Background

Self-injurious behavior (SIB) is a phenomenon whereby a person directly and deliberately damage themselves [1]. Various terms are used in the literature, such as, nonsuicidal self-injury (NSSI), deliberate self-harm (DSH), self-cutting, and self-harm [2,3,4,5,6]. SIB can be considered as a class of behaviors on a spectrum from NSSI—in which the person has no intent to die—to suicidal attempt (SA), which is a form of suicidal behavior [1]. Although NSSI and SA are two distinct behaviors, it is not always easy to decide whether there was a suicidal intent behind SIB, and in this case, it is difficult to separate them [1, 3]. Moreover, the comorbidity between NSSI and suicidal behaviors is approximately 50% in traditional and 70% in clinical populations [1, 7, 8]. Previous research has indicated that the higher risk of SA is associated with the following in regard to NSSI: greater frequency, more methods, and longer duration [1].

Several meta-analyses conducted in the past 20 years have focused on the prevalence of SIB [6, 9,10,11,12]; see Table 1).

Table 1 Previous meta-analyses
Full size table

Based on the previous meta-analyses, the prevalence of SIB shows a significant increase in the 1990s, but then a stagnation from 2005 [9, 10]. Gillies et al. (2018) found that the lifetime prevalence increased over time from 1990 to 2015, and Muehlenkamp et al. (2012) found no significant difference in the prevalence of NSSI and DSH between 2005 and 2011 [6, 10]. According to previous meta-analyses, between 1990 and 2015, the average lifetime prevalence of SIB among adolescents is between 16.9 and 19.7% [6, 9,10,11,12].

There are conflicting results about the gender differences in the prevalence of SIB [6, 9,10,11,12]. Some research has indicated that females have a lifetime prevalence of SIB that is two to three times higher than males [11], whereas other studies have found that the gender difference does not appear at all or, if it does, to a lesser extent [9, 10].

As shown in Table 1 the results are conflicting in regard to whether there is a difference in the prevalence of NSSI and DSH [6, 10, 12]. Muehlenkamp et al. (2012) did not find any significant difference between the two phenomena, however Gillies et al. (2018) and Lim et al. (2019) found a significantly higher prevalence of NSSI than DSH [6, 10, 12]. The lowest lifetime prevalence of NSSI was 18%, while the highest was 22.9% [6, 9,10,11,12]. At the same time, the lowest lifetime prevalence of DSH was 11.4%, while the highest was 16.1% [6, 9,10,11,12]. As mentioned above, unlike NSSI, DSH can be suicidal and nonsuicidal as well, but it must be a non-fatal self-harm [6, 12].

There are also conflicting results about the continental differences in the prevalence of SIB. While Swannell et al. (2014) didn’t find any significant continental differences in the prevalence of SIB, Gillies et al. (2018) did: Sweden had the highest, Norway had the lowest prevalence of self-harm in the meta-analysis [6, 9]. Lim et al. found that the lifetime prevalence of both NSSI and DSH were higher in non-Western countries than Western countries [12].

The above described previous meta-analyses highlight that these conflicting results can be due to the large differences among the included studies in methodological factors (e.g., sampling method, measurement, mean age of the sample), in the definitions of SIB (e.g., with or without suicidal intent), or in the place of data collection [6, 9,10,11,12].

Aims

Because of the heretofore conflicting results, our aim in this study was to follow up previous meta-analyses on the prevalence of SIB in adolescent community samples [6, 11]. We focused on data published between 2015 and 2020.

Our first hypothesis was that the prevalence of SIB did not change over time between the examined period (2015 and 2020) for both females and males. Our second hypothesis was that females reported a higher prevalence of a history of SIB than males.

Before addressing these hypotheses, we investigated the following questions: (a) What kinds of definitions of SIB are used? (b) What kinds of assessments are used to measure SIB? (c) What was the sampling method? (d) Where were the data of the included studies collected? (e) What was the mean age of the sample? (f) Was suicidal intent excluded from the definition of SIB?

Method

Literature search

We conducted a systematic literature search on June 18, 2020. We used six computerized literature databases: PubMed, Scopus, Web of Science, OVID Medline, PsycINFO, and EBSCO Discovery Service for Semmelweis University. Search terms were the following: “non-suicidal” OR nonsuicidal OR “self-injur*” OR “self-harm” OR “self harm” OR parasuic* OR “self-mutilation” AND prevalence AND adolescen* NOT “clinical trial” OR “case report” OR review. Two filters were added: (a) date between January 1, 2015, and June 18, 2020, and (b) English language. We used EndNote X9 software to remove duplicates and screen the 374 search hits.

