Risk and Protective Factors for Personality Disorders: An Umbrella Review of Published Meta-Analyses of Case–Control and Cohort Studies

The putative risk/protective factors for several personality disorders remain unclear. The vast majority of published studies has assessed personality characteristics/traits rather than disorders. Thus, the current umbrella review of meta-analyses (MAs) aims to systematically assess risk or protective factors associated with personality disorders. We searched PubMed–MEDLINE/PsycInfo databases, up to August 31, 2020. Quality of MAs was assessed with AMSTAR-2, while the credibility of evidence for each association was assessed through standard quantitative criteria. Out of 571 initial references, five meta-analyses met inclusion criteria, encompassing 56 associations of 26 potential environmental factors for antisocial, dependent, borderline personality disorder, with a median of five studies per association, and median 214 cases per association. Overall, 35 (62.5%) of the associations were nominally significant. Six associations met class II (i.e., highly suggestive) evidence for borderline personality disorder, with large effect sizes involving childhood emotional abuse (OR = 28.15, 95% CI 14.76–53.68), childhood emotional neglect (OR = 22.86, 95% CI 11.55–45.22), childhood any adversities (OR = 14.32, 95% CI 10.80–18.98), childhood physical abuse (OR = 9.30, 95% CI 6.57–13.17), childhood sexual abuse (OR = 7.95, 95% CI 6.21–10.17), and childhood physical neglect (OR = 5.73, 95% CI 3.21–10.21), plus 16 further associations supported by class IV evidence. No risk factor for antisocial or dependent personality disorder was supported by class I, II, and III, but six and seven met class IV evidence, respectively. Quality of included meta-analyses was rated as moderate in two, critically low in three. The large effect sizes found for a broad range of childhood adversities suggest that prevention of personality disorders should target childhood-related risk factors. However, larger cohort studies assessing multidimensional risk factors are needed in the field.

The putative risk/protective factors for several personality disorders remain unclear. The vast majority of published studies has assessed personality characteristics/traits rather than disorders. Thus, the current umbrella review of meta-analyses (MAs) aims to systematically assess risk or protective factors associated with personality disorders. We searched PubMed-MEDLINE/PsycInfo databases, up to August 31, 2020. Quality of MAs was assessed with AMSTAR-2, while the credibility of evidence for each association was assessed through standard quantitative criteria. Out of 571 initial references, five meta-analyses met inclusion criteria, encompassing 56 associations of 26 potential environmental factors for antisocial, dependent, borderline personality disorder, with a median of five studies per association, and median 214 cases per association. Overall, 35 (62.5%) of the associations were nominally significant. Six associations met class II (i.e., highly suggestive) evidence for borderline personality disorder, with large effect sizes involving childhood emotional abuse (OR = 28.15, 95% CI 14.76-53.68), childhood emotional neglect (OR = 22.86, 95% CI 11. 55-45.22), childhood any adversities (OR = 14.32, 95% CI 10.80-18.98), childhood physical abuse (OR = 9.30, 95% CI 6. 57-13.17), childhood sexual abuse (OR = 7.95, 95% CI 6. 21-10.17), and childhood physical neglect (OR = 5.73, 95% CI 3.21-10.21), plus 16 further associations supported by class IV evidence. No risk factor for antisocial or dependent personality disorder was supported by class I, II, and III, but six and seven met class IV evidence, respectively. Quality of included meta-analyses was rated as moderate in two, critically low in three. The large effect sizes found for a broad range of childhood adversities suggest that prevention of personality disorders should target childhood-related risk factors. However, larger cohort studies assessing multidimensional risk factors are needed in the field.

