2.1. Study Design
We analysed data from the UCL Bereavement Study [8
]. This was a UK-wide cross-sectional survey of young adults aged 18–40 working and/or studying at UK higher education institutions (HEIs) who had experienced the sudden bereavement of a close friend or relative. This study had focused on young adults due to concerns about their risk of suicide [18
] and the difficulties of engaging young suicidal men with services [19
]. Full details of sampling for this closed online survey have been described elsewhere, including the survey instrument (see Supplementary Materials
]. Sampling via institution-wide email lists (to all staff and students) avoided the biases associated with recruiting a help-seeking sample, and was felt to be the most efficient, comprehensive and pragmatic means of recruiting a hard-to-reach population of young adults [20
]. Of 5085 respondents to the survey, we included those who consented to participate, completed a stigma score, and specified their mode of bereavement (n
The study was approved by the UCL Research Ethics Committee in 2010 (ref: 1975/002). All participants provided online informed consent.
Our exposure measure was high perceived stigma of the bereavement, defined using the 10-item stigma subscale of the Grief Experience Questionnaire (GEQ) [21
]. The GEQ is a standardised, self-administered instrument for the assessment of the phenomenology of grief. It was originally developed in the U.S. using qualitative data from individuals bereaved by natural causes, accidental death, and suicide [22
], and subsequently validated [21
]. The stigma sub-scale includes items describing perceptions of others’ avoidance and lack of concern (see Box 1
), capturing perceived rather than personal stigma. Responses to items in each subscale are rated using a 5-point Likert-style frequency scale, generating subscale scores of 5 to 25 (at 0.5 intervals). The majority of studies measuring GEQ scores use GEQ subscales rather than overall GEQ scores, allowing them to delineate specific components of grief [8
]. Based on precedent [23
] and the normal distribution of stigma scores in this sample, we used the mean to dichotomise stigma scores, classifying them as low (5 to 12) or high (12.5 to 25) to aid clinical interpretation.
Box 1. GEQ stigma subscale items.
Since the death how often did you….
feel like a social outcast?
feel like no-one cared to listen to you?
feel that neighbours and friends did not offer enough concern?
feel avoided by friends?
think people were gossiping about you or the person?
think that others didn’t want you to talk about the death?
feel somehow stigmatised by the death?
feel like people were probably wondering about what kind of personal problems you and the person had experienced?
think that people were uncomfortable offering their condolences to you?
feel like the death somehow reflected negatively on you or your family?
Our primary outcomes were self-reported suicidal ideation (“Have you ever thought of taking your life, even though you would not actually do it?”) [26
] and self-reported suicide attempt (“Have you ever made an attempt to take your life, by taking an overdose of tablets or in some other way?”) [27
] post-bereavement. These standardised, validated measures were derived from the Adult Psychiatric Morbidity Survey (APMS) [28
], a regular national population survey in England, qualified by whether each was before or after the sudden bereavement, or both, to derive an incident measure.
Our three secondary mental health outcomes were post-bereavement non-suicidal self-harm (self-poisoning and self-injury without suicidal intent) using the standardised, validated APMS measure [27
] (adapted as above); post-bereavement suicidal and non-suicidal self-harm (aggregating the suicide attempt and non-suicidal self-harm measures, to correspond to that used in a major longitudinal study of self-harm in England [29
]); and post-bereavement depression, using the Composite International Diagnostic Interview (CIDI) screen for lifetime depression [30
], also validated for use in an online questionnaire [31
] (adapted for incident cases as above).
Our three self-reported support measures were level of current social support (using a standardised ordinal measure from the APMS [28
]); receipt of any formal bereavement support (using a binary measure developed for this study); and receipt of any informal bereavement support (using a binary measure developed for this study). Classification of formal and informal bereavement support was derived from similar British [32
] and international [33
] studies of service use . Self-help was excluded due to problematic categorisation in relation to formal versus informal bereavement support [34
]. Thus, formal support was defined as that received from healthcare or social services staff; psychological therapists or counsellors; voluntary sector helplines or counsellors; police officers; funeral directors; coroners’ officers; teaching staff; school or HEI counselling services; line managers, or employer counselling services. Informal support was defined as that received from friends; family; spiritual/religious advisors, or complementary and alternative medicine practitioners.
We selected nine confounding variables on the basis of existing literature and clinical judgement: age; gender; socio-economic status (using the UK Office for National Statistics Standard Occupational Classification [35
]); mode of sudden bereavement; kinship to the deceased; family history of suicide (excluding an index bereavement by suicide); pre-loss depression; pre-loss suicidal and non-suicidal self-harm; and years since sudden bereavement. Mode of bereavement was classified via self-report as bereavement by suicide, bereavement by sudden natural causes (e.g., cardiac arrest), and bereavement by sudden unnatural causes (e.g., accidental death). In the case of exposures to more than one mode of sudden bereavement, all those bereaved by suicide were classified as such, regardless of other exposures. Those bereaved by non-suicide death were asked to relate their responses to whichever person they had felt closest to, with exposure status classified accordingly.
Missing data for model covariates and outcomes were less than 7%.
2.3. Statistical Analysis
We investigated simple associations between the outcome variables and exposure using χ2 tests or one-way analysis of variance, as appropriate.
We investigated the relationship between outcomes and high stigma scores using multilevel regression models with HEI as random effect, to take into account the clustering effect at the HEI level. We used ordinal logistic regression to investigate the relationship between social support and high levels of perceived stigma scores. All multivariable models included the nine pre-specified confounding variables described above. Models were fitted using complete case analysis. We used the Bonferroni correction to set a significance threshold of p = 0.006 for multiple testing.
To test whether the effect of high stigma scores on primary outcomes varied by gender and by mode of bereavement, we added interaction terms to these models, using a less stringent p-value threshold (p = 0.1) to reflect the limited statistical power of interaction tests.
To test an additional research question about whether high perceived stigma helps differentiate those who attempt suicide after bereavement from those with suicidal ideation after bereavement, we ran our multivariable model for suicide attempt in the sub-sample of those who reported suicidal thoughts or attempts post-bereavement (n = 1510).
We ran a series of a priori defined sensitivity analyses to assess the robustness of our main findings when taking into account biases introduced by <7% missing data and by our sampling strategy. In the first and second analyses, we used best-case and worst-case scenarios to impute missing values by recoding all missing values on outcomes/covariates as positive (e.g., no suicidal ideation/attempt) or as negative (e.g., suicidal ideation/attempt) respectively [36
]. In the third and fourth, we used more stringent inclusion criteria: dropping the 10 HEIs that modified the stipulated recruitment method, and the 18 HEIs with participant numbers below the median cluster size. Finally, we conducted linear regression to test whether there was a linear association between stigma scores and outcomes.
All analyses were conducted using Stata version 12 (Stata Corp. 2011. Stata Statistical Software: Release 12. College Station, TX, USA).