DOI: https://doi.org/10.21203/rs.3.rs-32902/v1
A majority of the documented relationships between adverse childhood experiences (ACEs) and adult health outcomes are based in cross-sectional self-reported datasources such as the CDC’s Behavioral Risk Factor Surveillance System survey. Despite using the same or similar questions, there is a great deal of inconsistency in the operationalization of self-reported childhood adversity for use as a predictor variable. A cumulative risk model, resulting in a cumulative numeric ‘ACE Score’, is normative but there are differences in the way the ACE Score is calculated and used in statistical models. Alternative conceptual models are rarely used, even though predictor characterization directly impacts interpretations about the effects of adversity on outcomes. We investigated differences in model fit and performance between the cumulative ACE Score and a ‘multiple individual risk’ model that enters individual ACE events individually into prediction models.
We merged the 2011–2012 BRFSS data (N = 56,640) and analyzed 3 self-report outcomes. We compared model fit metrics and used Vuong’s test for model selection to arrive at a best fit model.
The multiple individual risk model was a better fit than the categorical ACE Score for the ‘lifetime history of depression’ outcome. For the outcomes of obesity and cardiac disease, the cumulative risk and multiple individual risks models were of comparable fit, but yield very different inferences.
Additional information-rich inferences about ACE-health relationships can be obtained from including a multiple individual risk modeling strategy, either in addition to or in place of a cumulative risk ACE Score.
A very large body of research documents relationships between Adverse Childhood Experiences (ACEs) and adult health outcomes. Much of the data in which these inferences are based comes from cross-sectional surveys containing adults’ retrospective self-reports of their ACEs and concurrent reports of their health status. We refer to this type of design and data structure as the ‘ACEs Framework’ (1) and to questionnaire responses over a specific set of adversity events contained in these datasets as srACEs (self-reported ACEs). While this tradition arguably began with the landmark 1998 Felitti et al Kaiser ACE’s Study (2), versions of the Kaiser group srACE questions are now used in several other large-scale health surveys, including the World Health Organization’s Adverse Childhood Experiences International Questionnaire (ACE-IQ) assessment (3) and the CDC’s Behavioral Risk Factor Surveillance System (BRFSS) survey (4). Between 2009 and 2019, the BRFSS srACEs were administered as both rotating core questions and as optional modules in 42 states, corresponding to several million cases of srACEs in the BRFSS datafiles, and over 200 BRFSS-based ACEs studies in the empirical literature.
Remarkably, these studies show a substantial degree of inconsistency in the operationalization of the srACEs as a predictor variable. While a ‘cumulative risk’ conceptual model guides most research, resulting in the operationalization of childhood adversity using the cumulative numeric ‘ACE Score’, there are differences in the way this cumulative score is calculated and used in statistical models. Additionally, alternative conceptual models to the cumulative risk model, which can yield different conclusions about the effects of adversity on outcomes, are rarely used and even more rarely directly contrasted to each other. Differences in operationalization may impede efforts to synthesize the literature and differences in conceptual models of how adversity impacts outcomes have high stakes as policy and intervention programming depend of this body of literature.
In this paper we analyze two conceptual models of the effects of ACEs on adult health, contrasting the cumulative risk model (using a cumulative ACE Score) with a ‘multiple individual risk’ model that allows for each ACE event to have its own relationship with the outcome (in a multivariable model that includes all the ACE event predictors). In order to undertake this comparison, we also analyze differences arising from the practical decisions that have to be made about which specific ACE questions to include from among those available and how to code individuals as ‘exposed’ when response options include information about the frequency of events.
In the ACEs framework literature, the dominant model of the effects of developmental adversity on later health is the cumulative risk model (5). This model holds that it is not so much specific events which are detrimental to health, but rather that it is an accumulation of events (regardless of which specific events they are) that confers risk for negative health effects. The cumulative risk model is a specific type ‘multiple risk’ model (5) where exposure to multiple risks are included in the same statistical model. In the ACEs literature this is widely implemented through the use of a cumulative numeric score (the ‘ACE Score’) that indicates the total number of exposures. This model answers the question ‘what is the impact of increasing numbers of events?’. There are multiple ways to implement this model, however. Here we contrast a continuous with a categorical variable approach.
