Setting and recruitment

The present study was based on data for the period 2017–2020 retrieved from the quality register of the Norwegian Network for Personality Disorders (Network), a nationwide clinical research collaboration [29, 43]. It included 15 different outpatient treatment units within the Network, which offer specialized treatment for adult patients with a variety of PDs or clinically relevant, subthreshold personality difficulties. Patients were referred to specialized PD treatment from regular mental health outpatient clinics, where an initial assessment of patients referred from general practitioners was performed. Patients with comorbid psychosis, bipolar I disorder, autism, mental disability, and severe substance use disorders are not considered eligible for the PD-treatment programs, but in practice, a minor proportion was nonetheless referred. The treatment units comprise multidisciplinary teams with different healthcare professionals including psychiatrists, psychologists, psychiatric nurses, social therapists, and occupational therapists. All units within the Network follow the same assessment procedures, using standard evaluation instruments and diagnostic interviews. Treatment approaches include specialized BPD programs (Mentalization-based treatment, Dialectical Behavioral Therapy, Schema-focused therapy) as well as psychodynamic group therapy, metacognitive interpersonal therapy, art therapy, body awareness therapy and groups focusing on psychoeducation [61].

Participants

Two different samples were used for the four separate regression analyses: One common sample for the number of PDs, the total number of PD criteria, and the number of BPD criteria, and a separate sample for the LPFS model from the AMPD.

Sample I

The sample which was used to study the number of PDs and the total number of PD criteria included 798 patients who had completed the cost interview (a specific interview of health and welfare service use and occupational activity), and were assessed for both PDs and comorbid other mental health and substance use disorders. In the sample, 24.6% were male and 75.4% were female, and the mean age was 30.0 (SD = 8.9, range 18–63 years). Table 1 presents the distribution of PD diagnoses in the sample.

Table 1 The distribution of personality disordersSample II

The sample used to study the LPFS model included 794 patients who had completed the cost interview, were assessed for comorbid other mental health and substance use disorders, and had completed a questionnaire for assessing LPFS. Sample I and II have 93% overlap, as they share 745 participants. The 7% discrepancy is due to the fact that some patients in the Network’s quality register who were assessed for PDs had not completed the questionnaire for assessing LPFS, and vice versa. In this sample, 23.9% were male and 76.1% were female, and the mean age was 30.04 (SD = 8.8, range 18–63 years).

Other mental health and substance use disorders (covariates) in the samples

Nearly, all (Sample I: 94.7%, Sample II: 94.2%) of the assessed patients were given at least one other mental health or substance use diagnosis (Sample I: mean = 2.02, SD = 1.30, Sample II: mean = 1.98, SD = 1.28). Most individual diagnoses were aggregated into categories, and the five most frequent categories were used as covariates in the regression analyses; Mood disorders, anxiety disorders, substance use disorders, eating disorders, and PTSD. The omitted diagnoses (somatoform disorder, dissociative disorder, ADHD, psychosis disorders, and autism spectrum disorder) had too few incidents (< 8%) to warrant inclusion as covariates in the analyses. Table 2 describes the number of patients in the different other mental health and substance use categories in both samples.

Table 2 Categories of other mental health and substance use disordersDiagnostic assessment: the number of PDs, the total number of PD criteria, and the number of BPD criteria

The measures of three of the severity indicators, the number of PDs, the total number of PD criteria, and the number of BPD criteria, as well as the covariates of other mental health and substance use disorders, were based on standardized, semi-structured diagnostic interviews: The Structured Clinical Interview for DSM-5 Personality Disorders for PD (SCID-5-PD) [16], and the Mini International Neuropsychiatric Interview (M.I.N.I.) [52, 53] for other mental health and substance use disorders. Table 3 presents the number of PDs per patient in the sample, whereas Table 4 presents the distribution of SCID-5-PD criteria in the sample. Table 5 presents the number of BPD criteria per patient in the sample.

