The following information should be cited as:  Fox, K. R., Franklin, J. C., Ribeiro, J. D., Kleiman, E. M., Bentley, K. H., Nock, M. K. (in press). Meta-analysis of risk factors for nonsuicidal self-injury. Clinical Psychology Review.

OVERVIEW

Nonsuicidal self-injury (NSSI) is defined as intentional self-harm performed without suicidal desire (e.g., self-cutting). These behaviors are highly prevalent and are associated with numerous negative outcomes, including future suicidal behaviors. Unfortunately, no intervention has been consistently shown to reduce NSSI compared to control treatments, suggesting a need to identify novel treatment targets.

Risk factors (especially certain types of risk factors) represent strong treatment targets, as these factors precede NSSI and divide people into high and low-risk groups. Unfortunately, very little research has been conducted on risk factors for NSSI. In fact, the majority of NSSI research has focused on correlates, or factors that co-occur with NSSI. While this research has helped our understanding of NSSI, correlates are not typically strong treatment targets. 

Over the past decade, there has been a greater focus on NSSI risk factors, or factors that precede NSSI and divide people into high and low-risk groups. To provide a summary of current knowledge about NSSI risk factors, we conducted a meta-analysis of published, prospective studies longitudinally predicting NSSI. This included 20 published reports across 5,078 unique participants. 

We addressed 3 primary aims with the present random-effects meta-analysis:

  1. What does this literature look like?
  2. What factors predict NSSI? Are there any particularly strong risk factors?
  3. What, if any, factors influence the associations between risk factors and NSSI?
 


How did we identify studies?

 

ANSWER: Before we talk about our findings, it's important that you understand how we found the studies included in our meta-analysis. 

To find eligible studies, we used a wide range of search terms, including suicidal and nonsuicidal behaviors. We did this for two reasons:

  1.  Words used to describe intentional self-harm without suicidal desire have changed a lot over time, and even now people use different terms for these behaviors (e.g., including deliberate self-harm, self-mutilation, and NSSI).
  2. People often study suicidal and nonsuicidal behaviors together because they often co-occur. We included suicide-related terms to make sure we found all longitudinal studies assessing NSSI, even if NSSI wasn't the primary outcome of interest.

Through this wide search, we identified 2,165 unique published reports. Of these reports, only 20 studies were included in the present meta-analysis. Studies were excluded for each of the following reasons:

  • Examination of suicidal thoughts/behaviors (n=520) 
  • Examination of a composite suicidal and nonsuicidal behaviors (n=33)
  • Analyses were not longitudinal (n=13) 
  • Necessary statistics were not provided (n=14) 
  • Major methodological flaws (n=4). 



What does this literature look like?

What types of variables have been used to predict Nssi?

Answer: The earliest NSSI risk factor study was published in 1991,  but the next qualifying study wasn't published until 17 years later.

Of the 20 unique published reports, 16 unique study samples were included. In other words, some published reports used overlapping or identical samples. 

From these 20 reports, we examined 168 prediction cases, or variables used to predict NSSI. We sorted these cases into 34 categories (see pie chart below). On average, there were seven prediction cases per study; 56 prediction cases were coded as “binary” (e.g., depression diagnosis) and 112 were coded as “continuous” (e.g., depressive symptoms). 

 

Categories

Only categories with 3+ prediction cases are pictured

 

What follow-up lengths have been used in these studies?

 

Study follow-up lengths ranged from .45 to 108 months, with a mean follow-up length of 20.65 months (median = 12 months). 

 

How has Nssi been measured?

 

NSSI Measurement Types: 15 different measures of NSSI were used across these 20 reports! 

NSSI Coding: In total, 102 prediction cases were coded as predicting a “binary” NSSI outcome (yes versus no NSSI engagement) and 66 as predicting a “continuous” NSSI outcome (continuous and ordinal measurement of NSSI frequency).  

 

What samples have been used in these studies?

 

Age: These published reports included a total of 5,078 unique participants ranging in age from 10-44 years (mean age = 21.23).

