Introduction Hypertension (HT) is common worldwide, affecting an estimated billion people, nearly three-quarters of whom live in low or middle income countries (LMICs) . HT is second, after smoking, as a contributor to the Global Burden of Disease in the latest (2010) analysis . In most individuals it is easily treated and controlled, with effective control reducing deaths and disability from a number of conditions, including cerebrovascular, cardiovascular, and renal disease . Yet in both developed and developing countries, a significant proportion of people with HT remain unaware of their diagnosis, and of those who are aware, only a minority are treated and have their HT successfully controlled . The reasons are many but, as with other chronic diseases, they include weaknesses in health systems, related to both structures and ways in which systems function ,. Health systems have been defined by the World Health Organization as “all the organizations, institutions and resources that are devoted to producing health actions”  and weaknesses may exist at the national, regional, district, community, and household level. Previous systematic reviews have examined the effects of health systems interventions delivered at the community or health facility level on HT care, such as educational interventions that target providers, organisational interventions strengthening collaboration between physicians and pharmacists, and using electronic records to improve management –. However, we are unaware of any previous systematic review exploring the effect of actions originating at national or regional health systems level, including health policies, programs, and interventions, on HT outcomes. Actions that have been hypothesized to influence HT care include strategies for procurement of essential medications, the existence of simple national guidelines for HT management, introduction of financial incentives for health care practitioners to diagnose or treat HT, and enhanced health insurance coverage . To address this gap, we systematically reviewed the literature examining the effect of national or regional health system arrangements on HT care and control, and make recommendations for future research and policy. Methods A protocol for this study has been published on the PROSPERO international prospective register of systematic reviews, with the record number PROSPERO 2012:CRD42012002864 . We used an established framework to illustrate the health system and its elements and guide our systematic review (Figure 1). This conceptual framework, which has been found useful in understanding the systems failings that impede effective management of non-communicable diseases ,, consists of four domains relating to key system level inputs that are required for effective chronic disease care: namely, physical resources (e.g., health facilities and diagnostic equipment), human resources (e.g., trained health care workers and managers), intellectual resources (e.g., treatment guidelines), and social resources (which draws on the concept of social capital and includes organizational measures to enhance collaboration). The existence of inputs is insufficient in itself, without effective systems to finance, deliver, and govern care; and these are also reflected in the framework. All of these domains influence the impact of the health system inputs on the health care outcomes of interest, which are HT awareness, treatment, control, and antihypertensive medication adherence in this case. The framework aims to capture the complex interactions and inter-relationships that exist between the elements within a health system, acknowledging that success of health systems does not simply require a “laundry list” of building blocks (as the WHO's 2007 framework is often perceived), but requires effective integration and alignment of these inputs ,. The framework also highlights the important role that context plays in shaping the relationship between health systems inputs and outcomes, recognizing the complex adaptive nature of health systems so that changes may yield different results in different settings . 10.1371/journal.pmed.1001490.g001 Figure 1 Schematic diagram of health systems conceptual framework. Inclusion Criteria We included studies that reported on the effects of national or regional health system level arrangements (factors, interventions, policies, or programs) on HT control and key upstream determinants of control: HT awareness, treatment, and medication adherence. Definitions of these outcomes are given in Box 1. We included studies looking at any adult population, including general populations, populations on treatment, and studies of people with specific co-morbidities, such as diabetes. The following types of studies were included: (1) Studies, such as controlled trials, cohort studies, and cross-sectional studies, which quantify the effects on HT outcomes of interventions, policies, or programmes, which are enacted at national or regional health system level, acting on one or more domains of the health-system. (2) Studies, such as qualitative studies, which report on the views and experiences of actors (e.g., patients, physicians, or policy makers) on national or regional health-system level barriers