Trial design and participants

We undertook a parallel group, single blind, treatment as usual (TAU)-controlled, randomised trial of participants with borderline personality disorder. We specifically wanted to recruit participants who experienced crises and engaged in self-harming behaviour. Follow-up was considered at 6-months post-randomisation. We recruited a sample of adults accessing community mental health teams (CMHTs) in south east London, UK. Inclusion criteria were:

1. (a) aged 18 years or older;

2. (b) meeting diagnostic criteria for borderline personality disorder (according to DSM-IV-TR criteria ⁶ and measured using the Structured Clinical Interview for DSM-IV (SCID-II) - Borderline Personality Disorder subsection); ^{Reference First, Spitzer, Gibbon and Williams7}

3. (c) had self-harmed in the previous 12 months (defined as at least one act with a non-fatal outcome in which the individual had initiated a behaviour (such as self-cutting), or ingested a toxic substance or object, with the intention of causing harm to themselves); ^{Reference Madge, Hewitt, Hawton, De Wilde, Corcoran and Fekete8}

4. (d) under the ongoing care of a CMHT;

5. (e) able to provide written informed consent.

Exclusion criteria were:

1. (a) currently an in-patient;

2. (b) primary diagnosis of a psychotic illness;

3. (c) unable to read or write in English;

4. (d) unable to provide written informed consent.

All potential participants were in the first instance identified and approached by their care coordinator, who informed them about the trial. After they had expressed an interest in participating, a member of the research team met with the participant to explain the trial further and obtain written informed consent.

Ethics and governance approvals

Data collection protocols were approved by the South London Research Ethics Committee (ref: 09/H0803/113) and the trial was registered with the International Standard Randomised Controlled Trial registry (ISRCTN12440268) prior to the commencement of data collection. ^{Reference Moran, Borschmann, Flach, Barrett, Byford and Hogg9} All participants provided written informed consent prior to entering the trial, including allowing members of the research team to access their electronic records. Progress of the trial, adherence to protocol and participant safety were overseen by a trial steering committee (chaired by Professor Mike Crawford, Imperial College London).

Randomisation and masking

After consent and baseline assessment, randomisation was conducted at the level of the individual and was stratified by alcohol use (as measured by scores on the Alcohol Use Disorders Identification Test (AUDIT); ^{Reference Saunders, Aasland, Babor, de la Fuente and Grant10} low <8; medium 8-15; high >15) and depression (as measured by scores on the Hospital Anxiety and Depression Scale (HADS) ^{Reference Zigmund and Snaith11} depression subscale; low <8, medium 8-10, high >10), both of which have been shown to be predictive of future self-harm. ^{Reference Kapur, Cooper, King-Hele, Webb, Lawlor and Rodway12,Reference Colman, Newman, Schopflocher, Bland and Dyck13} Randomisation was managed electronically by the Clinical Trials Unit at the King's College London Institute of Psychiatry, UK. The nature of the intervention meant that neither participants nor staff members could be masked to allocation; however, all follow-up data were collected by a research worker who was masked to treatment allocation and all data analyses were conducted by a statistician who was also masked to treatment allocation. The extent to which masking was achieved in the collection of outcome data was assessed at the end of the trial.

Intervention and control arms

Outcome measures

The primary outcome was the proportion of participants reporting self-harm at 6 months post-randomisation. Self-harm data were obtained from an established self-report questionnaire. ^{Reference Hawton, Rodham, Evans and Weatherall14} Items included ‘How many times in the past year [or ‘past six months’ at follow-up] have you deliberately tried to harm yourself?' Secondary clinical outcomes, all measured at baseline and follow-up, and their corresponding instruments were as follows.

1. (a) Depression and anxiety: HADS. ^{Reference Zigmund and Snaith11} This is a 14-item self-report scale for detecting states of depression and anxiety in out-patients, with higher scores indicating higher levels of depression/anxiety.

2. (b) Working alliance: Working Alliance Inventory (WAI). ^{Reference Horvath and Greenberg15} The WAI is a 12-item self-report instrument for measuring the perceived quality of working alliance between client and practitioner, with higher scores indicating a more positive perception of alliance.

3. (c) Satisfaction with services: Client Satisfaction Questionnaire (CSQ). ^{Reference Larsen, Attkisson, Hargreaves and Nguyen16} The CSQ is an eight-item measure of participants' level of satisfaction with treatment received, with higher scores indicating a higher level of satisfaction with services.

4. (d) Engagement with services: Service Engagement Scale (SES). ^{Reference Tait, Birchwood and Trower17} The SES is a 14-item self-report scale, completed by the participant's treating clinician - in our trial typically a care coordinator or key worker - to measure the participant's level of engagement with community mental health services. Higher scores reflect a greater level of difficulty engaging with services.