Table 3 The number of personality disorders for each patientTable 4 The distribution of SCID-5-PD criteriaTable 5 The number of BPD criteria per patient

Diagnostic inter-rater reliability, using the SCID-5-PD and M.I.N.I., was not directly investigated in this study. However, several measures were undertaken to address possible reliability issues. Within the Network, diagnostic assessments were performed in each unit by clinical staff who had received systematic training in diagnostic interviews and principles of the LEAD-procedure (Longitudinal, Expert, All-Data), [42, 56]. This means that diagnoses were based on all available information including referral letters, self-reported history and complaints, and overall clinical impression, in addition to the diagnostic interviews. All diagnoses were set or evaluated by a specialist in psychiatry or clinical psychology. In the study period, local training courses/workshops focusing on understanding and assessment of PDs, associated comorbidity, and use of structured interviews were conducted by an experienced psychiatrist (last author) at all units to ensure clinical competence and calibrate diagnostic evaluation. A total of 29 local workshops were held within the study period in addition to shorter clinical discussions on request [60]. Furthermore, in a former study within the Network, using the Structured Clinical Interview for DSM-IV (SCID-II – the previous version of SCID-5-PD), reliability was investigated and acceptable diagnostic reliability was indicated [17].

As first defined in the DSM-III (APA, 1980), PD-not otherwise specified (PD-NOS) is indicated when the general criteria for PD are fulfilled, but criteria are below the threshold of any specific PD. The diagnosis is either set directly by the clinicians or set by the researchers according to a given set of criteria. The operationalization of PD-NOS lacks precision, and former studies have suggested cut-offs ranging from 5 to 11 fulfilled PD criteria across categories [13, 41, 42, 66, 72]. In the current study, based on SCID-5-PD and DSM-5 terminology, we chose to categorize patients with eight or more fulfilled PD criteria and no specific PDs as unspecified PD (if not already given the diagnosis by the clinicians) [61].

Assessment of the LPFS model

LPFS reflects core dimensions of personality pathology, involving impairments in self and interpersonal functioning, with self-functioning represented by the domains of identity and self-direction, and interpersonal functioning represented by empathy and intimacy [79]. All four domains are described along a continuum that ranges from healthy functioning (level = 0), to extreme impairment (level = 4) [37, 38]. Each of these components are further subdivided into three indicators (subdomains), i.e., 12 indicators altogether, and the LPFS thus identifies five levels of functioning for each of these 12 indicators, offering a severity index for personality pathology [24, 69].

Since its publication, several instruments for assessing the LPFS have been developed, and in the present study we used the second version of the LPFS-Brief Form (LPFS-BF 2.0) questionnaire [69]. The LPFS-BF is a 12-item self-report, each item representing one of the 12 indicators of the LPFS, yielding a global estimate of impairment related to personality functioning [25, 69]. Initially developed as a quick screening tool related to the LPFS, this instrument is now included in the standard set of patient-reported outcomes for PD [47]. The first version was subsequently revised, by rewording some items and introducing a 4-point Likert scale from 0 (completely untrue) to 3 (completely true). The severity index will thus have a possible range of 0–36 points [69]. The LPFS-BF 2.0 showed acceptable construct validity and psychometric properties [5, 48, 69]. Initially, there were some concerns that the LPFS-BF yielded a two-factor solution [69, 78]. However, results of recent bifactor studies have indicated that the LPFS, as assessed by the LPFS-BF, can be considered as essentially unidimensional [48, 79]. Table 6 presents the distribution of LPFS-BF 2.0 scores in the sample.

Table 6 The distribution of LPFS-BF 2.0 total scoresCost measures

Societal costs are the sum of direct and indirect costs. Direct costs cover all actual costs of healthcare utilization, e.g., general practitioner visits, psychotherapeutic treatment, medication, inpatient treatment (both somatic and mental health). Indirect costs cover the lost productivity due to suffering from PD. Intangible costs, i.e., the psychological pain experienced by people with PDs, are not included in societal costs in this study, as such costs are very difficult to measure [28]. Hence, the societal costs in this study are the sum of direct healthcare costs and productivity loss. Calculations of healthcare costs and productivity loss for the total period of 6 months prior to evaluation were estimated using a bottom-up approach [28], i.e., taking the individual patients’ reported health service use and degree of absenteeism from the labor market, and multiplying it with the estimated unit cost of each specific cost element [60].

Clinicians performed the cost interview as a part of the pretreatment assessment, collecting data for the six-month period prior to assessment. Questions on health service use included: (1) general practitioner (GP) visits, (2) emergency health services (psychiatric emergency helpline, emergency room, psychiatric outpatient emergency service, and ambulant emergency service), (3) hospitalization (admission to medical hospital, admission to psychiatric hospital, admission to addiction clinics, and day-patient care), (4) outpatient treatment at mental health centers (individual- or group therapy), and (5) pharmacological treatment. The participants were also asked to which degree they were employed the last six months (range 0–6) [60].