Inclusion and exclusion criteria

To be included, studies had to report on the prevalence of SIB in adolescence in a community sample in a published article written in English. We used “adolescen*” among our search terms, and we included studies with an age range 11–18 years. However, in some articles this age range was wider (22 years being the oldest), so from these we included in the final analysis only those studies where the mean age of the sample was adolescence. For details, see Table 2.

Table 2 The included relevant studies
Full size table

When multiple studies reported on the same database, we included the ones with the largest sample size [13], the ones that provided data separately for males and females [14], and the ones that provided follow-up results [15, 16]. This led to the exclusion of six studies [17,18,19,20,21,22]. In addition, we contacted by email the authors of articles from which prevalence data could not be extracted. In case we did not receive sufficient statistics, we excluded the study (e.g., Carvalho et al., 2015). The methodology of this review follows the PRISMA guidelines [23].

Data extraction

Two authors (BFF, NK) coded the following information:

  1. (1)

    bibliographic information: authors, year of publication and data collection;

  2. (2)

    sample information: age range and mean age of sample, gender ratio, country, and continent the sample was recruited in, representativeness of the community sample, design;

  3. (3)

    measurement of SIB: measurement instrument, suicidal intent, terminology;

  4. (4)

    information for effect size: prevalence estimate and sample size.

Interrater reliability ranged from 73 and 100%. In case a consensus could not be reached between the two coders, the other two authors were consulted (ZKT, JB).

To test our hypotheses, we preferred to include the prevalence estimates separately for males and females if a study reported on those. For longitudinal studies, prevalence at all measurement points was coded; however, they were averaged to calculate an effect size for a study before we included the data in any analyses. We made an exception when prevalence estimates were available separately for males and females at one time point but not at another. In those cases, we chose to include only the estimates at the time point when they were reported separately for males and females.

During the coding, we had to impute some scores that were not reported in the primary studies. For studies that reported only the age range, we imputed the mean age as the geometric mean of the range. For studies that did not report the year of data collection, we subtracted 2 years from the year of publication (for a similar procedure, see Protzko et al., 2020) [24].

Statistical analyses

We used the Comprehensive Meta-Analysis software to conduct the analyses [25, 26]. We applied a random effects model. When a study reported results at more than one time point, we entered all in the software, which takes the average between multiple time points before entering a study in the grand average. We made an exception when conducting meta-regression analyses regarding the year of data collection and the mean age of the sample. In these cases, we only selected the first time point from these longitudinal studies to be included. In contrast, we considered estimates for males and females when reported separately in a study as independent effect sizes in all the analyses. Outliers were inspected based on a standardized residual exceeding ±3.29. We inspected the results according to several moderator variables. When inspecting results according to the different continents and suicidal intent, we conducted a subgroup analysis to statistically contrast them. We only included subgroups with at least four effect sizes in this analysis (for a similar procedure see Takacs and Kassai 2019) [27].

Results

Included studies

In sum, a total of 97 articles were included in this meta-analysis; we identified 178 effect sizes (see Figs 1, 2 and Table 2).

Fig. 1
figure 1

The selection process is summarized in the QUORUM flowchart

Full size image
Fig. 2
figure 2

The flowchart of inclusion and exclusion criteria

Full size image

There were six outlying effect sizes that we excluded. Altogether, we had data from 439,818 participants. The overall average SIB prevalence in the studies was 16.0% (95% confidence interval [CI] [14.7, 17.4], k = 172). This was a heterogeneous effect, Q(171) = 30,136.96, p < 0.001, I2 = 99.43 τ2 = 0.44.

In our assessment of publication bias, Egger’s test showed significant asymmetry (intercept = −2.88, p = 0.046), but the funnel plot showed a symmetric distribution based on visual inspection, which was confirmed by no imputed studies in the Duval and Tweedie’s trim-and-fill procedure.

Among the 97 included articles, 74 reported prevalence data for females and males separately. There were 79 effect sizes reported for females. Two effect sizes were outliers and thus were excluded. We found an average prevalence of 19.4% for females, 95% CI [17.5, 21.4], k = 77. This effect was heterogeneous, Q(76) = 8,660.74, p < 0.001, I2 = 99.12 τ2 = 0.29. There were 75 effect sizes reported for males. One outlying effect size was excluded. We found an average prevalence of 12.9%, 95% CI [11.3, 14.8], k = 74. Again, this was a heterogeneous effect, Q(74) = 10,315.75, p < 0.001, I2 = 99.2 τ2 = 0.43.

