Mental disorder,psychological problems and terrorist behaviour: A systematic review and meta-analysis

Inconsistent findings on mental health difficulties and risk of involvement in terrorism

There has been an increasing focus on the potential role of mental health difficulties in the process of violent radicalisation into terrorism. In part, this has been fuelled by studies appearing to show high prevalence rates in some samples of terrorists. However, findings are inconsistent, with some studies reporting higher rates than those observed in the general population, some lower, and others that are comparable to those observed in the general population.

What is this review about?

This review synthesises the prevalence rates of mental health difficulties in terrorist samples (Objective 1—Prevalence) and prevalence of mental health disorders pre-dating involvement in terrorism (Objective 2—Temporality). The review also synthesises the extent to which mental health difficulties are associated with terrorist involvement compared to non-terrorist samples (Objective 3—Risk Factor).

In addressing these objectives, the review offers an initial assessment of what we refer to as the mental health-terrorism hypothesis (that mental health difficulties are a risk factor for terrorist involvement). Tentative support for the hypothesis would be provided where studies, when collated, suggest higher rates of difficulties in terrorist samples than those expected in the community.

The review distinguishes between mental disorders, suspected mental disorders and psychological problems. These are collectively termed mental health difficulties.

What is the aim of this review?

This Campbell Collaboration systematic review focuses on mental health and its association with terrorist involvement. The review examines evidence based on 56 papers reporting on 73 terrorist samples.

The problem, condition or issue

In the period 1970–1990 there was extensive debate on the potential link between mental disorder and terrorism (Cooper, 1978; Ferracuti & Bruno, 1981; McCauley & Segal, 1987; Rasch, 1979; Shaw, 1986; Silke, 1998; Smith & Morgan, 1994; Wardlaw, 1982). Commentators proposed that some of those who became involved in terrorism had an underlying mental disorder that was causally linked to their violence propensity, including, for example, a range of personality disorders (Cooper, 1978; Lasch, 1979; Pearce, 1977). However, successive studies failed to support this link (e.g., Elliot & Lockhart, 1980; Ferracuti & Bruno, 1981; Lyons and Harbinson, 1986) leaving many convinced that those involved in groups like the Provisional IRA and Euskadi Ta Askatasuna (ETA) were no different to the general population in terms of mental health difficulties and were, instead, attracted to terrorism by virtue of a multitude of interacting psychosocial processes (e.g., Crenshaw, 1981; Heskin, 1984; Taylor & Horgan, 2006). The assertion that terrorism was the product of abnormality was ultimately deemed ‘unfair’ to the terrorist and abandoned by researchers and policy makers (e.g., Silke, 1998).

In recent times the debate has re-emerged, fuelled in part by a more nuanced approach where the focus has been on specific forms of terrorism (e.g., lone actor terrorist or right-wing terrorism) rather than terrorism in general. For example, those interested in so-called lone actor violence raised concerns about ‘fundamental errors’ in past research (Corner et al., 2016, p. 561) and presented plausible case formulations linking experiential stressors such as social isolation with mental disorder and violence (e.g., Corner & Gill, 2015). Cohort studies of lone actors emerged that appeared to show high rates of mental disorder, with 37% of Liem and colleagues' sample of European lone actor terrorists having ‘some indication of mental illness’, and 25% being clinically diagnosed with ‘a particular mental disorder’ (Liem et al., 2018, p. 60). Such findings were not limited to lone-actor terrorism (e.g., Weenink, 2015).

However, those who have looked more closely at this evidence base have acknowledged that the picture emerging is far from clear, with the lack of clarity attributable, in part at least, to methodological limitations in that literature (Gill et al., 2021; Jensen et al., 2020). To some extent these limitations centre on one core problem–the difficulty in determining to what extent, if any, the presence of a mental disorder confers risk of terrorist involvement (as opposed to being associated with increased risk). To credibly conclude that disorder (or, more broadly, psychological problems) can increase the risk of becoming involved in terrorism, we suggest that the existing literature needs to demonstrate that certain criteria are met.

The first criterion relates to the prevalence of disorder among terrorist samples, and is termed here ‘the prevalence problem’. Assessing prevalence of mental disorder among terrorist samples using appropriate comparators sheds light on the magnitude of the relationship (if any) between the risk and outcome. It is an important criterion for determining causality and one of several Bradford-Hill criteria for causation (Hill, 1965). Reporting and interpretating prevalence rates requires careful consideration of the distinction between point and period prevalence, and use of appropriate comparison groups. If mental disorder confers risk of terrorism, then the prevalence rate amongst terrorist samples should be higher than that expected in the general population (matched on key demographics such as age, gender and geographic location). A prerequisite for making such comparisons is synthesising the prevalence rates of mental disorders, and more broadly psychological problems, in terrorist samples.

The second criterion is that of temporality (see Lucas & McMichael, 2005), another Bradford-Hill criterion (and termed here ‘the temporality problem’). To demonstrate that mental disorder confers risk of involvement in terrorism, then studies must demonstrate that the disorder emerges before involvement in terrorism. If we cannot demonstrate temporality, then differentiating cause and effect is problematic—disorder could be the consequence of terrorist involvement rather than a cause of involvement. The question that must be asked of the literature, then, is: To what extent do findings in the literature support the assertion that earlier mental disorder confers risk of terrorist involvement later in life?

The third criterion that must be met to infer a causal link exists between mental disorder and terrorism is that of plausibility; offering plausible explanations for just how the disorder (or disorders) confers risk of terrorism involvement (the Bradford-Hill ‘plausibility problem’). Addressing the plausibility problem is hampered by the heterogenous nature of terrorism (the outcome of interest), controversy around the concept of ‘mental disorder’ (the potential risk factor of interest) and, finally, the complexity of the relationship that may exist between terrorism and mental health difficulties.

A final consideration for those conducting research on terrorism and mental disorder relates to the concept of ‘mental disorder’ itself. To illustrate the difficulty with the concept, it is worth referring to the recent contribution of Bakker (2019). Bakker's thesis, like many other clinical psychologists, is that the concept of ‘mental disorder’ is fundamentally flawed, a ‘medical nosology of diseases’ that does little to aid our understanding of clients or how best to intervene to alleviate their distress (p. 1). What is required, they argue, is a paradigm shift towards a focus on ‘psychological problems’ which Bakker defines as ‘a negative psychological-level state of affairs’ (p. 10). These psychological problems are transdiagnostic (appear across diagnoses and capture the common difficulties reported by patients) and may or may not require intervention. Bakker's thesis draws attention to two types of mental health difficulties—mental disorders and psychological problems.

The difficulty for the terrorism literature is that it is not always clear what the outcome of interest is, particularly when terms like disorder, mental health difficulties, psychological distress, etc. are used interchangeably. Yet, where a formal diagnosis is not made by a mental health professional, then the presence of a mental disorder cannot be inferred. At best, we can conclude that the individual has or had what Bakker refers to as a psychological problem.

This review seeks to contribute to our understanding of the potential link between terrorism and mental health difficulties by focusing on the problems of prevalence, temporality and plausibility, while being sensitive to the distinction between mental disorder and psychological problems. Specifically, the review will present a synthesis of the evidence on prevalence, temporality and plausibility drawing on the best evidence available. In reviewing prevalence, temporality and plausibility, we are presenting an initial test of what we refer to as the mental health—terrorism hypothesis: That mental health difficulties confer risk of involvement in terrorist behaviour.

Outcome—Terrorist behaviour

The outcome of interest in this systematic review is terrorist behaviour. While there is no universally accepted definition of terrorist behaviour (Ganor, 2002; Saul, 2019; Silke, 1996), there is at least some consensus that it refers to: (a) an act or campaign of actual or threatened violence that seeks to elicit the terror emotion in a target audience and; (b) with the intention of bringing about change in line with the world-view of the terrorist (e.g., Kruglanski & Fishman, 2006; Moghaddam, 2007). Terrorist behaviour intends to cause harm, physical or otherwise (Van Der Does et al., 2021).

One complexity encountered by those conducting research in the area of terrorism is the heterogeneity of the phenomenon itself (Herrington & Roberts, 2012; Monahan, 2012; Roberts & Horgan, 2008). For example, one study of 176 terrorist organisations identified 33 different ideologies as well as diversity in terms of size, organisational structure, geographic location, and lethality (Cook & Lounsbery, 2011). It cannot be assumed that individuals who become involved in different forms of terrorism do so through the same processes (Change Institute, 2008). The implications of this heterogeneity, for the proposed review, are that, when we refer to terrorism, we specify this behaviour in terms of the ‘type’ of terrorism waged.

The review excludes violent radicalisation as an outcome, understood within this study to refer to a process of growing acceptance of the legitimacy of violence to bring about societal change (McCormick, 2003). We justify this exclusion on the following basis: 1.

Violent radicalisation and terrorism are conceptually non-synonymous. Terrorist behaviour is action orientated whilst violent radicalisation is a process whereby individuals become increasingly committed to the use of violence, yet may not necessarily perpetrate violence (i.e., it can be cognitively orientated) (Sarma, 2017).

Potential risk factor—Mental health difficulties

The review informs our understanding of the potential link between mental disorder and terrorist involvement. However, we are acutely aware of the limitations of the largely disease-orientated focus of ‘mental disorders’, and for the reasons set out below extend our focus to include a more transdiagnostic lens by considering, broadly, psychological problems.

The American Psychological Association defines mental disorder as any condition ‘characterized by cognitive and emotional disturbances, abnormal behaviours, impaired functioning, or any combination of these’ (American Psychological Association, 2020). Diagnostic manuals, such as iterations of the Diagnostic and Statistical Manual of Mental Disorders (DSM; American Psychiatric Association, 2013) and International Classification of Diseases (ICD; World Health Organization WHO, 2019), present diagnostic criteria for a range of mental disorders, and diagnoses can be made based on the combination of symptoms and their severity.

While diagnostic systems convey a sense that mental disorders are discrete psychological experiences, Allsopp and colleagues note that diagnostic systems are characterised by: (a) varying diagnostic rules across presentations; (b) overlap in symptoms across diagnoses and; (c) a tendency for diagnostic labels to mask the root causes of distress and problematic behaviours (Allsopp et al., 2019). They also argue that diagnoses can distract from the real-world work of reaching an in-depth understanding of the individual (see also Galatzer-Levy & Bryant, 2013; Olbert et al., 2014).

The use of diagnostic systems in terrorism research is particularly problematic. Mental disorders can only be said to be present where they have been diagnosed by an appropriately trained mental health professional. The presence of a disorder, as characterised in DSM or ICD, cannot be reliably inferred from reports of symptoms present in open-source data alone (e.g., press coverage of trials of suspected terrorist offenders). It requires careful assessment by a professional, often in collaboration with the individual being assessed, of the presence or absence of various criteria. Where the results of such assessment are made available for the purposes of research, then we can have at least some confidence in the diagnostic process.