to HT awareness, treatment, control, or antihypertensive medication adherence. (3) Studies reporting on the impact of national or regional HT care policies or interventions that have relevance for other disease programs or for the design of the health system more broadly, such as those that require or lead to changes in primary care provision or other general aspects of the health system. Quantitative studies were included only if they reported a measure of association between the health system arrangement under investigation and at least one of the HT outcomes of interest (Box 1). There were no date or language restrictions. Studies that evaluated interventions, policies, or programs that are enacted at the individual level (e.g., provider or patient level) or organizational level of the health system (e.g., hospital or primary care organization), and do not require change at the level of the national or regional health system were excluded. Box 1. Definitions of Included Hypertension Outcomes (1) HT awareness. Defined as persons with clinically measured HT who have been diagnosed by a health care professional as hypertensive. (2) HT treatment. Defined as the use of at least one antihypertensive medication in an individual with known HT. (3) Antihypertensive medication adherence. Defined as consistently taking the antihypertensive medication regimen as prescribed by the health care provider. (4) HT control: defined as the achievement of BP below 140/90 mmHg (or other explicitly defined threshold) in individuals being treated for HT, or, alternatively, measured by the mean BP amongst individuals with HT. Search Strategy The search strategy and terms were developed collaboratively with an information specialist. Key words (MeSH terms) and free text terms were identified for each domain of our health systems framework and combined with search terms for HT outcomes to generate the search strategy for the electronic databases Medline, Embase, and Global Health (Text S2). To improve the likelihood of identifying studies from LMICs, modified searches were performed on the following databases: Latin American and Caribbean Health Sciences Literature (LILACS), Africa-Wide Information, Index Medicus for the South-East Asian Region (IMSEAR), Index Medicus for the Eastern Mediterranean Region (IMEMR) Western Pacific Rim Region Index Medicus (WPRIM). All databases were searched from inception to the present day on 8th May 2013. To identify further relevant studies, reference lists of included articles were hand searched and a forward citation search was performed on included studies using Web of Science. Study Selection Two reviewers independently screened the search results by title and abstract for potential eligibility. Full texts of potentially suitable articles were obtained and were further screened for inclusion by two reviewers. Disagreements in the screening of full texts were resolved by a third reviewer with expertise in health systems and this was required for four of the 122 screened papers. Data Extraction for Study Setting, Methodology, and Findings A data extraction form was developed in Microsoft Excel. Data were extracted from each study on study design, setting, health system domains investigated, study methods, and outcomes (Table S1). Where multiple analytical models were used for HT outcomes in a study, data were taken from the analytical model that had the highest level of control for other confounding factors. Data extraction was performed independently by two reviewers and compared and checked for disparities. Erroneous or inconsistent data were identified in one of the included papers, and we attempted to contact the authors of this paper for clarification. Clarification of these data was not forthcoming, so these data were excluded from the analysis. Risk of Bias Assessment Included studies were independently assessed for risk of bias by two reviewers. For observational study designs, risk of bias was assessed using a simple proforma for three domains: selection bias, information bias (differential misclassification and non-differential misclassification), and confounding (Text S3). Assessment of non-differential misclassification took into consideration the reliability of the measure used to report HT outcomes, which was particularly important for medication adherence, where a variety of methods were used for measurement. Risk of bias for each domain was assessed as either low, unclear, or high. Studies that had a low risk of bias in each domain, including a low risk of confounding, were classified as having a low overall risk of bias. For randomized studies the Cochrane risk of bias tool was used . Qualitative studies were evaluated for quality using an adapted version of a checklist used in a previous series of mixed methods systematic reviews incorporating both quantitative and qualitative studies (Text S4) ,. Assessment of Context and Complexity Considerations Due to the recognized importance of context and complexity to health systems research , we examined the extent to which included studies describe and explore these factors. We assessed to what extent studies had described the sociodemographic, political, or economic context in which they were conducted and