Factors associated with health service utilisation for common mental disorders: a systematic review

We searched four databases; MEDLINE, PsycINFO, EMBASE and Scopus. We combined two key concepts (CMD and health service utilisation) using keywords and subject headings in the respective databases. Results were retrieved on 5th May 2016. The search strategy can be found in the Additional file 2. We supplemented the database search by hand searching and reference searches. We only included articles published in English.

Since the population of interest is the general adult population, we included only population-based studies, defined as community-based epidemiological studies that are representative of the adult population. We excluded studies that focused only on specific sub-populations such as veterans, students, or prisoners, whose experiences may not be representative of the wider population and warrant separate reviews.

The primary outcome measure of interest was any contact with formal health services – including private, public, generalist and specialist – for mental health reasons (also referred to as “treatment-seeking”) by adults aged 18 and above with CMD. Reflecting the definition of the treatment gap, we focussed specifically on use of services as a binary outcome – i.e. any versus no use – rather than volume of treatment received or quality of care.

To be eligible for inclusion, the study must have tested the association between treatment seeking and any other factors. We therefore included only quantitative studies, published in peer-reviewed journals, and excluded narrative reviews and commentaries. We included only studies in which the analyses were restricted to those individuals who either met diagnostic criteria or screened positive for CMD using a standardised instrument.

For the purposes of this article, CMD is defined as those ICD-10 (International Classification of Diseases - 10th Revision) disorders measured by the Clinical Interview Schedule - Revised [33] – often considered the gold standard for measuring CMD [34, 35] – namely, depressive disorders, generalised anxiety disorder, panic disorder, phobias, obsessive compulsive disorder, and mixed anxiety-depression disorder.

We excluded papers that measured only intentions to seek help, or perceived barriers to care, since multiple studies have found that these are not closely correlated with behaviour [36‐40].

No restrictions were placed on geographic area or date of publication. Table 1 provides full details of the inclusion criteria applied.

The first author completed title and abstract screening for all references retrieved. Subsequently two researchers (GME and DK) independently screened a random sample of 10% of the references, and inter-rater reliability was calculated at 94%. Full texts were retrieved for all studies included after the title/abstract screening. The first author screened all full texts, while the second author (GME) screened a purposive sample of 10%. At both stages, disagreements were resolved through discussion.

We assessed the quality of the relevant evidence extracted from the included studies using the Mixed-Method Appraisal Tool (MMAT) [41], which has been shown to be quick and reliable to apply [42]. Table 2 sets out the criteria used. Extracted evidence for the purposes of this review was rated as poor, fair, good or excellent if 0–1, 2, 3 or 4 of these criteria were met, respectively. These ratings are not intended to reflect the study quality in relation to its own primary aims, but only of the quality of the evidence that related to this review.

Due to the number of different factors investigated, and heterogeneity in the measures used, it was not feasible to attempt a meta-analysis of the effect of each factor. Instead, the “best fit” framework synthesis method [43] was used to compare the fit of the data with an existing model of factors affecting health service utilisation. This technique was originally developed for the synthesis of qualitative research, but has since been applied to reviews of quantitative and mixed methods studies [44, 45].

The first author extracted the data from each of the included papers and coded these deductively using the Andersen framework described above [17]. Any data that did not fit any of the headings in the Andersen model headings were to be coded separately under a new theme in a subsequent inductive phase.

To avoid bias in the synthesis and interpretation of results due to pre-conceived ideas about which factors are associated with treatment-seeking, we created a priori definitions with which to categorise the associations found for each factor. These definitions (summarised in Table 3) were created for the purposes of the current review, and are intended to be conservative.

No prior studies were found to guide the operationalisation of these definitions, and therefore the cut-off points chosen are necessarily arbitrary. However, we have tried to be entirely transparent in how these have been applied, and present the full findings and quality ratings in the appendices provided so the reader can examine how the evidence relates to the conclusions drawn.

Figure 1 summarises the search process. After removing duplicates, 10,331 papers were retrieved. Fifty-two papers were found to meet the criteria at the full text screening stage. Of these, eleven were cohort studies while forty-one were cross-sectional.

Thirty-two (62%) of these studies reported data from North America, nine (16%) from Europe, two (4%) from Australasia, three (6%) from Africa, one (2%) from Asia, two (4%) from Latin America and three (6%) used international data from across world regions.

The study sizes varied considerably, from 56 participants to 18,972 participants with elevated levels of CMD symptoms.

In terms of quality, evidence from one study was rated as poor, evidence from 16 was classified as fair, evidence from 20 was classified as good, and evidence from 15 studies was rated excellent. Additional file 3 presents the characteristics of the included studies.

