Using decision tree to predict non-suicidal self-injury among young adults: the role of depression, childhood maltreatment and recent bullying victimization

ABSTRACT

Importance: Non-suicidal self-injury (NSSI) is a significant mental health issue requiring a deeper understanding of its underlying causes, such as childhood maltreatment, adult bullying victimization, and depression. Previous studies have not adequately addressed the cumulative risks of these factors on NSSI among college students. This population-based study investigates these cumulative risk factors.

Design, setting, and participants: The cross-sectional study included 63 university’s college students with a mean age of 19.6 years (N = 95,833).

Main outcomes and measures: Two Chi-Square Automatic Interaction Detection (CHAID) decision tree models were used to classify subgroups based on childhood maltreatment and adult bullying victimization experiences and to investigate their cumulative risks of NSSI. Recursive partitioning algorithms determined each predictor variable’s relative importance.

Results: The CHAID model accurately predicted NSSI behaviours with an overall accuracy rate of 77.8% for individuals with clinically relevant depressive symptoms and 97.2% for those without. Among depressed individuals, childhood emotional abuse was the strongest NSSI predictor (Chi-Square, 650.747; adjusted P < .001), followed by sexual and physical abuse. For non-depressed individuals, emotional abuse in childhood was the strongest NSSI predictor (Chi-Square, 2084.171; adjusted P < .001), with sexual and verbal bullying in the past year representing the most significant proximal risks.

Conclusions and relevance: Emotional abuse during childhood profoundly impacts individuals, increasing the risk of NSSI in both depressed and non-depressed individuals. Clinically relevant depressive symptoms have a moderating effect on the relationship between childhood maltreatment, adult bullying victimization, and NSSI. Identifying these factors can inform targeted interventions to prevent NSSI development among young adults.

HIGHLIGHTS

• Emotional abuse during childhood has a profound impact on individuals, increasing their risk of non-suicidal self-injury (NSSI), regardless of whether they are depressed or non-depressed.

• Among depressed individuals, childhood emotional abuse emerges as the strongest predictor of NSSI, followed by sexual and physical abuse.

• In non-depressed individuals, emotional abuse in childhood assumes a similar role as the strongest NSSI predictor, with sexual abuse and verbal bullying in the past year representing the most significant proximal risks.

Importancia: La autolesión no suicida (NSSI, en sus siglas en inglés) es un importante problema de salud mental que requiere una comprensión más profunda de sus causas subyacentes, como el maltrato infantil, la victimización por acoso escolar en la edad adulta y la depresión. Los estudios anteriores no han abordado adecuadamente los riesgos acumulativos de estos factores sobre las NSSI entre los estudiantes universitarios. Este estudio basado en la población investiga estos factores de riesgo acumulativos.

Diseño, contexto y participantes: El estudio transversal incluyó a 63 estudiantes universitarios con una edad media de 19,6 años (N = 95.833).

Resultados y medidas principales: Se utilizaron dos modelos de árbol de decisión de Detección Automática de Interacción Chi-Cuadrado (CHAID, en sus siglas en inglés) para clasificar los subgrupos en función de las experiencias de maltrato en la infancia y victimización por acoso en la edad adulta y para investigar sus riesgos acumulativos de NSSI. Los algoritmos de partición recursiva determinaron la importancia relativa de cada variable predictiva.

Resultados: El modelo CHAID predijo con exactitud las conductas NSSI con una tasa de exactitud global del 77,8% para los individuos con síntomas depresivos clínicamente relevantes y del 97,2% para los que no los tenían. Entre los individuos deprimidos, el abuso emocional en la infancia fue el mayor predictor de NSSI (Chi-cuadrado, 650,747; p ajustada < 0,001), seguido del abuso sexual y físico. Para los individuos no deprimidos, el abuso emocional en la infancia fue el predictor más fuerte de NSSI (Chi-cuadrado, 2084,171; P ajustada < 0,001), y el acoso sexual y verbal en el último año representó los riesgos próximos más significativos.

Conclusiones y relevancia: El abuso emocional durante la infancia afecta profundamente a los individuos, aumentando el riesgo de NSSI tanto en individuos deprimidos como no deprimidos. Los síntomas depresivos clínicamente relevantes tienen un efecto moderador en la relación entre el maltrato en la infancia, la victimización por acoso en la edad adulta y las NSSI. La identificación de estos factores puede servir de base para las intervenciones dirigidas a prevenir el desarrollo de NSSI entre los adultos jóvenes.

KEYWORDS:

• Non-suicidal self-injury
• depression
• childhood maltreatment
• bullying victimization
• decision tree

PALABRAS CLAVE:

• Autolesiones no suicidas
• maltrato infantil
• victimización por acoso escolar

1. Introduction

Non-suicidal self-injury (NSSI) refers to the intentional harm inflicted upon one’s own body tissues without suicidal intent (Nock et al., 2006). The prevalence of NSSI during adolescence is reported to be 16.9% (Gillies et al., 2018), while among college students, it is estimated to be 17.7% (Kiekens et al., 2021). In early adolescence, NSSI serves as a behavioural indicator for the late emergence of mental disorders like depression, anxiety, and eating disorders (Wilkinson et al., 2018). Additionally, NSSI is a significant predictor of suicidal ideation, suicidal attempt, and suicide death (da Silva Bandeira et al., 2022; Groschwitz et al., 2015; Hamza et al., 2012), making it increasingly important in clinical settings in recent years.

