treatment are different questions

A design is “sharp” when every patient receives the intervention specified by the threshold rule. It is “fuzzy” when adherence is imperfect.

For fuzzy designs, the authors distinguish three effects: how crossing the threshold changes the probability of receiving treatment; how it changes the outcome; and the effect of receiving treatment among patients whose care changes because they cross the cutoff.

That last quantity is the complier average treatment effect. Under additional assumptions, researchers estimate it by dividing the rule’s effect on the outcome by its effect on treatment receipt. It does not describe every patient who receives the intervention.

A scan-and-kidney-function example

The guide illustrates these distinctions using a 2021 study by Goulden and colleagues involving 156,028 emergency department patients tested for suspected pulmonary embolism.

Clinical guidance typically recommended contrast-enhanced computed tomographic pulmonary angiography, or CTPA, above a D-dimer level of 500 ng/mL. Some hospitals used cutoffs of 460 or 470 ng/mL. Crossing the applicable threshold substantially increased scan receipt, but did not determine it perfectly, making this a fuzzy RD design.

Across the full sample, patients above the cutoff averaged approximately 60 years of age, compared with 50 below it. Near the threshold, that age difference was much smaller, and other measured baseline characteristics were similar. This illustrates why a local comparison can be more credible than comparing everyone above and below a cutoff.

The estimated effect of crossing the threshold on long-term estimated glomerular filtration rate, or eGFR, was −0.1 mL/min/1.73 m², with a 95% confidence interval from −1.2 to 1.1. The estimated effect of receiving CTPA among compliers was −0.4, with a wider interval from −4.9 to 4.0 in the same units.

Both estimates were close to zero. Assuming the required RD and fuzzy RD assumptions hold, the authors interpret them as providing no evidence of long-term kidney harm among patients near the cutoff whose scan receipt was affected by the rule. They do not establish an absence of harm: the complier interval permits both decreases and increases in kidney function.

Where the approach can fail

The authors recommend checking whether baseline characteristics or the distribution of scores jump at the cutoff, whether conclusions survive reasonable window choices, whether apparent effects occur at artificial cutoffs, and whether excluding observations closest to the threshold changes results. Such checks can expose problems, but cannot prove validity.

Causal interpretation can fail if patients or clinicians influence scores, measurement timing or recording to obtain preferred care. Another intervention changing at the same threshold also makes attribution difficult.

Precision depends on having enough observations near the cutoff and, for fuzzy designs, a sufficiently large change in treatment receipt. Even a large overall sample may produce uncertain estimates.

Finally, RD answers a local question. Results near one threshold need not apply to substantially different patients or to people whose treatment would remain unchanged across it. Extending those conclusions requires additional strong assumptions.