However, the data on terrorist offenders does not always provide access to data gathered through structured clinical assessment by a suitably trained professional. Some of the most widely cited studies in the area, for example, are based on open-source information derived from the media coverage, etc., of terrorists, and where inferences are made as to the presence/absence of disorder based on difficulties experienced by the terrorist and publicly reported (e.g., Gill et al., 2014). Yet it is not always clear if these difficulties are any different to those experienced in the normal course of one's life.

Bakker (2019) has discussed this in the context of differentiating between ‘psychological problems’ that ‘do not self-perpetuate… and tend to ease without interventive therapy’ (p.11) and ‘clinical psychological problems’. The former, Bakker argues, are normal adaptive processes that might include an experience of depression after a bereavement, but which follows the normal course of recovery. The latter, however, might take the form of avoidance, be pervasive and enduring, impair one's quality of life, and require intervention to resolve.

In the review we attend to the literature on mental disorder and psychological problems. While our initial intention was to differentiate between psychological problems that are ‘clinical’ (required intervention) and those that were not clearly clinical (no evidence of a requirement of intervention), no papers that met our eligibility criteria made such a distinction. Synthesising the literature on ‘mental disorder’ provides coverage of the diagnostic literature. Attending to the broader literature on psychological problems captures a wider body of literature and means that our review is not constrained by the limitations of the psychiatric model (see ‘Inclusion Criteria’ for additional detail).

How mental disorder and psychological problems may be linked to terrorist behaviour

The link between mental disorder and forensic risk has been the subject of research for decades. Findings are unclear and inconsistent, with some studies appearing to link disorder to violence risk, while others have reported no link (e.g., Augestad Knudsen, 2020; Bhui & Jones, 2017). There appears to be three core complexities in this area that are directly relevant to an assessment of the link between terrorism and disorder.

The first is to do with the prevalence problem, as discussed earlier in this protocol, and which requires an assessment of the rates of mental disorders, and more broadly psychological problems, in terrorist samples. The second is temporality, initially discussed by Bradford Hill in their consideration of association and causality (i.e., one of the Bradford Hill criteria for causal inferences; see Lucas & McMichael, 2005). To argue that mental disorder confers risk of violence, research must demonstrate that the onset of psychological problems (Factor A) pre-dates involvement in violence, or at the very least detection for terrorist offences (Factor B). If temporality cannot be established, there is a risk of misinterpreting correlation (where Factor A is associated with Factor B) as causation (Factor A causes Factor B).

For example, high rates of mental disorder are often observed among incarcerated violent offenders (O'Sullivan et al., 2018). A typical study examining the link between mental disorder and violent offending in this population will involve those incarcerated completing a battery of measures that assess psychological wellbeing and severity of offence (‘index offence’). Where a relationship emerges, the temptation is to conclude that higher levels of psychological distress confer risk of serious offending (Factor A causes Factor B). The problem here, however, is that the research design deployed cannot test causal relationships and is limited to measuring associations. Because we do not have longitudinal data showing levels of psychological symptoms before involvement in violence, we cannot exclude the possibility that the distress arose due to the index offence, or during incarceration for that offence. In such a case, incarceration/offending (Factor B) could cause distress (Factor A). This resonates with the conceptual difficulties surrounding the suggestion that pre-existing Post-traumatic Stress Disorder (PTSD) increases risk of terrorist involvement among so-called ‘foreign fighters’ and where PTSD is assessed after they have returned from conflict (e.g., Al-Attar, 2019).

Even where temporality is determined, we also need to consider the ‘Third Variable Problem’. Here an apparent causal relationship between disorder and violence may be explained by a lurking third variable (Variable X) that influences both factors and causes them to co-vary. This might arise, for example, where experiences of discrimination and isolation drive both mental disorder and violent radicalisation. Here the apparent relationship between mental disorder and violent radicalisation may be wholly attributable to situational stressors and intervening to manage disorder may not reduce risk.

In reality, there is unlikely to be a large body of scientifically robust longitudinal evidence that addresses both temporality and the ‘Third Variable Problem’. This is particularly the case in the area of terrorism studies where the problem of concern is so difficult to expose to academic enquiry due to its low base rate in the population (Sarma, 2017). In such a situation, another of Hill's criteria becomes important—that of plausibility. That is, in assessing the relationship between mental disorder and violence risk, we must be able to present a plausible theoretical argument as to just how disorder confers risk. In clinical forensic practice, plausibility is addressed through careful assessment of forensic risk and the presenting of a theoretical argument (or formulation) that explains how the presence of risk factors may confer risk (Davies et al., 2013).

One way of presenting a formulation of risk is through the ‘4Ps’ Framework. The 4Ps Framework places a risk factor in a temporal space or chronology and proposes the nature of the relationship between the factor, other factors, and the outcome of interest. In doing so it differentiates between predisposing, precipitating, protective and perpetuating factors. It is widely used in both clinical and forensic psychology case formulation (e.g., Macneil et al., 2012).

The 4Ps framework can encourage a more nuanced consideration of the link(s) between terrorism and mental disorder. A predisposing risk factor is one that places the individual at risk of becoming involved in terrorism later in life. In the broader literature on clinical and clinical forensic psychology mental disorder is viewed as primarily a non-causative background predisposition for becoming involved in violent behaviour and which is part of a complex set of interacting risk factors that together lead to a scaffolding of risk (e.g., Van Dorn et al., 2012). For example, Markowitz (2011) adopts a Social Disorganisation Theory approach to formulating risk in proposing that people with long-term mental disorders are more likely to reside in disadvantaged communities characterised by socially disorganised neighbourhoods with a lack of health care, limited job opportunities, racial diversity and fragmented families. Crime flourishes in such criminogenic contexts, they argue, because there is a culture of acceptance of violent crime and poor social control over offending. This resonates with the terrorism literature, with some arguing that vulnerability to violent radicalisation is due, in part, to radicalising settings where some sections of the community endorse beliefs that justify terrorism (e.g., Schils & Pauwels, 2016).

A precipitating risk factor is one that apparently triggers a ‘crisis’. In the mental health literature, extreme situational stressors such as relationship break-down, bereavement or other acute trauma can result in a cascade of events leading to the negative outcome (e.g., Barber et al., 2014). For mental disorder to precipitate involvement in terrorism, research would need to demonstrate that involvement in terrorism was immediately preceded by the onset of an episode of psychological distress that causally led to involvement in violence. In the broader clinical forensic literature, this has typically been associated with the presence of ‘positive’ psychotic symptoms, such as irrational (delusional) beliefs about others who subsequently become the target of violent intent (Markowitz, 2011). Of course, it could be reasonably argued that in some cases where terrorist behaviour is precipitated by a mental disorder, and derives from disorder (e.g., persecutory beliefs), that the issue of concern is clinical forensic risk rather than terrorism.

Some forms of mental disorder may actually preclude an individual from becoming involved in terrorism (i.e., it is a ‘protective factor’). For example, there is evidence that organisations like the Provisional IRA sought to recruit the most psychologically robust individuals into their ranks as a way of reducing the potential for members to be compromised and turn informer, or to provide information while being questioned by the police and security services (Sarma, 2005). Here mental disorder actually protects against involvement.

Finally, perpetuating factors, in the context of terrorism, serve to maintain the problematic behaviour, and thus hamper the ability of the individual to disengage from terrorism. In their review of push and pull factors that influence the ability of extremists to disengage from terrorism, Jensen and colleagues (2020) noted that increased social mobility, onset of new intimate romantic relationships, children, and access to rehabilitation services can support disengagement. Conversely, mental health difficulties can impede the ability of individuals to develop relationships, access services, and be more socially mobile, hampering disengagement. They conclude that ‘[e]specially when co-occurring with substance abuse, mental illness can act as a strong barrier to disengagement, especially if it counteracts the feelings of disillusionment that otherwise may prompt one's exit’ (Jensen et al., 2020, p. 8).

Why it is important to do the review

As noted earlier, findings in the literature examining the link between terrorism and mental disorder are inconsistent (Ho et al., 2019). Where such inconsistencies are a feature of an evidence base, the cherry-picking of results to suit a specific position can impede a nuanced understanding. Systematic reviews provide a synthesis of the available literature in one accessible paper and in doing so reduce bias (White & Waddington, 2012).

Gill and colleagues (2021) provide a valuable review of the literature exploring the link between mental disorder and terrorism. They note the heterogeneity in prevalence rates in the literature and provide some plausible explanations for this heterogeneity. Our review, however, we will attend in particular to point and period prevalence rates and include studies published since 2020. We will also present comprehensive data syntheses for both mental disorder and psychological problems and where appropriate present sub-group analyses. Moreover, we will critically evaluate the appropriateness of the comparator populations being used as benchmarks for prevalence. This can help us more carefully consider the causal link, if any, between disorder/difficulties and terrorist behaviour.

We also considered Misiak et al. (2019) systematic review of the link between mental health and radicalisation and mass violence. While the review also presents a valuable contribution to the literature, it focuses primarily on the risk of radicalisation, with nine of the 12 studies included in their review using community samples and self-reported radical beliefs (i.e., not terrorist samples). As noted above, we cannot assume that the literature on violent radical beliefs is valid for our understanding of terrorist behaviour.

Findings from the current systematic review will support a more informed debate on the link between mental health difficulties and terrorist behaviour. For each paper included in the review, we isolate and specify the prevalence being reported (e.g., present at time of assessment (point); childhood etc.). We will synthesise studies that are sensitive to temporal sequencing (temporality), where difficulties are onset before involvement in terrorism, and studies that examine the extent to which difficulties are associated with involvement (i.e., compare rates across two groups).

Apart from supporting debate in the area, the findings will be of value to a range of professionals who are responsible for risk assessment, risk mitigation and psychological intervention. In relation to risk assessment, for example, two popular risk assessment tools, the VERA 2R (Pressman et al., 2016) and ERG 22+ (Lloyd & Dean, 2015), both contain items relating to psychopathology despite concerns that (a) the evidence supporting their inclusion is contested and (b) there is a need to disaggregate disorders into various forms or problem clusters to determine which, if any, may be linked to risk of terrorist behaviour (see Herzog-Evans, 2018). In supporting a more nuanced understanding of the link between disorder, psychological problems and terrorist behaviour our review will help guide the use of such tools and in doing so support decision making around psychological support, appropriate detention settings, and release from detention.

In relation to psychological intervention, there are multi-disciplinary teams working in most countries tasked with supporting individuals who may at risk of transitioning into terrorism to redirect their lives towards non-violence. In many cases these teams include health workers who are sensitive to psychological problems that are believed to exacerbate risk of becoming involved in terrorism—teams comprised of professionals who would benefit from a systematic review of the relevant literature. The work of such teams has been reviewed and discussed elsewhere (see Sarma, 2018, 2019a, 2019b).

