ABSTRACT
Phenylketonuria is an inborn error of metabolism that can have devastating effects on child intelligence. Children of mothers with phenylketonuria can suffer damage in utero, even if they do not have the full phenylketonuria genetic mutation. One crucial question is the relation between prenatal phenylalanine exposure by the fetus in utero and intelligence in childhood and later. Prior research and theory predicted a threshold effect, with prenatal phenylalanine exposure up to a threshold having no effect, but exposure above the threshold having a teratogenic effect. To investigate this issue, we estimated the relation between average prenatal phenylalanine exposure and child intelligence assessed at ages 4 years and 7 years. Two analytic models were used: two-piece linear splines, and multivariate adaptive regression splines (MARS), the latter a new data-mining approach. For intelligence at 4 years of age, the two-piece linear spline model identified a threshold around 4 mg/dL of exposure, whereas the MARS model with multiple predictors identified multiple thresholds, the first around 8.5 mg/dL. For intelligence at 7 years of age, the two-piece linear spline model identified a threshold around 6.8 mg/dL of exposure, whereas the MARS model with multiple predictors identified multiple thresholds, the first around 4.5 mg/dL and the second around 9 mg/dL. Discussion centered on the trade-off between simple models and simple policy implications of the two-piece linear spline results versus the more complex representations and greater explained variance provided by MARS analyses.