A cumulative ACE Score can be treated as an integer count variable (i.e. a continuous variable) in statistical models. This treatment implies that the ACE Score has a linear effect on the outcome (that each additional ACE has an equal impact) and the coefficient that results applies for every score interval of the ACE score predictor. As an example in the BRFSS literature, Nurius, Logan-Green and Green used a total ACE Score (0–8) and reported significant coefficients of − .19 (healthy days) and .23 (mental health symptoms) (6), implying a constant decrease of .19 healthy days and constant increase of .23 mental health symptoms for each additional ACE reported. An ACE Score characterized in this way serves as the primary illustration of a dose-effect relationship: one in which the dose-effect relationship is constant across levels of the ACE score.
The most common alternative characterization for the ACE Score is to use it as a categorical variable in prediction models; this approach allows each specific ACE count to have its own relationship with the outcome. In this treatment, the ACE Score-outcome relationship is not presumed to be linear, and instead the counts are entered into prediction equations as categories, producing a separate coefficient for each, compared to a reference category (usually zero). The most common approach is categories of 0, 1, 2, 3, and ≥ 4, (7, 8) although a top category of ≥ 5 is also common (9–11). Other researchers have combined counts into other categories such as 0, 1–3, 4–6, 7–9 (12), or used a wider range of the variable (1–8 compared to zero) as a categorical predictor (13, 14). Compared to a continuous ACE Score approach, this model is more flexible and yields a different inference about the dose-response relationship: that risk for the outcome increases monotonically, but not at the same rate for every additional ACE. This was the approach used in the first ACEs study publication (2), where a total of 17 individual questions were aggregated into 7 categories of events and a truncated categorical variable (0, 1, 2, 3, ≥ 4) was used in statistical modeling. Those results showed increasing odds (compared to those in a zero ACEs reference category) of 1.1–2.2 for current smoking; 1.5–4.6 for two or more weeks of depressed mood in the previous year; and 1.1–1.6 for ‘severe obesity’ (2). Results arrived at using this treatment of the ACE predictor are not only exceedingly common in the literature, but have been translated into public policy (15) as well as public-facing internet sources that refer to the risks associated with ‘4 or more ACEs’ (16, 17).
In contrast to the cumulative risk model, the presence or absence of multiple separate ACE events can be included together as separate predictors in a single regression; we refer to this as a ‘multiple individual risk’ model. This model is not the same as analyzing univariate (unadjusted) associations between single ACEs and outcomes, which is a common feature in studies that use a cumulative risk model (ACE score) in their primary analysis. Instead, this model answers the question ‘what is the impact of the occurrence of each specific event (given the presence/absence of the other events)’?. The model yields coefficients for each event separately, controlling for the other events in the model. It is therefore expected to function well for events which are highly correlated, as has been extensively supported for ACEs (18). Despite the high degree of information contained in this type of model, it appears only rarely in the ACEs framework literature. Our review found only one instance in the BRFSS data, in a study examining individual and cumulative effects of ACEs on adult mental health (19). In that study, only the specific ACEs that had a significant univariate relationship with the outcome were included in the ‘multivariate’ models, and they found that different sets of ACE events had significant associations with the mental health outcomes under study (20).
We presume that the absence of this type of model in the literature is due to the fact that, while the model itself is additive with respect to the joint effects of the events on the outcome, this model does not contain general summary information about cumulative effects. That is, when effect estimates for specific individual ACEs are estimated in the same model, effect estimates for a specific number of events are generally not estimated. Functionally then, models with individual events estimated in the same model do not yield specific information about a dose-response relationship.
