Adverse childhood experiences and impact on quality of life in adulthood: development and validation of a short difficult childhood questionnaire in a large population-based health survey

DCQ. The DCQ was developed for and used in a large Norwegian public health study (the HUNT study). In HUNT 3, the third wave of this study, a single item was used (Q3 in Table 1) that allowed respondents to give a general evaluation of their childhood. The results based on this query showed strong relationships between perceiving childhood as difficult and a range of health-related outcomes [27]. For the fourth wave of the study (HUNT 4), the HUNT team thus sought to place greater focus on the theme of childhood difficulties. The overall questionnaire is quite long so that any addition had to be brief, preferably no more than two extra questions. Based on what was identified as the most important concepts, a question about communication and conflict level in the family (Q1 in Table 1) was included, as was a question about childhood trauma (Q2 in Table 1). To avoid potential embarrassment, the latter question was formulated indirectly to assess whether respondents struggled with difficult memories about adverse experiences during childhood because of loss, violence, or abuse. The items were scored on a 5-point ordinal scale (from “not at all” to “very much”), and higher scores represented greater perceived difficulties in childhood. The overall HUNT 4 survey was pilot tested in the municipality of Selbu. Respondents (N = 31) provided written comments to the questions, and six of these respondents were interviewed in detail by phone. A special focus was whether the questions were comprehensible and whether the respondents considered them to be inappropriate or too sensitive in nature. The items in the DCQ seemed to work well, and the pilot respondents offered no negative opinions.

SCL-5. To measure mental distress, we used a short version of the Hopkins Symptom Checklist, the SCL-5 [28]. The SCL-5 performs almost as well as more extended versions (SCL-10 and SCL-25), and the reliability of the scale was excellent in the Norwegian validation study (Cronbach’s alpha = 0.87) [29]. Participants reported whether at any time in the previous week they had felt sad or depressed, hopeless about the future, tense or keyed up, constantly fearful and anxious, or worried. Response categories are “not bothered,” “a little bothered,” “bothered quite a lot,” and “extremely bothered,” scored 1–4. Higher scores indicate greater mental distress.

Loneliness scale: The Three-Item Loneliness Scale is a short version of the UCLA Loneliness Scale [30]. In a validation study, the reliability of the scale was acceptable (Cronbach’s alpha = 0.72) and appeared to measure overall loneliness quite well. This scale contains questions about how often the respondent feels a lack of companionship, left out, and isolated from others. The scale was developed for use in telephone surveys and has a simplified set of response categories. In the present online survey, the 5-point response categories ranged from “never” to “very often,” and higher scores indicate greater loneliness.

CAST-6. The questionnaire included the CAST-6 scale [31], used to assess whether or not participants perceived their parents’ alcohol consumption as problematic. CAST-6 is a brief version of the 30-item Children of Alcoholics Screening Test (CAST) [32] and has demonstrated high internal consistency and test–retest reliability. In a previous validation study, the Cronbach’s alpha ranged from 0.86 to 0.92 for students and patients being treated for substance use disorder [31]. The validity of the CAST-6 has been confirmed in adult populations [31, 33] and among adolescents [34]. Items include questions about whether respondents have ever thought of one of their parents as having a drinking problem, encouraged one of their parents to quit drinking, argued with a drinking parent, heard parents fight when drinking, or felt like hiding or emptying a parent’s bottle of alcohol, and whether they had wished their parent would stop drinking. Response categories for the six items are yes/no. Responses are summed (scale 0–6), and for this study, a score of ≥ 3 was defined as a parental alcohol problem. The original measure used the term “adult children of alcoholics” (ACOA). For simplicity, we also use this term, although we note that “alcoholics” is now outdated, and first-person language is preferred (“children of parents who had problematic alcohol use”) [35].

