Setting
Norway is a sparsely populated country of five million inhabitants. Four Regional Health Authorities commission 21 health trusts to deliver specialist mental health care from hospitals and CMHCs. In 2016, 71 CMHC areas – the unit used in our analyses – ranged in size from 9,000–125,000 inhabitants aged 18 years and above. Their combined catchment areas covered the entire national population. The NMHA permits involuntary care for observation and treatment of both inpatients and outpatients [
2]. A small number of patients are also committed to involuntary care by the criminal courts. Compared internationally, the Norwegian population rates of involuntary care are toward the higher end, with 151 involuntary hospitalized patients per 100,000 adult population in 2015 [
4] and a point prevalence of 47.4 patients under a CTO per 100,000 adult population in 2012 [
7]. In 2015, the review board rejected 66 of 7824 instances of initiated involuntary care after mandatory document control. In addition, 21 patients successfully appealed against an involuntary observation, and 156 against involuntary care [
32].
Study design and data
The study is a retrospective, longitudinal register study, with data from 2014–2018, where the involuntary care ratio in the area a patient lives is studied as a predictor of subsequent outcomes. We registered the analysis plan prospectively (Clinicaltrials.gov identifier NCT04655287).
We obtained patient data on patients aged 18–65 from the Norwegian Patient Registry (NPR), to which all Health Trusts are required to report all specialist inpatient or outpatient health service use. This means that those diagnosed with an SMD in one year but with no specialist service contact in the following year would not be counted for that second year. This registry provides reliable, patient-identifiable, national data with a good degree of completeness [
33], including valid diagnoses for severe mental disorders [
34]. From 2015 onward, data completeness for inpatient status (voluntary/involuntary) is considered adequate [
35]. The
population at risk for SMD and involuntary care was defined as all adults ages 18–65 in Norwegian municipalities and city districts (hereafter called local authorities), and data were acquired at Statistics Norway’s online table generator [
36,
37]. We acquired data on suicides for men and women ages 18–65 years in CMHC catchment areas from the Norwegian Cause of Death Registry (NCDR). These could only be released aggregated to local authority level and could not be linked to individuals from the NPR data. Data availability precluded analyses of the effect on health and safety of others.
Variables
We based our operationalization of
severe mental disorder (SMD) on the NMHA [
39], which restricts the major diagnostic prerequisite for involuntary care to psychoses or psychotic symptoms. The law discourages involuntary care for mental health disorders without such symptoms [
39]. Therefore, we defined an SMD as a diagnosis on the schizophrenia spectrum (F20-29) or bipolar disorder with psychosis (F30-31) as either the primary or secondary diagnosis, as classified following the ICD-10 system [
40]. For patients with more than one recorded diagnosis during a year, we used the following hierarchical order to select each patient’s diagnosis: bipolar, schizophrenia spectrum, substance abuse (F10-19), personality disorder (F60-69), and depressive (F32) and other disorders.
For the purposes of the present study, involuntary care was defined as involuntary inpatient or outpatient care or observation sanctioned by the NMHA, including those in forensic care and those sentenced to care by a criminal court. A patient with such a care episode was classified as under involuntary care for that year.
Mental health inpatient days for a patient was the term we used for the combined number of days in hospital for that patient during the calendar year in question, regardless of care formality and including inpatient days in mental health institutions outside the patient’s CMHC area.
Area of residence represented the local authority in which the patient’s registered address belonged, aggregated to the appropriate CMHC area. For those who moved during the period, we used the address at the beginning of the last episode of mental health care during the year in question. For analyses that studied effects over time, the area of residence in the index year was used as the area of residency throughout.
Urbanicity was classified as one of five levels of urbanicity as described in Table
1, which were assigned to each local authority area. This was based on Statistics Norway’s classification [
41] as modified in a previous study [
42]. An urbanicity value was assigned to each patient based on his or her area of residence.
