Study design and participants
This study was a part of the research project “Adolescent Health and Risky Behaviors in Anhui Province.” The study protocol has been described in previous studies [16, 17]. Briefly, a 3-stage, random, cluster sampling approach was employed to select participants in Anhui Province. Three cities (i.e., Xuancheng, Hefei, and Huaibei) of Anhui Province were randomly selected in the first stage. Xuancheng, Hefei, and Huaibei are located in the south, middle and north of Anhui, respectively. In the second stage, one regular middle school was randomly chosen from each of those three cities. Then, ten target classrooms from 7th to 9th grade within each school were randomly selected.
A total of 5832 students were invited to participate in this study. After subtracting absent students and removing incomplete questionnaires, we finally obtained an effective sample of 5724 students. The schools from the three cities accounted for 35.8%, 30.4%, and 33.8% of valid questionnaires, respectively, with similar response rates (98.0%). Among the students with valid responses, 3006 (52.5%) were boys, 2718 (47.5%) were girls, and the mean age was 13.5 ± 1.0 years.
Procedures
All the procedures used in this study were approved by the Biomedicine Ethical Committee of Anhui Medical University (NO. 20180083). The investigators went to the contacted schools in November 2020 after unified training. With the assistance of the teachers in each class, the investigators distributed the questionnaires, explained the purpose of this survey and obtained informed consent from all of the students in the selected classes. It took approximately 25 minutes to ensure that every student completed an anonymous questionnaire independently, and all questionnaires were collected on the spot.
Measures
Childhood Trauma Questionnaire-Short Form (CTQ-SF)
The CM of students was assessed by using the Chinese version of the Childhood Trauma Questionnaire Short Form (CTQ-SF), which has been proven reliable for retrospectively measuring CM [18]. The CTQ-SF contains 25 items and is divided into 3 types of abuse (physical, emotional, and sexual abuse) and 2 types of neglect (physical and emotional neglect); each item is scored on a 5-point Likert scale (1 = never, 2 = rarely, 3 = sometimes, 4 = often, and 5 = very often). The CTQ-SF cut-off scores used in this study were as follows: physical abuse ≥ 8; emotional abuse ≥ 9; sexual abuse ≥ 6; physical neglect ≥ 8; and emotional neglect ≥ 10 [19]. Previous studies have shown that the Chinese version of the CTQ-SF has high validity and reliability in measuring childhood abuse and neglect in adolescents [20]. In this study, the total score of the 25 items was calculated, with a higher score indicating more serious CM. The CTQ-SF had a Cronbach’s alpha of 0.89 in this study.
Self-Harm Questionnaire
Based on previous studies [9], we developed the Chinese version of the Self-Harm Questionnaire to assess the SH of students. The questionnaire consists of 39 items, including five core dimensions: (1) highly lethal SH; (2) less lethal SH with visible tissue damage; (3) SH without visible tissue damage; (4) self-harmful behaviors with latency damage; and (5) psychological SH. The detailed items for each SH subtype are displayed in Additional file 1. In this study, participants were asked whether they had self-harmed in the past 6 months. The response option was dichotomized as follows: 0 = no and 1 = yes. The Self-Harm Questionnaire had a Cronbach’s alpha of 0.94, which indicated that it had good internal consistency in this study.
We utilized LCA to classify the subtypes of SH among Chinese early adolescents. Three latent phenotypes were considered to be the best-fitting pattern of SH based on a series of goodness-of-fit indices (see Additional file 2 for detailed information). The probability plots for the three latent phenotypes of SH are shown in Fig. 1. The three latent phenotypes for SH were low SH (n = 3309, 57.8%), medium SH (n = 1660, 29.0%), and high SH (n = 755, 13.2%).
Connor-Davidson Resilience Scale (CD-RISC)
The psychological resilience of students was measured by using the Chinese version of the Connor-Davidson Resilience Scale (CD-RISC) translated by Yu et al. [21]. The CD-RISC comprises 25 items and three factors: tenacity, strength, and optimism. It utilizes a 5-point Likert scale for each item measuring how the respondent has felt over the past month (0 = not true at all, 1 = rarely true, 2 = sometimes true, 3 = often true, and 4 = true nearly all the time). Previous studies have shown that the revised Chinese version of the CD-RISC has high validity and reliability in representing students’ resilience [22]. In this study, the total scores range from 0 to 100, with higher scores indicating greater psychological resilience. The CD-RISC had a Cronbach’s alpha of 0.94 in this study.
Covariates
Based on our previous studies [9, 23], we collected data on several relevant sociodemographic characteristics, including gender (boy or girl), age, family structure (nuclear family, large family, single-parent family, or others), self-perceived family economic status (bad, general, or good), relationship with mother (poor, general, or good), relationship with father (poor, general, or good), and number of friends (< 3, 3–5 or ≥ 6). These variables have been shown to influence SH; they were thus included as covariates in the analysis (for details, see Additional file 3).
Data analysis
All statistical analyses were performed using SPSS for Windows (version 23.0; SPSS Inc., Chicago, IL, USA). Descriptive statistics were used to describe the characteristics of the participants analysed. A univariate logistic regression model was used to explore the associated factors of SH. Associations among CM, phenotypes of SH, and psychological resilience were estimated by using multivariate logistic regression models that subsequently controlled for possible influencing factors. Moreover, based on the results of univariate and multivariate models, the mediation effect of psychological resilience in the CM-SH association was further evaluated by using a path model. This analysis process was performed with a robust maximum likelihood estimator in Mplus 7.4. The level of significance was set at p < 0.05.
LCA is a person-centered approach that describes population heterogeneity in terms of differences across individuals on a set of behaviors or characteristics rather than describing the variability of a single variable [24]. The best-fit number of classes was determined based on a series of indices when conducting LCA: Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), adjusted Bayesian Information Criterion (aBIC), Lo-Mendell-Rubin (LMR), Bootstrapped Likelihood Ratio Test (BLRT), and Entropy [25]. The BIC and LMR help determine the optimal number of classes [26]. When the number of classes was specified, the entropy value showed the accuracy of the class assignments, and the value was close to 1, indicating better classification [27].