the wider health system setting. We also assessed whether studies demonstrated a consideration of the complexity of health systems, including addressing inter-relationships between different health systems domains, for example, those between financing arrangements and retention of skilled health care workers, as well as interactions with contextual factors, such as the level of poverty or literacy amongst the population being served. This process was performed by one reviewer and checked for consistency by a second reviewer. Data Synthesis and Analysis A narrative synthesis was performed, with studies categorized according to the health system domain they investigated and the setting in which the study was performed. For making causal inferences about reported associations between health systems arrangements and HT outcomes, randomized controlled trials (RCTs) were considered the strongest study design, followed by cohort studies and then case-control studies. Cross-sectional studies and ecological studies, alone, were not considered appropriate for causal inference. Meta-analysis was not conducted as we judged that the included studies were heterogeneous in important aspects, including: populations (different ages and settings), study designs (cross-sectional, case-control, cohort), variable definitions (including different definitions of exposures and outcomes), comparisons (e.g., different type of insurance schemes), and analytical strategies (adjustment for different confounders). Results The screening process is described using an adapted Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flowchart (Figure 2) . 5,514 articles were screened by title and abstract for inclusion. The full text of 122 of the 5,514 articles was obtained and assessed for eligibility. 53 studies met eligibility criteria for this review. Full details of the included studies, including study design, setting, key findings, and risk of bias assessment can be found in Table S1. 51 of the included studies were quantitative and two were qualitative ,. Of the 51 quantitative studies, one was a RCT ; 12 were cohort studies –, two of which were retrospective ,; three were case-control studies –; 32 were cross-sectional studies; and three were ecological studies –. 42 of the 53 studies (79%) were carried out in countries classified by the World Bank as high-income countries, 36 of which were in the US. Six studies were carried out in upper middle-income countries ,–, three in lower middle-income countries ,,, and one in a low-income country . Table 1 describes the health systems factors investigated, classified into the domains of the conceptual framework (Figure 1). 10.1371/journal.pmed.1001490.g002 Figure 2 PRISMA flowchart. 10.1371/journal.pmed.1001490.t001 Table 1 Health system arrangements investigated by included quantitative studies, classified by health system domain. Health System Framework Domaina Health System Factor Being Investigated Number of Studies Number of Studies and Study Designs Setting of Studies (Countries) Physical resources Distance to health facilities 1 Cross-sectional (1) Ethiopia (1) All physical resources studies 1 Cross-sectional (1) Ethiopia (1) Human resources Level of training/specialism of treating physician 2 Cross-sectional (2) US (1), Mexico (1) Supply of health professionals 2 Cross-sectional (1) Mexico (1) All human resources studies 3 Cross-sectional (3) US (1), Mexico (2) Intellectual resources All intellectual resources studies 0 0 studies n/a Social resources All social resources studies 0 0 studies n/a Health system financing Health insurance status 21 Cohort (2)Case-control (3)Cross-sectional (16) US (20), Mexico (1) Medication costs or medication co-payments 14 Cohort (7)Case-control (1)Cross-sectional (6) US (9), Finland (1), Brazil (1), Israel (1), China (1), Cameroon (1) Co-payments for medical care 3 RCT (1)Cohort (1)Case-control (1) US (2), Hong-Kong (1) Physician remuneration model 2 Cross-sectional (1)Ecological (1) US (1), Canada (1) All financing studies 38 1 RCT (1)Cohort (10)Case-control (3)Cross-sectional (23)Ecological (1) US (30), Canada (1), Mexico (1), Hong-Kong (1), Israel (1), Finland (1), Brazil (1), China (1), Cameroon (1) Governance and delivery Care delivered by private or public provider 3 Cross-sectional (3) US (1), Greece (1), South Africa (1) Routine place of care 6 Cross-sectional (6) US (6) Routine treating physician 7 Case-control (1)Cross-sectional (6) US (7) Either a routine physician or place of care 1 Case-control (1) US All governance and delivery studies 16 Case-control (2)Cross-sectional (14) US (14), Greece (1), South Africa (1) a Some studies separately assess more than one health system arrangement. Effect of Health System Arrangements on Hypertension Outcomes Physical resources One study examined the effect of health system factors relating to physical resources (Table 2). This study was conducted in a low-income country, Ethiopia, and examined the effect of distance that patients were required to travel to health facilities providing HT care . The study was cross-sectional in design and had a low risk of bias for all methodological domains assessed. The study reported a moderate positive association between a shorter distance of travel to a health facility and antihypertensive medication adherence (odds ratio [OR] for medication adherence in those with a travel time to health facilities of less than 30 min versus travel time of more than 30 min, 2.02, 95% CI 1.19–3.43). 