Table 1
Classification of urbanicity
1 | Norway’s capital and four largest regional centers, as defined by Statistics Norway |
2 | Remaining local authorities with 20,000 + inhabitants, where 80% or more live in densely populated areasa |
3 | County centers in one of the 19 counties not classified as 1 or 2, plus urban local authority areas that are a continuation of a densely populated area with urbanicity level 1 or 2 (typically suburbs) |
4 | Remaining local authorities with 5,000 + inhabitants, of which 60–79% live in densely populated areas |
5 | Remaining local authority areas |
Area deprivation. We used Statistics Norway’s continuous living condition index for 2008 as a measure of deprivation in the local community. The index is based on averaged deciles of unemployment, welfare benefits, educational level, mortality, etc., resulting in a number between 1 and 10 [
43] and was updated every eight years until 2008.
CMHC areas consist of one or more local authorities. Several CMHCs and local authorities merged during the study period (2014–2018). In order to have a fixed area structure in the analyses, we analyzed all data using the area structure from 2016. In two cities, CMHC areas deviated from local authority borders, and in these two cases, we combined CMHC areas, thereby reducing the number of areas from 71 to 69.
The standardized involuntary care area ratio (SIAR) was used as the primary covariate in our analyses. First, raw involuntary care rates were calculated by dividing the number of people experiencing involuntary care during a given year by the population at risk during that year. We prepared the standardization by estimating a linear regression model with the raw rate of persons experiencing involuntary care in the local authority as the outcome and with age (six groups), sex (two groups), urbanicity (five classes), and deprivation (range) as covariates. We did not include SMD rates into the standardization to reduce the risk of bias from differing diagnostic thresholds. We reasoned that a lower threshold for setting SDM diagnoses would mean more patients with less severe symptoms – and therefore unlikely to be treated involuntarily – would be diagnosed. If used in standardization, areas with lower threshold for setting SMD diagnosis could therefore have their rate of involuntary care adjusted downwards, whereas for areas with a higher diagnostic threshold this would adjust their level upwards. If service capacity or paternalism is associated with the threshold for involuntary care and SMD diagnosis setting, including SMD rates in the standardization could bias the analyses in the direction of null-results.
The variables that were significant in the regression model, which were age, sex, and urbanicity, were then used for indirect standardization. For each local authority, we calculated the expected number of patients under involuntary care for each age and sex stratum in the area’s urbanicity and aggregated this to the CMHC area population. We then divided the observed number of persons under involuntary care by the expected number of persons under involuntary care. The resulting SIAR values (range 0.58–1.46) have 1 as the reference, where SIAR values below 1 indicate fewer persons than expected under involuntary care in the CMHC area. Each patient was assigned a SIAR value based on his or her area of residency in 2015.
Statistical analyses
To answer our research questions about possible negative effects of low levels of involuntary care, we predefined five models [
44] (see Table
2). The first four were estimated using longitudinal data at the patient (Models 1–3) or CMHC area level (Model 4). In Model, 5 which used area suicides as outcome, we aggregated suicide numbers for five years in order to increase power. It is therefore a cross-sectional analysis.
Table 2
Overview of statistical models for adverse effects of low standardized ratios of involuntary carea in CMHCb areas
Model 1 (Case fatality) | Individual patients with SMDc in 2015 | Increase in number of deaths among SMDc patients, 2015–2018 | Cox regression, with patient’s survival time as outcome, adjusted for age and sex |
Model 2 (Inpatient days) | Individual patients with SMDc and no involuntary care in 2015 | Increase (or lower decrease) in the number of mental health inpatient days from 2015– 2016 and/or 2017 | Linear mixed model with random effects for CMHCb and change in inpatient days from 2015–2016 and 2015–2017 as outcome |
Model 3 (Involuntary care) | Individual patients with SMDc and no involuntary care in 2015 | Increase in number of patients transitioning into involuntary care in 2016 and/or 2017 | Cox regression, with death as a competing risk and random effects for CMHCsb, with time to an incident of involuntary care during the next two years as outcome |
Longitudinal data at the CMHC level |
Model 4 (SMDc patients) | CMHCb areas in 2015 | Increase in number of persons diagnosed with SMDc in 2016 and/or 2017 | Linear mixed model with time, standardized CMHCb ratios of involuntary care in 2015 and interaction between the two as covariates; outcome was yearly number of patients with SMDb in the area |
Cross-sectional data at the CMHC level |
Model 5 (Suicides) | CMHCb areas in 2014–2017d | Higher suicide ratios in 2014–2018 | Correlation between mean of the standardized yearly ratios of involuntary care in 2014–2017 with the similarly standardized 5-year ratio of suicides in the CMHCb areas in 2014–2018d |
Data were not available in advance to prespecify any cut-off value of high/low SIAR or to decide between non-linear and linear associations between SIAR and the outcomes. Therefore, we examined each outcome for non-linear associations and for adequacy of a linear model with suitable tests and estimated the latter when appropriate.