The factors that contribute to NSSI are diverse, involving various sociodemographic and educational aspects, negative life events, family adversity, and psychiatric and psychological factors (Hawton et al., 2012). For instance, factors like low socioeconomic status, adverse childhood experiences, and bullying can increase the risk of NSSI. Additionally, psychiatric factors such as depression also play a role in the complexity of NSSI (Hawton et al., 2012; Wang et al., 2022). Among these, child maltreatment and bullying have become global public issues. According to world health organization (WHO), child maltreatment is the abuse and neglect that occurs to children under 18 years of age (Child Maltreatment, n.d.). It includes all types of physical and/or emotional ill-treatment, sexual abuse, neglect, negligence and commercial or other exploitation, which affects one in three children globally and doubles the risk of developing mental illness (Chandan et al., 2019). Similarly, data from 68 low-income and middle-income countries indicate an overall prevalence of bullying, regardless of age and sex, of 34·4% at least once in the past 30 days (Han et al., 2019). Bullying is a form of aggressive behaviour, typically characterized as ‘behavior intended to inflict injury or discomfort upon another individual’ (Olweus, 2013), it includes three criteria of intentionality, repetitiveness, and imbalance of power, and is considered a risk factor for mental health issues (Lereya et al., 2015). Recent studies demonstrate that experiencing child maltreatment and being bullied can modify the hypothalamic-pituitary-adrenal axis and other stress response systems (Turecki et al., 2012) or lead to long-term increases in inflammatory processes from childhood to young adulthood (Copeland et al., 2014). Notably, both child maltreatment and school bullying significantly increase the risk of self-harm. Extensive meta-analyses consistently establish a link between childhood maltreatment, including sexual abuse, physical abuse, neglect, and emotional abuse, and non-suicidal self-injury (NSSI) (Liu et al., 2018). Moreover, frequent exposure to bullying is strongly associated with higher rates of self-harm, even after controlling for emotional and behavioural issues, low IQ, and family environmental risks (Fisher et al., 2012). Besides, various forms of bullying victimization, such as cyber-bullying and traditional bullying, have also been found to elevate the risk of major depressive disorder, self-harm, and suicidality (Islam et al., 2020).

A study was conducted to investigate whether exposure to both maltreatment and bullying, or bullying alone, has distinct effects on mental health consequences. The findings revealed that being bullied has similar, and in some cases worse, long-term adverse effects on young adults’ mental health compared to being maltreated (Lereya et al., 2015). However, further exploration is needed to understand the cumulative impact of these factors. For instance, the study employed binary logistic regression analyses to examine the association between maltreatment, being bullied, and mental health problems. However, it did not consider the interaction between different types of childhood abuse and peer bullying during adulthood. It is likely that a combination or interaction of these factors contributes to the risk of NSSI. What’s more, previous research has shown that depression plays a significant role in mediating the relationship between bullying victimization and self-harm (Moore et al., 2022), as well as between childhood maltreatment and NSSI (Holden et al., 2022). However, further investigation is needed to gain insights into the complex interactions that exist between these risk factors and depression.

To explore how these factors interact and create profiles that predict an individual's level of NSSI risk, a decision tree approach may be suitable. This technique has been used as a tool in recent years to explore psychological outcomes, such as depressive disorder (Zhu et al., 2022) and suicide attempts (Eagle et al., 2022). To achieve this, the present study will (1) employ logistic regression models to assess the associations between childhood maltreatment, recent campus bullying, and the likelihood of NSSI, and (2) use a decision tree algorithm to examine the interaction and cumulative effect of childhood maltreatment and recent bullying victimization on the likelihood of NSSI. The primary objective of this study is to offer customized assistance to individuals who are at risk for NSSI, regardless of whether they already have depression.

2. Methods

2.1. Data source and participants

This study is part of a large-scale, cross-sectional study conducted in 2021 (26 October to 18 November) with all the university and college students in northeast China. The researcher created a Quick Response code (i.e. QR code) and distributed it to all universities (n = 63). All the participants have completed the informed consent online before filling out the questionnaire. Participants were informed of their right to withdraw from the survey at any time. All participants received information regarding accessing mental health support if they experienced any emotional distress during or after the study. Participants deemed to be at high risk for suicide were provided with information on suicide prevention hotlines as well as advised to seek professional help. Additionally, help-seeking information from local hospitals’ mental health departments was made available. To compensate participants for their involvement, cash rewards were offered. Please refer to Appendix 1 for the Checklist for Reporting Results of Internet E-Surveys (CHERRIES).

The study included participants who met the following inclusion criteria: (1) education level at or above the first year of university; (2) were able to understand the contents of the questionnaire; (3) answered at least three out of the four attention check questions correctly; (4) had a BMI within the normal range (16 < BMI < =40 (Weir & Jan, 2023; WHO expert consultation, 2004); (5) did not exhibit logical contradictions or omit answers, nor selected options that were unrelated to the current study objectives (such as addictive behaviours); (7) did not show a clear pattern in their option selections (such as almost always choosing the first option). The study initially included 96,218 participants, but 385 of them were excluded from the analysis due to providing logically contradictory answers. Therefore, the final sample size was 95,833 participants, 41.6% of whom were male. Among the participants, age data was missing for 54 individuals, and median age (19.00 years) was used to impute these missing values.