Table 4 shows the number of studies that investigated each of the factors in the Andersen model.

Compared to other factors, we identified the highest number of studies on the association between socio-demographic factors (classified according to the Andersen model as “predisposing” factors) and treatment-seeking for CMD. We also found a large number of studies that investigated symptom severity, symptom profile and comorbidity (termed “need” factors in the Andersen model) as correlates of treatment-seeking. Fewer of the included studies examined enabling factors such as insurance, household wealth and social support. There was a lack of published evidence on some factors implicated by the Andersen model, such as psychological factors (e.g. beliefs and attitudes, classified as “predisposing” factors) and health systems factors (e.g. the availability and accessibility of services).

Almost all of the factors identified were individual rather than contextual level factors. No factors were identified that could not be accommodated by the model.

A summary of findings for each factor group is presented below. For more detailed results see the Additional file 4.

There was a clear discrepancy in the quantity of research identified between high-income and low-and-middle-income countries, with just six of the included studies originating from LMIC and one international study that included data from both HIC and LMIC [93]. Five of the LMIC-only studies were from middle-income countries; two from South Africa [64, 75], and one from Brazil [57], Mexico [69] and China [67]. The only study from a low-income country was from Ethiopia [62].

Evidence from three out of six LMIC studies was rated as good or excellent. On average LMIC studies were smaller than HIC studies, with a mean of 1742 participants with high CMD symptoms, compared to 3374 for HICs.

The LMIC studies identified predominantly reported on the effect of predisposing factors, such as age, gender, and education levels, and on measures of income or wealth.

There was insufficient published evidence from LMIC to compare the factors associated with treatment-seeking for CMD between HIC and LMIC.

The majority of studies used secondary datasets, which limited the choice of variables to those that are typically collected as part of multi-purpose epidemiological surveys. The frequent use of cross-sectional data also limits our ability to disentangle the direction of causation when associations are found. The majority of studies used multivariate logistic regression models for analysis. The use of hierarchical models, or structural equation modelling that explicitly recognises the potential interactions between some of these factors, may have led to differing conclusions. Although several studies cited the Andersen model to justify their choice of variables, there seems to be little agreement as to how the model should be operationalised and much heterogeneity in the measures used, making it difficult to compare the results across studies. In particular, agreement is needed on how variables indicating level of “need for care” should be measured in the context of CMD, so that is it possible to control for this consistently when investigating whether the use of health services is equitable. Finally, many of the included studies did not correct for multiple testing when investigating multiple associations simultaneously, and as such their findings should be viewed as hypothesis-generating rather than hypothesis-testing.

This review furthers our understanding of the treatment gap for CMD by summarising patterns of treatment-seeking. Need factors were most consistently found to be associated with treatment-seeking for CMD symptoms. Enabling factors were not found to be consistently associated with treatment seeking for CMD. The evidence on predisposing factors was inconsistent, although there was weak evidence for an association with demographic factors, specifically age, gender, ethnicity, education level and marital status. Finally, the current results suggest that urban or rural residence is not associated with treatment-seeking.

With regard to the second objective, there was insufficient published evidence from LMIC to draw any firm conclusions about whether the factors associated with health service utilisation for CMD differ from high-income countries.

This review has several strengths: It employed a broad search strategy, informed by previous reviews [22‐24], since the literature on this topic spans several disciplines with varying terminology. It followed an a priori protocol, had screening verified at multiple stages by a second researcher, and employed a widely recognised theoretical framework to analyse the results. Compared to the most recent review in this area [25], we searched a larger number of databases in order to make the review as comprehensive as possible.

This review adds to previous research by considering the wider category of CMD rather than a single diagnostic category, which several researchers have argued is a more appropriate grouping for community and primary care settings [26, 28‐31]. It was also deliberately more liberal in terms of its definition of CMD symptoms, since it is generally accepted that CMD symptoms are better conceptualised as a spectrum rather than a dichotomy between those who meet diagnostic criteria and those who do not [27]. Since conducting full diagnostic interviews in large population studies is often not feasible, it was hoped that this broader definition would lead to the inclusion of studies from a wider range of settings.

Other related reviews have been restricted to young adults [98] or to one country only [99]. While the results reported here are broadly consistent with the findings of these reviews, this study extends previous research by (a) comparing results across settings; (b) including only population-based studies to ensure the generalisability of findings; (c) examining a set of symptoms that typically present together in community settings, making the results a stronger basis for informing interventions at the population level; and (d) separating service utilisation by adults from that of children or adolescents, since in many countries services are delivered separately for these two groups, and decisions regarding treatment-seeking may follow different paths for minors (defined here as those aged under 18).