Criteria for considering studies for this review

Search methods for identification of studies

Our search strategy aligned with Cochrane Training and our past reviews for both the Cochrane Collaboration (Doody et al., 2019) and the Campbell Collaboration (Carthy et al., 2020). It was also informed by Kugley et al. (2017) guidance on information retrieval for Campbell systematic reviews.

Data collection and analysis

Deviations from the protocol

There were three deviations from the protocol. First, in our protocol we stressed the importance of distinguishing between point and period prevalence, and different points and periods, when reporting and pooling prevalence rates. For example, we envisaged that studies would report life-time, 12- and 24-month prevalence rates (i.e., different periods) and at different points (i.e., right now vs. at time of engagement vs. at time of disengagement, etc.). However, of those studies reporting period prevalence, all focused on life-time prevalence rates of mental health difficulties. For point prevalence, eligible studies focused on difficulties present at the time of an attack (‘detection’), while being processed through the courts, or while incarcerated. Thus, the review reports on a more restricted set of points and periods than initially envisaged in the protocol.

Second, Objective 2 deals with temporal sequencing and we had initially envisaged isolating studies that reported on rates of difficulties that were onset before the involvement in terrorist behaviour. Such evidence, we stressed, places the hypothesised risk factor in the correct temporal sequence to the outcome and provided a purer test of the mental health-terrorism hypothesis. However, studies that dealt with temporal sequencing, with one exception, focused on the presence of difficulties before detection rather than involvement—the limitation being that it is possible that such difficulties were onset between the point of engagement and their detection. The exception is Corner and Gill (2020), with Corner providing the authors with data on mental health difficulties before engagement.

Third, we had anticipated differentiating between diagnosed disorders and psychological problems. However, when we extracted the data, it became clear that studies had adopted a number of different approaches to coding for diagnosed disorders and which set more or less stringent thresholds for assessment and diagnosis. To be sensitive to this we differentiate between ‘diagnosed disorders’, ‘suspected disorders’ and ‘psychological problems’ as follows: 1.

If a study referred to mental disorder or mental illness and where all individuals in the sample have a ‘confirmed diagnosis’, ‘clinical diagnosis’, ‘formal diagnosis’, or diagnosis made by a trained professional (or within a mental health report), then it was coded as relating to ‘mental disorder’. As examples, Bubolz and Simi (2019) refer to mental disorder as being present where ‘a medical practitioner had ever diagnosed the person’ (p. 6), Capellan (2015) refers to ‘formal’ and ‘confirmed diagnosis’ (p. 400) and Corner and Gill (2021) required an ‘official psychiatric consultation or diagnosis’ (p. 705).

Table 5 sets out our interpretation of the nature of the mental health difficulties being reported in each paper, with an accompanying justification.

Description of studies

Figure 1 (PRISMA flowchart) presents the results of our search, screening and identification process for all review objectives. The systematic search of electronic databases identified 196,888 records, unrestricted by publication date. After de-duplication, 109,027 records were progressed to title/abstract screening in DistillerSR. An initial stage of de-duplication was conducted in EndNote, followed by another stage of de-duplication using the Detect Duplicates function in DistillerSR (which can identify additional duplicates compared to EndNote). A total of 108,763 records were excluded at this stage as duplicates, ineligible types (e.g., book reviews), or because they were not about terrorism and mental disorder, leaving 264 records for full-text review.

OPEN IN VIEWER

Figure 1

PRISMA flow diagram. ¹LOTE = Language other than English. ²Studies could be included under more than one review Objective and so studies listed by Objective will not sum to total studies included in the review.

The full-texts of all records were obtained and screened in DistillerSR. These were combined with full-texts of records harvested from past reviews, the grey literature, hand searches of journals, forward citation searching, and consultation with expert groups (n = 32). Of the 296 documents, the full text of 14 could not be sourced and are listed in the ‘Studies awaiting classification’ reference section. The remaining full texts were reviewed, with two excluded due to being duplicates of other studies. Others were ineligible as they were not published in English (n = 12), were not focused on terrorist behaviour (n = 33), were not empirical studies (n = 69), did not include a measure of mental health (n = 8), or did not measure mental health in a terrorist sample (n = 102). This left 56 papers reporting on 73 terrorist samples (studies). For clarity, henceforth we use the term ‘paper’ to refer to the original full-text record and ‘study’ to refer to each unique terrorist sample isolated.

All 56 papers, reporting 73 studies, were included in Objective 1 (Prevalence), though not all were included in the data syntheses reported due to overlapping datasets. Of these, 9 papers (10 studies) reported prevalence rates where the difficulties were onset before either terrorist involvement or the index offence and were eligible for Objective 2 (Temporality). Finally, 9 papers were eligible for synthesis under Objective 3 as they examined the association between mental health difficulties and terrorist behaviour (i.e., comparing terrorists and non-terrorists). In the remainder of this section, we deal with each objective in turn.

Objective 1—Prevalence

Risk of bias in included studies for Objective 1

We used the JBI Checklist for Prevalence Studies (JBI, 2020), for all papers that were eligible for consideration for our synthesis of prevalence rates of disorder among terrorist samples. For each paper, specific reasons for risk ratings for each item are reported. We use the Risk of Bias (ROB) assessment to draw attention to the methodological limitations and other sources of bias in this body of literature in the review rather than use it to isolate a set of high quality papers for inclusion in the meta-analysis and narrative synthesis.

Our rationale for this is that research involving terrorist samples is inherently challenging, particularly in relation to data quality—accessing terrorist samples for the purpose of building profiles is very difficult, often necessitating that data sets are based on the compilation of data from a range of open-sources. If we were to exclude studies where samples were not demonstrably representative of the wider terrorist population, or where mental disorders were not assessed through clinical interview or valid psychometric assessment, the consequence would be a limited review. We would also note, again, that while the presence of mental health difficulties was reported in all studies, it was not a specific objective, or primary focus, of some studies and rather one of many characteristics of terrorists reported on—therefore, for some studies our ROB is focused on an aspect of these studies that was not necessarily a primary focus. We also acknowledge that our ROB is based on information published in each study, and which may, in part, reflect the conventions of the relevant journal rather than the ROB of the study itself.

Table 8 summarises our risk of bias assessment for all papers across the 9 items of the JBI checklist, as well as the tenth item on plausibility. The ROB was completed in duplicate with just one discrepancy across the two reviewers (κ = 0.98) and which was resolved through discussion. We have sought further information on 5 papers to allow us complete both the risk of bias assessment and extract data for the data synthesis. Excluding these from the 56 papers, left 51 for full risk of bias assessment. As anticipated, 98% of papers (50 of 51 papers) were assessed as having high risk of bias as each received a rating of ‘High’ for at least one item. Only one paper (Duits et al., 2022) had all items as rated ‘Low’, thus achieving the low risk of bias designation.

OPEN IN VIEWER

Table 8

Risk of bias assessment according to the JBI checklist for prevalence data

Study	1. Representative	2. Recruitment	3. Sample Size	4. Description of participants	5. Coverage	6. Standard criteria	7. Measurement	8. Statistics	9. Missing data analysis	10. Plausibility	Risk of bias rating
Bakker (2006)	Low	Low	Low	Low	High	Unclear	High	Low	High	High	High
Bergen et al. (2017)a	NA	NA	NA	NA	NA	NA	NA	NA	NA	NA	?
Böckler et al. (2018)a	NA	NA	NA	NA	NA	NA	NA	NA	NA	NA	?
Bronsard et al. (2022)	Low	High	Low	Low	Low	Low	Low	Low	Low	High	High
Brym and Araj (2012)a	NA	NA	NA	NA	NA	NA	NA	NA	NA	NA	?
Bubolz and Simi (2019)	High	High	High	Low	Low	Low	Low	High	Low	Low	High
Candilis et al. (2021)	Low	Low	Low	Low	Low	Low	Low	Low	Low	High	High
Capellan (2015)	Low	Low	High	Low	Low	Low	Low	Low	High	Low	High
Capellan and Anisin (2018)	Low	Low	Low	High	Low	NA	NA	Low	Low	Low	High
Capellan et al. (2019)	Low	Low	Low	Low	High	Low	Low	High	High	High	High
Chermak and Gruenewald (2015)	High	Low	Unclear	Low	High	Low	Unclear	High	High	High	High
Cherney et al. (2022)	Low	Low	Low	Low	Low	Low	Low	Low	Low	High	High
Clemmow et al. (2020)	Low	High	Low	Low	High	Low	Unclear	High	High	High	High
Corner and Gill (2015)	High	Low	Unclear	High	High	Low	High	High	High	Low	High
Corner et al. (2019)	High	Low	Unclear	High	High	Low	Unclear	High	High	Low	High
Corner and Gill (2020)	High	Low	Unclear	High	High	Low	Unclear	High	High	Low	High
Corner and Gill (2021)	High	Low	Unclear	Low	High	Low	Unclear	High	High	Low	High
de Roy van Zuijdewijn and Bakker (2016)	High	Low	Unclear	Low	High	High	High	High	High	Unclear	High
Dhumad et al. (2020)	High	Unclear	Unclear	Low	Low	Low	High	High	Low	Low	High
Duits et al. (2022)	Low	Low	Low	Low	Low	Low	Low	Low	Low	Low	Low
FBI (2019)	High	High	Unclear	Low	High	Low	Low	Low	High	High	High
Freilich et al. (2019)	Low	Low	Unclear	Low	Low	Low	Unclear	High	Low	High	High
Gibson (2018)	High	High	High	Low	High	High	High	Low	High	High	High
Gill et al. (2014)	Low	High	Unclear	Low	High	Low	Unclear	High	High	High	High
Gill (2015)	Low	Low	Unclear	Low	High	NA	NA	High	High	Low	High
Gill et al. (2021)	Low	Low	Unclear	Low	High	Low	Unclear	High	Low	High	High
Gill et al. (2022)	High	Low	Unclear	High	High	Low	Low	Low	High	High	High
Gottschalk and Gottschalk (2004)1	NA	NA	NA	NA	NA	NA	NA	NA	NA	NA	?
Gruenewald et al. (2013)	Low	Unclear	Unclear	Low	High	Unclear	Unclear	Low	High	High	High
Hamm, Spaaij, Hamm and Spaaij (2017)	Low	Low	Low	Low	Low	Low	High	Low	Low	High	High
Haugstvedt and Koehler (2021)	Low	Low	Low	High	High	Unclear	Unclear	High	High	Low	High
Horgan (2016)	Low	Low	Low	Low	High	NA	Unclear	Low	High	Low	High
Khazaeli and Khoshnood (2019)	Low	Low	Low	Low	High	NA	NA	High	High	High	High
Jensen and Kane (2021)	High	Low	Low	Low	Low	Low	Low	Low	High	High	High
Kupper and Meloy (2021)	High	High	High	High	High	Unclear	High	Low	High	High	High
LaFree et al. (2018)	Low	Low	Low	Low	Low	Low	Unclear	Low	High	Low	High
LaMontagne (2019)	Low	Low	Low	High	High	NA	NA	Low	Low	High	High
Lankford (2013)a	NA	NA	NA	NA	NA	NA	NA	NA	NA	NA	?
Liem et al. (2018)	Low	Low	Low	Low	Low	Low	Low	Low	High	High	High
Lucas (2021)	Low	Low	Low	Low	High	High	High	High	High	Low	High
Lyons and Harbison (1986)	High	High	High	Low	Low	Low	Low	Low	Low	High	High
Merari and Ganor (2020)	Low	High	High	Low	Low	Low	Low	Low	Low	Low	High
Merari et al. (2010)	Low	Low	Low	Low	Low	Low	Low	Low	Low	Low	High
Merari (2021)	Low	High	High	Low	Low	Low	Low	Low	Low	Low	High
Meloy et al. (2019)	High	High	High	Low	High	High	High	Low	High	High	High
Perry et al. (2018)	Low	Low	Low	Low	Low	NA	NA	Low	High	High	High
Pfundmair et al. (2022)	High	High	High	Low	High	NA	NA	Low	High	High	High
Pitcavage (2015)	High	Low	Low	High	High	NA	NA	Low	High	Low	High
Pendergast (2020)	High	Low	Low	Low	High	NA	NA	Low	High	High	High
Simi et al. (2015)	High	High	High	Low	Low	Low	Low	High	Low	Low	High
Simi et al. (2016)	High	High	Unclear	Low	Low	Low	Low	Low	Low	Low	High
Thijssen et al. (2021)	Low	Low	Low	Low	High	Low	Low	Low	High	High	High
van Leyenhorst & Andreas (2017)	High	Low	Low	Low	High	Low	Low	Low	High	High	High
Weenink (2019)	Low	Low	Low	Low	High	NA	NA	Low	High	High	High
Weenink (2015)	Low	Low	Low	Low	High	Low	Low	Low	High	High	High
Zeman et al. (2018)	High	High	High	Low	Low	Low	Low	High	High	Low	High

