The role of neuroticism in self-harm and suicidal ideation: results from two UK population-based cohorts

Generation Scotland:Scottish Family Health Study (GS:SFHS) is a population- and family based epidemiological adult (age 18+) cohort recruited February 2006–March 2011, which has been described elsewhere [80, 81]. GS:SFHS had a higher proportion of females (59%) and was of older age (mean 49 males, 49 females) compared to the Scottish population (mean 37 males, 39 females, 2001 census) [80]. GS:SFHS participants were typically healthier and more affluent that the general Scottish population, nevertheless 32.9% of individuals lived in areas with worse than average socioeconomic deprivation [80]. 99% of the study group was of white ethnicity (Scottish population 98%). Sociodemographic information on age, sex, educational attainment and relationship status were collected by questionnaire on enrolment.

Neuroticism was measured using the Eysenck Personality Questionnaire-Revised Short Form (EPQ-SF) [30]. The neuroticism subsection of the EPQ-SF consists of 12 ‘Yes/No’ questions (e.g., ‘Are you a worrier?’). Scores range from 0 to 12, with higher scores indicating greater neuroticism. This scale has been concurrently validated with other quantitative measures of neuroticism [39] and has high-reported reliability (α-coefficients 0.85–0.88) [30].

Trained researchers elicited lifetime history of major depressive disorder (MDD) using the screening questions from the Structured Clinical Interview for DSM-IV Disorders [80] and, if either screening question was positive, going on to administer the mood sections of the SCID. The screening questions were: “Have you ever seen anyone for emotional or psychiatric problems?” and “Was there ever a time when you, or someone else, thought you should see someone because of the way you were feeling or acting?”. A diagnosis of MDD was made according to DSM-IV criteria and all interviews were conducted by a trained researcher (2011 cases identified, 12.7% of cohort). Individuals with a history of bipolar disorder were excluded.

Cognitive testing included the digit symbol substitution test from the Wechsler Adult Intelligence Scale III [91], logical memory from the Wechsler Memory Scale III [92] and verbal fluency [58]. From these tests, a measure of cognitive function (g) was derived as the first unrotated principal component, explaining 44% of the variance in scores [59]. Socioeconomic deprivation was determined using the Scottish Index of Multiple Deprivation 2009 (SIMD) [77]. This measure employs 6976 geographical area-based data-zones across Scotland which are then ranked in order of deprivation, ascertained through weighted scores in seven domains including employment, education, health, housing and crime, with data-zone 1 the most deprived and 6976 the least deprived.

All Scottish citizens registered with a general practitioner are assigned a unique identifier, the Community Health Index (CHI). This was used to deterministically record-link GS:SFHS participants to the Scottish Morbidity Records to obtain information about hospital admissions (SMR01) and psychiatric hospital admissions (SMR04) associated with self-harm. Written informed consent was obtained from 98% of GS:SFHS and only those who consented were linked. Self-harm cases were identified by matching to admissions codes with E950–E959 (ICD-9) or X60–X84, Z915, E98 and Y1–Y3 (ICD-10) [5]. Scottish NHS data on mortality was also linked, to exclude any GS:SFHS participants who died during follow-up.

In 2014, GS:SFHS participants were recontacted for a follow-up assessment of mental health [66]. Suicidal Ideation was elicited using two questions from the General Health Questionnaire-28 [38]. Participants were asked “During the past few weeks…Have you thought of the possibility you might make away with yourself?” and “Have you found the idea of taking your own life kept coming into your mind?”. Participants who answered ‘Definitely have’ or ‘Has crossed my mind’ to either question were defined as suicidal ideation cases [n = 3503, cases = 158 (4.7%)].

Stressful life events were ascertained using the List of Threatening Experiences (LTE), whereby respondents self-reported their experiences from a list of 12 common threatening life events, occurring in the preceding 6 months [10, 11]. Examples of LTE include “Serious injury or assault to yourself”, “Made redundant or sacked from job” and “marital difficulties or break off of a steady relationship” (for full list see Supplementary Materials). For each event endorsed, contextual threat was rated on a scale from 3 (“very bad”) to 1 (“not too bad”). The LTE has demonstrated high test–retest reliability and good agreement with informant information (Cohen’s κ 0.63–0.90) [11].