Choosing a predictor characterization directly impacts interpretations about the effects of adversity on outcomes. In terms of the model comparisons in this study, the cumulative risk model and the multiple individual risk model yield very different inferences. Only in the first case would we be able to infer that an accumulation of adverse events (regardless of which specific events) is related to outcomes in a dose-response manner. However, only in the second case are we able to infer that one or some specific ACE events are a strong predictor, compared to other ACE event types.
Similarly, treating the ACE score as continuous vs categorical in the CR framework also has implications for interpretation. Finding that each additional ACE event contributes in a constant linear way to risk for negative outcomes (e.g. Dube et al. 2001; Nurius 2012) is a very different conclusion than finding that change in risk for an outcome is smallest across intervals at the low end of ACE Scores, but that increases in risk for the outcome are accelerated across intervals at the higher end of the score range (e.g. Anda et al. 1999).
However, even though these modeling choices can result in very different conclusions, there is only very limited evidence that directly contrasts them. Some exceptions include a comparison of a latent class predictor characterization (LCA) to the cumulative risk ACE Score in predicting outcomes in college students (24) which found that LCA performed similarly to the cumulative ACE Score. In contrast, Schilling et al found that a cumulative risk approach produced different predictions than treating the same data with a cluster analysis approach (25).
The cumulative risk model is a straightforward and easy to understand explanatory model that has helped to publicize the negative health effects of adversity, but it has both statistical and theoretical shortcomings (5, 26). In this study, we investigated differences in model fit and performance based on operationalization of an ACE predictor variable in a cumulative risk model (with ACE score as continuous or categorical) vs a multiple individual risk model when applied to three commonly studied health outcomes.
We merged data from the 2011 and 2012 publicly-available Behavioral Risk Factor Surveillance System (BRFSS) cross-sectional, random-digit-dial telephone surveys conducted by health departments in all 50 US states in collaboration with the Centers for Disease Control (27). Respondents are English and Spanish speaking adults aged 18 years or older, who are non-institutionalized, and live in a household with a working landline telephone or included cell phone. Only the states that administered the ACEs module in each year were included (16 states total).
The BRFSS survey contains 11 srACEs, prefaced with “Before the age of 18”… :
Did you live with anyone who was depressed, mentally ill, or suicidal?
Did you live with anyone who was a problem drinker or alcoholic?
Did you live with anyone who used illegal street drugs or who abused prescription medications?
Did you live with anyone who served time or was sentenced to serve time in a prison, jail, or other correctional facility?
Were your parents separated or divorced?
How often did your parents or adults in your home ever slap, hit, kick, punch or beat each other up?
Before age 18, how often did a parent or adult in your home ever hit, beat, kick, or physically hurt you in any way? Do not include spanking.
How often did a parent or adult in your home ever swear at you, insult you, or put you down?
How often did anyone at least 5 years older than you or an adult, ever touch you sexually?
How often did anyone at least 5 years older than you or an adult, try to make you touch them sexually?
How often did anyone at least 5 years older than you or an adult, force you to have sex?
For items 1–5, response options are ‘yes’ and ‘no’. For questions 6–11 response options are ‘never’, ‘once’, and ‘more than once’. Although it is possible to code ACE predictors that incorporate the frequency information, it is uncommon in the literature. Instead, investigators routinely define a cut-off to determine an exposure. In most cases responses of ‘once’ are sufficient, but in some cases ‘more than once’ is used (28). In many published studies in the BRFSS data this decision is not noted (7, 11, 29). In order to ensure that the model comparisons between the cumulative risk (CR) and multiple individual risk (MIR) models were fair, we first arrived at the best fitting model within each category, by creating different versions of the ACE predictor based on the permutations possible for exposure coding (‘once’, vs ‘more than once’) and number of items (9 questions vs 11) and iteratively arrived at the best-fit model for each outcome separately, through within-category pair-wise comparisons.
We selected three dichotomous outcomes: lifetime history of depression: (Ever told) you that you have a depressive disorder, including depression, major depression, dysthymia, or minor depression?, obesity status (calculated in the data from self-reported height and weight) and presence of cardiac disease (coded in the data by any affirmative response to ‘having had a heart attack’ or ‘having had angina’).