INTRODUCTION
Personality disorders are defined as "an enduring pattern of inner experience and behavior that deviates markedly from the expectations of the individual's culture, is pervasive and inflexible, has an onset in adolescence or early adulthood, is stable over time, and leads to distress or impairment." It has been reported that globally, personality disorders have a prevalence around 3 to 10% in the global population (1), and much higher in people affected by other mental disorders (2), and so it is considered a global mental health priority (3). Their peak age at onset is at age 20.5 years (4).
In the last 50 years, a lot of interest went into personality disorders, promoting their status from an unreliable and not so valid diagnosis before the 1960s, to a condition with clear diagnostic criteria in particular after introduction of DSM-III (5). Then, DSM-based criteria have raised criticism, and alternative diagnostic frameworks have been proposed (1). They are now classified differently in Diagnostic and Statistical Manual (DSM), 5th version (6) and in International Classification Diseases (ICD)-11 (7). DSM-V classifies personality disorders in three clusters (A, B, and C) and "other personality disorders." Cluster A includes paranoid personality disorder, schizoid personality disorder, and schizotypal personality disorder; cluster B includes antisocial personality disorder, borderline personality disorder, histrionic personality disorder, and narcissistic personality disorder; cluster C includes avoidant personality disorder, dependent personality disorder, and obsessive-compulsive personality disorder (6). ICD-11 applies a different approach, with the aim to identify fewer categories, which overlap less and ultimately have greater clinical utility (1). Specifically, ICD-11 categorizes personality disorders (6D10) into mild personality disorder, moderate personality disorder, and severe personality disorder, in addition to which Prominent personality traits or patterns (6D11), namely, Negative affectivity, Detachment, Dissociality, Disinhibition, Anankastic, or Borderline pattern in personality disorder or difficulty, must be further specified (7). Structural brain alterations, namely, bilateral gray matter reductions in concentrations in ventral cingulate gyrus, medial temporal lobe, and fronto-limbic structures, are associated with different personality disorders defined according to several versions of DSM (8)(9)(10)(11). Beyond cross-sectional associations with phenotypes and biomarkers, several risk factors for personality disorders have also been described. An overview of systematic reviews (12) focused on systematic reviews on risk factors for personality disorders. However, it only focused on parenting style, and it also focused on qualitative reviews rather than quantitative meta-analyses (12). Also, evidence from meta-analyses is frequently biased, and the credibility of the claimed associations between putative risk or protective factors for personality disorders remains unknown. To fill this gap in the literature, we conducted an umbrella review focused on environmental risk and protective factors for personality disorders, to identify and measure possible methodological limitations and sources of bias in the published and unpublished evidence, which might have underestimated or inflated claimed associations, as previously shown in several previous umbrella reviews on risk factors for mental disorders or obesity (13)(14)(15)(16)(17). Therefore, the aim of this umbrella review was to grade the evidence from meta-analyses of cohort and case-control studies on protective and risk factors for personality disorders accounting for several sources of bias and applying established quantitative criteria.

METHODS
This umbrella review adhered to state-of-the-art methods of previously published or planned umbrella reviews (15,(18)(19)(20)(21)(22)(23), and according to the Meta-analysis of Observational Studies in Epidemiology (MOOSE) and the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) were followed in conducting and reporting this umbrella review (Supplementary Tables 1, 2) (24,25). The study followed an a priori protocol, available on request. MS, AFC, PF-P designed the study, prepared the search key, and drafted the protocol. ED run the statistical analyses. Four investigators (MS, ED, AM, and PK, all MDs) divided into two couples independently performed literature screening and data extraction, including quality assessment of included meta-analyses.

Literature Search Strategy
We conducted a systematic search in PubMed and PsycINFO from inception to August 31, 2020. We included meta-analyses of case-control and cohort studies that assessed risk or protective factors for personality disorders, defined according to ICD or DSM, any version (6,7,26). The search strategy was "(personality disorder) AND meta-analysis." No restrictions regarding year of publication, language, country, ethnicity, or any other characteristic were applied during the search process. We also hand searched references of included meta-analyses and other relevant articles. When authors did not agree regarding screening or data extraction, a third author (MS) resolved any conflict, reaching a consensus with the two authors.