Coping styles were elicited using the Coping Inventory for Stressful Situations (CISS) [28]; Cosway, [21], a 48 item self-report questionnaire enabling responders to rate on a 5-point scale their engagement in coping styles in response to stress, including task-, avoidance- and emotion-oriented coping. The CISS shows robust validity and reliability (alpha reliability coefficients (Cronbach’s alpha) of 0.82–0.90 for the main factors) [21]. History of MDD was re-ascertained using the Composite International Diagnostic Interview-Short Form (CIDI-SF) self-report questionnaire [55], with 605 cases identified (18.1% of sample). Bipolar disorder cases were excluded. Unlike the main GS:SFHS cohort, only one member from each family was analysed (i.e., unrelated sample).

UK Biobank is a population-based cohort of adults aged 40–69 years recruited across the UK from 2006 to 2010, which has been described elsewhere [85]. During baseline assessment [82] participants provided sociodemographic information via a touch-screen questionnaire, including educational attainment and whether they lived as a singleton or couple. This study included a subset of 35,227 (7.0%) of UKB with complete case information for the variables of interest. Individuals in UKB who were also present in GS:SFHS (n = 201) were excluded.

Self-harm was ascertained through the touch-screen questionnaire. Participants were asked “Have you deliberately harmed yourself, whether or not you meant to end your life?”. A follow-up question enquired “Following any time when you took an overdose or deliberately tried to harm yourself did you (tick all that apply)”. Participants who ticked “see anyone from psychiatric or mental health services, including liaison services” and/or “need hospital treatment (e.g., A&E)” were included as cases in this study (772 cases, 2.2% of sample). The other answers, which were not included as cases, were “use a helpline”, “see own GP”, “receive help from friends/family” and “prefer not to answer”.

Neuroticism was assessed using the EPQ-SF [30], administered via the touch-screen questionnaire. Lifetime history of depression was ascertained by touch-screen questionnaire relating to lifetime experience of depressive symptoms and contact with mental health services [82].

Cognitive testing was administered via three touch-screen tests: (1) a symbol matching task over 12 trials (reaction time); (2) 13 logic/reasoning questions over 2 min (verbal-numerical reasoning); (3) card pair matching task (visuo-spatial memory). From these tests a single measure of cognitive ability (g) was extracted as the first unrotated principal component, explaining 42% of the variance.

Socioeconomic deprivation was measured via the Townsend Deprivation Index, a census-based measure incorporating unemployment, non-home ownership, household overcrowding and non-car ownership [50]. Each small postcode-based geographical area is assigned a Townsend Score, with zero indicating mean deprivation, negative scores indicating relative affluence, and positive scores indicating relative deprivation.

All analyses were carried out using R version 3.2.3. Complete case analysis was employed in both cohorts (see Supplementary Table S4 for analysis of complete case versus whole-cohort variables). Generalised linear models with logit-link function (logistic regression) were used to identify predictors of self-harm in UK Biobank. In the GS:SFHS self-harm study, additional adjustment for inter-relatedness of the family-based cohort was performed using a Bayesian mixed model approach, with pedigree fitted as a random effect, using an inverse pedigree matrix within the R package MCMCglmm. This implements a Markov Chain Monto Carlo estimator, with a “threshold” family probit link function which produces similar results to a logit function, optimised to pedigree based mixed effects models. In the GS:SFHS and UKB multivariable analyses of hospital-treated self-harm, predictor variables are reported unstandardized.

In the GS:SFHS suicidal ideation follow-up study, an unrelated sample was used and multivariable logistic regression was employed. In this analysis, continuous variables were scaled to have a mean of zero and standard deviation of one, to facilitate interpretation of the CISS and LTE predictor variables. During fitting of models, interaction terms for neuroticism and depression, and neuroticism and coping styles, were tested to investigate potential moderation on neuroticism.