  • 53 adult and 115 adolescent prediction cases were included

Sample Type: Prediction cases were drawn from 52 general (e.g., community, school-based) samples, 87 clinical, and 29 NSSI history samples.


 

Question 2. What is the overall effect size for risk factors of NSSI?

Analyses produced an overall weighted mean odds ratio (OR) of 1.59 (95% CI: 1.50 to 1.69). Although this OR was significant, the meta-analysis was limited by publication bias. Accounting for this bias, the overall OR was reduced to 1.16. 

Put in simpler terms, prevalence data suggest that about 1 (well, 0.9) in 100 adults will engage in NSSI in a given year. An odds ratio of this magnitude suggests that, on average, using these variables we can increase these odds to be 1.4 in every 100 adults. This estimate would be higher in adolescent and clinical samples, but still low in an absolute sense. This is especially concerning for clinicians, who must assess short-term risk (week, days, hours).

 

AND ARE THERE ANY ESPECIALLY STRONG RISK FACTORS?

Among categories drawn from at least 3 prediction cases (and 3 unique samples), significant ORs ranged from 1.05 (affect dysregulation) to 5.95 (history of NSSI engagement).

The chart below lists the top five strongest NSSI risk factors. Remaining risk factor categories had an average OR of 1.36.

Top 5 Strongest Categories

*Cluster b personality had a very large confidence interval compared to other risk factor categories. This finding should be considered with caution, as two out of three studies examining this factor were non-significant, and each of the studies had very large confidence intervals around this effect.

Question 3: what (if anything) affects the association between risk factors and NSSI?

Given that risk factor magnitude may change across different conditions, we examined whether four factors affected overall risk factor magnitude: NSSI measurement type, severity of sample, age of sample, and prediction case measurement type.

NSSI measure type

Cases predicting continuous NSSI frequency generated significantly stronger weighted mean ORs than cases predicting binary NSSI frequency.

Sample Population

Cases drawn from general (community) samples were slightly weaker than those drawn from samples with a history of mental illness and samples with a history of NSSI.

Sample Age

Cases drawn from adolescent samples resulted in significantly weaker effects than cases drawn from adult samples. 

Prediction Case Measure type

Binary prediction cases (i.e., cases drawn from variables with a scale ranging from 0 to 1) were significantly stronger than continuous prediction cases.

Meta-Regression

We used meta-regression to look at the unique importance of each moderator. Results suggested that NSSI measurement type and prediction case measurement type, but not sample population or age, were significant moderators. These findings suggest:

  • Differences in odds ratio magnitudes are difficult to interpret without considering the scale used for prediction cases. Risk factor categories drawn from primarily continuous prediction cases will likely have lower odds ratio magnitudes than risk factor categories drawn from primarily binary prediction cases.*
  • Continuous NSSI frequency results in stronger overall prediction.


*This does not mean that binary measures are “better” NSSI predictors! This difference is due to a mathematical artifact: ORs reflect increased odds for each unit change on a given measure. With binary measures, there is only 1 total unit. With continuous measures, there can be a wide range of scores, and for each unit change on that measure, the odds of NSSI increase by that OR magnitude.
 

What can we take away from this research?

  1. Summarizing data from longitudinal studies of NSSI, results suggest that we can significantly, but weakly, predict NSSI.
  2. A prior history of NSSI was the strongest NSSI risk factor; remaining factors were not particularly strong. Given the need to predict NSSI over short-term periods, these variables are likely not clinically useful.
  3. Measures of NSSI were highly disparate across studies, varying in terms of the types of behaviors assessed, questions used to assess NSSI, and coding of NSSI engagement. 
  4. Follow-up lengths across these studies were pretty long on average, typically lasting one year.
  5. Very few studies were conducted among participants with a history of NSSI. 

Future Directions: Future research examining NSSI risk factors should seek to standardize NSSI measurement and to conduct shorter-term longitudinal studies exploring both traditional and novel risk factors for these behaviors, especially among participants with a history of NSSI. Using more consistent assessment, novel risk factors, and shorter follow-up periods may yield stronger results.