Causal reasoning and data analysis: problems with the abusive head trauma diagnosis

Expert witnesses need an admissible instrument to provide an opinion and testify that a child’s head trauma was caused by abuse. Because physicians rarely observe the alleged abuse, it is unknown whether any given diagnosis of abusive head trauma (AHT) is correct. Yet AHT diagnoses are currently used to identify, interrogate, arrest and charge suspects. Maguire and co-authors developed a statistical model to make an AHT diagnosis. The model estimates the probability that a child with certain clinical features was abused. We review the Maguire model and determine that it is an inappropriate foundation upon which to base an opinion that will be used in a criminal prosecution. It suffers from numerous statistical and data problems. More importantly, it does not consider the possibility that even if a child was abused, the condition was caused not by the abuse but by an underlying (possibly unknown) medical issue or accident. No studies have suggested that, instead, a model should estimate the probability that the child’s features can be attributed to abuse. Unless attribution is incorporated into the analysis, the data are improved or revised, and the statistical issues are resolved, a diagnosis supported by such a model should be discounted.