Eligibility Criteria
We only included systematic reviews that also conducted a quantitative meta-analysis pooling data from case-control or cohort (either retrospective or prospective) studies reporting on environmental factors that may affect the risk of the disorders of interest, as per the ICD or DSM criteria. Specifically, we included disorders that corresponded to ICD-10 "F60 Specific personality disorders" (26) and "Personality disorder" in DSM-5 (6). Risk or protective factors of interest were deemed eligible, regardless of the direction of the association (protective or risk factor). No language restriction was applied.
Meta-analyses of studies that included other-than-human population, having a cross-sectional design, as well as focusing on genome-wide associations or on single nucleotide polymorphism were excluded. Also, systematic reviews without a quantitative meta-analytic data synthesis, narrative reviews, and commentaries/letters to the editor were not included in the present umbrella review.
Finally, if multiple meta-analyses investigated the same risk or protective factor and the same outcome, only the metaanalysis with the largest number of studies pooled to measure the association was retained.

Data Extraction
We extracted information into a standardized pre-defined template. The list of variables of interest included PMID/DOI of the included study, first author, year of publication, design of included studies (cohort, case-control), number of included studies in the meta-analysis, specific population cohort (i.e., general population, primary school, secondary school, university students, hospital sample, or a sample with a specific somatic, mental, or somatic/mental comorbid condition, etc.) as well as the reference/comparison population (i.e., no risk factor in cohort studies, no disorder in case-control studies), tools for the definition of both population and risk/protective factor (DSM, ICD, clinical records, rating scales), specific protective or risk factor, outcome (ICD or DSM code if available, or definition of specific disorders as reported by authors given inclusion criteria were met), and its risk estimate. We assessed the methodological quality of included meta-analyses as independent couples of two investigators (JD, PK, AM, and MS, all MDs) by means of the AMSTAR (A Measurement Tool to Assess Systematic Reviews) version 2 (27). If needed, we contacted authors to ask for data.

Data Analysis
For each association (i.e., between each specific risk or protective factor and personality disorder), we pooled effect sizes of individual studies reported in each meta-analysis, as well as cases developing personality disorder and total sample size, and recalculated the pooled effect sizes with its 95% CIs, using random-effects models (28). We transformed the effect sizes or modified the direction of associations reported in original publications only for the associations with continuous or correlational data (e.g., Hedges g, beta coefficients) to present comparable estimates (i.e., equivalent odds ratio-eOR) (22). Heterogeneity was measured with the I 2 statistic (29). In addition, 95% prediction intervals for the effect sizes were computed to estimate the possible range in which the effect sizes of future studies were anticipated to fall (30). We also examined small-study effect bias, testing whether smaller studies generated larger effect sizes compared with larger studies (15,(18)(19)(20)(21)(22)31). Specifically, as indicators of small-study effect, we used both the Egger regression asymmetry test (p ≤ 0.10) and whether the random-effects summary effect size were larger than the effect size of the largest study contributing to that association (18,20,21,31). We finally measured the presence of excess significance bias by assessing whether the observed number of studies with nominally statistically significant results was different from the expected number of studies with statistically significant results (32,33). The expected number of statistically significant studies per association was calculated by summing the statistical power estimates for each component study. The power estimates of each component study depend on the plausible effect size for the examined association, which we assumed to be the effect size of the largest study (i.e., the smallest SE) per association (33). For excess significance bias, a p ≤ 0.10 was considered statistically significant (32). All analyses were conducted in Stata/MP, version 10.0 (StataCorp LLC).

Assessment of the Credibility of Evidence, and Sensitivity Analyses
In line with former umbrella reviews (14,16,17,34). eligible associations for personality disorders were classified into five levels according to the strength of the evidence of potential environmental risk/protective factors: convincing (class I), highly suggestive (class II), suggestive (class III), weak (class IV), and not significant (NS) ( Table 1). Sensitivity analyses were conducted removing the >1,000 cases criterion.

Search Results
Out of 571 initial hits, 436 were assessed at title and abstract level, and 415 were excluded. The remaining 21 were assessed at full-text level, and 16 were excluded, ultimately including 5 meta-analyses. The list of excluded meta-analyses after full-text assessment, with specific reason for exclusion, is reported in online Supplementary Material. Finally, five meta-analyses were included in this umbrella review (35-39) ( Table 2). The study selection flow is reported in Figure 1. The whole dataset with results from individual studies is available on reasonable request.

