Our primary goal in this research was to evaluate the fit and performance of a ‘multiple individual risk’ model, where all ACE events are separately entered into a prediction model, in contrast to a ‘cumulative risk model’ approach for predicting adult health outcomes. This research was motivated by observation that the cumulative risk model, while a statistically powerful and parsimonious approach (5), may not be necessarily the best characterization of the impacts of childhood adversity on adult health for all outcomes because it obscures the relative contributions of individual adversity event types.
In contrast, a multiple risk model, while sacrificing information about the general impact of an accumulation of events, will yield information about the relative strength of the associations between individual event types and outcomes. The multiple individual risk model is also more sensitive in that it can allow frequency and severity of specific events to be considered in a statistical model when such information is available, while in a cumulative risk approach a threshold has to be defined for ‘exposure’. Timing, frequency and severity of adverse events are known risk factors for several adult outcomes (33).
Despite the additional information gained from application of a multiple individual risk model, it is virtually absent from the literature, despite two facts that support its use. First, there is a long history of research into the effects of specific abuse types. For example, there is substantial theoretical and empirical support for childhood sexual abuse specifically (compared to other childhood adversities ) as most strongly predictive of several outcomes including suicidality (34), cardiopulmonary symptoms, and obesity (35). The same is supported for the importance of childhood neglect (36). Importantly this earlier body of research most often did not model the co-occurrence of other individual risks. Second, it is well documented that adversity co-occurs in the lives of children and young adults. This finding has been consistent from within the ACEs framework literature (18, 37) and preceding it (38).
We found that the multiple individual risks model was a significantly better fit to the data for the lifetime history of depression outcome only. In addition to the significant difference in fit found via hypothesis testing, the MIR model accounted for 21% more variability in the outcome by R2, and an increase in model predictive performance of 17% by the c-statistic. In the case of the other two outcomes, the multiple individual risks model and the cumulative risk model (with categorical coding) were population distinguishable, but not of different fit, and inspection of the other model fit indices reveal little difference in their performance.
This is an intriguing finding that may reflect the fact that among the outcomes we analyzed, current depression may be most strongly related to biased recall for childhood events (39). Also, obesity and cardiac disease can be construed as more ‘biological’ outcomes than depression, and it may be the case that it is, in fact, an accumulation of adversity that predicts ill physical health, but that specific individual events are more strongly predictive of mental health outcomes. This possibility goes unexamined when the cumulative ACE Score is analyzed without a multiple individual risk model analyzed as well.
In the course of the model comparisons in this study, we arrived at a statistically best fit model within each category. For comparisons between models with 11 items (with the 3 sexual ACEs counted separately) and models with 9 items, we found that in all but one case an 11 items model fit better. The exception was in the case of the continuous variable treatment for the depression outcome, which we suspect may be an artifact of the need to include a quadratic term in that model. We also found that coding individuals as exposed who responded that the reported events happened ‘more than once’ was the best fit for the cardiac disease outcome only, for the other two outcomes the response of ‘ever’ happened was the best fit.
For all three outcomes the continuous score treatment (in the cumulative risk model) performed worst. Given the additional statistical and theoretical assumptions required to employ a continuous cumulative risk model, it seems an untenable approach. Overall, we conclude that utilizing the available ACE event predictors with as much information as possible by using all 11 is a reasonable approach in large sample data sources.
Taken together, we interpret these results as suggesting that investigators working with large srACEs data sources should empirically derive the number of items, as well as the exposure coding strategy, that are a best fit for the outcome under study. These analytic processes should be reported in order to improve the rigor and reproducibility of findings. From the perspective of information gained, these analytic choices can be seen not just as initial steps in data analysis, but also that their result confers additional information about the relationship between adversity and outcomes. Additionally, we suggest that unadjusted univariate associations between ACEs and outcomes (which are often reported in research publications) be supplemented with estimation of the ‘multiple individual risk’ model in studies that implement a cumulative ACE Score. This process yields additional information about ACE-health relationships. It may be premature to assume in all cases that the ACE-health effect is cumulative, rather in some cases it may be that the individual additive effects of specific events are a better predictor.