10.1371/journal.pmed.1001490.t002 Table 2 Summary of findings of studies examining the associations of arrangements relating to human or physical resources with hypertension outcomes. Health System Arrangement Study Setting and Sample Size Study Design Findings (95% CIs Given in Brackets Where Available). ORs Are Adjusted for Confounding Unless Stated Otherwise. Risk of Bias Assessment Physical resources Distance to health facility Ambaw et al. 2012  Ethiopia - University hospital, mixed rural and urban populationn = 384 Cross-sectional OR for medication adherence travel time to health facilities 30 min 2.02 (1.19–3.43) Low risk of bias. Human resources Grade of treating physician Federman et al. 2005  US - All male Veterans Affairs populationn = 15,893 Cross-sectional OR for BP control (baseline = 1 for resident). Mid level doctor 1.12 (0.98–1.28), attending 1.23 (1.08–1.39) Unclear risk of non-differential misclassification. Physician specialism Mejia-Rodriguez et al. 2009  Mexico - Regional Family medicine unitsn = 4,040 Cross-sectional OR for uncontrolled HT in those treated by non-specialists versus specialists 1.43 (1.20–1.71) Unclear risk of non-differential misclassification. Per capita supply of health professionals Bleich et al. 2007  Mexico - Nationally representative samplen = 2,130 Cross-sectional OR for HT treatment 1.04 (0.85 to 1.26) and control 0.81 (0.61–1.09) in areas with high versus low supply of health professionals. Unclear risk of non-differential misclassification. Human resources Three studies examined the effect of health system factors relating to human resources, none of which had a low risk of bias (Table 2). Two of the three were conducted in an upper-middle income country (both in Mexico), and one was conducted in a high-income country (US). One US cross-sectional study evaluated the effect of the treating physician's seniority on HT control . This study found a small positive association between seniority of treating physician and HT control. The adjusted OR for HT control was 1.23 (95% CI 1.08–1.39) for patients treated by an attending level physician compared to those treated by a resident level physician. One Mexican cross-sectional study evaluated the impact of being treated by a specialist on HT control . This study found a moderately increased risk of uncontrolled HT in hypertensive individuals treated by non-specialist physicians (general practitioners) compared to those treated by specialists (adjusted OR 1.43, 95% CI 1.20–1.71). Another Mexican cross-sectional study evaluated the effect of the density of health professionals and did not find an association with HT treatment or control . Intellectual resources None of the included studies evaluated the effects of health system factors relating to intellectual resources on HT outcomes. Social resources None of the included studies evaluated the effects of health system factors relating to social resources. Health Systems Financing 38 quantitative studies analyzed the effects of health systems financing on HT outcomes (34 of these studies were conducted in high-income countries and four in middle-income countries). Four different health system arrangements were analyzed, with 21 studies assessing effects of health insurance coverage, 11 examining the effects of medication co-payments or costs, three analyzing co-payments for medical care, and two looking at physician remuneration models. Twenty of 21 studies evaluating health insurance coverage were conducted in the US and one in Mexico (Table 3). Seven of the 21 studies had a low risk of bias. Two were cohort studies, three were case-control studies, and 16 were cross-sectional studies. 19 of the 21 studies evaluating health insurance reported direct comparisons of HT outcomes in insured and uninsured patients, while two studies only compared private and public insurance schemes. Two cohort studies, both set in the US, compared uninsured patients with insured patients ,. One of the cohort studies had a 9-y follow-up and found that being uninsured was associated with an increased risk of both unawareness of HT (relative risk [RR] of unawareness in uninsured versus insured patients 1.12, 95% CI 1.00–1.25) and inadequate control of HT (RR of inadequate control in uninsured versus insured patients 1.23, 95% CI 1.08–1.39) . The other cohort study had a follow up period of 30 mo and found that medication adherence was lower in uninsured patients compared to insured patients (OR for medication adherence for uninsured versus insured 0.426, 95% CI 0.282–0.757) . One of two US set case-control studies comparing HT outcomes in uninsured and insured patients reported that insurance was associated with an increased likelihood of HT control (OR for HT control in insured versus uninsured 2.15, 95% CI 1.02–4.52) . The other case-control study reported a non-significant association between being uninsured and having severe uncontrolled HT (OR for severe uncontrolled HT in uninsured versus insured 1.9, 95% CI 