Model 1 – Cox proportional hazards (PHs) regression analysis. We assessed the effect of SIAR on case fatality over time, adjusted for age and sex, by following all SMD patients over four years, which allowed for observing delayed deaths. Schoenfeld’s residuals were used to assess the PHs assumption, while potential non-linear associations were tested by martingale residuals. Cluster effects on CMHC, Health Trust, and regional health authority levels were assessed by intra-class correlation (ICC), which showed no effects or negligible effects, and hence, no adjustment was needed. Models 2 and 3 assessed whether SMD patients in voluntary care in 2015 deteriorated over the subsequent two years in areas with low SIAR. As we here wanted test deterioration (into more inpatient days in Model 2 and into involuntary care in Model 3), we included only those in voluntary care in the baseline year. It is also among voluntary patients with SMD that we could expect to find those whose needs for protection against deterioration may not be met where there is a high threshold for using involuntary care. Model 2 was a linear mixed model, which assessed the trend in inpatient days as a function of SIAR. The model contained random intercepts for CMHCs and fixed effects for time dummy, SIAR, and interactions between time and SIAR. A model with SIAR as a non-linear covariate was considered; however, no non-linear associations were detected. The residuals were inspected graphically to assess the assumptions of normality and homoscedasticity. Post-hoc analyses were performed to explore the interaction further. Model 3 – a Cox PH model – assessed time to involuntary care as a function of SIAR. Death was included as a competing risk. The model assumptions were assessed in the same way as for Model 1.
Model 4 – a linear mixed model – assessed whether low SIAR in a CMHC area was followed by an increased number of people treated for SMDs in the area in the following two years, suggesting differential rates of improvement or deterioration. As this is an area characteristic it was analyzed at area level, and different baseline levels of SMDs were controlled for by investigating changes in SMD rates. The model included fixed effects for time as a second-order polynomial to account for non-linear effects, SIAR, and interaction between the two. The model without the interaction term was chosen based on the Bayesian information criterion (BIC). The model assumptions were assessed in the same way as for Model 2.
Model 5 was based on cross-sectional data for SIAR and suicides; scatter plot and correlation analysis was employed to assess whether a low SIAR is associated with more suicides. For the smallest CMHC areas, the expected number of suicides per year (based on the national count of ca. 600 suicides per year [
45]) was around 1 per year. Therefore, we aggregated suicide numbers for 2014–2018 in order to increase power. This required a cross-sectional design, and we therefore observed and averaged involuntary care for the years 2014–2017, to observe suicides that might happen some time after discharge from involuntary care. We considered an area level measure of suicide to be a relevant outcome for the SMD population, even if it might introduce some noise in that it includes suicides that are not SMD related.
For all analyses, we set the threshold for significance at 0.05. We prepared data for analyses using R 3.6.1 [
46] with data.table 1.12.8 [
47], tidyverse 1.3.0 [
48], and lubridate 1.7.4 [
49], and we used Stata 16 for standardization and regression models. Figures were drawn with Stata, ggplot2 [
50] and MATLAB R2020b.
Changes to the analysis plan
For technical reasons, we needed to make minor alterations to the registered analysis plan for Model 5. Due to several municipality mergers, the NCDR could not map data on suicides from 2019 to the local authority structure of 2018 and earlier; therefore, we moved the periodization from the planned period of 2015–2019 to 2014–2018. In addition, because the NCDR could not allocate suicides to city districts, we merged several CMHCs in cities for the model with suicides as the outcome. Due to the registry’s privacy requirements, a minimum cell size was required in order to release area data for suicides per urbanicity stratum, and therefore, we combined level 4 (the smallest stratum) and level 5. In 2014, four health trusts had < 85% data completeness on care formality. Thus, we ran Model 5 with and without data from nine CMHC areas in these four health trusts.