Coefficients were expressed as odds ratios with 95% credible intervals and 95% confidence intervals as applicable. p values were reported after False Discovery Rate adjustment [7]. Group differences between numeric variables were ascertained using Cohen’s t test and Cohen’s d measure of effect size, and differences between proportions were assessed using z test and Cohen’s h. For all experiments, we have reported all measures, conditions, data exclusions and the determination of sample sizes and further information is available in Supplementary Table S4.

As presented in Table 1, there were 339 (2.1%) GS:SFHS individuals identified with previous self-harm requiring hospital admission. Self-harm cases were slightly younger (mean age 44.7 versus 47.1, p = 0.0005, Cohen’s d = 0.16), predominantly female (66.7% versus 58.4%, p 0.002, Cohen’s h = 0.17), with lower mean cognitive ability scores, greater prevalence of depression history (47.5% versus 12%, p < 2.2 × 10⁻¹⁶, h = 0.81) and with higher mean neuroticism (mean 6.4 versus 3.7, p < 2.2 × 10⁻¹⁶, d = 0.89). Self-harm cases were more likely to be from more deprived areas as measured by SIMD (mean 1964 versus 1823, p < 2.2 × 10⁻¹⁶, d = 0.58). The proportion of graduates was lower in self-harm cases (17.1% versus 33.9%, p < 10.0 × 10⁻¹¹, h = 0.39). A greater proportion of self-harm cases reporting being single (51.9% versus 31.7%, p < 3.6 × 10⁻¹⁵, h = 0.31).

The most predictive factor for previous self-harm (Table 2) was history of major depressive disorder [OR 5.6 95% credible interval (CI) 3.5–8.9, p_FDR 0.0004]. Neuroticism was positively associated with self-harm risk by an odds ratio of 1.2 (95% CI 1.1–1.2, p_FDR = 0.0003) per EPQ-SF unit. No significant interaction terms were found during model fitting. The significant effects of neuroticism were found in both male-only and female-only combined models (see Supplementary Table S1). Figure 1 displays the increased risk of self-harm per unit of EPQ-SF neuroticism score predicted by our model for both cohorts.

The age groups 25–34, 35–44, and 45–54 were positively associated with self-harm whereas age groups 64–74 and 75+ were negatively associated, compared to the reference category of 55–64. Gender did not show a significant association in the combined model. Having a higher SIMD score (less deprived) was associated with decreased risk of self-harm (per quintile unit OR 0.8 95% CI 0.7–0.9, p_FDR = 0.0004). Having no qualifications and being single increased risk. Cognitive ability showed an inverse association with self-harm (per unit OR 0.8; 95% CI 0.7–0.9, p_FDR = 0.0005).

There were 772 (2.2%) individuals self-reporting self-harm requiring hospital or psychiatrist review in UKB (Table 1). Self-harm cases were slightly younger (UKB’s minimum age is 40), predominantly female (70.5% versus 54.0%, p < 2.2 × 10⁻¹⁶, h = 0.34), and with higher mean neuroticism (mean 5.6 versus 3.3, p < 2.2 × 10⁻¹⁶,d = 0.83) and higher prevalence of history of depression (90.5% versus 33.3%, p < 2.2 × 10⁻¹⁸, h = 1.3). Cognitive ability scores were not significantly different for any of the tests. Self-harm cases were more likely to be from deprived areas (more positive scores) as measured by the Townsend index (mean − 0.5 versus − 1.7, p < 2.2 × 10⁻¹⁶, d = 0.44). Educational attainment was not significantly different between the two groups (χ = 7.43, p value 0.28). The proportion of the self-harm group living as single was 39.6% versus 22.5% for those reporting no history of self-harm.