2.2. Measures

2.2.1. Sociodemographic characteristics

Participants’ sociodemographic characteristics, including age, sex assigned at birth, gender identity, ethnicity, only child, smoking and sleep quality was collected.

2.2.2. Other covariances

The associations between concussion and negative outcomes such as self-harm, psychiatric hospitalization, and suicide have been emphasized in studies and meta-analyses (Fralick et al., 2019; Ledoux et al., 2022). Hence, the current study also included the collection of sport-related concussion history as a covariate to control for potential confounding factors. Additionally, considering the suggestion that sensitivity to physical pain could act as a barrier to NSSI in many individuals (Lalouni et al., 2022), data on pain tolerance was likewise obtained. The regression analysis incorporated sociodemographic characteristics, concussion history, and pain tolerance as covariates.

2.2.3. Non-suicidal self-injury

NSSI was measured through the adapted Clinician-Rated Severity of Non-suicidal Self-Injury (CRS-NSSI). The CRS-NSSI is a single-item measurement of the presence and severity of NSSI, with good reliability (Weighted Cohen’s κ = 0.91) (Somma et al., 2019). The participants in the study were asked a two-part question to assess their history of non-suicidal self-injury (NSSI). Firstly, they were asked whether they had ever intentionally hurt themselves without the intention of suicide. If they answered ‘Yes’, they were then asked a second question about the frequency of their NSSI behaviour in the past year, with options ranging from 0 days to 12 or more days (1 = 0 day, 2 = 1∼4 days, 3 =  5∼7 days, 4 = 8∼11 days, 5 = 12 days and more). Based on the responses to these questions, the participants’ NSSI behaviours were dichotomously categorized as either present or absent in their lifetime. Additionally, their past year NSSI behaviour was defined as a dichotomous variable based on their response to the second question, with a response of 0 days being categorized as ‘No’ and any other response being categorized as ‘Yes’. The study also defined the severity of past year NSSI as a multi-categorical variable based on the response to the second question.

2.2.4. Childhood maltreatment

The Chinese version of Childhood Trauma Questionnaire (CTQ) was used to assessing childhood trauma in the current study (Cronbach’s α of subscales were 0.51∼0.71, total scale was 0.60) (Fu & Yao, 2005). It is a 28-item questionnaire that consists of five subscales, including emotional abuse (EA), emotional neglect (EN), physical abuse (PA), physical neglect (PN), and sexual abuse (SA), which are rated on a Likert scale. The CTQ cut-off scores were as follows: PA ≥ 8, SA ≥ 6, EA ≥ 9, PN ≥ 8, and EN ≥ 10. The reported sensitivity and specificity for these cut-off scores were 89% and 97%, respectively (Tietjen et al., 2010).

2.2.5. Bullying victimization

Bullying is an aggressive behaviour aimed at causing harm or discomfort to another individual. In our study, we specifically examine school bullying. To assess school bullying victimization, participants were asked a single question: ‘In the past year, have you experienced any bullying in school?’ (1 = Verbal bullying, 2 = Physical violence, 3 = Sexual harassment, 4 = Cyber bullying, 5 = No bullying experiences). We then created four dichotomous variables to define each type of bullying victimization. For ‘Verbal bullying victimization’, we defined a ‘Yes’ response as reporting verbal bullying in the past year and a ‘No’ response as not reporting verbal bullying (including they selected ‘No bullying experiences’). The same approach was used to define ‘Physical violence victimization’, ‘Sexual harassment victimization’, and ‘Cyber bullying victimization’.

2.2.6. Depressive symptoms

The severity of clinically relevant depressive symptoms was measured using the Chinese version of the nine-item Patient Health Questionnaire (PHQ-9), the cut-off scores as 10 (Manea et al., 2012; Zhang et al., 2013).

2.3. Statistical analysis

The analysis in this study involved both univariate and multivariable logistic regression to evaluate the association between various exposures and the likelihood of reporting at least 1 NSSI behaviour within the past year. Univariate logistic regression was used to assess the relationship between a single exposure (e.g. verbal bullying victimization, sex, concussions) and NSSI behaviour. The odds ratios (OR) and 95% confidence intervals (CI) were calculated using a single multivariable forward stepwise logistic regression, with a cutoff for inclusion set at P < .05. The logistic regression was conducted by ‘autoReg’ package in R (Keon-Woong Moon, 2023. version 0.3.4). Multinomial logistic regression was also performed to examine the predictive effects of various factors (e.g. sex, depressive symptoms, pain tolerance, concussion, types of childhood maltreatment and bullying victimization) on the severity of NSSI. The ‘nnet’ package in R was utilized for this analysis (Venables & Ripley, 2002. Version 7.3–18), and the reference group for severity was ‘Never NSSI’. Bonferroni corrections were applied to prevent false positives.