Descriptive Results of the Included Associations
The characteristics of included studies are described in Table 2. The five eligible meta-analyses corresponded to 56 with available data for synthesis, between 26 potential putative risk/protective factors and 3 personality disorders (antisocial, borderline, and dependent personality disorder) (Tables 2, 3). The eligible metaanalyses were published between 1999 and 2020. The median number of studies per association was 5 [interquartile range (IQR) = 3 -14]. The median number of cases per association was 214 (IQR = 98-2,420) and the number of cases was >1,000 in 14 associations, while in 2 associations the number of cases was not reported (Tables 3, 4). All eligible meta-analyses used summary-level data from published literature. No protective factors were identified.

Quality Assessment of Included Articles
Based on the AMSTAR2 assessment, two meta-analyses (40%) met the moderate quality level and three (60%) were of low quality ( Table 2).

Associations for Antisocial Personality Disorder
A total of 14 of the 56 associations examined associations for antisocial personality disorder. Six of those presented a nominally  statistically significant effect (p ≤ 0.05) and met the class IV evidence criteria while only 1 association reached p < 10 −6 . None of those associations were supported by class I, II, and III evidence. After excluding the criterion of 1,000 cases, one factor, namely, disorders/pathology during childhood, was upgraded to the class II evidence while two others, namely, psychobiological factor and behavioral factor, were upgraded to the class III evidence (Tables 3, 4).

Associations for Borderline Personality Disorder
A total of 22 of the 56 associations examined associations for borderline personality disorder. Twenty-two of those presented a nominally statistically significant effect (p ≤ 0.05), while nine of those associations reached p < 10 −6 . None of those associations were supported by class I or III evidence ( Table 3). Six associations were supported by class II evidence ( Table 3)   involving childhood emotional abuse, childhood emotional neglect, childhood any adversities, childhood physical abuse, childhood sexual abuse, and childhood physical neglect. Sixteen other associations were supported by class IV evidence ( Table 3) and 12 were non-significant (Table 4). After excluding the criterion of 1,000 cases, all factors with class II evidence remained at the same level while two others, namely, sexual abuse overall and sexual abuse by non-relatives, were upgraded to the class III evidence.

Associations for Dependent Personality Disorder
A total of 8 of the 56 associations examined associations for dependent personality disorder. Seven of those presented a nominally statistically significant effect (p ≤ 0.05) and met the class IV evidence criteria while only two associations reached p < 10 −6 . None of those associations were supported by class I, II, and III evidence. After excluding the criterion of 1,000 cases, two factors, namely, any childhood maltreatment and any childhood abuse in other clinical populations, were upgraded to the class II evidence, while another one, namely, any childhood abuse vs. no childhood abuse, was upgraded to the class III evidence (Tables 3, 4).