Authors contacted for more detail on study to enable completion of ROB assessment.

The first two items on the checklist pertain to the representativeness of the sample (Item 1) and the appropriateness of the recruitment of the sample (Item 2). In relation to representativeness, we assessed each paper to the extent that ‘the sample frame was appropriate to address the target population’, with guidance suggesting that the sample frame should represent the target population (JBI, 2020). Importantly, the guidance suggests that a ‘sample frame may be appropriate when it includes almost all the members of the target population’. Hence, where papers sought to understand a specific form of terrorism within a specific jurisdiction and all known cases were sampled, a ‘low-risk’ rating was applied. Twenty-nine papers were identified as having low-risk based on this criterion, with 22 rated as high-risk. For example, Bakker's (2006) population of interest was Jihadi terrorists in Europe. They sampled 242 terrorists active in Europe between 2001 and 2006 and the information provided in the paper that all those identified as meeting their inclusion criteria were included. Based on the assumption that the sample was therefore likely to be representative, a rating of low-risk of bias was allocated to the Bakker paper for Item 1. For other papers, however, authors acknowledged that the sample may not be representative (e.g., van Leyenhorst & Andreas, 2017), the full sample frame was not sampled (or it was not clearly stated the full sample frame was sampled) (e.g. Chermak & Gruenwald, 2015) or there was insufficient information to assess representativeness and it was not discussed by authors (e.g. Bubolz & Simi, 2019). In all such cases a rating of ‘high-risk’ was applied.

When assessed in terms of the appropriateness of the recruitment of their samples (Item 2), 34 papers were assessed as low risk, 15 as high-risk and 3 as unclear. Based on guidance, sampling was assessed as low risk where random sampling was used (e.g., LaFree et al., 2018) or where everyone, or almost all individuals, in a sampling frame were recruited. This meant that even though some data sets may not have been representative of the global target population, once everyone within the data set was recruited we assessed this as low-risk of bias.

Item 3 on the JBI checklist assesses the appropriateness of the sample sizes included. We acknowledge that sample size of individual studies is not a primary concern when they are included in a meta-analyses, but retained the item here for completeness and to help in our assessment of each included study.

As studies included do not report sample size calculations, and following advice in the JBI checklist, we calculated a threshold using the following formula:

n = ( ( Z 2 ) ( P ) ( 1 − P ) ) / d 2 , $n=((Z^2)(P)(1-P))/d^2{,}$

where n is the sample size, Z the Z statistic for the desired confidence interval (in this case 0.95%, and the Z value of 1.96), P the expected prevalence of the proportion (the lifetime rate of disorder in the general population of 25% (p.25) is used), and d the precision (set as half the Confidence Interval, or 0.05). This yielded a target population size of 294. We accept that this sets a very high threshold for terrorism research, and apply the sample size calculation as a benchmark only. We coded papers that met or exceeded n = 294 as low-risk and those lower than n = 294 as high-risk. Papers where the full, or almost all, of the sample frame were recruited were also coded as low-risk regardless of the sample size (as arises, e.g., in the van Leyenhorst and Andreas paper, 2017). Twenty-five papers were coded as low-risk and 11 as high-risk. Fifteen papers were marked as unclear, primarily because there was a lack of information in the papers as to how they determined their target sample size.

Item 4 relates to descriptions of the sample and setting and here we coded that criterion as met where the study reported demographic characteristics of age (mean and/or range) and gender, and the form of terrorism at a minimum. Nine of the papers were scored as high-risk having not reported either age or gender. All other studies were assessed as low-risk.

Items 5 and 9 related to data-analysis and response rates. We used both items to assess the risk of bias that can arise when missing values are not appropriately reported and managed. To be assessed as low-risk, papers had to report the number of missing values for the mental health difficulties (Item 5). For Item 9, to be assessed as low-risk they had to compare those for whom data is missing with those for whom data is present (Item 9). The rationale for this has been discussed in the literature, with the omission of missing cases from the denominator in calculating proportions leading to an inflation in prevalence rates (e.g., Freilich et al., 2019; Gill et al., 2022).

While almost all papers calculated prevalence rates with missing cases included in the denominator, they do not report what proportion of the denominator is based on ‘missing data’ vs mental health difficulties reported to be absent (coded as high-risk for Item 5). In addition, some papers do not compare missing cases with non-missing cases, and which would inform our understanding of data-coverage bias (Item 9). Twenty-one papers were assessed as low risk for Items 5 and 17 for Item 9. Papers based on interview or clinical assessment tended to score as low-risk for these items as they explicitly probed both presence and absence of difficulties leading to no or low missing data (i.e., Bronsard et al., 2022; Bubolz & Simi, 2019; Candilis et al., 2021; Dhumad et al., 2020; Lyons & Harbinson, 1986; Merari & Ganor, 2020; Simi et al., 2015; Simi et al., 2016).

Item 6 assesses the validity of the measurement of mental disorder, and was coded as low-risk where there was reference to ‘confirmed’ or ‘formal’ etc., diagnosis. Item 7 assessed the risk of bias that may arises when assessment of diagnosis is made by individuals who are not trained in diagnostic interviews and systems. This item was coded as low-risk when individuals were trained. For both items, we coded studies as not-applicable where they were not assessing mental disorder and were, instead, focused on psychological problems (10 papers for Item 6, and 9 papers for Item 7). However, they were coded high-risk if they purported to be reporting on mental disorders but do not refer to how individuals were assessed (Item 6) or who conducted the assessment (Item 7). Thirty-three papers were low-risk for Item 6, and 21 papers were low risk for Item 7.

Item 8 assesses the appropriateness of the statistical analysis. As we are primarily interested in the positive events and total sample size for our meta-analysis of prevalence rates, once both are reported, or sufficient data is present to allow their calculation (e.g., percentage and denominator), studies were assessed as low risk. Nineteen papers were assessed as high-risk as a result of not including a sufficient amount of statistical information. However, 32 papers included the required information and were thus assessed as low-risk. As an exemplar, the most comprehensive reporting of data relevant to this review was provided by Duits et al. (2022) who reported numerator, denominator and proportion/percentages. They compared two samples of extremists (younger vs older) and also presented odds ratio, 95% Confidence Intervals and results of inferential statistics.

Item 9 focused on missing data. A large number of ‘missing data’ or ‘unknowns’ can reduce the study's validity. As a result, to reduce risk of bias, authors should compare those in the study (found cases) with missing data/unknown cases to assess nonresponse biases. Thirty-four papers were marked as high-risk as a result of not considering the missing data in their study.

Finally, as part of our ROB, we also explored the extent to which studies presented a theoretical account of the link between mental health difficulties and terrorist behaviour (plausibility). Again, it is relevant to note that mental health was not the primary focus of many of the studies included. This said, 41% of the papers (k = 21) made some attempt to explain how difficulties may be linked to terrorism and we return to these explanations in the Discussion section of the review.

Synthesis of results for Objective 1

As noted earlier, we have classified the included studies as reporting on diagnosed mental disorder, suspected disorder, and psychological problems, with some studies reporting on more than one of these categories. In our synthesis, we start by reporting the pooled estimates of life-time rates of mental disorder in terrorist samples. We then report pooled life-time estimates for suspected disorder, psychological problems and for any form of mental health difficulty (disorder, suspected disorders and psychological problems). Next, we report the point-prevalence rates of mental health difficulties (i.e., ‘now’).

Finally, we present two moderation analyses. The first explores the extent to which rates of mental health difficulties are moderated by studies' source of data and the second examines type of terrorism as a moderator. The latter, in particular, is important as one of the criticisms of the literature has been the failure to disaggregate different forms of terrorism when considering the link between mental health difficulties and terrorist behaviour.

Lifetime rates of mental disorder

As a benchmark against which to consider the pooled life-time prevalence rate of mental disorder we refer to Steel and colleagues' (2014) meta-analysis of lifetime prevalence rates of mental disorder among 16-65 year olds globally, and where included studies applied a structured psychiatric diagnostic interview. They concluded that 29.2% [95% confidence interval (CI) = 25.9%–32.6%] of people experience a mental disorder at some point in their lives. That sample was based on a pooled sample of 452,595 individuals from 82 countries.