All unit costs were measured in €, yearend 2018. Unit cost for GP was estimated based on a public report that estimated the total cost of all GPs [27] in 2017, adjusted by the official consumer price index (CPI) [57], divided by the total number of consultations by GPs during 2018 [58]. Unit cost for psychiatric emergency helpline was calculated based on the annual report 2018 from “Mental Helse”, a typical helpline provider in Norway [35]. The total cost of the service was divided by the total number of telephone calls (answered), chat-service and mail service. Emergency room unit cost was set to the price for same day consultation with specialist medical doctor at a private healthcare center in Oslo [40]. Psychiatric emergency outpatient service was set at the same unit cost as standard outpatient consultation, while emergency consultation at the patients home out of an outpatient clinic were given an ambulatory fee add-on [20]. Unit costs related to treatment at outpatient mental health centers, medical and psychiatric hospitals, addiction clinics, and day-patient care were obtained from reports published by the Norwegian government [21, 22]. Calculations of medication unit costs were based on information from the Norwegian Medicines Agency, and cost per daily dose of typical drugs per medication class were used to calculate cost per month [34].

The human capital approach was used to calculate the productivity loss, as most cost-of-illness studies have used the human capital approach to estimate productivity loss [46, 64]. This method measures lost productivity as the patients’ absence from work due to illness, valued at the market wage. As the patients did not report their individual gross income, and the marked wage of patients with PDs is not available in public registers, the patients’ unit cost had to be estimated. As many patients with PDs struggle to stay in the workforce and achieve higher levels of education, the average monthly wage for the total population probably is an overestimation of the wage-level of patients with PDs (only 12% of both samples report they have been in ordinary employment during the whole 6 months, while 73% of both samples report no connection to the labor marked during the same period). The unit cost of lost productivity was, thus, set to be equal to the average monthly sickness benefit [39], which is 58% of the average monthly wage in Norway [59].

All unit costs, mean health service costs, mean productivity loss, and mean societal costs in the period six months prior to assessment are reported in detail by Sveen and colleagues in their cost-of-illness study of treatment-seeking patients with PDs, using data from the quality register for the same period as the present study [61].

Ethics

All participating patients from each treatment unit gave their written consent to use anonymous clinical data for research purposes. Anonymized data were collected and transferred to the quality register. The collection procedures were approved by a local data protection officer at each contributing unit. Data security procedures for the quality register were approved by the data protection officer at the research center of the Network at Oslo University Hospital. Because the data were anonymous, formal approval from the Norwegian State Data Inspectorate and Regional Committee for Medical Research and Ethics was not required [61].

Statistical analysis

Four separate multiple regression main effect analyses, one for each severity indicator, were performed in order to investigate how well each of them could predict societal costs, while controlling for the effects of the five categories of comorbid other mental health and substance use disorders. All four regression models thus included six independent variables. Due to the exploratory nature of the current investigation, no general adjustments for multiple comparisons was strictly required, and an alpha level of 0.05 was used to determine statistical significance for all analyses [10]. Tables 7, 8, 9 and 10 present exact p values, and power analyses were conducted post hoc. The correlation matrix between all the independent variables as well as the Tolerance and Variance Inflation Factor (VIF) coefficients gave no indication of a multicollinearity problem in any of the models.

Table 7 Regression model 1: Number of PDs as severity indicatorTable 8 Regression model 2: Total number of PD criteria as severity indicatorTable 9 Regression model 3: Number of BPD criteria as severity indicatorTable 10 Regression model 4: LPFS-BF 2.0 total scores as severity indicator

Societal costs data in the present study were non-normally distributed. Most patients had similar health service costs, but a small proportion of patients had very high costs due to inpatient admissions. As many as 73% of the patients had been out of the workforce during all 6 months, incurring a large productivity loss, while only 12% had no productivity loss. The residuals were non-normally distributed as well. They were not improved by preliminary trials of log transformations. However, when the sample size is sufficiently large, as in the present sample, the Central Limit Theorem ensures that the distributions of parameter estimates will approximate normality when the errors are independent and identically distributed with finite variance, regardless of the shape of the population distribution [44, 45, 50]. Supplementary subgroup analyses of health service costs and productivity loss were performed to further explore the impact of the severity measures on costs. As the residuals of the health service costs model approximated the normal distribution when it was log transformed, but no improvement of the residuals of the productivity loss model was made by such a transformation, a log transformed model was used in the analysis of health service costs, because a transformation which could improve the normality of the residuals significantly is recommended, also when the sample size is large [44]. Statistical analyses were performed using SPSS version 29, except for the power analyses, which were performed using the R package «pwr» (version 1.3-0).