5. (e) Well-being: Warwick-Edinburgh Mental Well-Being Scale (WEMWBS). ^{Reference Tennant, Hiller, Fishwick, Platt, Joseph and Weich18} The WEMWBS is a measure of subjective mental well-being over the preceding 2 weeks and focuses entirely on positive aspects of mental health. Higher scores indicate a higher level of well-being.

6. (f) Social functioning: Work and Social Adjustment Scale (WSAS). ^{Reference Mundt, Marks, Shear and Greist19} The WSAS is a five-item self-report instrument to assess impaired functioning, with higher scores indicating a higher level of impairment.

7. (g) Perceived coercion: Treatment Experience Scale (TES). The TES was adapted from the Admission Experience Survey, ^{Reference Gardner, Hoge, Bennett, Roth, Lidz and Monahan20} a 16-item instrument designed to assess the perceived level of coercion experienced by service users during hospital admission. Respondents endorse each item as either ‘true’, ‘false’ or ‘don't know’.

8. (h) Health-related quality of life: EuroQoL 5-dimensions (EQ-5D). ^{Reference Kind and Spiker21} The EQ-5D assesses respondents' subjective health-related quality of life across five life domains: mobility, self-care, usual activities, pain/discomfort and anxiety/depression. A higher score indicates a better health-related quality of life.

9. (i) Resource-use: Adult Service Use Schedule (AD-SUS) adapted for use in this trial based on previous research involving people with personality disorders. ^{Reference Byford, Knapp, Greenshields, Ukoumunne, Jones and Thompson22} The AD-SUS, completed by participants in interview at baseline and 6-month follow-up, collected data on use of all hospital and community health and social services. To enhance accuracy, in-patient psychiatric admission data were additionally collected from electronic clinical records of the local National Health Service (NHS) trust (South London and Maudsley). This data replaced self-reported contact data for this NHS trust, although self-reported contacts with other trusts were retained. The economic evaluation took a health and social care perspective, in line with National Institute for Health and Clinical Excellence guidelines. ²³ Unit costs for the financial year 2009-2010 were applied, and these are detailed in online supplement DS2.

Sample size and power calculation

Sample size calculations are not required for most pilot studies, because the aim is to gather information about recruitment processes, consent and attrition rates and trial procedures. Nevertheless, we wanted to know whether it was feasible to recruit and retain a pre-determined number of people with borderline personality disorder into a trial of JCPs and, for this reason, we undertook a power calculation in order to give us a target sample size to aim for. The trial was powered to detect a threefold difference in the proportions of participants who had self-harmed during the follow-up period (36% in the TAU group v. 12% in the JCP + TAU group). The predicted TAU proportion of self-harm (36%) was the same proportion of self-harm as that observed in a previous RCT of cognitive therapy to reduce repetition of self-harm. ^{Reference Tyrer, Thompson, Schmidt, Jones, Knapp and Davidson24} On the basis of these figures, an overall sample of 114 (randomised 1:1 to JCP + TAU:TAU) would provide 80% power to detect an observed difference between JCP + TAU and TAU alone, based on a two-sided test at the 5% significance level. The target sample size was increased to 120 in order to allow for attrition and loss of data on self-harm. This sample would also be large enough to provide 80% power to detect a constant hazard ratio between the groups of 0.29 with proportions of episodes in the two groups as stated above, based on the log-rank statistic assuming no accrual rate, a fixed time of follow-up and an estimated 10% rate of drop out.

Statistical analyses

All analyses were based on the intention-to-treat sample using a statistical analysis plan finalised by the trial statistician (J.M.H.) and approved by the principal investigator (P.M.) in advance of conducting any analyses. All analyses were performed with Stata version 11.0 for Windows.

Analysis of outcome variables

We summarised continuous variables as mean (s.d.) and categorical variables as n (%). We assessed the primary outcome with a logistic regression model with treatment and stratification factors; alcohol misuse (AUDIT) and depression (HADS) as covariates. Model assumptions were checked by the use of diagnostic plots. Models were undertaken with the assumption that data were missing at random. Categorical data were compared using Fisher's Exact test. We analysed secondary outcomes in a generalised linear model (GLM) framework; covariates in the model were treatment group, baseline value of outcome, alcohol misuse and depression. For the frequency of self-harm at 6-month follow-up, a negative binomial distribution was specified with a log link. Logistic regression was utilised for binary outcomes and clinical scales were analysed using the assumption of a normal distribution. Results of the treatment effects were summarised as odds ratios (ORs, logistic and ordinal logistic regression), incidence rate ratios (RRs, negative binomial distribution GLM) and effect sizes (Gaussian models) at 6-month follow-up with two-sided 95% confidence intervals.