We conducted a random-effects meta-analyses using the total sample size for the studies (k = 18, n = 3520), regardless of type of terrorism, and with the number of positive cases of mental disorder as a proportion of all cases (i.e., denominator = all cases, including missing cases, in all samples). As anticipated, between-study variability was high. The homogeneity test suggested that there was significant heterogeneity in the effect sizes across the included studies (Q ₁₇ = 334.4, p < .001) and 94.9% of the variability in effects is due to true differences across the studies rather than sampling error (I ² = 94.9; τ ² = 1.16). The prevalence estimates across the studies ranged from 0.5% (Chermak & Gruenewald, 2015) t0 42.4% (Cherney et al., 2022). The pooled prevalence was 17.4% (95% CI = 11.1%–26.3%) (see also Figure 2), which is lower than our benchmark from Steel et al. (2014).

OPEN IN VIEWER

Figure 2

Forest plot of lifetime prevalence of mental disorders in terrorist samples

We completed a sensitively analysis for the assessment of diagnosed mental disorder by first exploring the relative weights of each study, with no study responsible for more than 6.04% of the full weight and all studies contributing between 3.29% and 6.04%. As such, no study dominated the analysis. To complete the assessment we examined the pooled prevalence effects if any one study was removed from the data synthesis using the ‘one study removed’ option in CMA. As Chermack and Gruenewald (2015, 3 samples) and LaFree et al. (2018) both contribute data from large US datasets, we were particularly sensitive to the effects of removing the LaFree et al. data set. Doing so did not change the reported pooled prevalence considerably, increasing it slightly to 19.4% (95% CI = 13.5%–27.1%). Regardless of what study was removed, the point prevalence does not drop below 16.4% or exceed 19.4%.

We assessed publication bias through a contour enhanced funnel plot and Eggers' regression text. The funnel plot was symmetrical, with some effects falling outside the funnel. Egger's test for a regression intercept was p = 0.37 (two-tailed) suggesting no evidence of publication bias however (intercept = 2.41) even though the Trim and Fill test suggested the addition of four additional effects to be imputed (high, left and outside the funnel) to the left of the funnel to obtain symmetry in the plot (Figure 3).

OPEN IN VIEWER

Figure 3

Funnel plot for publication bias of studies reporting on lifetime prevalence of mental disorder in terrorist samples

Life-time rates of suspected disorder

We next replicated the analysis with studies that reported suspected mental disorder. In the data synthesis, seven papers with eight samples in total were included (n = 585). Again, a random-effects meta-analysis was run. There was evidence of heterogeneity in the effects reported across studies (Q ₇ = 57.69, p < 0.001) with almost 90% of the variability due true differences across the studies rather than sampling error (I ² = 87.87; τ ² = 0.68). The pooled prevalence was 23.2% (95% CI = 13.8%–36.4%), lower than the benchmark 29.2% from Steel et al. (2014) (see Figure 4). We assessed the impact of removing any one study from the analyses and noted that removing Corner and Gill (2015) led to an increase in the pooled prevalence rate to 29.5% (95% CI = 20.0%–41.2%). As the total number of studies was less than 10, in line with Cochrane guidelines, we did not assess publication bias (Page et al., 2022).

OPEN IN VIEWER

Figure 4

Forest plot of lifetime prevalence of suspected disorders in terrorist samples

Life-time rates of psychological problems

Next, we synthesised all studies and samples that reported on psychological problems. Some studies that reported diagnosed disorder or suspected disorder provided separate data for psychological problems. In such cases, we used the data relating to psychological problems. As a benchmark we draw on Winning et al. (2015) longitudinal study of psychological distress in adults from the 1958 British Birth Cohort Study, which reported that 50.7% of their sample of 6714 were distressed in childhood (24.07%, n = 1683), adulthood (14.97%, n = 1005) or both (10.68%, n = 717). We sought out other benchmarks but noted that psychological distress is typically measured as a point-prevalence (‘right now’) reflecting the episodic nature of some forms of distress (e.g., following bereavement) and because the lifetime prevalence rate could be expected at close to 100%.

Eighteen papers with 21 samples/studies were included in the meta-analysis (n = 3,806). There was heterogeneity in effects across the 21 studies (Q ₂₀ = 424.47, p < 0.001) with 95.3% of the variability due to true differences across the studies rather than sampling error (I ² = 95.29; τ² = 0.82). The pooled prevalence was 28.5% (95% CI = 21.0%–37.5%) which is less than both the benchmark for mental disorder and psychological distress (50.7%) (see Figure 5). Removing the FBI (2019) study from the data set reduced the pooled prevalence to 26.6% (95% CI = 19.6%–35.1%) and removing LaMontagne's (2019) left-wing sample increased the pooled prevalence to 30.8% (95% CI = 23.0%–39.9%). Freilich et al. (2019) study used the ECDB data set and LaMontagne's (2019) three samples were based on the PIRUS data set. Omitting Freilich and colleagues' data did not change our conclusion that the pooled prevalence of life-time rates of psychological problems does not exceed those expected in the general population (pooled prevalence = 29.4%, 95% CI = 21.3%–39.1%).

OPEN IN VIEWER

Figure 5

Forest plot of lifetime prevalence of psychological problems in terrorist samples

A funnel plot to assess publication bias was approximately symmetrical (Figure 6). The Trim and Fill test suggested that four inferred studies could be imputed left and outside of the funnel plot. Eggers' test suggested no evidence of publication bias however (point of intercept = 3.95, p = 0.09, two-tailed).

OPEN IN VIEWER

Figure 6

Funnel plot for publication bias of studies reporting on lifetime prevalence of psychological problems in terrorist samples

Life-time rates of any mental health difficulties

Finally, we combined all eligible studies, regardless of their reporting on disorder or problems, in one overall assessment of reported mental health difficulties in the identified literature. As some studies provide data for multiple mental health risks (e.g., disorder and problems), and other studies report different risks used overlapping datasets, we selected studies as follows. First, where one study reports two prevalence figures (e.g., psychological problems and diagnosed disorder; psychological problems and suspected disorder; suspected disorder and diagnosed disorder) we included only one effect in the synthesis with the most liberal risk selected. This meant that we prioritised risks in the following order: (1) Psychological Problems, (2) Suspected Disorder, (3) Diagnosed Disorder.

Second, if two studies reported results using overlapping datasets, but reported on different risks (e.g., study 1 = problems, study 2 = diagnosed disorder), we again prioritised problems over suspected disorder, problems over diagnosed disorder, and suspected disorder over diagnosed disorder. We do so as this strategy is most responsive to a lenient assessment of mental health issues in terrorist samples (we minimise the requirement for a clinical diagnosis).

Based on this approach, we excluded Chermack and Gruenewald's (2015) sample of right wing terrorists, for which they report mental disorder, in favour of Gruenewald et al.'s (2013) sample of right wing terrorists (for whom they report rates of suspected disorder). Both draw on the ECDB data set. Weenink (2015), which reported rates of diagnosed disorder, was displaced by Weenink (2019), which included but expanded up on their 2015 sample, but reported rates of psychological problems.

Thirty-two papers reported on 37 samples for the analysis (n = 5082). Again, there was evidence of considerable heterogeneity in the data (Q ₃₆ = 548.66, p < 0.001) with 93% of variability in effect sizes attributable to true differences across studies rather than sampling error (I ² = 93.44, τ ² = 0.75). The pooled prevalence was 25.5% (95% CI = 20.2%–31.6%). Removing any one of the studies (i.e., sensitivity analysis) would not have changed the conclusion reached that the pooled prevalence of any mental health difficulty reported here does not exceed that expected in the general population. Removing the FBI (2019) study would lead to slight reduction in the pooled prevalence (24.4%, 95% CI = 19.4%–30.3%) and removing LaMontagne's (2019) left-wing sample would lead to an increase in the pooled estimate to 26.6% (95% CI = 21.3%–32.8%). As LaMontagne's (2019) three samples derive from the PIRUS data set and Chermack and Gruenewald's (2015) two samples are based on the ECDB, we re-ran the analysis excluding the latter. Again, there was no meaningful change in the pooled estimate (pooled prevalence = 27.4%, 95% CI = 21.8%–33.8%).

Publication bias for all studies reporting mental health difficulties (i.e., k = 37) was assessed through a contour enhanced funnel plot and Eggers' regression text. The funnel plot was symmetrical. A Trim and Fill test suggested the addition of one additional effect to be imputed (left of the plot) though Egger's test for a regression intercept (intercept = 1.04) was p = 0.47 (two-tailed) suggesting no evidence of publication bias.

Moderator analysis of life-time rates of mental health difficulties

On reviewing the eligible studies, and noting the considerable heterogeneity in all analyses, we considered potential moderators that might explain, in part at least, the evident heterogeneity. Both moderators were anticipated a priori (see Sarma et al., 2022). The first was data source. It may be the case that the heterogeneity noted in the data reflects variation in the way that different data sources capture the presence of mental health difficulties. In particular, we noted that some of the studies were based on either interviews with terrorists or closed sources (or a mixture of open and closed sources). Others were based on open source information including media reports. We therefore coded studies as Interview/Closed Source or Open Source and repeated the analysis using the data set with all studies reporting mental health difficulties to determine if the effect sizes varied across the two sets of studies. Nine studies drew on closed sources or interviews and 28 studies used open-sources (k = 37, n = 5,082).

As evident from the Figure 5, the mean life-time prevalence rate of mental health difficulties for samples using closed source information or interviews was 39.1% (95% CI = 28.9%–50.4%), higher than that for samples based on open source information (21.2%, 95% CI = 16.5%–28.1%). The difference was statistically significant (Q ₁ = 8.24, p = 0.004). There was evidence of considerable heterogeneity across both sets of studies (Interview/Closed Q ₈ = 71.76, p < 0.001, I ² = 88.86, τ ² = 0.41; Open Source Q₂₇ = 379.52, p < 0.001, I ² = 92.89, τ ² = 0.72).

Next, we explored the extent to which different forms of terrorism might moderate the life-time prevalence rates of mental health difficulties reported in the studies. Past research has proposed that there may be higher rates of mental health difficulties among lone-actors. For this reason, we ran a moderator analysis with Lone-Actor versus all other forms of terrorism (including mixed samples). Again, studies reporting any form of mental health difficulty were included (k = 37, n = 5082).

Samples reporting rates for lone actors (k = 14) reported higher rates (36.1%, 95% CI = 29.1%–43.7%) than those reporting rates for other forms of terrorism (k = 23) (20.5%, 95% CI = 15.0%–27.5%). This difference was statistically significant (Q ₁ = 9.45, p = 0.002). Again, there was considerable heterogeneity across both sets of studies (lone actor Q ₁₃ = 60.74, p < 0.001, I ² = 78.60, τ ² = 0.27; Other Q ₂₂ = 376.21, p < 0.001, I ² = 94.15, τ ² = 0.75) (Figures 7–13).