The study utilized Chi-Square Automatic Interaction Detection model (CHAID) to assess the interaction between sex, clinically relevant depressive symptoms, pain tolerance, concussion, different types of childhood maltreatment and bullying victimization, and reporting of NSSI behaviour within the past year. CHAID model is a decision tree model that uses a series of merging and splitting steps based on χ² test to generate a decision tree by identifying patterns among variables (McCarty & Hastak, 2007). CHAID has been repeatedly used in studies with clinical applications (Eagle et al., 2022; Zhu et al., 2022). The strongest factor associated with the outcome becomes the first cut point in the tree, and each subgroup is then identified recursively further down the decision tree until no further statistically significant splits can be made (Alpha for Splitting and Merging were 0.05). Referring to the model parameters of previous studies, Bonferroni corrections at each level for multiple comparisons. The data set was stratified by depressive symptom (yes or no) for CHAID analyses, resulting in 2 decision trees (depression [DEP], without depression [NO-DEP]). The pruning criteria were applied to limit the size of the tree and prevent overfitting. The tree growth was limited to 3 layers, and the groups smaller than 100 were not split further (parent branch), and no group smaller than 50 was formed (child branch). The model was created and validated using 10- fold cross-validation. The accuracy of the final decision tree model was evaluated using the confusion matrix, which showed the proportion of participants with each outcome variable that was correctly or incorrectly classified. The AUC, accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were reported for the model (Table 4). Analyses were conducted using R-4.2.3 and SPSS Statistics version 27 (IBM).

3. Results

3.1. Sociodemographic characteristics of the participants

Among the 95,833 participants analyzed, 8.3% (7930) reported a lifetime history of NSSI. The majority (4976, 62.7%) had done so within the past 12 months. Among those who had engaged in NSSI within the past 12 months, 54.3% reported 1–4 days, 17.9% reported 5–7 days, 6.9% reported 8–11 days, and 20.9% reported 12 or more days.

Of the participants included, those who reported engaging in NSSI in the past year demonstrated significant differences, with certain groups more likely to engage in NSSI, including those who were assigned female at birth, sole children in their families, transgender individuals, smokers, experiencing disrupted sleep, suffering from anxiety and depression, experiencing child maltreatment, bullying at school, and concussion at any point in the past year (Table 1). Females reported NSSI at a significantly higher rate than males (χ² = 162.60, P < .001). Among participants who reported NSSI, 81.5% identified as cisgender, 3.4% identified as transgender, and 6.9% identified as nonbinary or genderqueer (χ² = 1017.88, P < .001). Only child status was also more common among participants who reported NSSI (χ²= 41.46, P < .001). The smoker’s rate was significantly higher among participants who had reported NSSI (χ² = 366.24, P < .001), as were the rates of concussion (χ² = 600.44, P < .001). Participants who had reported NSSI were also more likely to have experienced child maltreatment and school bullying (χ² values ranged from 656.06 to 5618.36) (Table 1 and Figure 1).

Figure 1. Co-occurrence between different types of child maltreatment and bully victimization and the ROC curves of the CHAID model for DEP and No-DEP models. A: All the bullying means the self-report bullying victimization in past year. B: EA. Emotional Abuse. PA. Physical Abuse. SA. Sexual Abuse. EN. Emotional Neglect. PN. Physical Neglect. The CTQ cut-off scores were as follows: PA ≥ 8, SA ≥ 6, EA ≥ 9, PN ≥ 8, and EN ≥ 10. C: the cut-off scores of PHQ-9 as 10.

Table 1. Characteristics of the study population by past year NSSI.

Characteristic	Levels	No-NSSI (N = 90857)	NSSI (N = 4976)	t/χ^2 (p)
Age	Mean ± SD	19.6 ± 1.7	19.6 ± 1.8	0.33 (1.00)
Sex	Male	38251 (42.1%)	1639 (32.9%)	162.60 (<.001)
	Female	52606 (57.9%)	3337 (67.1%)
Only Child	No	47975 (52.8%)	2394 (48.1%)	41.46 (<.001)
	Yes	42882 (47.2%)	2582 (51.9%)
Ethnicity	Others	9505 (10.5%)	556 (11.2%)	2.47 (.116)
	Han	81352 (89.5%)	4420 (88.8%)
Gender Identification	Cisgender	82327 (90.6%)	4057 (81.5%)	1017.88 (<.001)
	Transgender	1253 (1.4%)	168 (3.4%)
	Nonbinary/ genderqueer	1440 (1.6%)	344 (6.9%)
	Uncertainty	1271 (1.4%)	170 (3.4%)
	Questioning	4566 (5%)	237 (4.8%)
Smoking	No	79313 (87.3%)	3874 (77.9%)	366.24 (<.001)
	Yes	11544 (12.7%)	1102 (22.1%)
Sleep	Mean ± SD	7.2 ± 2.3	5.3 ± 2.6	50.07 (.005)
Pain Tolerance	Mean ± SD	17.6 ± 3.1	17.2 ± 3.3	–7.75 (.005)
Depression Score	Mean ± SD	4.8 ± 4.1	11.1 ± 5.9	74.09 (.005)
Anxiety Score	Mean ± SD	3.4 ± 3.6	8.4 ± 5.1	67.03 (.005)
CTQ: Emotional Abuse	No-EA	80607 (88.7%)	2576 (51.8%)	5618.36 (<.001)
	EA	10250 (11.3%)	2400 (48.2%)
CTQ: Physical Abuse	No-PA	85728 (94.4%)	3710 (74.6%)	2965.68 (<.001)
	PA	5129 (5.6%)	1266 (25.4%)
CTQ: Sexual Abuse	No-SA	78876 (86.8%)	3203 (64.4%)	1931.51 (<.001)
	SA	11981 (13.2%)	1773 (35.6%)
CTQ: Physical Neglect	No-PN	54879 (60.4%)	1955 (39.3%)	870.47 (<.001)
	PN	35978 (39.6%)	3021 (60.7%)
CTQ: Emotional Neglect	No-EN	56711 (62.4%)	1662 (33.4%)	1667.16 (<.001)
	EN	34146 (37.6%)	3314 (66.6%)
Past Year Verbal Bullying	No	85963 (94.6%)	3987 (80.1%)	1716.29 (<.001)
	Yes	4894 (5.4%)	989 (19.9%)
Past Year Physical Bullying	No	89983 (99%)	4728 (95%)	656.06 (<.001)
	Yes	874 (1%)	248 (5%)
Past Year Sexual Bullying	No	90443 (99.5%)	4807 (96.6%)	669.84 (<.001)
	Yes	414 (0.5%)	169 (3.4%)
Past Year Cyber Bullying	No	90061 (99.1%)	4684 (94.1%)	1043.01 (<.001)
	Yes	796 (0.9%)	292 (5.9%)
Past Year Concussion	No	84692 (93.2%)	4177 (83.9%)	600.44 (<.001)
	Yes	6165 (6.8%)	799 (16.1%)
NSSI Severity	Never	87903 (96.7%)	0 (0%)	95833.00 (<.001)
	Before Past Year	2954 (3.3%)	0 (0%)
	Past Year 1–4 Day	0 (0%)	2700 (54.3%)
	Past Year 5–7 Day	0 (0%)	890 (17.9%)
	Past Year 8–11 Day	0 (0%)	344 (6.9%)
	Past Year >=12 Day	0 (0%)	1042 (20.9%)

