The present study had a prospective, longitudinal design and was part of a large research project, which is following a community cohort of compulsory school students in a Swedish middle-sized municipality (around 40,000 inhabitants).

Participants

The community cohort in this study comprised all compulsory school students in Grade 7 and 8 (N = 1,064) in 2007. Students attending schools for students with learning disabilities were not included, however. Normally, students in Sweden start in the 7th grade the year when they turn 13. The study includes three data collection waves. Of the students in the cohort, 991 (93%; 50.3% girls) participated in the data collection at Time 1 (T1) in 2007. In a new data collection one year later, at Time 2 (T2) 984 (out of 1,098 eligible) students participated (90%; 51.1% girls). The total number of eligible students at T1 and/or T2 was 1,109, which was the target sample for the data collection at the 10-year follow-up (T3) in 2017. Of the individuals in this sample, 557 participated (response rate: 50.2%; 59.2% women). In the present study, we included those participants (n = 475) who had data on NSSI from all three waves of the project. Table 1 presents the descriptive statistics at T3 for this longitudinal sample. Most participants were married/cohabiting or were in a relationship (63.2%) and were part or full-time employees (61.5%) at T3.

Table 1 Demographic Characteristics at T3 of the Longitudinal Sample (N = 475) MeasuresNon-suicidal self-injury (NSSI)

The Deliberate Self-Harm Inventory, short 9-item version (DSHI-9r) was used to assess self-harm across three waves. DSHI-9r is a shortened and modified version of Gratz’s Deliberate Self-Harm Inventory (DSHI; [47]), adapted to Swedish adolescents [48, 49] and then revised [29]. Respondents were instructed to rate how often they had deliberately engaged in nine different self-harm behaviours (i.e., cutting, minor cutting causing bleeding, burning, punching/banging oneself, biting, carving, severe scratching, sticking sharp objects into skin, and preventing wounds from healing) during the past six (T1 and T2) or twelve (T3) months, on a scale from 0 (“never”) to 6 (“more than five times”). The scores on the nine items are summarized into a total NSSI score. The DSHI-9r shows good test–retest reliability [49]. In the present study, Cronbach’s alpha for DSHI-9r were 0.90 (T1), 0.89 (T2) and 0.81 (T3).

RepNSSI

Repetitive NSSI was defined as reports of at least 5 instances of self-harm during the past six (T1 and T2) or twelve (T3) months. A dichotomous measure of presence or not of repetitive NSSI was labeled repNSSI and computed from the total NSSI score. Total scores of 0 to 4 rendered a value of 0 for repNSSI, and total scores of 5 or higher rendered a value of 1. Please note that a total score of 5, for example, could be achieved in different ways: by reporting one instance of self-harm with each of five different methods, by reporting five instances of self-harm with one and the same method, or with some other combinations.

Individual developmental pathways of repNSSI

The individual developmental pathways were represented by three-digit value patterns formed by the individuals’ values in the measure of repNSSI at the three different time points. The pathway of an individual not reporting repNSSI at any of the three time points was represented by the value pattern 000; an individual reporting repNSSI only at the first time point was represented by the value pattern 100; and so on. With dichotomous measures of repNSSI at three time points there were eight (2 × 2 × 2) theoretically possible value patterns. All eight value patterns and their observed frequencies in the longitudinal sample are presented in Table 2, together with the mean values (and SDs) on total NSSI at T1, T2, and T3 for the eight groups with the different value patterns.

Table 2 Participant Categorization According to Individual Longitudinal Value Patterns (N = 475): (a) NSSI Total at T1, T2 and T3 (Mean and SD); (b) Configural Frequency Analysis (Observed and Expected Frequencies); and (c) Gender Distribution for Each Value Pattern Predictors measured at T1 and T2 Psychological difficulties

Participants completed the SDQ-s [50] at T1 and T2. SDQ is a widely used screening instrument for psychological difficulties among children and adolescents and contains five subscales with five items each. Four of these measure difficulties: emotional symptoms, hyperactivity-inattention, conduct problems, and peer problems; and the fifth subscale measures prosocial behaviour. Each item is rated on a 3-point scale (0 = true, 1 = somewhat true, and 2 = certainly true) with a time frame of the last 6 months. Five items on the difficulties scales are worded positively and reversed before scoring. In the present study we only used the Total Difficulties score which is the sum for the four difficulties scales. The SDQ was translated into Swedish by Smedje et al. [51], and the self-report version was empirically validated by Lundh et al. [52], who reported a test-retest reliability of 0.72. In the present study, the internal consistency of the total difficulties scale was α = 0.76 and α = 0.75 at T1 and T2, respectively.