Sex, age (5 categories), race (Black, White, and Other), marital status (3 categories), education (4 categories), income group (5 categories), and insurance status (has insurance vs not) were included as covariates in all models.
We included 56,640 cases with no missing data on any covariates or outcomes, and with no more than 1 missing ACE (82.1% of cases in the merged 2011–2012 data). Cases missing one ACE were imputed as ‘no’ (< 1% of cases). This analytic decision was made in order to ensure that model comparisons were made between models fitted in the same data. Data were screened to ensure that at least 20 cases were present in the cross-tabs of the ACE scores and the outcomes as well as the covariates.
Because of the substantial reduction in the total number of cases, we omitted the survey design variables from our modeling, as survey weights are calculated based in the full dataset. The dichotomous outcomes were modeled with logistic regression. Data were analyzed in R using the R Studio® IDE,(30) and the package ‘nonnest2’(31) for model comparisons.
The model comparisons of interest in this study are between the cumulative risk model (with the ACE Score used as either a continuous variable (CrCn) or a categorical variable (CrCat)), and the multiple individual risk (MIR) model. In order to make fair comparisons between those models, we first arrived at the best fitting model within each category, as described above. Within the CRCn models we allowed for non-linearity by estimating a model that included a quadratic term. Within the ‘multiple individual’ risk models, variance inflation factors were obtained for all predictors to assess multi-collinearity.
After the best-fit model was obtained for the MIR, CRCn, and CRCat model categories separately, we estimated a ‘covariate-only’ baseline model for each outcome. Then, the best fit models within each category were compared to the baseline model, and to each other.
We used the two-step approach introduced by Vuong (32) for hypothesis testing of differences in non-nested (or partially non-nested) models. Nested model comparisons using the likelihood ratio test are common, for example in instances of comparing two regression models where the second contains all the predictors from the first, except one. Non-nested models are defined as pairs (or sets) of models where one model cannot be obtained by introducing a restriction or constraint on the other model. Because all models compared here include the same set of covariates, but different characterizations of the ACE predictor, they are classified as partially non-nested. The Vuong approach first tests for model distinguishability via the Ω test (the ratio of the log-likelihoods of the models, obtained from the Kullback-Leibler information criteria). Distinguishability implies a population-based (not just sample-based) difference in fit (31). If significant, indicating distinguishability, it is followed by Vuong’s closeness test (a z-test of the difference in model predicted probabilities ) to test for differences in the fit of distinguishable models (32). All data and R scripts are available on request from the first author.
Full results for all models and model comparisons, including the comparisons used to arrive at the best fit model within each category, can be found in the Appendix, Tables A – C.
Descriptive results for the best fit model within each of the three model categories and for each outcome are shown in Table 1. Fit indices for descriptive comparisons between the models in Table 1 include: 1) the Akaike information criterion (AIC), a log-likelihood based measure of ‘unexplained information’ in a model, where smaller values are preferred, 2) the concordance statistic (c-stat), a measure of predictive accuracy of the model, and 3) a pseudo R2 as an estimate of total variability explained by the model.