However, the current review nonetheless has several limitations that must be acknowledged. One is that it was not possible to assess the power of each study to detect an association, meaning that in studies where no association was found with a given factor, this cannot be interpreted with confidence to indicate a lack of association rather than a lack of statistical power. Secondly, it is possible that some studies in which this was not the primary research question may have been erroneously excluded if associations with treatment-seeking were not reported in the title or abstract of the paper. This is more likely to be the case when no associations are found, leading to potential selection bias. For reasons of feasibility, the search was restricted to studies published in English.

We were not able to present data on the amount of variance explained by the factors included in the studies reviewed, since this was not reported in the majority of these studies. Nor was it possible to discuss the confounding factors controlled for in every analysis, due to the large number of studies included. To definitively assess the causal effect of any one factor on treatment-seeking for CMD a meta-analysis of that specific association would be recommended; this was not the purpose of the current review, which set out to summarise associations, not to make causal claims.

The inclusion of multiple measures of CMD symptoms also means that these will not be exactly comparable across studies. Furthermore, when the quality of studies was assessed, we considered the measure of CMD used and the measure of treatment-seeking from the formal health sector; however, due to the number of factors investigated it was not possible to assess the appropriateness of measures used for each of these factors. Finally, as mentioned in the methods section, although the consistency of evidence for each factor was graded according to pre-defined criteria, other ways of operationalising levels of evidence are possible, which could lead to more or less conservative conclusions. Full details of all studies and the criteria applied are presented in the appendices.

Our findings are consistent with previous research pointing to need factors as the strongest determinants of health service utilisation for mental disorders [100‐103]. This is also consistent with the finding from the World Mental Health Surveys (WMHS) – which included both LMICs and HICs and measured substance use disorders and bipolar disorder as well as CMD – that low perceived need was the most common reason cited for not seeking treatment [104].

The same associations with female gender, middle age, higher levels of education, and being unmarried were found in the WMHS [8].

The fact that the evidence included in the current study did not support an association with economic factors was surprising, given the evidence that socio-economic factors affect the type of provider contacted [63, 68, 77], the quality of care received [63], adherence [105‐109] and response to treatment [110, 111]. However, a recent analysis of WMHS data by Evans-Lacko et al. (2017) found that differences in treatment rates in the WMHS by socio-economic status were predominantly accounted for by education rather than income [112].

Thus our findings on treatment-seeking for CMD are largely in keeping with the largest international study of mental disorders and service utilisation to date. The WMHS did not investigate rural/urban residence, or any of the other factors included in the Andersen model besides those listed above.

Need factors, reflecting the extent to which CMD symptoms interfere with people’s lives and whether outside help is needed, appear to be central to explaining treatment-seeking behaviour. This suggests that many of those who do not seek care from formal health services for their CMD symptoms fail to do so not because of limited supply, but because of lack of demand for services.

Whether meeting criteria for a disorder is a good indication of a “need for health services” is an ongoing debate in the context of mental health care [113]. The limited demand for interventions for CMD, compared to the number of people who meet criteria for CMD, can be conceptualised as a lack of education or awareness about mental health issues, indicating a need for information, education and communication campaigns. On the other hand, it may be an indication that current diagnostic categories are overly broad, and include a large number of people who do not require formal medical care. Patel (2014) has argued that current prevalence estimates should not be regarded as the number of individuals in need of care, since a large proportion of these individuals do not require formal interventions through the health system [31]. Measures of functioning or quality of life may represent better indicators of “need for care” than meeting diagnostic criteria (it is notable that the latter concept was not investigated by any of the studies included here).

Patel’s argument that increasing the supply of mental health services will not alone make a substantial impact on the treatment gap for mental disorders is supported by the current findings that; (a) lack of perceived need is a major determinant of failure to seek help from health services, and (b) that enabling factors do not appear to be a major determinant of treatment-seeking (discussed below). Many individuals with less disabling symptoms are likely to view informal support – such as social interventions in the community, or advice on self-care, listed at the bottom of the World Health Organization (WHO) Service Organization Pyramid [114] – as more appropriate for their needs. As such, encouraging these individuals to seek care through the health system may not be the best use of resources.

The lack of evidence for an association between enabling factors and health service utilisation, even in settings with weak public health systems, such as South Africa, and without universal health coverage, like the USA, was surprising. Of course, absence of evidence is not proof of a lack of association, especially given that the studies included here were not explicitly powered to detect this relationship. There is also the potential for information bias, given the sensitivity of financial topics, since most studies used self-reported data.