Note. Between-group differences were examined using χ² test for categorical variables and independent-sample t test for the continuous variable. P values were adjusted for Bonferroni.

3.2. Logistic regression analyses to predict NSSI in the past year versus no NSSI

After adjusting for socio-demographic factors, our univariate analysis found significant links between all variables and NSSI in the past year. The strongest associations with NSSI were found for sexual bullying in the past year (OR, 7.68; 95% CI, 6.41–9.21), emotional abuse (OR, 7.33; 95% CI, 6.91–7.77), cyber bullying in the past year (OR, 7.05; 95% CI, 6.15–8.09), physical abuse (OR, 5.70; 95% CI, 5.32–6.12), physical bullying in the past year (OR, 5.40; 95% CI, 4.68–6.24), verbal bullying in the past year (OR, 4.36; 95% CI, 4.04–4.70), sexual abuse (OR, 3.64; 95% CI, 3.43–3.87), emotional neglect (OR, 3.31; 95% CI, 3.12–3.52), physical neglect (OR, 2.36; 95% CI, 2.22–2.50), and depressive symptoms (OR, 1.24; 95% CI, 1.24–1.25).

In the multivariable analysis, controlling for all other variables, emotional abuse had the strongest association with past year NSSI (B = 0.76; OR, 2.13; 95% CI, 1.97–2.30), followed by female gender (B = 0.78; OR, 2.18; 95% CI, 2.02–2.35), sexual bullying in the past year (B = 0.51; OR, 1.66; 95% CI, 1.31–2.10), sexual abuse (B = 0.50; OR, 1.64; 95% CI, 1.53–1.77), and physical abuse (B = 0.46; OR, 1.59; 95% CI, 1.45–1.74). Other forms of child maltreatment and school bullying were also significantly associated with NSSI, with odds ratios ranging from 1.01 to 1.33. Table 2 shows the results of the univariate and multivariable logistic regression analyses used to predict past year NSSI.

Table 2. Univariable and multivariable logistic regression model to identify variables associated with reporting NSSI in the previous year.

Dependent: Past Year NSSI		No-NSSI (N = 90,857)	NSSI (N = 4976)	OR (univariable)	OR (multivariable)
Depression Score	Mean ± SD	4.8 ± 4.1	11.1 ± 5.9	1.24 (1.24–1.25, p < .001)	1.17 (1.17–1.18, p < .001)
CTQ: Emotional Abuse	No-EA	80607 (88.7%)	2576 (51.8%)
	EA	10250 (11.3%)	2400 (48.2%)	7.33 (6.91–7.77, p < .001)	2.13 (1.97–2.30, p < .001)
CTQ: Physical Abuse	No-PA	85728 (94.4%)	3710 (74.6%)
	PA	5129 (5.6%)	1266 (25.4%)	5.70 (5.32–6.12, p < .001)	1.59 (1.45–1.74, p < .001)
CTQ: Sexual Abuse	No-SA	78876 (86.8%)	3203 (64.4%)
	SA	11981 (13.2%)	1773 (35.6%)	3.64 (3.43–3.87, p < .001)	1.64 (1.53–1.77, p < .001)
CTQ: Physical Neglect	No-PN	54879 (60.4%)	1955 (39.3%)
	PN	35978 (39.6%)	3021 (60.7%)	2.36 (2.22–2.50, p < .001)	1.01 (0.94–1.09, p = .730)
CTQ: Emotional Neglect	No-EN	56711 (62.4%)	1662 (33.4%)
	EN	34146 (37.6%)	3314 (66.6%)	3.31 (3.12–3.52, p < .001)	1.46 (1.35–1.57, p < .001)
Past Year Verbal Bullying	No	85963 (94.6%)	3987 (80.1%)
	Yes	4894 (5.4%)	989 (19.9%)	4.36 (4.04–4.70, p < .001)	1.49 (1.35–1.64, p < .001)
Past Year Physical Bullying	No	89983 (99%)	4728 (95%)
	Yes	874 (1%)	248 (5%)	5.40 (4.68–6.24, p < .001)	1.33 (1.10–1.62, p = .004)
Past Year Sexual Bullying	No	90443 (99.5%)	4807 (96.6%)
	Yes	414 (0.5%)	169 (3.4%)	7.68 (6.41–9.21, p < .001)	1.66 (1.31–2.10, p < .001)
Past Year Cyber Bullying	No	90061 (99.1%)	4684 (94.1%)
	Yes	796 (0.9%)	292 (5.9%)	7.05 (6.15–8.09, p < .001)	1.48 (1.23–1.78, p < .001)

