Service coverage

Service coverage was defined as the proportion of people with a disorder contacting a mental health service (from service utilisation data) among those estimated to have the disorder (population prevalence) during a 12-month period. This specific measurement has been termed ‘contact coverage’ by Tanahashi et al. and reflects the actual contact between the service provider and the user (Tanahashi, Reference Tanahashi1978). Unless otherwise specified, we use the more familiar term ‘service coverage’ to represent 12-month contact coverage throughout the remainder of this paper.

$${\rm Total\; service\; coverage} = \displaystyle{{{\rm total\; treated\; cases}} \over {{\rm expected\; cases}}}$$

Treated cases

We drew upon 12-month service utilisation data from the Mental Health Atlas 2017; specifically, data from specialist (inpatient and outpatient) mental health facilities (not primary healthcare) for psychosis, bipolar disorder and depression. Data, including the number of people receiving care from inpatient and outpatient services, population representativeness and population size, were extracted from section 8 of the Mental Health Atlas Questionnaire 2017 into Microsoft Excel 2016 for analysis (online Supplementary Fig. S1). Data were excluded if they did not report on service utilisation of both inpatient and outpatient facilities for a specific disorder, or if countries did not report representativeness of the population that the data were drawn from.

Data from section 7 of the Mental Health Atlas Questionnaire 2017, relating to the number of inpatient and outpatient visits, were used as a validity check of the service utilisation data (online Supplementary Fig. S2). If the number of visits (from section 7) per individual case (from section 8) was less than 1, data were excluded from further analysis.

Total treated cases were calculated according to the following formulas:

$${\rm Total\;treated\;cases} = {\rm inpatient\;cases} + {\rm outpatient\;cases}$$

$${\rm Total\;treated\;cases\;per}\;100, \;000\;{\rm of\;population} = {\rm \;}100, \;000{\rm \;} \times \displaystyle{{{\rm \;\;total\;treated\;cases}\;} \over {{\rm total\;population\;\;}}}$$

Expected cases

Expected prevalent cases of individual disorders were estimated using the disorder-specific prevalence rate estimates of the Global Burden of Disease Study (GBD) 2016 (Vos et al., Reference Vos, Abajobir, Abate, Abbafati, Abbas, Abd-Allah, Abdulkader, Abdulle, Abebo and Abera2017) and total population sizes. GBD 2016 prevalence estimates were derived by first conducting systematic reviews of the literature to identify studies reporting the prevalence, incidence, remission and/or excess mortality for individual disorders. These estimates were entered in DisMod-MR 2.1 for analysis, a Bayesian meta-regression tool (Flaxman et al., Reference Flaxman, Vos and Murray2015). If no raw epidemiological data were available for a particular location, data from surrounding locations were used to estimate prevalence. Within DisMod-MR 2.1, regions and super-regions were defined according to GBD 2016's classification of broad geographic regions or continents. Additional information on DisMod-MR 2.1 can be accessed elsewhere (Vos et al., Reference Vos, Abajobir, Abate, Abbafati, Abbas, Abd-Allah, Abdulkader, Abdulle, Abebo and Abera2017).

$${\rm Expected\;cases} = {\rm prevalence\;rate\;} \times \;{\rm total\;population}$$

Disorder-specific prevalence rates (point prevalence for depression and schizophrenia, and 12-month prevalence for bipolar disorder) as estimated by GBD 2016 were adjusted in two ways to ensure consistency with service utilisation data for each country that provided Mental Health Atlas 2017 data. Firstly, point prevalence rates of depression were adjusted to 12-month prevalence rates using the inverse of a covariate coefficient which adjusted studies reporting 12-month prevalence to point prevalence. Given the understanding that mild cases of depression are rarely treated in specialist services, we further adjusted depression prevalence estimates to reflect the proportion of cases which are moderate-severe by applying the GBD severity splits [moderate proportion = 0.17, 95% confidence interval (CI) 0.13–0.22; severe proportion = 0.1, 95% CI 0.03–0.2] (Vos et al., Reference Vos, Abajobir, Abate, Abbafati, Abbas, Abd-Allah, Abdulkader, Abdulle, Abebo and Abera2017).