OPEN IN VIEWER

Figure 7

Forest plot of lifetime prevalence of any mental health problem in terrorist samples

OPEN IN VIEWER

Figure 8

Funnel plot for publication bias of studies reporting on lifetime prevalence of any mental health difficulty in terrorist samples

OPEN IN VIEWER

Figure 9

Moderating effects of data source on life-time prevalence of any mental health difficulties in terrorist samples

OPEN IN VIEWER

Figure 10

Moderating effects of terror type on life-time prevalence of any mental health difficulties in terrorist samples

OPEN IN VIEWER

Figure 11

Forest plot of studies reporting point-prevalence (now) rates of disorder or suspected disorder in terrorist samples

OPEN IN VIEWER

Figure 12

Forest plot of studies reporting lifetime rates of psychological problems where problems were onset before detection or involvement

OPEN IN VIEWER

Figure 13

Forest plot of studies reporting lifetime prevalence rates of any mental health difficulties where these were onset before detection or involvement

Current mental health difficulties

Focusing on the 5 papers (7 studies, n = 384) reporting point prevalence of mental health diffiutlies, the pooled estimate of current mental disorder or suspected disorder was 14.4% with very wide confidence interval (95% CI = 4.8%–35.7%). There was heterogeneity evident across the studies (Q ₆ = 53.33, df = 6, p < .001) with 88.75% of the variability attributable to true differences in effects across studies rather than sampling error (I ² = 88.75; τ² = 2.05). None of these studies were included in the data syntheses reported earlier (i.e., analyses relating to life-time prevalence rates). We did not assess publication bias given the low number of studies. In our sensitivity analysis we noted that the removal of Bakker (2006) led to an increase in the point-prevalence rate to 22.5% (95% CI = 10.8%–41.0%). That study was the only study to use open-source information in assessing diagnosed disorder and may point towards the potential for open sources to underestimate prevalence rates.

Objective 2—Temporality

Objective 3—Mental health difficulties as a risk factor for terrorism

Synthesis for objective 3—Mental health difficulties as a risk factor

We considered the appropriateness of presenting a data synthesis of the eligible studies for Objective 3. A particular concern is that, while all the studies used a comparison group of non-terrorist offenders, these non-terrorist offender samples are non-homogenous. Dhumad et al.'s (2020) comparison group is comprised of incarcerated murderers from Iraq while Liem et al.'s (2018) comparison group were non-terrorist homicide offenders. Other studies drew on non-terrorist lone-actors, described as ‘grievance motivated lone actors’ (Kupper & Meloy, 2021), ‘school shooters’, ‘rampage shooters’ and ‘disgruntled mass shooters’ (Capellan et al., 2019), non-ideological mass shooters (Capellan, 2015) and solo mass murderers (Horgan et al, 2016). We acknowledge that this non-homogeneity raises questions about the value of a data synthesis via meta-analysis and so present single odds ratios for the studies.

Three studies reported point-prevalence rates of diagnosed disorders: Bronsard et al. (2022), Lyons and Harbinson (1986), and Dhumad et al. (2020). Bronsard et al. (2022) compared point-prevalence rates of disorder among 15 minors convicted for ‘criminal association to commit terrorism’ (p. 1) with 101 teenagers who had been convicted of non-terrorist activities and were detained in Closed Educational Centres. The odds of the extremist sample having a mental disorder was less than that of the non-extremist sample [odds ratio (OR) = 0.05, 95% CI = 0.01–0.18, p < 0.001]. In fact, the non-terrorist sample was 20 times more likely to have a diagnosed disorder. Similarly, Dhumad et al. (2020) compared rates of Antisocial Personality (ASPD) in a terrorist sample (n = 160) with a sample of convicted murders (n = 65) in Iraq and reported that the terrorist sample was less likely meet the diagnostic requirements for ASPD (OR = 0.51, 95% CI = 0.27–0.96, p = .04). The sample of murderers was almost twice as likely as the terrorist sample to be diagnosed with ASPD. Lyons and Harbinson (1986) also compare point-prevalence rates between terrorists and non-terrorist offenders but there was a discrepancy in the reporting of the numerator for the comparison group. They report that 60 of the non-political offenders had a mental disorder (a reported rate of 59%) but there were only 59 offenders in this sample. We did not attempt to contact the authors due to the time lapse since publication and an inability to identify an email address for either author. As we were unable to compute odds ratio for this study, we report the authors conclusion that terrorist offenders are less likely to have a mental disorder than non-terrorist offenders, and more likely to come from a stable family background.

The remaining studies report life-time prevalence rates of mental health difficulties. Capellan advised that the samples in their three studies overlap (Capellan & Anisin, 2018; Capellan, 2015; Capellan et al., 2019), and we, therefore, selected the most recently published study (Capellan et al, 2019). As the corresponding author provided additional data for this review (for prevalence of diagnosed disorder and psychological problems), the data reported here is not consistent with the data reported in the published journal article. In that study, the team compared ideologically motivated mass shooters with three types of non-ideologically motivated mass shooters (school shooters, disgruntled mass shooters and rampage shooters) but, for our analyses, we compare ideologically motivated mass shooters (terrorists) with the other groups combined (i.e., non-terrorists). We also included Kupper and Meloy's (2021) comparison of psychological problems among lone actor terrorists and grievance-motivated lone actors, Dhumad et al.'s (2020) comparison of conduct disorder among convicted terrorists and murderers in Iraq, Liem and colleagues' (2018) comparison of psychological problems among 98 European lone actors and 300 non-terrorist homicide offenders, and Horgan et al.'s (2016) comparison of suspected disorder across 71 lone actor terrorists and a comparison group of 115 solo mass murderers.

The effect sizes for the remaining studies are summarised in Figure 14. Due to the variation in comparison groups, it was not considered appropriate to synthesise the effects of these studies and draw broad conclusions in relation to Objective 3. Instead, the effect sizes are displayed in Figure 14 for summary purposes only, with no pooled estimate provided. For four of the studies the ORs ranged from 0.68 to 0.73 with 95% CIs overlapping 1 (i.e., the difference in odds across the groups in these studies was non-significant). The exception was Liem et al.'s (2018) study where the terrorist sample had a greater odds of having a life-time history of psychological problems than non-terrorist homicide offenders (OR = 3.13, 95% CI = 1.87–5.23, p < 0.001). Again, it is important to emphasise that these effect sizes should be interpreted in relation to each study's respective comparison group, rather than a homogenous ‘non-terrorist’ sample.

OPEN IN VIEWER

Figure 14

Summary of studies comparing life-time prevalence rates of mental health difficulties in terrorist samples and non-terrorist offender samples

In Liem et al.'s study, 37% of the terrorist sample (n = 36/98) had some indication of a mental illness with no missing values reported. For the common homicide offender, they report indications of mental illness for 47 cases in the sample, and which they report as 48% excluding 202 missing values (but we calculate as 15.66% based on the full sample of 300 offenders). The terrorist sample was drawn from the EU-funded CLAT project, with the homicide offender sample drawn from the European Homicide Monitor (EHM). As such, findings may, at least in part, reflect differences in data coverage across the two data sets.

Finally, Dhumad et al. used a second comparison group drawn from the community. As all other studies used offender populations as their comparison group, it is important to distinguish it from the previous studies. The terrorist sample had greater odds of a life-time history of childhood conduct disorder (OR = 2.24, 95% CI = 1.25–4.01, p = 0.007) than community controls, but the odds of a current diagnosis of ASPD did not differ across the groups (OR = 1.96, 95% CI = 0.94–4.09, p = 0.07).

Summary of main findings

The link, if any, between mental health difficulties and terrorist behaviour has perplexed researchers, policy makers and practitioners for several decades. Anecdotal evidence from end-users would suggest that mental health difficulties do present among those who become involved in terrorism, as well as those classified as ‘at risk’ of becoming involved. This has led to the establishment of multidisciplinary teams comprised of law enforcement and health workers, with the latter expected to consider the psychological vulnerabilities of those being assessed and offer the appropriate supports (see Sarma, 2018, 2019a, 2019b). This is done with the dual intention of alleviating distress and protecting society.

Yet, across the empirical literature, evidence on the link between terrorism and mental health difficulties has been inconsistent and, at times, conflicting. This review has presented a synthesis of the body of empirical literature that has reported prevalence rates of mental health difficulties in terrorist samples (Objectives 1 and 2) and compared these rates with rates in non-terrorist samples (Objective 3).

For Objective 1 we isolated studies reporting life-time rates of diagnosed mental disorder, suspected disorder, and psychological problems in separate analyses. In doing so, we anticipated that we would see marked increase in rates from reported problems to suspected disorder and diagnosed disorder (we expected diagnosed and suspected disorders to be less prevalent than psychological problems). We did see such a trend, through it was less marked than anticipated.

Of the 18 studies that reported on life-time prevalence rates of mental disorder, the pooled prevalence was 17.4% (95% CI = 11.1%–26.3%), with an increase to 23.2% (95% CI = 13.8%–36.4%) for the eight studies that reported life-time rates of suspected disorder, and 28.5% (95% CI = 21.0%–37.5%) for the 21 studies reporting psychological problems. When we collapsed all studies reporting problems, disorder and suspected disorder into one analysis, the pooled prevalence rate for the 37 studies included was 25.5% (95% CI = 20.2%–31.6%). At the upper-end, then, our pooled estimate of life-time prevalence rate of mental health difficulties (Objective 1) was 28.5%.

We anticipated that isolating studies that reported prevalence rates of difficulties that emerged before involvement in terrorism would be sensitive to temporal sequencing and thus complement our test our review (Objective 2). However, the studies captured typically focused on the point of detection/arrest in temporal sequencing rather than the point of involvement. As we cannot rule out the possibility that some individuals develop mental health difficulties between the point of engagement and point of detection, our analysis is less sensitive to temporal sequencing that we had intended. Our estimates here, based on a relatively small number of studies, were 30% (k = 5, 95% CI = 17.7%–45.9%) and 27.8% (k = 9, 95% CI = 20.9%–35.9%), respectively.

In our protocol, we had envisaged adopting a range of benchmarks for disorder and problems for different jurisdictions. However, given that the range of estimates reported here are lower than anticipated, we consider it sufficient to note the following: In Steel and colleagues' meta-analysis (2014) of lifetime prevalence rates of mental disorder among 16-65 year olds globally, and where included studies applied a structured psychiatric diagnostic interview, they concluded that 29.2% (95% CI = 25.9%–32.6%) of people experience a mental disorder at some point in their lives. That sample was based on a pooled sample of 452,595 individuals from 82 countries. We also note that Winning et al. (2015) longitudinal study of psychological problems (distress) in adults from the 1958 British Birth Cohort Study reported that 50.7% of their sample of 6,714 were distressed at some point in their lives.