Note. Abbreviation: OR odds ratio. Adjusting for age, sex assigned at birth, gender identity, ethnicity, only child status, smoking, sleeping. P values were unadjusted.

3.3. Multinomial logistic regression analyses to predict NSSI severity

To investigate the impact of various variables on the severity of Non-Suicidal Self-Injury (NSSI), we conducted a multinomial logistic regression model while controlling for socio-demographic factors. Our analysis revealed that clinically relevant depressive symptoms and all types of child abuse and emotional neglect were significantly associated with different levels of NSSI severity, including individuals who engaged in NSSI for 1–4 days, 5 or more days, or those who had ceased NSSI in the past year. The odds ratios for these associations ranged from 1.023 to 2.649, as presented in Table 3. In addition, past year verbal bullying was identified as a significant predictor of NSSI engagement, showing odds ratios ranging from 1.513 to 1.586 for different levels of NSSI severity. Additionally, past year sexual bullying and cyberbullying were found to be significantly associated with more severe NSSI frequency (≥ 5 days in the past year) with odds ratios of 1.919 for sexual bullying and 1.674 for cyberbullying.

Table 3. Multinomial logistic regression to predict the severity of NSSI (Reference Group: NSSI Severity = ‘Never’).

	Past Year no NSSI				Past Year 1–4 Day				Past Year >=5 Day
Variables	B	95%CI	p^a	OR	B	95%CI	p^a	OR	B	95%CI	p^a	OR
Depression	0.110	0.10–0.12	.000	1.116	0.154	0.15–0.16	.000	1.166	0.196	0.19–0.21	.000	1.217
CTQ: Emotional Abuse	0.567	0.47–0.67	.000	1.764	0.683	0.58–0.78	.000	1.980	0.974	0.86–1.08	.000	2.649
CTQ: Physical Abuse	0.319	0.20–0.44	.000	1.376	0.342	0.22–0.46	.000	1.407	0.684	0.56–0.80	.000	1.981
CTQ: Sexual Abuse	0.483	0.39–0.57	.000	1.620	0.561	0.47–0.65	.000	1.753	0.524	0.42–0.63	.000	1.688
CTQ: Physical Neglect	–0.002	–0.09–0.08	1.000	0.998	–0.032	–0.13–0.06	1.000	0.968	0.071	–0.04–0.18	1.000	1.074
CTQ: Emotional Neglect	0.380	0.29–0.47	.000	1.463	0.385	0.29–0.48	.000	1.469	0.419	0.30–0.53	.000	1.520
Past Year Verbal Bullying	0.414	0.29–0.54	.000	1.513	0.461	0.34–0.58	.000	1.586	0.446	0.32–0.58	.000	1.563
Past Year Physical Bullying	–0.082	–0.36–0.20	1.000	0.922	0.280	0.04–0.52	.856	1.323	0.281	0.03–0.53	.936	1.324
Past Year Sexual Bullying	0.279	–0.06–0.62	1.000	1.322	0.465	0.17–0.76	.081	1.592	0.652	0.36–0.94	.000	1.919
Past Year Cyber Bullying	–0.203	–0.49–0.09	1.000	0.816	0.207	–0.03–0.45	1.000	1.231	0.515	0.29–0.74	.000	1.674

Note. Abbreviation: OR odds ratio. Adjusting for age, sex assigned at birth, gender identity, ethnicity, only child status, smoking, sleeping. p^a means P values were adjusted for Bonferroni.

3.4. CHAID decision trees

The CHAID model achieved an overall accuracy rate of 77.6% for DEP and 97.2% for NO-DEP. Table 4 displays the area under the curve (AUC), sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), while Figure 1 presents the ROC curves for the models. The DEP model generated 12 nodes (Figure 2A), with emotional abuse during childhood as the first splitting variable, followed by sexual and physical abuse as the second, and cyber and verbal bullying in the past year as the third. Physical and emotional neglect during childhood, and sexual and physical bullying in the past year were not included in the model. The No-DEP model generated 14 nodes (Figure 2B), with emotional abuse during childhood as the first splitting variable, followed by sexual and verbal bullying in the past year as the second, and physical and sexual abuse, and emotional neglect as the third. Physical neglect during childhood and cyberbullying in the past year were not included in the model.

Figure 2. Past year NSSI with and without depression symptom. N, Node, CTQ-EA, emotional abuse, CTQ-PA, physical abuse, CTQ-EN, emotional neglect, CTQ-SA, sexual abuse, All the p values were adjusted by Bonferroni corrections.

Table 4. Model parameters of chi-square automatic interaction detection Decision Tree.