| Depression | Cardiac Disease | Obesity | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Model Type | R2 | AIC | c-stat | R2 | AIC | c-stat | R2 | AIC | c-stat |
| Multiple Individual risk | .197 | 47899 | .745 | .200 | 27970 | .798 | .042 | 67013 | .609 |
| Best-fit | 11 items, ‘once’ | 11 items, ‘more than once’ | 11 items, ‘once’ (by AIC) | ||||||
| Cumulative Risk – Categorical | .178 | 48638 | .734 | .200 | 27974 | .797 | .041 | 67035 | .608 |
| Best-fit | 11 items, ‘once’+ quadratic | 11 items, ‘more than once’ | 11 items, ‘once’ (by AIC) | ||||||
| Cumulative Risk – Continuous | .176 | 48645 | .732 | .199 | 27979 | .796 | .041 | 67037 | .608 |
| Best-fit | 9 items, ‘once’ + quadratic | 11 items, ‘more than once’ | 11 items, ‘once’ (by AIC) | ||||||
| R2 is Nagelkerke. (by AIC) indicates that the model comparisons within that model category were not significantly different from each other, and the best fit model was chosen as that with the smaller AIC. For the depression outcome only, the CR continuous model included a quadratic term. | |||||||||
All model comparisons between the covariate-only baseline model and models including any ACE predictor found significantly better fit for models including any ACE predictor. For the depression outcome, the best fitting model was the MIR model, with a large magnitude of difference between the models both in terms of Vuong’s test, and reduction in AIC. For the cardiac disease and obesity outcomes, the MIR model was ‘distinguishable’ but of equal fit to the CRCat model, and both were of significantly better fit than the CRCn model. The results of the between-category model comparisons are shown in Table 2.
| Outcome | Model | AIC | Covar | MIR | CRCat |
|---|---|---|---|---|---|
| Depression | Covar | 51519 | 1 | ||
| MIR | 47899 | .067*** | 1 | ||
| -29.58*** | |||||
| CRCat | 48638 | .052*** | .019*** | 1 | |
| -26.76*** | 11.34*** | ||||
| CRCn | 48645 | .051*** | .018*** | .001*** | |
| -26.66*** | 12.21*** | 2.51*** | |||
| Cardiac | Covar | 28232 | 1 | ||
| Disease | |||||
| MIR | 27968 | .006*** | 1 | ||
| -7.72*** | |||||
| CRCat | 27974 | .006*** | .001*** | 1 | |
| -7.61*** | NDF | ||||
| CRCn | 27979 | .005*** | .001*** | .001** | |
| -7.49*** | 2.75** | 2.24*** | |||
| Obesity | Covar | 67186 | 1 | ||
| MIR | 67013 | .004*** | 1 | ||
| -6.92*** | |||||
| CRCa | 67035 | .003*** | .001*** | 1 | |
| -6.54*** | NDF | ||||
| CRCn | 67037 | .003*** | .001* | .001*** | |
| -6.10*** | 3.30*** | 2.39** | |||
| Covar = covariate only model with no ACE predictors. Within-cell values are the test statistics for Ω (top) and Vuong’s test (bottom). Positive values of Vuong’s indicate the model in the column was better fitting, negative values indicate the model in the row was better fitting. NDF = Non Different Fit, *p < .05, **p < .01, ***p < .001 | |||||
For all three outcomes, every level of the cumulative ACE Score predictor was significant compared to the zero category (Table 3). Also, in all cases, each of the unadjusted relationships between ACEs and outcomes were significant (unadjusted for other ACEs but including the covariates).
As can be seen in Table 3, inferences about the relationships between the srACEs and outcomes are different under the MIR model, and patterns of relationships in the adjusted models suggest that different specific srACEs are related to each outcome. For the depression outcome, 8 of the 11 ACEs had a significant relationship. For the obesity outcome, only 4 of the srACEs were significantly related, including the 3 sexual abuse questions.