QoL. We used three items representing subjective QoL, all of which have been suggested as part of a “minimum” list for measuring QoL in national public health surveys in Norway [36]. The items represent three dimensions of subjective QoL: cognitive (satisfaction with life), affective (positive feelings such as happiness), and eudaimonic (whether life is perceived as meaningful) [37, 38]. Conceptually, the first two items address classical subjective well-being (hedonic). The third item builds on the Aristotelian view that a “good life” comprises not only desire fulfillment but also whether a person lives up to their perceived potential or in accordance with their perceived purpose in life. The items are rated on 0–10-point scales, with higher scores representing more satisfaction, happiness, and perceived meaningfulness in life. To our knowledge, these items have not been validated as an expression for QoL in a latent variable analysis before using a structural equation modeling (SEM) framework. Because we used this scale as an outcome measure in our mediation analysis and associations are bi-directional, our findings indirectly assessed the convergent validity for the QoL scale. In other words, the analysis indicated whether the associations of QoL with the DCQ scores and the mediators were in the expected direction. We also report the internal consistency of the scale.

Descriptive statistics are used to show sample characteristics. The constructs used in the following analyses were handled as latent variables in a SEM framework, so that the questions were modeled as reflective indicators of their respective constructs [39]. Mplus version 8.4 was used for the analyses. Because of the ordinal nature of the DCQ, we used the maximum likelihood estimation with robust standard errors (MLR) and report standardized factor loadings (beta = β) [40]. The default procedure in Mplus, full information maximum likelihood, handled missing values. The significance level was set at p < 0.05. The root mean square error of approximation (RMSEA) and the comparative fit index (CFI) were used to assess goodness of fit, with a cut-off value for a good model fit of 0.05 or less for RMSEA and 0.95 or higher for CFI [41, 42].

We assessed the discriminant validity of the DCQ with an MGCFA, examining whether the DCQ discriminated between a group with expected higher negative childhood experiences (ACOA) versus a group without these experiences (non-ACOA). To legitimately be able to compare means between the two groups, we first examined whether the DCQ measured the construct similarly in the two groups to ensure that the construct was invariant (equivalent) across groups [43]. First, we established a configural (baseline) model. The reliability of the scale is reported with the composite reliability (CR) value [44], and a CR value of > 0.70 is considered acceptable. For convenience, we also added McDonald’s omega coefficient [45].

We then implemented a simultaneous MGCFA with increasing equality constraints on measurement parameters to examine if the scale had strong measurement invariance (scalar equivalence) across groups. Scalar equivalence implies that the scales have the same factor structure (configural equivalence) and equivalent factor loadings (metric equivalence) across groups, as well as invariant intercepts, with equivalent origin on the scales in the different groups [46]. If scalar invariance is not achieved, intercepts of the non-invariant items could be freed to assume partial scalar equivalence, which is still considered sufficient for comparing latent means if at least two items are metrically invariant per factor [47]. The testing implies that the more restricted models are compared with the less restricted models, and chi-square differences between the nested models illustrate whether differences in the chi-square value (∆χ2) relative to the change in degrees of freedom (∆df) are significant [40]. A non-significant ∆χ2 value indicates that constraining does not worsen the fit of the model and that the null hypothesis of measurement invariance can be retained [48]. However, Cheung and Rensvold found in a simulation study that the chi-square test might be too conservative, especially in large samples, running the risk of being too likely to deem a measure as non-invariant [49, 50]. They recommended using the change in CFI (∆CFI), which sample size does not affect. A value of ΔCFI less than or equal to − 0.01 indicates that the null hypothesis of invariance should not be rejected [49]. Latent mean differences are reported as unstandardized β values with 95% confidence intervals (CIs).

We assessed the convergent validity of the DCQ by examining whether the score was related in the expected direction with constructs from the literature that are known to correlate with adverse childhood experiences (mental distress, loneliness, and QoL). Results of this latent regression analysis are reported with unstandardized β and 95% CI values to facilitate interpretation of the associations between constructs. The R square (R²) value was used to assess the percentage of the variation in the latent constructs explained by the model.