	AUC (95% CI)	Sensitivity	Specificity	PPV	NPV	Accuracy (95% CI)	Balanced accuracy
Depression	0.665 (0.654–0.677)	0.036	0.993	0.617	0.778	0.776 (0.769–0.784)	0.515
No Depression	0.670 (0.689–0.710)	0.020	0.9995	0.551	0.972	0.972 (0.971–0.973)	0.510

Note. Abbreviations: AUC, area under the curve; CI, confidence interval; PPV, positive predictive value; NPV, negative predictive value.

3.5. Decision tree for DEP

The results of our study showed that emotional abuse experienced before the age of 16 was the most significant variable associated with past-year NSSI, with a prevalence of 43.1% in those who also experienced physical abuse (χ², 650.747; adjusted P < .001). Additionally, respondents who experienced emotional and physical abuse and reported cyberbullying in the past year had a higher prevalence of NSSI (61.7% vs. 41.1%, χ², 23.438; adjusted P < .001). For those who experienced emotional abuse but not physical abuse, the prevalence of NSSI increased when they also reported experiencing verbal bullying in the past year (prevalence 41.6% vs. 29.0%, χ², 32.847; adjusted P < .001).

3.6. Decision tree for no-DEP

Our study found emotional abuse before the age of 16 to be the strongest predictor of past-year NSSI, with a prevalence of the behaviour in individuals who also suffered sexual bullying being 39.6% (χ², 2084.171; adjusted P < .001). Respondents who experienced emotional and sexual bullying in the past year and reported physical abuse during childhood also presented higher NSSI rates (prevalence 55.1% vs. 14.5%, χ², 23.327; adjusted P < .001). Those who experienced emotional abuse, but not sexual bullying reported higher NSSI prevalence when they also experienced emotional neglect in childhood (prevalence 12.4% vs. 5.6%, χ², 94.941; adjusted P < .001).

4. Discussion

The current study explored how distal factors, such as different types of childhood maltreatment, and proximal factors, such as different types of bullying victimization interact to influence college students’ NSSI behaviours by a decision tree model. After controlling for demographic factors, the study found that emotional abuse during childhood had a significantly stronger connection to adult recent NSSI than any other types of childhood abuse and recent bullying victimization in both logistic regression, multivariate logistic regression, and in the results of decision trees.

Emotional abuse, despite its prevalence in nearly half of the sample (Vachon et al., 2015), has often been overlooked by clinicians and researchers due to the misconception that it is the least harmful form of abuse. However, recent studies have shed light on the significant impact of emotional abuse. Turner et al. found that emotional abuse is the type of maltreatment most strongly associated with thoughts of self-harm in youth (Turner & Colburn, 2022). Additionally, a meta-analysis by Liu et al. provided compelling evidence that childhood emotional abuse has a greater association with NSSI compared to other subtypes of abuse, among both adolescents and adults (Liu et al., 2018). Our current study’s findings align with these previous studies, highlighting the importance of emotional abuse in contributing to NSSI behaviour (Liu et al., 2018; Turner & Colburn, 2022). These findings further underscore the significance of childhood emotional abuse in relation to the risk of NSSI, surpassing other types of childhood maltreatment and recent bullying victimization. Some theoretical approaches, such as the Developmental Psychopathology Network (Yates, 2004) and the Biosocial Model proposed by Marsha Linehan (Linehan, 1993), establish a connection between childhood traumatic experiences with primary caregivers and the development of dysfunctional emotion regulation, which can lead to the adoption of maladaptive coping skills like NSSI. Additionally, childhood emotional maltreatment may indirectly impact NSSI behaviours while also potentially directly influencing dysfunctional behaviours in adulthood. Recent evidence suggests that maltreatment can modify the developmental trajectories of the brain, affecting sensory systems, network architecture, and circuits involved in threat detection, emotional regulation, and reward anticipation (Teicher et al., 2016). These alterations in emotional and reward systems may contribute to the occurrence of NSSI (Cummings et al., 2021). Moreover, studies have demonstrated that poor emotion expressivity, rather than emotion coping, mediates the relationship between childhood experiences of emotional abuse and NSSI (Thomassin et al., 2016). These findings emphasize the importance of recognizing and addressing childhood emotional abuse, as it has the potential to influence an individual’s emotions and behaviour throughout their lifespan.

Moreover, our study found that, apart from physical neglect, other forms of childhood trauma are significant predictors of recent NSSI behaviours and the severity of NSSI independently. Notably, unlike childhood maltreatment that have a considerable influence, proximal events such as bullying victimization – especially past-year sexual harassment and cyberbullying – better predicted severe NSSI behaviours (greater than 5 times) rather than less severe ones (less than 5 times). Furthermore, past-year physical bullying victimization did not predict any degree of NSSI behaviour.