Terms and definitions of SIB in the included studies

The terminology of SIB was not uniform across the included studies. All the studies defined SIB as a deliberate damage to oneself, but not all of them defined it as a nonsuicidal intent. Seventy-two articles (73.5%) made a clear distinction between suicidal and nonsuicidal intent.

There were 11 different terms for SIB in the included 97 papers. The most frequently used term was NSSI; this expression appeared in 60 articles (see Table 3).

Table 3 Prevalence differences in the terms of SIB
Full size table

Measurements of SIB in the included studies

Among the included studies we found diagnostic interviews, self-reported questionnaires, and single-item questions to measure SIB. Two studies measured NSSI based on DSM-5 criteria [122]. The most frequently used questionnaire was the Deliberate Self-Harm Inventory [123], which was mentioned in 13 articles. The Inventory of Statements About Self-Injury [124] was used in five studies, and the Functional Assessment of Self-Mutilation [125] also was used in five. Effect sizes based on a single item to assess SIB found an average prevalence of 11.6%, 95% CI [9.3, 14.5], k = 31. We found of 14.8%, 95% CI [12.8, 17.2], k = 60, in studies that used nonvalidated questionnaires. Questionnaires that had been validated for other constructs showed an average prevalence of 14.7%, 95% CI [9.8, 21.5], k = 6. Finally, questionnaires that had been validated for SIB showed the highest average percentage: 18.9%, 95% CI [16.9, 21.1], k = 77. For results separately for males and females, see the Supplementary Materials. Only one study used a diagnostic interview and reported on two effect sizes. The average of these showed a similar estimate as the grand average (14.2% (95% CI [7.7, 24.8], k = 2), more specifically, 18.7% for females and 10.2% for males).

Sampling

Of the 172 effect sizes, 99 were based on convenience sampling. These showed an average prevalence of 15.2%, 95% CI [13.4, 17.2]. Eighteen effect sizes were based on samples that applied randomization, showing a prevalence of 24.7%, 95% CI [18.9, 31.6]. For 55 sample sizes, the sample was representative of the population. Representative samples showed a pooled prevalence of 15.1%, 95% CI [13.2, 17.4]. A similar pattern was noted for females and males (see the Supplementary Materials).

Place of data collection of the included studies

From the 98 included articles, we found three collaborations in which data were collected in multiple countries; for the rest, the data were collected in single countries. When we inspected the results over all the effect sizes, we noted differences according to the continent on which the data had been collected. There were three effect sizes in two publications from South America that showed an average prevalence of 33%, 95% CI [13.7, 60.3], and we found two effect sizes in one publication from Africa that showed an average prevalence of 24.4%, 95% CI [19.1, 30.7], and data for two effect sizes published in the same article were collected in North America and Australia and Oceania as part of an international cooperation that showed a prevalence of 2.6%, 95% CI [1.3, 4.9]. These categories were excluded from the subgroup analysis as they contained less than 4 effect sizes. After we excluded these, we noted a significant difference between the prevalence estimates from the different continents (see Table 4), Q(3) = 10.97, p = 0.012. More specifically, prevalence estimates from Asia (19.5%, 95% CI [17.1, 22.2], k = 51) were significantly larger than those from the other three continents (14.6%, 95% CI [13.1, 16.2], k = 114), Q(1) = 11.20, p = 0.001. As shown in Table 4, the effect of continent was similar when we inspected effect sizes for female and male samples separately.

Table 4 Prevalence differences in continental distribution
Full size table

Mean age of the included samples

For assessing the effects of the mean age of the samples, we chose to focus on the first measurement point in the 17 longitudinal studies. In this analysis, seven outliers appeared that were then excluded. For an additional six effect sizes we could not extract the sample’s age, and thus those were also excluded from this analysis. This resulted in 165 effect sizes. The mean age of the sample ranged from 11.00 to 18.53 years. The mean age of the sample did not have a significant effect on the effect size (coefficient = 0.067, p = 0.12). For results separately for males and females, see the Supplementary Materials.

To make sure that longitudinal studies from which we chose to include the first estimate in this analysis did not influence the results by possibly reporting on substantially younger samples, we also ran the regression model on the cross-sectional studies only as a sensitivity analysis. This resulted in 134 effect sizes to be included. Again, the mean age of the sample did not have a significant effect on these prevalence estimates (coefficient = 0.058, p = 0.24).