| Depression - MIR | Unadjusted association | Cardiac Disease - MIR | Unadjusted association | Obesity - MIR | Unadjusted association | |
|---|---|---|---|---|---|---|
| ACE Score Categorical1 | 1–11, all significant2 | 1–11, all significant2 | 1–11, all significant2 | |||
| DEPRESS | 2.89*** | 4.08*** | .95 | 1.26*** | 1.01 | 1.14*** |
| DRINK | 1.20*** | 1.97*** | 1.05 | 1.30*** | 1.02 | 1.12*** |
| DRUGS | .96 | 2.14*** | 1.30*** | 1.73*** | .76 | 1.14*** |
| PRISON | 1.04 | 1.95*** | 1.36*** | 1.82*** | 1.08 | 1.21*** |
| DIVORCE | .93* | 1.41*** | 1.05 | 1.23*** | .99 | 1.07*** |
| PUNCH | .98 | 2.01*** | 1.13* | 1.52*** | 1.02 | 1.19*** |
| HURT | 1.29*** | 2.51*** | 1.18** | 1.59*** | 1.09* | 1.27*** |
| SWEAR | 1.46*** | 2.37*** | 1.19*** | 1.46*** | 1.11*** | 1.21*** |
| TOUCH | 1.74*** | 3.11*** | .87 | 1.56*** | 1.18** | 1.43*** |
| TTHEM | 1.13* | 3.02*** | 1.46*** | 1.93*** | 1.04* | 1.49*** |
| HVSEX | 1.28*** | 3.60*** | 1.27* | 2.07*** | 1.08 | 1.53*** |
| The srACEs are noted with the variable names in the BRFSS codebook and data. Italicized coefficients are non-significant. | ||||||
| 1Model results are from the best-fit model for that category. 2All p-values < .001, compared to a zero reference category. | ||||||
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.
The primary purpose of this study was to compare predictor characterization, not to draw conclusions about the effects of ACEs. We therefore only included cases with complete data on all the predictors and the outcomes in order to avoid inconsistent listwise deletions across models, resulting in a loss of 17.9% of the data. Because of this decision we were unable to estimate the models using the survey design weighting appropriate for drawing population-true point estimates or relational inferences. Therefore, our model results in terms of the point estimates of ACE predictors should be interpreted with this caveat. Similarly, we used the same set of covariates in every model, even when they may not have been significantly related to the outcome or may have been collinear with each other or with the ACEs predictors. Model fit may have been influenced by this decision, but we know that the differences between models was attributable to differences in ACE predictor characterizations, not to variations in covariates or the distribution of unequal survey weights.
In this work, we have highlighted only two possible models for the effects of adversity on outcomes (CR vs MIR) but there are numerous others. Some research frames the srACEs in a psychometric context, treating them as indicators of underlying latent variables and applying techniques like exploratory and confirmatory factor analysis (40, 41). Some researchers working in the ACEs Framework have begun conceptualizing how ACEs might be related to outcomes by aggregating people, applying models like cluster analysis (25), latent class (42) or recursive partitioning (43) to classify people into groups, rather than classify ACEs into scores. Results obtained from these differing conceptualizations also differ in what they imply about how adversity and trauma impact individuals, and researchers are advised to include these modeling considerations in their discussions. Model fit approaches such as we utilized here can guide researchers in choosing an operationalization specific to the data.
There has been some recent criticism of the ‘ACE Score’ (44) some of it from within the original Kaiser ACEs Study team (26). Such criticism tends to focus on using a ‘crude’ measure in policy-making. However, these large data sources represent a significant investment of research resources. The BRFSS effort specifically surveys over 450,000 individuals each year, with a yearly budget over $18 million (4). These data sources should be leveraged to produce the most robust inferences possible, given their limitations. Given the magnitude of the population health and policy implications of childhood abuse and trauma, as well as the investment of research resources in the collection of the data, we suggest that additional modeling approaches could be considered as we work toward a more nuanced understanding of the effects of childhood adversity on adult health.
The Thomas Jefferson University IRB reviewed and exempted this research.
Not applicable.
The data supporting the results in this study are publically available through the CDC’s website at https://www.cdc.gov/brfss/index.html. The specific analytic files and R scripts used to analyze the data are available upon request from the first author.
All authors declare they have no competing interests.
This work was unfunded.
ML conceived of the study and was the primary analyst. SK was the secondary analyst and BG also provided statistical expertise. All authors participated equally in drafting and revising the manuscript, and all authors approved the final version.
We are grateful to the students who worked on portions of this project including Justin Wlodarczyk, MPH, John H King, MPH, Zachary Fusfeld, MPH and Daniel Casey. Additionally, we offer grateful acknowledgements to Drs. Amy T. Cunningham, PhD, MPH and Lee Pachter, DO, as well as Se Ryeong Jang, MPH for their reviews of pre-submission versions of the manuscript.