The most predictive factor in the multivariable logistic model was self-reported history of depression (Table 2, OR 12.7 95% confidence interval (CI) 9.9–16.4, p_FDR 5.4 × 10⁻⁸⁶). The odds of self-harm were significantly positively associated with increasing neuroticism scores, OR 1.1 95% CI 1.1–1.2, p_FDR 9.8 × 10⁻¹⁷ per EPQ-SF unit. No significant interactions were found during model fitting. The significant effects of neuroticism were found in both the male-only and female-only models (Supplementary Table S1).

Being female was also associated with somewhat higher risk (OR 1.3 95% CI 1.1–1.5, p_FDR 0.005), as was being in the 35–44 and 45–54 age groups, whereas the 65–74 age group was protective. No educational factors were significant in the multivariable analysis. Being single and higher Townsend scores (more deprived) were associated with higher odds of self-harm. Cognitive ability (g) was not significantly associated (p_FDR = 0.051, OR 1.1, 95% CI 1.0–1.2).

In Table 2, comparison is also made to UKB participants who self-reported any self-harm, irrespective of whether hospital attention was sought (1578 cases, 4.4%). In this group self-harm was also positively associated with neuroticism scores (OR 1.1, 95% CI 1.1–1.2, p_FDR 3.4 × 10⁻⁴¹ per EPQ-SF unit). Positive association was also found for history of depression, being female, younger age group, increasing Townsend deprivation score, and being single. However, in this group increasing cognitive function score increased odds of self-harm (OR 1.1, 95% CI 1.0–1.1, p_FDR 0.004 per unit g).

In the GS:SFHS recontact study (N = 3342) there were 158 individuals with self-reported suicidal ideation(4.7%) (Table 3). Of these 21 (13.3%) had a record-linkage based history of self-harm compared to 1.9% in the control group. History of self-harm was the most predictive factor for suicidal ideation in the multivariable model (OR 3.5, 95% CI 1.9–6.2, p_FDR 7.5 × 10⁻⁵) followed by history of depression (OR 3.2, 2.3–4.7, p_FDR = 7.8 × 10⁻¹⁰). Scores in the List of Threatening Experiences increased odds of suicidal ideation (1.3, 1.2–1.5, p_FDR = 2.4 × 10⁻⁵ per standard deviation unit).

Neuroticism was positively associated with suicidal ideation in the multivariable model (OR 1.6, 1.3–1.8, p_FDR = 5.8 × 10⁻⁸ per standard deviation unit). However, this association attenuated to non-significant OR 1.1 (0.9–1.4, p = 0.44) when coping styles were added to the model (Table 3). In the full multivariable model including coping styles, EoC was positively associated with suicidal ideation (OR 2.4, 1.9–3.0, p_FDR = 1.7 × 10⁻¹²) and ToC was negatively associated (OR 0.8, 0.8–0.9, p_FDR = 0.03), while AoC was not significantly associated. The correlation matrix revealed that EoC and neuroticism were significantly correlated, r = 0.50 p < 2.2 × 10⁻¹⁶ and task-oriented coping were moderately negatively correlated [r = − 0.18 p < 2.2 × 10⁻¹⁶, Table S2 (Supplementary)]. In moderation analysis no significant (p ≤ 0.05) interaction terms were found for neuroticism*ToC, neuroticism*AoC or neuroticism*EoC on suicidal ideation, controlled for age, sex and depression status.

This study had a number of strengths for establishing the association of neuroticism to hospital-treated self-harm. We have employed two large, population-based cohorts which both have phenotypic information for major covariates of self-harm, allowing comparison between the groups while both using the same EPQ-SF measure of neuroticism. By utilising self-report in one cohort, and health data record-linkage in the other, our study design obviates some of the biases which can arise from utilising either method alone. GS:SFHS encompasses the range of adult age groups, and UKB focuses on middle-age to older adults, thus our findings are a significant contribution to self-harm research where many of the available studies are for teenagers or young adults. By extending our analysis to suicidal ideation, we were also able to demonstrate an association with neuroticism and correlated coping styles (emotion- and task-oriented coping), the latter of which are potentially modifiable by clinical intervention.