In our analysis for Objective 1 we exceeded Steel et al.'s (2014) estimate only twice. Studies that reported rates based on closed sources or interviews with terrorists reported a life-time prevalence rate of 39.1% (95% CI = 28.9%–50.4%), though this was for any mental health difficulty (including psychological problems). Similarly, studies that reported rates for lone actor terrorists reported a life-time prevalence rate of 36.1% (95% CI = 29.1%–43.7%), again based on any mental health difficulties. In both moderator analysis, however, the inclusion of diagnosed disorder, suspected disorder and psychological problems in the prevalence rates sets a less conservative threshold for mental health difficulties than that of Steel et al., where a confirmed psychiatric diagnosis (determined through psychiatric interview) was required.

We could also refer to other estimates, including an estimate by Carlson and colleagues (2019) which specifically estimated the current rate (i.e., point-prevalence at any point in time) in conflict-affected populations at 22.1%. This finding is relevant as it demonstrates that the point-prevalence rates of mental disorder or suspected disorder identified in our review (14.4%, 95% CI = 4.8%–35.7%) is no greater than we would anticipate in the general population.

We also conducted a summary of single effects for the studies that compared terrorist samples with other samples (Objective 3). In our synthesis of five studies reporting life-time prevalence rates of mental health difficulties, the studies trended towards terrorist samples being broadly comparable to their respective, comparison group. Two studies (Bronsard et al., 2022; Dhumad et al., 2020) provided data comparing point-prevalence (current) rates of disorder between terrorist and non-terrorist samples, with results from each study suggesting that the odds of the terrorist samples having a disorder was lower than that of non-terrorist offenders. One study compared a terrorist sample with a community sample (Dhumad et al., 2020). In that study, the terrorist sample had a greater odds of having had a conduct disorder in childhood but were there was no difference in the odds of meeting the diagnostic criteria for ASPD.

In summary, our findings in relation to prevalence rates of mental health difficulties (Objective 1 and Objective 2) have provided unremarkable pooled estimates for terrorist populations as a whole, with some evidence that lone actor terrorists have higher rates of difficulties that other terrorist populations in line with earlier research (Corner & Gill, 2015; Gruenewald et al., 2013; Liem et al., 2018). It is not clear, however, whether these rates, which relate to any mental health difficulty, exceed those expected in the general population. It is important to note that the dearth of studies drawing from comparative general populations is a major limitation in the literature as a whole. Studies comparing terrorist samples with other samples have relied primarily on serious criminals as a comparison group and, perhaps unsurprisingly, do not provide evidence that mental health difficulties are associated with terrorist involvement.

At the outset of this review, we referred to the Bradford-Hill criteria for determining causation. We viewed these criteria as important in the context of terrorism research because they point to the kinds of evidence that one would expect to see if there was a causal link between the risk and the outcome, even in the absence of prospective longitudinal case-control studies. Here we suggested that if there was a causal link between mental health difficulties and terrorism, then we could reasonably expect to see higher prevalence rates of difficulties in terrorist samples than the general population or a comparison group (prevalence), and that this would be particularly marked where studies were sensitive to temporal sequencing (temporality). The evidence from this review did not support either criterion. As discussed further below, this may reflect the methodological limitations of the body of evidence and the narrow scope of the review.

The review was also interested in the Bradford-Hill criterion of plausibility – that there should be a strong theoretical link between the risk and outcome. We considered this as part of our Risk of Bias assessment. In general, there is a sense of convergence in the literature included in the review where authors are arguing that the link between mental health and terrorism is (a) complex, (b) indirect and involving interaction with other psychosocial experiences and (c) highly individualised (e.g., Corner et al., 2019). At most mental health difficulties present a background non-determinative predisposing factor nested among other factors and that may present a vulnerability to terrorism through the depletion of resilience, impairment of judgement, restriction of opportunity, and an almost unlimited range of other life experiences and situational factors. Such complexity is not uniquely a feature of trajectories into terrorism—it characterises almost all forms of serious offending where no single risk factor or process is determinative.

Bubolz and Simi (2019) illustrate this approach to formulation in their examination of mental health problems among former members of white supremacist groups, suggesting that these psychological problems may be rooted in traumatic experiences, often during childhood and adolescence and compounded by experiences of rejection in adulthood. Ultimately this may create a susceptibility to being drawn into extremist groups. Jensen and Kane (2021) reach a similar conclusion based on their research with QAnon followers as do Cherney and colleagues (2022) following their analysis of young people who radicalised to violence in Australia. They suggest that background predisposing factors like mental health difficulties and ‘deviant conduct’ interact with exposure to radicalising content and networks to ‘set up a potential pathway to violent extremism’ (p. 112). Others observed interactions between mental health difficulties and past non-terrorist criminality (Thijssen et al., 2021), exposure to military-related trauma (Haugstvedt & Koehler, 2021) and grievances such as poverty and oppression (Candilis et al., 2021).

With one exception, studies included in the review didn't propose that mental health difficulties precipitated terrorist involvement. Clemmow et al. (2020) do appear to allude to such a process however. They viewed lone actor terrorism as heterogeneous and develop a typology of lone-actor violence based on person-exposure patterns. Of four types that they isolate, one is labelled ‘susceptible’ and, according to the authors, may radicalise quickly due to high impulsivity and psychiatric disorder which presents a ‘core cognitive susceptibility to environmental influence’ (p. 468).

The possibility that mental health difficulties are protective against terrorist involvement was also considered in the studies, with most echoing what King et al. (2018) refer to as a selection effect, whereby terrorist organisations prefer recruits who can be trusted, are reliable and can take direction. This said, Bubolz and Simi (2019) found that white supremacist organisations are open to individuals with mental health problems and, in contrast with other forms of terrorism, ‘may actually prefer’ (p. 10) to recruit such individuals (p. 19).

In summary, the findings across our three review objectives converge with the theoretical positions set out in the papers. We don't see indications/tentative evidence of a causal relationship between mental health difficulties and risk of engagement in the pooled prevalence rates. Instead, our authors are arguing for more nuanced ways of thinking about the role mental health may play in the lives of those who may be at risk of involvement in terrorism.

Overall completeness and applicability of the evidence

We searched a broad range of platforms and databases, unrestricted by year of publication. Anticipating that some studies reporting relevant data would not have mental health as a primary focus, and thus may not refer to mental health in the title or abstract, we also completed an extensive supplementary search of the grey literature, contacting leading authors and expert networks, by hand searching specialist journals and by harvesting records from included studies. We believe that we have captured the bulk of studies that meet our eligibility criteria.

Clarification or additional data was requested from corresponding authors representing more than half of all included studies. This aided our decision making with regard to prioritising and excluding non-independent studies (overlapping data set). It also helped us determine how authors managed missing data and allowed us to standardised our approach to calculating effect sizes (i.e., by counting positive cases as the numerator and all cases as the denominator).

Despite these strengths, we would acknowledge that our review had a number of limitations which hampers its applicability. In particular, we expect that there are additional, unpublished and published studies that we have not captured in our search, and acknowledge that the exclusion of studies not published in English undermines the overall completeness of the review. We also acknowledge that our review was narrow in scope and that there are other sources of evidence, including qualitative studies, that offer an equally valuable, and complementary, lens through which to explore the role of mental health difficulties in terrorism.

Quality of the evidence

Notwithstanding the barriers to conducting research with terrorist samples, and the high threshold set by the JBI checklists deployed in our risk of bias assessment, there were a number of limitations in the body of evidence included. We highlight three here. First, the majority of studies included were based on open-source datasets that are almost entirely reliant on media and court reporting. While such datasets are valuable for dealing with many research questions, we would question to what extent open sources accurately capture the presence and absence of mental health difficulties and, by extension, the validity of interpretations of disorder or clinical psychological problems by researchers or clinicians based on such data.

Bakker (2006) captures the limitation of open sources well, reporting that ‘it is impossible to gather complete information from open sources for all variables… This means that our generalisations sometimes required some crude judgement on our part’ (p. 43). Others have been more vocal in their concerns, with Brym and Araj (2012) referring to researchers making inferences as to the characteristics of terrorists based on ‘scraps of biographical evidence pasted together from newspaper reports and other published sources’ (p. 433) and Kupper and Meloy (2021) acknowledging that their own source of data (manifestos) offered ‘very limited ability’ to identify the presence or absence of mental health difficulties (p. 14). Conversely, whilst acknowledging the limitations of open sources in terrorism research, Clemmow and colleagues (2020) note that systematic approaches to trawling open sources, extracting data and coding that data can mitigate against some concerns. Similar defences have been mounted by others (e.g., Cherney et al., 2022), and we see evidence of this rigour in most of the studies included in this review. The detailed accounts of the processes used to build the datasets in the eligible studies would suggest a more systematic process than suggested by Brym and Araj.

Some have argued that one way of overcoming this limitation is through in-depth research with the terrorists themselves (e.g., Bubolz & Simi, 2019), an approach adopted by a number of studies included in this review. However, even here there have been concerns about the reliability of data based on retrospective recall of life experiences (Simi et al., 2016).

A second limitation of the evidence base is uncertainty as to the representativeness of the samples recruited, with many of the authors of the papers included here acknowledging that their findings may not generalise to other groups within their jurisdictions or apparently similar forms of terrorism in other parts of the world (e.g., Candilis et al., 2021; Pfundmair et al., 2022).

Third, few studies considered specific diagnoses and links to terrorism risk (for exceptions see Bronsard et al., 2022; Thijssen et al., 2021). The potential value of doing so is illustrated by Thijssen et al.'s (2021) study with incarcerated terrorist offenders. Having reported rates of diagnosed disorders that are in-line with that expected in the general population, they then noted that rates of schizophrenia and autism spectrum disorder (ASD) were higher than expected. With regard to ASD, they suggest that individuals with ASD may have social and communication deficits and be detail-orientated, both of which may attract them to on-line environments and extremist content.

Fourth, we believe it is important to highlight the two different approaches to treating missing values adopted in the studies included and the consequences for reported effect sizes. Some papers based prevalence estimates for mental health difficulties based on positive cases against all cases for which the mental health status was known (i.e., confirmed positive or negative cases, but excluding unknown cases, and sometimes referred to as ‘valid precent’). Others reported rates based on positive cases against all cases. Our synthesis of all studies was based on the latter approach.

To illustrate the consequence of these divergent approaches it is worth considering LaFree and colleagues (2018) study of 1473 extremists in the US and based on open-source data (where significant levels of missing values for mental health difficulties might be expected). Researchers were instructed to code disorder as a missing value if there was no reference to mental disorder in the sources. In their main analysis they report that this left them with 284 individuals for whom they had data, and 43% of whom had a mental illness. However, elsewhere they note that if they include all cases in the denominator, then the ‘effective rate of mental illness in the data set drops to 8.4%, which is comparable to rates of mental illness in the general population’ (p. 246).