Suicidal intent

For 125 effect sizes, suicidal intent was excluded. Those showed a pooled estimate of 18.3%, 95% CI [16.7, 19.9]. This was significantly higher than what was found in studies that did not exclude suicidal intent (11.3%, 95% CI [9.3, 13.7], k = 47), Q(1) = 20.52, p < 0.001. This pattern was also confirmed in only-female and only-male samples. For results separately for males and females, see the Supplementary Materials.

Prevalence of SIB

We found 92 effect sizes reporting on lifetime prevalence of SIB, 72 effect sizes that estimated 1-year prevalence, and 17 that estimated 6-month prevalence. An average of 17.9%, 95% CI [16.3, 19.5], was found overall when lifetime prevalence was assessed. This estimate was 22.9 (95% CI [20.9, 25.0], k = 42) for females and 13.7% (95% CI [11.2, 16.8], k = 39) for males.

An overall average prevalence of 13.4%, 95% CI [11.5, 15.6] was found when assessing prevalence in the last year. This estimate was 15.9% (95% CI [12.9, 19.4], k = 32) for females and 10.7% (95% CI [8.7, 13.2], k = 32) for males. An overall prevalence of 16.2%, 95% CI [11.0, 23.3] was estimated when we considered only the last 6 months, 18% (95% CI [8.9, 33.2], k = 7) for females and 13.8% (95% CI [6.2, 27.9], k = 7) for males.

Year of data collection

In regard to assessing the effects of the year of data collection, we chose to focus on the first measurement point in the 17 longitudinal studies. In this analysis, seven outliers appeared that were then excluded. Data for the primary studies were collected between 1998 and 2018. The year of data collection had a significant, positive effect on the 171 effect sizes (coefficient = 0.035, p = 0.008); that is, more recent studies found larger prevalence. For results presented separately for females and males, see the Supplementary Materials.

For further investigation, we restricted the year of data collection to 2013 and onward so we could assess the effect in the time constraints that corresponds to the time constraints of year of publication of the present meta-analysis (2015 and onward). Data for 119 effect sizes were collected in or after 2013. When we considered only these studies, the effect of year of data collection was not significant on the effect sizes (coefficient = −0.015, p = 0.72). The same was found for the 53 effect sizes for females (coefficient = −0.005, p = 0.92) and for the 51 effect sizes for males (coefficient = −0.05, p = 0.49). Scatterplots are shown in the Supplementary Materials.

Risk of bias

Risk-of-bias criteria was based on the Cochrane Risk of Bias Tool [126], adapted for the studies (cohort, cross-sectional, and longitudinal).

Discussion

Because previous meta-analyses have yielded conflicting results on the prevalence of SIB in community adolescent samples [6, 11], we found it important to complete a follow-up meta-analyses with clear methodology on recently published data. In the present meta-analysis, we found that the prevalence of SIB in adolescents was 16% in studies published between 2015 and 2020. This result is comparable to the estimate of 16.9% found in a previous meta-analysis [6]. Regarding methodological differences, as can be expected, a slightly higher estimate was found when considering lifetime prevalence (17.9%) as compared with the 1-year (13.4%) or 6-month prevalence (16.2%). We also noted a significantly higher prevalence when suicidal intent was excluded (18.3%) than when it was not excluded (11.3%), and the largest prevalence was found when measurement instruments were used that had been validated for self-injurious behaviors (18.8%). In addition, methodologically more rigorous studies that focused on representative samples found an average SIB prevalence of 15.1%. This is an interesting issue, while self-harm without suicidal intent should be a subgroup of self-harm covering forms both with and without suicidal intent. Hence the first number should always be lower than the second number. A possible explanation could be that the studies used the same term but actually employ different criteria. In addition, differences in the prevalence of NSSI and DSH may also result from measurement differences between the two types of SIB. Previous meta-analyses have reported higher prevalence rates for multi-item instruments [6, 10], and 65.5% of NSSI measurements consisted of multiple items, compared to 60% of DSH measurements consisting of a single item. In a meta-analysis made by Swannell et al. (2014), checklist versus single-item measurement explained the 41% of variance between studies [9]. Our review shows that, among adolescents, there are no significant changes in the prevalence between ages 11.0 and 18.5 years. This result is comparable to Lim’s meta-analysis [12] but does not align with Gillies and colleagues’ (2018) study. We found similar prevalence estimates among studies that used convenience and representative samples. However, and surprisingly, studies that used a random sample found larger estimates. This is puzzling and needs further research.

