Abstract
<title>Abstract</title> <p>Background. Real-world evidence increasingly informs comparative effectiveness, yet many clinical outcomes are recorded by estimation or template rather than measured. Value heaping is well described as a whole-sample phenomenon; heaping that differs by exposure group, and its capacity to create rather than blur an association, has received little attention. We show that exposure-linked recording of a continuous outcome can manufacture a large, robust-looking treatment effect, and derive a checklist to detect it. Methods. In an 8-year single-institution electronic medical record cohort in which the study drug was self-paid (an unambiguously labeled, non-random exposure), we analysed vaginal deliveries with a recorded blood-loss value (n = 3,575). We compared the recorded (visually estimated) blood-loss field, a binary postpartum-hemorrhage threshold, and an instrument-generated outcome (peri-operative hemoglobin change), applying seven pre-specified diagnostics: per-arm digit preference, within-year trajectories, threshold-exceedance, continuous–binary–objective triangulation, exposure-versus-time decomposition (year adjustment and propensity weighting), and specification stability. A simulation under a known null tested whether differential recording alone reproduces the pattern. Results. A conventional log-linear analysis showed an apparent 62% reduction (ratio 0.38, 95% CI 0.36–0.39), varying across estimators (49–75%). About 60% of treated cases were recorded as exactly 30 or 50 mL versus 2.4% of controls; the treated median was fixed at 50 mL every year from 2019 while controls remained 150–200 mL, never converging. The effect was keyed to the exposure label, not time — calendar-year adjustment and propensity weighting left it unchanged. The between-arm exceedance gap fell from 59.9 percentage points at ≥ 100 mL to 1.0 at ≥ 500 mL (not significant) and reversed at ≥ 1000 mL. The same-patient binary threshold showed no meaningful difference, and an external objective measure (cesarean hemoglobin) at most a far smaller signal; unlike the heaped field, hemoglobin change was smooth and unimodal. A null simulation reproduced the entire signature with no true effect. Conclusions. Exposure-linked differential recording of a continuous outcome is a distinct, under-recognized bias that can fabricate a large effect surviving routine adjustment, because it is differential outcome measurement error, not confounding. Simple pre-specifiable diagnostics detect it; inference should rest on objective or binary measures.</p>