Secondly, schizophrenia prevalence rates were adjusted to non-affective psychoses prevalence rates based on a 0.49 ratio derived from the literature (Kendler et al., Reference Kendler, Gallagher, Abelson and Kessler1996; Widerlöv et al., Reference Widerlöv, Lindström and Van Kitarring1997; Cho et al., Reference Cho, Kim, Jeon, Suh, Chung, Hong, Bae, Lee, Park, Cho, Lee and Hahm2007; Perala et al., Reference Perala, Suvisaari, Saarni, Kuoppasalmi, Isometsa, Pirkola, Partonen, Tuulio-Henriksson, Hintikka, Kieseppa, Harkanen, Koskinen and Lonnqvist2007; Perala et al., Reference Perala, Saarni, Ostamo, Pirkola, Haukka, Harkanen, Koskinen, Lonnqvist and Suvisaari2008; Phillips et al., Reference Phillips, Zhang, Shi, Song, Ding, Pang, Li, Zhang and Wang2009; Chang et al., Reference Chang, Hayes, Perera, Broadbent, Fernandes, Lee, Hotopf and Stewart2011; Binbay et al., Reference Binbay, Alptekin, Elbi, Zagli, Drukker, Aksu Tanik, Ozkinay, Onay and Van Os2012; Jorgensen et al., Reference Jorgensen, Allebeck and Dalman2014; Moreno-Kustner et al., Reference Moreno-Kustner, Mayoral, Navas-Campana, Garcia-Herrera, Angona, Martin and Rivas2016; Chang et al., Reference Chang, Wong, Chen, Lam, Chan, Ng, Hung, Cheung, Sham, Chiu, Lam, Lee, Chiang, Chan, Lau, Lee, Leung, Leung, Lau, Van Os, Lewis and Bebbington2017). Non-affective psychoses point prevalence rates were not adjusted to 12-month prevalence rates as no statistically significant difference was observed using a GBD modelling covariate.

If a country did not report its population size or its population size was dramatically different from United Nations (UN) population estimates, UN population estimates were used (Global Burden of Disease Collaborative Network, 2018).

Case definitions

GBD prevalence estimates and Mental Health Atlas service utilisation data adhere to ICD-10 case-definitions for major depression, schizophrenia and non-affective psychosis, and bipolar disorder. GBD prevalence estimates adhere to the DSM-IV and ICD-10 case definitions for major depression (ICD-10: F32.0–9, F33.0–9), schizophrenia (ICD-10: F20; adjusted to non-affective psychosis as per below) and bipolar disorder (ICD-10: F31.0–F31.6, F31.8–F31.9, F34.0–F34.1).

The Atlas questionnaire requests that service coverage data are reported for non-affective psychosis (ICD-10: F20–29), bipolar disorder (ICD-10: F30–31) and depression (ICD-10: F32.1-3–F33.1-3). For this reason, we have avoided using the diagnostic term ‘major depression’ and instead used ‘depression’ with the understanding that some depression cases may not reach diagnostic thresholds. We also refer to the term ‘psychosis’ for the remainder of the manuscript as representing ‘non-affective psychosis’.

Service utilisation data adjustments

To prevent double-counting of individuals who had used both inpatient and outpatient facilities, total service coverage rates were adjusted to appropriately reflect the total number of unique cases of individuals utilising mental health facilities. Countries were categorised into inpatient-prioritised countries (i.e. most of the population are treated in inpatient facilities) or outpatient-prioritised countries (i.e. most of the population are treated in outpatient facilities) based on reported data. In inpatient-prioritised countries, the data were not adjusted. In outpatient-prioritised countries, it was assumed that all individuals utilising inpatient facilities also utilised outpatient facilities.

The total unique cases adjustment was based on question 9.3 of the Mental Health Atlas 2017 questionnaire, which provides a follow-up rate ranging from 1 to 4, as follows:

1. 1. 25% or less of discharged inpatients received a follow-up outpatient visit within one month

2. 2. 26–50% of discharged inpatients received a follow-up outpatient visit within one month

3. 3. 51–75% of discharged inpatients received a follow-up outpatient visit within one month

4. 4. More than 75% of discharged inpatients received a follow-up outpatient visit within one month

The adjusted outpatient estimate was calculated by averaging the follow-up range, multiplying this by the inpatient cases, and subtracting this product from the reported outpatient utilisation value (see formula below). For countries that did not report a follow-up rate, the median follow-up rate for the respective GBD world region was used (Institute for Health Metrics and Evaluation, 2019a). If the adjusted outpatient utilisation value was negative, the original reported outpatient utilisation value was used.

