A rigorous reanalysis of ten NHANES cycles (1999–2018) covering 16,035 adults—representing 207.7 million Americans—finds that three urinary phthalate metabolites genuinely associate with obesity: MBzP (OR 1.098), MEHP (OR 0.857), and MiNP (OR 0.823). But the true predictive lift is modest: the entire phthalate exposure block adds only ΔAUC +0.016 above a demographics-only model. Three analytic artifacts routinely amplify this signal into something unrecognizable. The most striking is proxy-mediated leakage—imputing missing exposure values using waist circumference (ρ = 0.948 with BMI) while excluding BMI itself creates imputed exposure scores that correlate with obesity at |ρ| > 0.86, versus a true measured correlation below 0.15. Adding tautological body-composition predictors inflates ΔAUC by an additional +0.345. An audit of 210 published NHANES machine-learning obesity studies found that none reported any leakage check, only 2.9% described imputation procedures, and 14.3% reported the survey design.
This preprint, not yet peer-reviewed, is methodologically paradigm-shifting rather than merely incremental. It demonstrates that a structural feature of NHANES—measuring phthalates in only one-third of participants—creates conditions where analytic choices can manufacture associations from noise. For adults concerned about endocrine-disrupting chemicals, the take-away is nuanced: phthalate-obesity links are real but small, and the alarming effect sizes dominating headlines likely reflect pipeline contamination, not biology. Independent replication with pre-registered leakage controls is urgently needed before policy or behavioral guidance changes.