Our first hypothesis was only partially confirmed. When we considered all data that were published between 2015 and 2018, we found that there was a significant increase between 1998 and 2018 in the prevalence of SIB. However, when we restricted our analysis to the time frame between 2013 and 2018 (to reflect the publication time window of 2015 and 2018), we found no change in prevalence, as we had expected. Previous meta-analyses have found mixed results regarding this question. Muehlenkamp and colleagues (2012) did not find any significant difference in the prevalence of SIB between 2008 and 2015, whereas Gillies and colleagues (2018) found an increase between 1990 and 2018. Our results are in line with both previous findings in that they show an increase before 2013, but no change since then. This finding can be important to both decision makers and professionals for the appropriate planning of prevention programs.

Our second hypothesis was confirmed; we found a substantial difference between the estimates for females (19.4%) and males (12.9%), with nonoverlapping confidence intervals. A similar pattern was observed when we considered only lifetime prevalence, with 22.9% for females and 13.7% for males. These results are comparable to Bresin and Schoenleber’s (2015) meta-analysis, in which the prevalence was significantly higher among females. Studies that excluded suicidal intent found an average of 21% for females and 16.5% for males. Similarly, estimates based on measurement instruments that were validated for SIB showed 21.9% for females and 15.7% for males. On the other hand, studies with representative samples showed slightly lower estimates: 18.2% for females and 10.9% for males. Thus, the patterns were very similar for females and males when the effect of methodological differences in the primary studies were assessed. However, the cultural difference between countries in Asia and those on other continents was more articulated for males (Asia = 19.5% vs. other = 10.8%) than for females (Asia = 22.3% vs. other = 18.5%). Nock and Prinstein (2005) found that NSSI often is connected to psychological distress [127], and adolescent girls usually have more psychological distress than men [128]. These results highlight that it is necessary to pay more careful attention to NSSI by female adolescents and that perhaps further attention should be given to Asian male populations.

We found some differences in the prevalence estimate as a function of methodological differences among the primary studies; however, we should note that moderators might be confounded.

The pooled estimate from Asian countries (19.5%) was significantly higher than that from other continents (14.6%). Again, this confirms earlier meta-analytic results estimating a relatively large prevalence in Asian countries (Lim et al. 2019). This difference was even more articulated for males. So, it may be that the differences in SIB between Asian and non-Asian countries are somehow connected to gender. To understand this result, further research should focus on the transcultural aspects of SIB.

We did not find a difference between NSSI (18.7%) and DSH (15.1%), unlike Gillies’s results [6], but we found a substantial difference between NSSI (18.7%) and self-harm (12.7%). In contrast to females, we found a substantial difference between NSSI (17.1%) and DSH (6.3%), and between NSSI (17.1%) and self-harm (9%) among males.

Our review highlights that the highest prevalence rates were found when SIB was measured with a validated questionnaire as compared with studies that used single-item or nonvalidated questionnaires, a pattern that was also confirmed separately for male and female samples. This result is likely due to the fact that validated questionnaires are more sensitive than single-item measures [6, 10].

Our results are limited by the heterogeneity of the primary studies, that is, in regard to the sample and the measurement instruments and the conceptualization of SIBs. The findings of the present meta-analysis confirm that these differences among the primary studies have an important effect on the prevalence estimates. There is currently no consensus in the literature about the conceptualization of SIB [6, 10], which makes our work more difficult when evaluating the data. However, to provide the most precise estimate, we pooled the studies that used representative samples that reported on lifetime prevalence of SIBs excluding suicidal intent measured by a validated measurement instrument and found similar estimates. In addition, we did this to avoid a confound effect of these moderators. Moreover, although overall we found a relatively large number of studies that reported on prevalence of SIBs, it is questionable whether nonsignificant results in subgroup and meta-regression analyses are truly due to an absence of an effect or whether they are instead due to a lack of statistical power.

To our best knowledge, this is the most recent meta-analysis on the prevalence of SIB among adolescents. An overall prevalence of 16% was found, which means that one in six adolescents has a history of self-harm. Moreover, a larger estimate was found for females as compared with males: every fifth adolescent girl reported having conducted self-harm. It is interesting that estimates were largest in Asian countries with males, approaching a 20% prevalence. Further research should focus on the transcultural aspects of self-harm to understand this difference. All these results have public health importance in drawing the attention of clinicians and decision makers to adolescents who engage in SIB. Clinicians need to be aware of the high prevalence and risk factors (e.g., female gender, Asian populations) of SIB in adolescence. Prevention and intervention are very important in this age group.