Adjusted service coverage was calculated as follows:

$${\rm Total\;unique\;treated\;cases} = ( {\rm outpatient\;cases}-( {{\rm follow\;}\hbox{-}{\rm up\;rate} \times {\rm inpatient\;cases}) } ) + {\rm inpatient\;cases}$$

$${\rm Adjusted\;total\;treated\;cases\;per\;}100, \;000 = {\rm \;}\displaystyle{{{\rm total\;unique\;cases\;}/{\rm \;total\;population}} \over {100, \;000}}$$

$${\rm Adjusted\;total\;service\;coverage} = {\rm \;}\displaystyle{{{\rm total\;unique\;treated\;cases\;}} \over {{\rm expected\;cases}}}$$

Service coverage thresholds

Although Question 8 of the Mental Health Atlas requests data on number of cases, health information systems frequently only record number of patient visits. With this knowledge, we designed validity check to determine whether countries reported cases or visits by applying lower and upper service coverage thresholds derived from the published literature (Table 1). For psychosis, the lower threshold was based on WMHS data showing 12-month service use split by severity of mental disorder. For the lower threshold, we took the WMHS country with the lowest service coverage estimate for severe disorders and applied the lower confidence interval of this estimate (Wang et al., Reference Wang, Aguilar-Gaxiola, Alonso, Angermeyer, Borges, Bromet, Bruffaerts, De Girolamo, De Graaf and Gureje2007). This was done for each income group. For depression, the upper threshold was chosen using the same process – by applying the upper confidence interval of the highest service coverage estimate for severe disorders, by income group. For bipolar disorder, the upper limit was also based on WMHS data, which specifically examined 12-month treatment rates for bipolar spectrum disorders (encompassing type I, II and subthreshold bipolar disorder) split by income group (Merikangas et al., Reference Merikangas, Jin, He, Kessler, Lee, Sampson, Viana, Andrade, Hu, Karam, Ladea, Medina-Mora, Ono, Posada-Villa, Sagar, Wells and Zarkov2011). Service coverage estimates were excluded if they fell outside of these imposed thresholds.

Table 1.

Applied lower and upper service coverage estimate thresholds

Note: missing values mean no thresholds were set.

^a Taken from ‘severe’ category of Table 3: 12-month service use by severity of mental disorders in the WMH surveys (29)

^b The country listed has the lowest coverage in the WMH surveys for that income group. This is used as the lower threshold for psychosis coverage estimates.

^c The country listed has the highest coverage in the WMH surveys for that income group. This is used as the highest threshold for depression coverage estimates.

^d Taken from 12-month treatment from ‘any mental health service’ of Table 6: Lifetime and 12-month treatment of DSM-IV/CIDI Bipolar Spectrum Disorders (30).

^e For bipolar disorder, there are only global values from the WMH surveys, which are used as the highest threshold for bipolar coverage estimates.

No upper threshold was set for psychosis with the acknowledgement that there is potential for service coverage in specialist mental health services to be high. Conversely, for bipolar disorder and depression, no lower limits were set in order to capture the low rates of service coverage expected for these disorders.

Calculation of uncertainty and meta-analysis

We calculated the standard error (s.e.) around each countries' service coverage estimates using the following steps:

• – Estimate the s.e. of the treated prevalence $\left({{\rm S}{\rm .E} . = \sqrt {{{p( {1-p} ) } \over n}} } \right)\;$

• – Estimate the s.e. of the GBD prevalence $\left({{\rm S}{\rm .E} . = {{{\rm upper\;CI}-{\rm lower\;CI}} \over {3.92}}} \right)$

• – Estimate the proportion of cases that are treated $\left({{\rm S}{\rm .E} . = {{P_{{\rm treated}}} \over {P_{{\rm disorder}}}}} \right)$

• – Estimate the s.e. of the service coverage estimate $\left({{\rm S}{\rm .E} .\left({{{P_{{\rm treated}}} \over {P_{{\rm disorder}}}}} \right) = \sqrt {{{P^2_{{\rm treated}} } \over {P^2_{{\rm disorder}} }} \times \left({{{{\rm S}{\rm .E} .^2_{{\rm treated}} } \over {P^2_{{\rm treated}} }} + {{{\rm S}{\rm .E} .^2_{{\rm disorder}} } \over {P^2_{{\rm disorder}} }}} \right)} } \right)\;$

Country-level data were aggregated by World Bank income group (The World Bank, 2018), WHO region (World Health Organization, 2019) and GBD super region (Institute for Health Metrics and Evaluation, 2019b). Random-effects meta-analyses were performed using the metafor package in R (Viechtbauer, Reference Viechtbauer2010). Data were logit transformed prior to analysis using the delta method. Baujat plots were inspected for outliers. Exclusion of influential estimates did not change the final results of the meta-analysis, so they were retained in the final model.