Depressive symptoms

A Depression Index (DI) was constructed [30] by selecting depression-relevant items from the 11-page questionnaire used at T1 and T2, according to their correspondence with the DSM-IV criteria for major depression [53]. Because the items came from different instruments with different response formats, the scores on each item were transformed to z-scores, before computing the total DI score which was used in the present study. Items referring to positive feelings were reverse scored. The total DI included all items from six subscales: Dysphoric Relations to Parents (10 items); Negative Self-Image (6 items); Dysphoric Relations to Friends (6 items); Fatigue/Somatic Complaints (5 items, including a question about poor sleep, see below); Sadness/Loneliness (4 items); and Difficulties in Concentration (4 items). Test-retest correlations between Time 1 and Time 2 were r = .71 for the total DI. In the present study, Cronbach’s alpha for the DI was 0.91 at both T1 and T2.

Poor sleep

Poor sleep was assessed by means of one single question, “Do you sleep well?”, with a Likert response format and five response alternatives: 1 = always, 2 = most often, 3 = sometimes, 4 = seldom, and 5 = never. A pilot study with 80 adolescents who answered this question on two occasions, with a mean test interval of 7 weeks and 4 days, showed a test-retest correlation of r = .64 [31]. Because the periods between assessments were longer than is usual in studies of test-retest reliability (which should ideally not be more than about 1 month) the test–retest coefficient obtained was assumed to set a lower boundary for the true test-retest reliability of this measure.

Psychological outcomes measured at T3

To investigate the psychological adjustment in young adulthood among groups of individuals following different developmental pathways, we used a set of measures of both positive and negative adjustment at T3.

Life satisfaction

The Satisfaction with Life Scale (SWLS; [54]) contains 5 items (e.g., “I am satisfied with life”). Participants indicate how much they agree or disagree with each of the 5 items using a 7-point scale that ranges from 1 (Strongly disagree) to 7 (Strongly agree). Cronbach’s alpha for the scale was 0.92.

Flourishing

The Flourishing Scale (FS; [55]) is a brief 8-item summary measure of the respondent’s self-perceived success in important areas such as relationships, self-esteem, purpose, and optimism. Participants indicate how much they agree or disagree with each of the 8 items (e.g., “I lead a purposeful and meaningful life”) using a 7-point scale that ranges from 1 (Strongly disagree) to 7 (Strongly agree). The possible range of scores is from 8 (lowest possible) to 56 (highest possible). A high score represents a person with many psychological resources and strengths. Cronbach’s alpha for the scale was 0.88.

Resilience

Resilience was assessed with the Brief Resilience Scale (BRS; [56]) that assesses one’s ability to bounce back or recover from stress (e.g., “I tend to bounce back quickly after hard times”). Participants indicate how much they agree or disagree with each of the five items using a 5-point scale that ranges from 1 (Strongly disagree) to 5 (Strongly agree). The total score is calculated by averaging the item scores. Windle et al. [57] showed that the BRS has sound psychometric properties that are on par with longer measures of resilience. Cronbach’s alpha for the scale was 0.81.

Stress, anxiety and depression

The Depression, Anxiety and Stress Scale (DASS-21; [58]) was used to assess depression (7 items; e.g., “I felt downhearted and blue”), anxiety (7 items; e.g., “I felt I was close to panic”) and tension/stress (7 items; e.g., “I found it hard to wind down”). Participants responded to each item on a 4-point scale that ranges from 0 (never) to 3 (almost always). Cronbach’s alphas for depression, anxiety, and stress were 0.90, 0.79, and 0.87, respectively.