There are also some important limitations to our work. The cohorts we use are population-based, but are not fully representative, as UKB includes adults of ages 40–69 and GS:SFHS has an older mean age that the Scottish population. Additionally, the use of GP registration as an inclusion criteria for our GS:SFHS study (by enabling record-linkage via CHI number) leads to potential selection bias in our identification of self-harming individuals, although in the UK 96% of individuals are registered with a GP [80] indicating that such biases are likely to be small. The prevalence of self-harm we record should thus be used with caution and should not be taken as a reliable population estimate. Nevertheless, it is sobering that prevalence of hospital-treated self-harm was relatively high (2.1% for GS:SFHS and 2.2% for UKB). Since self-harm is more common in younger people, the true population prevalence is likely to be greater still. We have also adopted a cross-sectional design and thus causality between factors such as neuroticism and self-harm; and neuroticism, coping style and suicidal ideation; is suggested rather than conclusively demonstrated by our models.

The type of self-harm we have studied is self-harm involving hospital care. We used a general definition of self-harm as the data available to us did not allow distinction between nonsuicidal self-injury and suicide attempts, as this information is not available in the routinely collected administrative hospital data linked to in GS:SFHS (and was not part of the self-report question in UKB). This could limit the transferability of our results to other studies, although as discussed, the extent to which such distinctions of suicidal intent can be accurately made in practise is controversial.

In GS:SFHS we defined self-harm cases via admission to medical or psychiatric hospital, as ascertained by record-linkage. We therefore have not included a number of self-harm cases that were managed in the Emergency Department, where available data is incomplete [61]. This represents approximately 50% of self-harm cases presenting to hospital, although there are wide variations between hospitals [20]. A recent study has found that routine hospital data underestimates rates of self-harm by approximately 60% compared to combined survey-hospital database methods [15], as—for example—self-harm which is assessed in the Emergency Department, but which does not lead to hospital admission, may not be included. However, hospital admission self-harm is itself an important variable, as cases that are admitted are likely to be more serious and can therefore be expected to be of greater risk of further self-harm and completed suicide [37]. The UKB self-report variable was for self-harm requiring any hospital or psychiatric management (including Emergency Department) and therefore, while highly correlated with the GS:SFHS variable, was more general in its scope. The overall prevalence of self-harm in GS:SFHS and UKB was similar (2.1% and 2.2%, respectively). This might seem surprising as one might expect the more general self-harm definition in UKB to return a higher prevalence. This could be explained by the fact that the UKB cohort had no individuals younger than 40 and this has decreased the overall self-harm prevalence, since younger age groups are at relatively higher risk.

We employed a complete case design in our multivariable analyses in GS:SFHS and UKB. Potentially, this could have biased our results compared to the whole samples, although comparison (Supplementary Table S4) indicated that there were no significant and large-effect differences in major variables studied through the complete case approach. Nevertheless, this method could have introduced biases in ways we did not measure.

In summary, our findings must be seen in the context of self-harm with a high propensity to cause physical harm warranting medical attention. However, the UKB cohort did include a variable for any self-harm regardless of hospital attendance and we also included this multivariable analysis (Table 2). Neuroticism was found to be associated in this group also, with similar effect size (OR 1.1, 95 CI 1.1–1.2, p 3.4 × 10⁻⁴¹ per EPQ-SF unit).

With regard to our analysis of suicidal ideation and coping-style, neuroticism as a trait was measured during GS:SFHS enrolment, which was some years before the recontact when coping style and suicidal ideation were measured. However, as discussed, neuroticism is considered to be a relatively stable trait and would not be expected to change significantly over this time period. We also controlled neuroticism by age at enrolment rather than age at recontact within the models. Another important consideration is the extent to which neuroticism and emotion-oriented coping are separate constructs or both emanant from innate responses to stress. While we found the correlation of neuroticism and EoC to be significant (0.5), it was evidently not complete. There is also evidence that coping style is amenable to clinical treatment in prevention of suicide [36], whereas personality traits are understood as more therapeutically static.