However, the hypothesis that economic factors do not play a major role in determining whether people with CMD initially seek care from health services is backed up by findings from Evans-Lacko et al. (2017) [112], as well as Andrews et al. (2001), who found no association at the ecological level with health spending or out-of-pocket costs [112, 115]. Furthermore, Andrade et al. (2014) found that attitudinal barriers (most commonly, wanting to handle the problem alone) were reported much more often than structural barriers (which are linked to enabling factors), with the exception of severe cases. It is possible that the inclusion in this review of individuals with milder conditions, for whom low perceived need primarily inhibits treatment-seeking, might be obscuring the real impact of enabling factors such as cost and travel distance on the sub-group with severe CMD, who are most in need of care. Future research could usefully examine the extent to which supply side factors such as the availability, affordability and accessibility of care affect service utilisation by those with the most severe needs.

If equitable access to health care is defined as equal utilisation by those with equal need for care [116], then there is some evidence pointing to the need to target underserved groups such as men, ethnic minority groups, the elderly and young adults, at least in HIC. However, the extent to which need factors such as symptom severity and disability were controlled in these analyses varied between studies, so we cannot definitively rule out the possibility that these differences can be explained by variability in need for treatment.

Attempts to address these inequities have been made in HIC through strategies such as enhancing cultural competence in mental health services [117] and targeting underserved groups through social marketing [118], with some success [119, 120]. Evaluations of these interventions typically measure adherence/attrition, patient satisfaction or attitudes towards seeking care rather than treatment-seeking behaviour, so their effectiveness in reducing mental health care inequities is still to be determined.

Regarding geographic location, some studies indicate that this may affect the type of provider chosen and the quality of care received [54, 121]. It is possible that the initial decision of whether or not to seek treatment is made independently of location of residence, but the subsequent decision of where to seek treatment, and the health system’s response, is influenced by geography. This wants further investigation (see “Unanswered questions and future research”, below).

Finally, although it was not the topic of this review, there was some evidence to suggest that the factors associated with health service utilisation for CMD may vary between the specialist and generalist sectors [64, 77, 87], which has been highlighted in other studies [100, 112]. This warrants further investigation as it has important implications for service planning. Thornicroft and Tansella (2013) advocate a stepped care model of mental health services, with the majority of services delivered through primary care in low-resource settings [122]. However, it remains to be investigated which balance leads to the most equitable use of services for CMD, and whether some groups are more likely to seek treatment through primary care in LMIC.

This review identified three major gaps in our knowledge: Firstly, a lack of research from LMIC; secondly, a dearth of research on contextual factors, particularly health systems factors; and thirdly, an absence of studies that are explicitly powered to test associations between the factor of interest and treatment-seeking for CMD.

The first of these gaps directly relates to the second objective of this review. Although we have drawn some tentative conclusions above, the generalisability of these findings to LMIC is questionable at best, since nearly 90% of the studies identified were from high-income countries. In contrast, 85% of the world’s population is expected to live in LMIC by 2030 [123], making this is an extremely important omission.

Not only was there a noticeable lack of population-based studies from LMIC, but those studies that were identified were less consistent in their findings than those from HICs. This may be in part due to the reduced statistical power of studies from areas where treatment rates are low, meaning that larger sample sizes are needed to detect an association. More research is urgently needed in LMIC – especially in those countries for which no population-based studies were identified – to determine whether the same factors are associated with treatment-seeking for CMD in non-Western settings, using large enough samples to detect an association.

Secondly, there was also a notable lack of published evidence on several contextual factors, in particular health systems factors that are likely to affect treatment-seeking. This includes the availability of services, the geographical accessibility of those services, and characteristics of services such as opening times, which are central to several models of access to health care [15, 124, 125]. This is a crucial gap, as such evidence could usefully inform service planning to expand access to care.

The extent to which distance affects treatment-seeking has particular relevance to debates around decentralisation and integration of mental health care [126]. Facility-based studies have pointed to distance and travel time as a potentially important determinant of health service utilisation [127‐131], which contrasts with the lack of evidence supporting an association with urban/rural residence found in this review. However, these studies cannot disentangle geographic differences in prevalence from differences in treatment seeking behaviour. Furthermore, unless they assess the use of all health facilities in a given area – both public and private – it is not clear if distance affects whether affected individuals seek any care, or if it merely influences the choice of provider among those who do decide to seek treatment. This review showed that there is a lack of population-based data on the influence of geographic accessibility on the uptake of health services for CMD, with the exception of crude comparisons of rural and urban areas, for which no association was found with treatment-seeking.

Finally, none of the studies included here justified their sample size with regard to the relationship between treatment-seeking and the factors investigated. It is therefore possible that the lack of associations identified in some of the studies included here are the result of under-powered studies, rather than a genuine lack of association. To build the evidence base in this area and confirm the hypotheses generated by the current review, future studies should ensure that they have sufficient statistical power to detect an association with the factors investigated.