Emotion dysregulation

The Brief Difficulties in Emotion Regulation Scale (DERS-16; [59]) was used to assess participants’ difficulties to regulate emotions, from several aspects including lack of emotional clarity (e.g., “I have difficulty making sense out of my feelings”), difficulties engaging in goal-directed behaviors (e.g., “When I am upset, I have difficulty getting work done”) and controlling impulses (e.g., “When I am upset, I become out of control”), ineffective emotion regulation strategies (e.g., “When I am upset, I believe that I will remain that way for a long time”), and non-acceptance of emotional responses (e.g., “When I am upset, I feel ashamed with myself for feeling that way”). Participants estimated how often each of the 16 statements applied to them using a 5-point scale, ranging from 1 (almost never) to 5 (almost always), setting the total score at a minimum of 16 and a maximum of 80. The Cronbach’s alpha for DERS-16 was 0.95.

Procedure

Data collection at T1 and T2 was conducted in collaboration with the municipal body of the selected area and each of the regular schools therein. The headmaster of each school was contacted and agreed to their school’s participation in the study. Data were collected in school settings. Teachers were present but did not take part in the administration, which was conducted by research assistants from Lund University. The students were told that they could feel free to refrain from participation, and that they should not write their names anywhere on the questionnaire to ensure confidentiality. To match the data files from T1 to T2 a pseudo-anonymization procedure was used, which meant that a numeric code was used throughout the research project to designate the identity of the participant on all study documents. The code key was preserved separately from other documents and data files, in a secure place.

To conduct the follow-up at T3, participants’ names from the code lists from the two prior surveys (in accordance with the ethical approval to save the name list of participants for 10 years) were sent to the Swedish state’s personal address register (SPAR) to identify their present locations. After we had received current personal addresses of the participants, letters describing the purpose and procedure of the follow-up were sent to all eligible participants. The eligible participants could complete either a confidential web-survey designed using the Lund University survey system, Survey & Report, or a paper-and-pencil questionnaire. After completion of the survey, participants received two cinema tickets or four lottery tickets as compensation.

Attrition analysesAttrition between T1/T2 and T3

Attrition analyses were conducted comparing the responders (n = 541) and non-responders (n = 529) at T3 in terms of all variables measured at T1 and T2 [60]. We found some significant differences ranging from very small to small in effect size (Cohen’s d/Cramer’s V = 0.02‒0.21) but could not identify any clear patterns that differentiated the responders from the non-responders on the variables observed at T1 or T2.

Regarding the variables relevant for the current study, significantly more women than men responded to the survey at T3 (T1 & T2: 51%, T3: 58.4%; χ2(1) = 29.30, p < .001). No significant differences between the responders and non-responders were found on SDQ total, the depression index, sleep problems or NSSI, neither at T1 nor at T2. At T1, the NSSI means±SDs for the longitudinal sample (N = 475) and the T3 non-responders were, respectively, 3.63±8.92 and 3.04±6.30, t(910.1) = 1.19, p = .235. At T2, the corresponding means±SDs were 3.77±8.50 and 3.33±7.98, t(944) = 0.82, p = .412. Due to the low number of T1/T2 variables reliably associated with attrition and thus viable as predictor/auxiliary variables in multiple imputation [61], non-responders at T3 were excluded from analysis in the current study.

Internal missingness

At each wave, participants having no more than three missing values on the DSHI-9r were included for data analysis. Missing values were interpreted conservatively as absence of the self-injurious behaviour asked for (i.e., imputing 0).Footnote 2 With these imputations we had available data on DSHI-9r for 983 participants at T1, 979 at T2, and 556 at T3. In total, 896 participants had data on DSHI-9r at both T1 and T2, and 475 had data on DSHI-9r at all three waves. These 475 participants comprised the analytic longitudinal sample for this study.