Divergent approaches to managing missing data is not in itself an issue of quality (just reporting standards). What would improve the quality of the data, however, is consistent analysis of missing values to assess bias. Few studies included in the review presented such an analysis, though there are notable exceptions (e.g., Bronsard et al., 2022).

Potential biases in the review process

Our exclusion of studies that were not published in English (e.g., Alberda et al., 2018) introduces a potential source of bias in the review. It is hard to predict what effect this exclusion may have had on our synthesis. One review of 50 meta-analyses concluded that when non-English language papers were included in those reviews the estimates reported in the meta-analyses changed, but in inconsistent directions (Neimann Rasmussen & Montgomery, 2018). In five studies, the estimates became more positive and, in 16, the estimates became less positive. While acknowledging that there can be significant resource requirements to running searchers in multiple languages (and screen and extract (reliably) data from those records), the authors concluded that ‘non-English studies are important to include to avoid bias in reviews’ (p. 4).

Second, we did not screen titles and abstracts in duplicate which has been found to introduce biases. Waffenschmidt et al. (2019) presented a synthesis of papers that compared the reliability of single versus conventional double screening in systematic reviews. They noted that even among experienced reviewers, the median proportion of missed records was 3% when not screened in duplicate. In the current review, the collective experience of the reviewers, as well as time constraints associated with the large number of records returned from our search, informed our decision not to screen the titles and subtracts in duplicate. SC and KC are experienced reviewers with content expertise. They have screened for Campbell Collaboration and Cochrane reviews in the past, including one review in the area of terrorism where they screened together (Carthy et al., 2020). SC has a PhD in the area of psychology and terrorism and KC is a forensic psychologist who has co-authored reports with KS on the link between mental disorder and terrorism.

Third, two key decisions made by the research team shaped the review and analyses. First, we selected studies that met our conceptualization of terrorism. In the main this was uncomplicated, as the samples recruited engaged in behaviours that could be broadly considered as terrorism. However, this was not always the case. For example, we included Jensen and Kane's (2021) QAnon sample in our review. The sample is a sub-sample (n = 51) of QAnon followers and excludes those involved in the Capital Hill riots on January 6th, 2021. The QAnon movement is viewed as an emerging and evolving threat by the US authorities who have expressed concern that more followers will move from radical on-line behaviour (‘digital soldiers’) to violence. However, some have cautioned that most followers pose little threat of violence and that researchers run the risk of exaggerating the threat they pose by conflating radical thought and radical action (Moskalenko & McCauley, 2021). Others may have excluded this sample. In this case, we sought to be inclusive and selected the study as being eligible for our review.

Fourth, we classified studies as reporting on mental disorder, suspected disorder and psychological problems. We have sought to make this process rigorous and transparent in our reporting but recognise that we could have set more stringent requirements in classifying a study as reporting on diagnosed disorder by requiring a confirmed clinical assessment. We could also have set a less stringent requirement by not disaggregating those with diagnoses and suspected diagnoses. Others may have approached this differently.

Finally, we acknowledge the importance of moving beyond data based on general terrorist samples towards a more nuanced assessment of rates of mental health difficulties across different forms of terrorism. In our disaggregation of forms of terrorism in our data synthesis, we compared studies that reported on lone-actor terrorist samples with studies reporting on all other terrorist samples. However, as the literature base builds, future syntheses will be in a position to disaggregate the analyses further and, thus, consider rates of mental health difficulties across different forms of terrorism (e.g., Jihadi vs. lone actor vs. right-wing vs. separatist, etc.).

Disagreements or agreements with other studies or reviews

Our conclusion that the evidence reporting period and point prevalence of mental health difficulties does not support the mental health—terrorism hypothesis is in-line with much of the literature. Our estimates differ from an earlier review, but not remarkably so. Gill et al. (2021) completed a review and meta-analysis of 19 studies that reported rates of mental disorder in terrorist samples. They reported a pooled prevalence rate of confirmed diagnosis of 14.4%, with higher rates for studies using closed data (16.96%) than those using open sources (9.82%). They also noted that studies comparing lone actors with other terrorist samples reported higher rates in the former than the latter. They conclude that a) the evidence does not support the presence of e a unique profile of psychopathological traits that may leave an individual vulnerable to radicalisation but b) lone- and group-terrorists may appear to be two distinct groups of people in terms of their drivers and criminogenic needs.

Our review reports higher rates of mental disorder than reported in Gill and colleagues' (2021) synthesis and observes the same trend with regards of lone actors compared to other forms of terrorism. The higher rates reported here may reflect a more inclusive approach where we permitted terms like ‘confirmed diagnosis’, and reference to ‘disorder’ and diagnostic criteria as evidence that authors were counting medically diagnosed disorders.

Authors' conclusions

Based on our synthesis of this evidence base, the review does not lend support to the assertion that terrorist samples are characterised by higher rates of mental health difficulties than that observed in the general population. Accepting that there are methodological limitations in the studies extracted, the pooled life-time prevalence estimate for mental disorder is close to the prevalence rate seen in the general population and the confidence intervals of that estimate are within the range expected for the general population. The exception may be lone actor terrorism, where rates were significantly higher than all other forms of terrorism combined. However, even here the rates are broadly in line with what is expected in the general population.

Findings and conclusions are tentative, reflecting limitations in the body of evidence. We do not discount the possibility that as more evidence emerges our understanding in the area will evolve and conclusions may change. We accept that regardless of findings deriving from group-based statistics, mental health problems may emerge as relevant in an individual case when that individual is understood in context, as part of a broader psychosocial picture. This is in line with current thinking in the area (e.g., Thijssen et al., 2021).

Implications for policy and practice

In considering the implications of our review findings for policy and practice we are acutely aware of the narrow focus of our review, and issues with the reliability and validity of the body of evidence as a whole. We also acknowledge that in restricting the review to prevalence and case control studies we have taken just one of many lenses that could be used to consider the mental health—terrorism hypothesis, and that there is equally important social scientific literature that adopts qualitative approaches. These too provide valuable context and insights in the area. Against the backdrop of these limitations, we tentatively suggest that following implications of our findings.

The first implication relates to the focus of the debate around mental health and terrorism among policy, practice and research stakeholders, and which seems to all-too-frequently view mental health difficulties as a causal risk factor for violence and assume that if we could only deal with these mental difficulties we would see a reduction in risk. This case is not supported in the evidence reviewed here because there is little evidence from prevalence studies that mental health problems are more prevalent in terrorist samples than in the general community (terrorism in all its forms, as opposed to specific forms of terrorism). We note the emerging consensus that mental health difficulties are one of many indirect non-causal predisposing risk factors for terrorist involvement just as such difficulties are a background predisposing risk factor for a whole host of other negative outcomes in life including delinquent and offending behaviour (e.g., Malvaso et al., 2016), interpersonal violence (Coid et al., 2021) and cardiovascular disease (Suls & Bunde, 2005).

There may be exceptions for specific forms of terrorism including lone actors, or specific sub-types of lone actors. This points to the importance of disaggregating terrorism into more granular sub-types and continuing to seek out any evidence that mental health difficulties play a more direct or determinative role in specific forms of terrorism (e.g. Gill & Corner, 2017). Until such a case is made, however, we should neither assume that a more direct link exists nor discount that it may.

If it is accepted that the best evidence points towards mental health difficulties as a being an indirect background predisposing risk factor for terrorism involvement, then there are some obvious implications for both risk assessment of those believed to be at risk of terrorism involvement/recidivism and for broader preventative measures. Mental health should remain a focus of risk assessment. However, the focus should not be just on a possible direct link between a mental disorder and violence propensity (not supported by the evidence reviewed here), but also the implications of such difficulties for management and supervision (e.g. engagement and compliance with interventions to reduce risk), homelessness, social isolation, personal support, stress and coping, employment and a host of other difficulties that are associated with psychological problems and also offending behaviour. This observation resonates with the conclusions reached by authors of the studies included here who argue in favour of non-causal links between mental health and terrorism (e.g., Bubolz & Simi, 2019; Jensen & Kane, 2021). It also resonates with the broader practice of forensic risk assessment, and central to popular risk assessment systems including the HRC-20 (Douglas et al., 2013).

That rates of mental health difficulties are in line with those in the general population suggests that there is value in considering public health primary prevention approaches to countering terrorism which has been recommended by others (e.g., Bhui & Jones, 2017). Here the focus is upstream of the problem of concern, where resources are invested in communities to bolster resilience and access to healthcare etc., to reduce rates of psychological problems or ability to cope with problems and with the potential consequence that communities become resilient to crime in all its forms.

Implications for research

Despite a relatively large and growing literature on terrorism and mental disorder, it remains an area of research that is difficult to expose to academic enquiry. The barriers to conducting this type of research are many and varied. Indeed, the limitations identified in the evidence-based are predominantly born out of these challenges, rather than any oversight on the part of researchers. This said, the review has highlighted a number of key limitations in the literature on terrorism and mental disorder that can serve to either, (a) enhance our ability to critical evaluate research findings or (b) improve the way we do research in the first place. These follow from the conclusions of our Risk of Bias assessment and discussion of the quality of the evidence encountered during this review.

First, databases constructed from open sources such as media reports and court reports, present a readily available source of potentially rich data on the characteristics of terrorists and their behaviours. However, the value of open sources in terms of classifying cases as either ‘having’ or ‘not having’ a diagnosis of a mental disorder is entirely contingent on such information being made open-source in the first place, as well as its accuracy. There is a need to complement studies using such databases (which represented the majority of our studies), with studies that use alternative sources of data. These alternatives will also have strengths and limitations, but collectively the findings from studies incorporating different methodologies allow us to measure the phenomenon through different tools, thus having more confidence in their generalisability.

Second, there is a need for some reporting standards to be adopted in the field. The review raised concerns about the treatment of missing values and the potential for the exclusion of missing cases from the denominator in calculating prevalence rates to dramatically inflate these rates. At the same time, including missing cases may serve to artificially decrease the prevalence rates. The obvious solution is to report both prevalence estimates as a standard, highlight the difference between the two and analyse any bias in patterns of missing data.

Third, we will need increasingly nuanced research to explore non-causal processes that may lead to terrorism. Research needs to increasingly disaggregate terrorism into all its forms and sub-forms, and also consider the links between specific forms of mental health difficulties and terrorist behaviour. In this way, disaggregation of both risk and outcome are required. Although we have observed the disaggregation of mental health difficulties into formal diagnoses and other, less clinical constructs (e.g., psychological problems), disaggregation of the former is lacking, limiting our ability to understand the nature of the relationship between specific, types of disorder and terrorism involvement. In terms of the outcome, we have observed some disaggregation. However, to date, studies have relied primarily on serious criminals for comparison groups and, perhaps unsurprisingly, do not provide evidence that mental health difficulties are associated with terrorist involvement