Commentary on Witnessing Abusive Head Trauma: Accidents Show Higher Rates of Intracranial Pathologies Than Shaking—Caution Is Warranted

Abstract

As the topic of abusive head trauma (AHT) has become a cause célèbre in some nonpediatric academic circles, we read the research article by Brook [1] with great interest. The research article reports on an effort to compare some of the findings typically associated with AHT between cases in which shaking was witnessed or admitted to those in which accidents were confirmed. Initially, it appears to be similar to prior studies comparing witnessed accidents with witnessed/admitted abuse [2, 3] but particularly focused on shaking as a reported mechanism. The author takes a unique approach in attempting to answer the question about the rate of findings between accidental and inflicted head injuries due to shaking. The author's chosen methodology contains several flaws, which are quite puzzling. We chose to specifically highlight two fundamental flaws that we feel undermine meaningful conclusions: the study designs of the chosen data sets and the populations being compared between the categories of cases.

The first concern is the data sets that were constructed. The author used four independent sources from which he has grouped participants into three categories: unconflicted witnessed accident, witnessed shaking, and unconflicted witnessed shaking. The first data set assembled (unconflicted witnessed accident) was of participants extracted from the PediBIRN study (www.pedibirn.com) and as reported in Hymel et al. [4]. PediBIRN is a prospective, cohort study of children younger than 3 years admitted to pediatric intensive care units for traumatic head injury. Children who are injured in accidents and admitted to an intensive care unit represent the most severe accidents only. The second data set assembled (unconflicted witnessed shaking) was from the Swedish National Patient Register and reported by Thiblin et al. [5]. This is a retrospective case series of infant victims of abuse identified by ICD-10 codes. In this data set, shaking is not an explicit variable, and the author only includes those participants in which shaking was expressly reported in the medical records. Records that did not include a reference to shaking (i.e., missing data) were treated the same as those in which shaking was specifically denied (i.e., shaking = “no”), likely resulting in undercounting of shaking. The author apparently includes the same 36 participants identified by Thiblin et al. [5]. If so, it is important to note that Thiblin et al. [5] has since received an “Expression of Concern” (similar to a “black box warning”) by the editors of PLoS One, in part due to “suitability of the study design to support the conclusions” [6]. The last category assembled (witnessed shaking) was from two separate sources, both reported in Feldman et al. [7]. The ExSTRA study was a prospective, multicenter, observational study of children younger than 10 years referred for evaluation for suspected physical abuse, from which Brook sampled 17 participants with witnessed shaking. The second was “Helfer Society data,” from which the author sampled 6 participants from a haphazard sample obtained from a professional society listserv, identified by a reported history of shaking without impact. The three groups for comparison (unconflicted witnessed accident, witnessed shaking, and unconflicted witnessed shaking) were constructed from four different data sources. This is the first fatal flaw: combining participants from studies with different methodologies, that have been assembled by different criteria, from populations with different base rates, and treating them as if the same.

This study approach violates two tenets of clinical research. First, the directionality of case-control or cohort studies cannot be reversed, and second, rates cannot be compared across study designs [8]. It is important to emphasize that (a) within a particular study design, the direction of association is not reversible, and (b) rates of either exposures or outcomes are not comparable across different study designs. In this study, the “unconflicted witnessed accident” group [4] was assembled from a prospective cohort study of children in intensive care units. The “witnessed shaking” group [7] was a combination of a prospective, multicenter, observational study and a haphazard series from a professional listserv. The “unconflicted witnessed shaking” group [5] was assembled from a retrospective case series. The full data set represents a blending of prospective and retrospective, case series, case–control and cohort series, and haphazard sampling. This mélange of study designs cannot simply be placed side by side and compared. While these different study methodologies can be complementary, their data cannot be simply combined or compared.

The second concerning methodological flaw is related to the populations reported in each of the data sets. As noted, the four data sets represent different study designs (with different inclusion and exclusion criteria) and sample distinct populations. PediBIRN represented children under 3 years, admitted to pediatric intensive care units with suspected traumatic head trauma. It is important to note that accidentally injured children were “preselected” for increased severity, as the vast majority of children injured in accidents are not admitted to an intensive care unit. ExSTRA was a group of children under 10 years of age being evaluated for suspected physical abuse. Many of these children were seen in outpatient settings. Thiblin et al.'s [5] study included infants on a national register for various forms of maltreatment over 20 years, presumably across different care settings. Lastly, the Helfer Society was a nonsequential, haphazard series of undefined age or time span. The rates of either exposures or outcomes are not comparable given the clearly different populations. To illustrate the importance of representative populations, combining the four data sets as done in the current study would be akin to a study on lung cancer and combining data on adult cigarette smokers, teen nonsmokers, toddlers, and retired coal miners.

Given the concerns noted earlier, we caution the reader not to draw meaningful conclusions based on this study. While the author does mention “not uniform” data sets within the Limitations section of the manuscript, we believe that this minimizes the significant implications of different data, variables, study designs, and analyses. The author notes that his findings are consistent with Thiblin et al. [9], a study that suffers from similar methodological shortcomings [10]. The author's conclusions that intracranial findings typically associated with AHT are more common in accidental injuries as compared with reported shaking injuries are simply unsupported by the methodology utilized.

Christopher Greeley: conceptualization, methodology, writing – review and editing, writing – original draft. Jim Anderst: conceptualization, writing – original draft, methodology, writing – review and editing.

The authors declare no conflicts of interest.