DISCUSSION
This is the first umbrella review pooling data from five metaanalyses on risk factors for personality disorders. Findings show that out of 56 associations between 26 potential environmental factors and antisocial, dependent, borderline personality disorder, despite 35 (62.5%) of the associations were nominally significant, only 6 (8.92%) associations met class II evidence for borderline personality disorder, involving childhood emotional abuse, childhood emotional neglect, childhood any adversities, childhood physical abuse, childhood sexual abuse, and childhood physical neglect. All other significant associations were classified as weak (class IV evidence). These results likely reflect the epidemiological distribution of borderline personality disorders, which represents the most common personality disorders both in clinical populations (36) and in the young general population, with a lifetime prevalence cumulating to about 10% in university students (40). The relatively high prevalence of this condition is likely to have facilitated etiopathological research in this field and, consequently, accumulation of established evidence, reviewed in the current study. At the same time, it is also the most reliable diagnosis within personality disorders, with superior diagnostic reliability (Kappa 0.54) compared with other personality disorders, and similar diagnostic reliability to that observed in bipolar I disorders (0.56) or schizophrenia (0.46) (41). Given the severe individual and societal burden and impact of borderline personality disorder and the limited effect of psychological interventions (42), knowledge into risk factors associated with this condition may advance clinical care. The findings that childhood emotional abuse, emotional/physical neglect, physical/sexual abuse, and adversities in general emerge as robust risk factor for this condition align with multiple salient clinical features such as affect instability, emotion regulation difficulties, and maladaptive coping strategies including substance misuse and frequent self-harm (43,44). Indeed, it has been suggested that several core experiences of borderline personality disorders may be understood as complex post-traumatic stress disorders (45). The current finding that individuals with a diagnosis of borderline personality disorders are consistently more likely to report childhood adversity than non-clinical controls is consistent with this strong clinical narrative linking childhood adversity and this condition, and show that evidence survives several stringent additional criteria making it into class II. Furthermore, the magnitude of these associations appeared as very large, with ORs superior to 5 in all cases: emotional abuse, 28.15; emotional neglect, OR 22.86; adversities, OR 14.32; physical abuse, OR 9.30; sexual abuse, OR 7.95; physical neglect, OR 5.73. Although it is important to note that these ORs relate to case-control studies not surviving prospective analyses, their large magnitude holds clinical relevance. For example, these findings indicate that childhood trauma should be systematically ascertained during the diagnostic assessment of suspected cases or during their initial clinical management. Notably, these ORs were not controlled against each other, albeit being likely correlated. The next generation of research should then develop a multivariable assessment interview to collect these multiple exposures in the same individuals assessed for a potential borderline personality disorder, thus allowing multivariable association analyses, yet accounting for multicollinearity of different risk factors. Another future development of research may involve exploring the transdiagnosticity of childhood trauma as potential risk factor for other mental disorders such as eating disorders, depressive disorders, anxiety disorders, and even psychotic disorders. These considerations are particularly relevant in the context of potential screening (46) and preventive interventions (47,48) and mental health promotion initiatives (49,50), because it could potentially be possible to target multiple outcomes/mental disorders by reducing childhood trauma.
Interestingly, childhood trauma exposure has been linked with neurobiological modifications in personality disorders, particularly borderline personality disorder. For example, variations in volumes of the main brain regions involved in BPD (especially amygdala and hippocampus) have been associated to adverse childhood experiences and trauma exposure (51)(52)(53)(54). Also, there is evidence of an association between childhood trauma and alterations of the cortisol circadian rhythm and levels, indicating a deregulation of the HPA axis responsiveness, which could also affect hippocampal volumes (51,55,56). Yet, such alterations are not exclusively present in BPD, as well as childhood traumatic events are risk factors for other mental disorders (57), including dependent personality disorder. Hence, the specific component justifying such a high association between BPD and childhood adversities remains unknown.
The present work also has several limitations. First, none of the findings met class I evidence. However, this is due to the available evidence. Indeed, findings inform the field that more cohort studies assessing multidimensional risk factors for personality disorders are needed. Similarly, the lack of evidence on protective factors is also due to lack of eligible meta-analyses reporting on protective factors. Second, compared with umbrella reviews assessing credibility of evidence on risk factors for other mental or physical disorders (14-17, 58, 59), we included a limited number of meta-analyses. Again, this indicates that more research efforts should focus on this clinically relevant field. Third, despite the large ORs, carefully designed longitudinal research, including examination of dose-response relationships, is required before definitive conclusions can be drawn regarding any causal role played by childhood adversity in the development of borderline personality disorders.
In conclusion, this umbrella review shows that risk factors for borderline personality disorder occur during childhood, and primary prevention strategies should encompass a multidisciplinary mental health promotion activity going beyond health professionals, to protect children from any maltreatment, neglect, or abuse. Risk factors for other personality disorders have been poorly identified so far, and more longitudinal studies should be conducted to inform prevention strategies.

DATA AVAILABILITY STATEMENT
The raw data supporting the conclusions of this article will be made available by the authors, upon request.

AUTHOR CONTRIBUTIONS
AFC, MS, PF-P, and ED designed the study. GC, PK, ED, and MS, extracted the data. ED run the analyses. MS and ED drafted the first version of the manuscript. All authors contributed to the manuscript, revised it, and approved the final version and protocol.