The percentages of participants with internal attrition ranged from 1.3 to 3.2% in the T1 predictor variables, 1.5‒2.9% in the T2 predictor variables, and 1.1‒3.6% in the T3 outcome variables. Little’s [62] Missing Completely At Random (MCAR) test was non-significant (χ2[874] = 815.77, p = .921), suggesting that the internal attrition was MCAR, thereby justifying the inclusion of participants with missing data in the analyses after imputation [63]. Missing data in the different variables were imputed with the Expectation-Maximization algorithm in IBM SPSS.

Statistical analysesConfigural frequency analysis

A First-Order Configural Frequency Analysis (CFA; [19]) was used to explore whether the observed frequencies of different individual developmental value patterns were significantly higher – or significantly lower – than expected by chance in a comparison model. The phenomenon that we wanted to investigate with this analysis was the self-reinforcing property of NSSI, or at least of repNSSI, which should manifest itself in a tendency towards stability of repNSSI.

If NSSI to some extent is self-reinforcing, the engagement in NSSI at one time point would increase the risk of engaging in NSSI also at a subsequent time point. Engaging in NSSI would increase the risk of doing it again. This would mean that presence or absence of repNSSI at the later time point is not totally independent of whether repNSSI was present or not at the previous time point; on the contrary, there would be an association between the time points as concerns presence or not of repNSSI. In general, this type of association over time should increase the frequencies of stable repNSSI pathways. With three time points, however, some theoretical pathways would involve both stability and change, and we had no theoretical expectations or earlier findings on which we could set up specific hypotheses for these patterns. Therefore, the analyses we made were exploratory.

To analyse whether there was an association between presence/absence of repNSSI over time that influenced the observed frequencies of some of the value patterns, we used a comparison model of independence between time points. Based on the independence model, we could compute what frequencies to expect by chance for the various value patterns if the time points were totally independent of each other. Making the computations of expected frequencies, we used the observed marginal frequencies, that is, the observed frequencies of presence and absence of repNSSI at each separate time point. RepNSSI was more than twice as common at T2 (21.7% of the longitudinal sample) than at T3 (9.9%), for example, and such differences in marginal frequencies were included in the comparison model.

For each theoretically possible value pattern, the significance of the difference between observed and expected frequencies was tested with a two-tailed test, according to the binomial distribution. For this we used the computer program ROPstat [64]. Patterns that are significantly more common than expected are called types, and patterns that are significantly less common are called anti-types. It should be noted that whether a pattern is a type or whether it is an antitype is not related to its observed frequency per se. A pattern that is very frequent may be an antitype, and a very low-frequent pattern may be a type. A more elaborate presentation of the CFA analysis is made in Appendix A.

Gender distribution among developmental pathways

The gender distribution among developmental pathways was explored with exact single-cell tests, based on the hypergeometric distribution, of each cell in an 8 × 2 cross-tabulation between developmental pathway groups and gender [64, 65].

Logistic regression analysis

Logistic regression analyses (fulfilling the assumptions of log-odds linearity, and non-multicollinearity) were conducted to predict membership in (a) the Late Onset RepNSSI vs. No RepNSSI groups and (b) the Stable Adolescence-Limited RepNSSI vs. Prolonged RepNSSI groups. The regression analyses were performed separately for predictors measured at T1 and predictors measured at T2. The odds ratio (OR) was used as an index of effect size. The size of the OR was interpreted according to Chen et al. [66], who calculated odds ratio equivalents to Cohen’s d and suggested that OR = 1.68 should be considered as a small effect size (corresponding to d = 0.2), OR = 3.47 as a medium effect size, and OR = 6.71 as a large effect size.

Non-parametric tests of group differences at T3

We also wanted to compare young adult adjustment among the groups following the late onset, cessation, and prolongation pathways as well as the group with stable absence of repNSSI. These groups had unequal sample sizes and non-homogeneous variances, however. Therefore, we used the non-parametric Kruskal-Wallis’ test for the overall comparison of the groups on different psychological outcomes in young adulthood. Post hoc tests were performed with the Games-Howell test, which is a non-parametric test that does not assume homogeneity of variances or equal sample sizes. Effect sizes were estimated with Glass’s delta using the sample standard deviation of the comparison group. All analyses were carried out using IBM SPSS, version 27.