Based on current research and previous meta-analyses, it seems that physical neglect during childhood and physical bullying victimization during adulthood do not have an independent impact on NSSI. Actually, when comparing the impacts of physical-related trauma, such as physical abuse and neglect, with emotion-related trauma on self-harm risk, the previous findings are mixed. In a meta-analysis conducted by Liu et al., which included individuals who had experienced only physical neglect during childhood and found less significant relationship with NSSI while physical neglect and other types of childhood maltreatment can overall increase the risk of NSSI (Liu et al., 2018). The variation in the impact of physical neglect and emotion abuse on NSSI may be partially attributed to cultural factors. Specifically, within the traditional Chinese conception of education, physical neglect and even mild corporal punishment are sometimes regarded as effective strategies for fostering children’s resilience. This cultural perspective may contribute to a decreased sensitivity among Chinese individuals towards the potential consequences of physical abuse. However, it is crucial to acknowledge that many countries have implemented legal bans on all forms of physical punishment for children. Additionally, several intervention strategies have been proposed to reduce its prevalence among parents (Gershoff et al., 2017). This is because physical abuse alone may not be a significant predictor of NSSI, but it does increase the risk of NSSI when combined with other forms of abuse. Our decision tree model provides evidence to support this finding. The decision tree model has identified distinct patterns for depression and non-depression models. In the depression model, emotional abuse emerges as the most significant predictor. Among individuals who experienced childhood emotional abuse and were depressed, 35.8% reported a history of NSSI in the past year. However, if they also experienced physical abuse during childhood, the probability of NSSI increased to 43.1%. Similarly, in the non-depression model, individuals who were currently non-depressed but had experienced childhood emotional abuse and recent sexual bullying, when combined with childhood physical abuse, had an increased risk of NSSI from 14.5% to 55.1% compared to individuals without physical abuse. Therefore, it is important to recognize that physical punishment, when combined with other risk factors such as depression and emotional abuse, significantly increases the likelihood of NSSI.

Moreover, our findings indicate that childhood emotional abuse has a significant impact on individuals, increasing the risk of NSSI for both depressed and non-depressed individuals. Additionally, our results suggest that clinically relevant depressive symptoms play a moderating role in the relationship between childhood trauma, recent bullying victimization, and NSSI. Specifically, in depressed individuals, childhood sexual and physical abuse pose a greater risk compared to recent bullying victimization. On the other hand, in non-depressed individuals, verbal bullying and sexual harassment in the most recent year present higher risks compared to childhood sexual and physical abuse. This finding is consistent with previous research indicating that emotional abuse may directly affect an individual’s emotional regulation abilities, while sexual abuse and physical neglect may indirectly influence NSSI through the development of depression and anxiety (Brown et al., 2018). A recent review assessed universal strategies for preventing NSSI and found that none of the evaluated programs were effective in reducing the incidence or frequency of NSSI (Bürger et al., 2023). Our findings highlight the necessity of implementing targeted prevention interventions customized to individuals with varying depressive states. For individuals exhibiting symptoms of depression, it may be crucial to prioritize addressing their experiences of childhood maltreatment and difficulties with emotion regulation. Conversely, for non-depressed individuals, the implementation of more comprehensive anti-bullying school policies may prove to be effective.

5. Limitations

This study had several limitations. Firstly, the use of self-reporting which created potential issues with critical variables like bully victimization, not being objectively examined. Furthermore, the study exclusively focused on ‘campus bullying’, which may limit the generalizability of the findings to other bullying situations, such as community bullying. Additionally, data on childhood traumatic events relied on recall, which may have introduced recall bias. Secondly, the cross-sectional design of the study makes it difficult to ascertain causality among different variables. Follow-up studies with a longitudinal design would be needed to investigate how childhood abuse, followed by bullying victimization, contributes to the initiation and continued exacerbation of NSSI. Thirdly, this study only focused on potential risk factors for NSSI, and no protective factors were considered. A stable social support system and timely psychological support, for example, can effectively reduce the occurrence of NSSI. Including protective factors in the model would have provided insight into which factors exacerbate the occurrence of NSSI in the presence of childhood abuse and which become buffers to prevent NSSI from occurring. Fourthly, the present study may be susceptible to selection bias due to its exclusive focus on samples from northeastern China, leading to a decreased lifetime prevalence of NSSI (8.3% compared to the 16.9% prevalence among adolescents). Subsequent studies should strive for more balanced and geographically diverse sampling. Lastly, while the CHAID method has its advantages, the large quantity of terminal nodes and relatively small number of participants in each node could cause information overload. Additionally, due to the low number of cases of NSSI in the sample (current prevalence rate is 5.2%), the model showed low sensitivity, which led to decreased accuracy.

6. Conclusions

In this study, we examined the correlation between clinically relevant depressive symptoms, childhood maltreatment and recent bullying victimization, and likelihood of engaging in NSSI using logistic regression and CHAID decision trees. Our findings reveal that childhood emotional abuse has the strongest predictive power for NSSI. Additionally, physical, and sexual abuse, verbal bullying, and cyberbullying are also significant determinants. Moreover, we observed that clinically relevant depressive symptoms modify the impact of recent and childhood factors on NSSI behaviour. These results underscore the necessity of assessing multiple factors such as childhood maltreatment, contemporary bullying, and depressive symptoms when evaluating the severity of NSSI. Consequently, our study underscores the necessity of implementing tailored prevention interventions customized to individuals with varying depressive states.

Contributors

CRS, WHG were responsible for the conception, organization, and execution of the study, and WHG was responsible for the statistical analysis and verification of the underlying data and the manuscript preparation. CRS, XSC, WYY, WSH were responsible for the manuscript revision. CRS was responsible for project supervision. All authors had full access to all the data in the study and confirmed their responsibility for the decision to submit it for publication.

Data sharing statement

The dataset for this specific manuscript is available from the corresponding author upon request.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

Dr Runsen Chen was supported by the Research fund of Vanke School of Public Health, Tsinghua University [grant number 2021PY001].