Most of the participants (76%) were between 25 and 66 years old, slightly more than one-half (53%) were female, about one-half had a higher educational level (at least a bachelor’s degree), and 8 in 10 lived with a spouse or partner (Table 2). According to the CAST-6, about one in six was defined as a child of a parent(s) who had a problematic alcohol use (ACOA).

The covariance coverage of the indicators used in the following analyses were good, above 98.8% in all cases. The first step was specification of the baseline model of the DCQ for each group (Fig. 1). This one-factor solution had “perfect” goodness-of-fit values (RMSEA = 0.00 and CFI = 1.00). Although a model with three indicators is “just-identified” and goodness-of-fit evaluation does not apply, the model can still be evaluated in terms of interpretability and strength of its parameter estimates (e.g., magnitude of factor loadings) [51]. The model had factor loadings from 0.69 to 0.89 in both groups, and no error correlations emerged. The baseline model for the ACOA group is shown in Fig. 1. The reliability of the scale was excellent (CR and omega coefficient = 0.86).

The MGCFAs specifying the χ2 and CFI values for the configural, metric, and scalar models are shown in Table 3. In case of the metric and scalar models, these values have positive degrees of freedom, as more restrictions were set on the model. Thus, they are over-identified models, and comparison between them is legitimate [51]. The ∆χ2s between the nested models were significant as expected because of the large number of participants; thus, further investigation was based on ∆CFIs. The scalar model had a 0.03 lower CFI value compared with the metric model, and its RMSEA was too high, indicating a problem. The modification indices indicated that the intercept of Q1 (Table 1) was non-invariant and that the ∆χ2 would improve considerably if this intercept was freed. The final partial scalar model had excellent fit and was deemed as invariant because the ∆CFI was ≤  − 0.01 (Table 3), so latent means could be compared. The latent mean of the DCQ scale was 0.86 (95% CI 0.82–0.90, p < 0.001) higher in the ACOA group versus the non-ACOA group. To examine the magnitude of this value, we calculated an effect size d for the latent mean differences following the procedure proposed by Hancock [52]. The effect size d for the ACOA versus the non-ACOA group was 1.15 (see calculation in additional file). According to interpretative guidelines, 0.2, 0.5, and 0.8 are cut-off values for small, medium, and large effect sizes of latent means, respectively, which indicates a large effect size of differences between groups [52].

Continuing to the convergent validity analyses, we tested the relationship between the DCQ and adulthood QoL in a simultaneous mediation analysis that included mental distress (SCL-5) and loneliness as mediating constructs. The DCQ scores had a direct negative association with QoL (β =  − 0.04, 95% CI − 0.07 to − 0.02, p = 0.001; Fig. 2). The DCQ scores were associated in the expected direction with both mental distress (β = 0.31, 95% CI 0.30–0.33, p < 0.001) and loneliness (β = 0.38, 95% CI 0.36–0.39, p < 0.001), i.e., the level of mental distress and loneliness increased when the DCQ score increased. In turn, mental distress and loneliness had a substantial negative influence on QoL (β =  − 1.56 and β =  − 0.81, respectively; Fig. 2). The specific indirect effect of the DCQ score in association with QoL (shown only in the Mplus output, not in the figure) via mental distress was β − 0.49 (95% CI − 0.51 to − 0.46, p < 0.001). The specific indirect effect of this score in association with QoL via loneliness was β − 0.31 (95% CI − 0.33 to − 0.29, p < 0.001). Thus, the DCQ accounted for 31% of the influence of the mental distress construct (− 0.49/ − 1.56) on QoL and 38% of influence of the loneliness construct on QoL (− 0.31/ − 0.81). The total influence of the DCQ score in association with QoL was − 0.84 (95% CI − 0.87 to − 0.80, p < 0.001), and of this value, − 0.80 was indirect and − 0.04 direct. The model as a whole had excellent goodness-of-fit indices (RMSEA = 0.05 and CFI = 0.98) and explained 18% of the variance in mental distress, 19% of the variance in loneliness, and 65% of the variance in QoL (R²). The reliability of the QoL scale was excellent (CR and omega coefficient = 0.90).