0.8–4.6). . Fifteen cross-sectional studies reported comparisons of HT outcomes in insured and uninsured patients. Eight of these 15 studies reported that insurance was associated with improved HT treatment, control or medication adherence ,–. The seven other cross-sectional studies that compared HT outcomes in insured patients and uninsured patients, reported no significant negative or positive associations between insurance status and HT outcomes –. Two further studies looking at health insurance status compared HT outcomes in patients with public and private health insurance. A case-control study set in the US found increased odds of HT control in patients with private insurance compared to patients with public insurance (OR for HT control 3.40, 95% CI 1.25–9,28) . A cross-sectional study, also set in the US, found no significant association between private or public insurance and HT awareness or treatment, but did report significantly lower levels of systolic blood pressure (BP) in patients with private insurance compared to public insurance (p 0.05), other government insurance 86.9% (77.3–92.8; p>0.05), uninsured 60.2% (46.0–72.8; p 0.05), other government insurance 74.5% (64.0–82.8; p>0.05), uninsured 42.6% (28.7–57.7; p 0.05 systolic BP). High risk of sample bias. Unclear risk of non-differential misclassification bias. High risk of confounding. Bleich et al. 2007  Mexico. Nationally representative samplen = 2,130 Cross-sectional OR for HT control with seguro popular (insured) versus uninsured, for treatment 1.50 (1.27–1.78), and for control 1.35 (1.00–1.82) Unclear risk of non-differential misclassification bias. Brooks et al. 2010  US. Framingham cohortn = 1,384 Cross-sectional Men and women treated less when uninsured (OR 0.19 [0.07–0.56] and 0.31 [0.12–0.79], respectively). Men less controlled when uninsured (OR 0.17 [0.04–0.68]), not women. Low risk of bias. Duru et al. 2007  US. Nationally representative samplen = 3,496 Cross-sectional OR for HT control (ref 1.0 for private insurance), Medicare = 0.80 (0.61–1.05), Medicaid 0.75 (0.47–1.20), no insurance 0.63 (0.44–0.92). Low risk of bias. Ford et al. 1998  US. Nationally representative samplen = 1,724 Cross-sectional Found no differences in HT awareness, treatment, or control with no health insurance, Medicaid only, or other health insurance compared to those insured fully. High risk of non-differential misclassification bias. He et al. 2002  US. General populationn = 4,144 Cross-sectional OR of HT control with government insurance = 1.08 (0.70–1.68); private insurance = 1.59 (1.02–2.49), versus no insurance. Low risk of bias. Hill et al. 2002  US. Inner-city African American men presenting to the emergency departmentn = 309 Cross-sectional No significant association between health insurance status and HT control. Unclear risk of sample bias. Hyman and Pavlik 2001  US. Nationally representative samplen = 10,576 Cross-sectional OR for uncontrolled HT with insurance versus without 1.30 (0.79–2.13) Low risk of bias. Kang et al. 2006  US. Low SES Korean-American elderlyn = 146 Cross-sectional OR of HT treatment with any insurance 2.41 (0.91–6.39), Medicare 2.06 (0.66–6.42), Medicaid 3.21 (0.89–11.61), private insurance1.46 (0.29–7.39) versus none. No association between insurance type and control. High risk of sample bias. Unclear risk of non-differential misclassification bias.High risk of confounding Moy et al. 1995  US. Nationally representative samplen = 6,158 Cross-sectional OR of non-treatment of HT with Medicare or Medicaid versus private 1.19 (0.99–1.41), Uninsured versus private 1.49 (1.18–1.89) High risk of non-differential misclassification bias. Unclear risk of differential misclassification bias. Nguyen et al. 2011  US. Population sample from NYCn = 1,334 Cross-sectional Public versus private insurance. OR for HT awareness 1.2 (0.4–4.1), treatment 1.1 (0.4–3.6). Average SBP lower with private insurance versus public. Low risk of bias. Shea et al. 1992b  US. Hospital-based African American and Hispanic inner-city populationn = 207 Cross-sectional Health insurance was not significantly associated with medication adherence in a multivariable model. High risk of sample bias. Unclear risk of non-differential misclassification bias. Turner et al. 2009  US. Mostly African American women in Philadelphian = 300 Cross-sectional OR: In the past year had to go without usual BP medications because not covered (yes) 1.29 (0.26–9.49) versus no High risk of sample bias. Wyatt et al. 2008  US. African American population from Jackson, MSn = 4,986 Cross-sectional No association reported between health insurance status and HT awareness, treatment, or control. Unclear risk of sample bias. RR, risk ratio; SES, socioeconomic status. Fourteen quantitative studies measured the association of medication co-payments or costs with HT control or treatment adherence, nine of which were set in the US, and one in each of Cameroon, China, Finland, Israel, and Brazil (Table 4). Two of the 14 studies had a low risk of bias. Seven of the 14 studies were cohort studies, one was a case-control study, and six were cross-sectional studies. All seven cohort studies reported associations between increased medication costs or co-payments and reductions in HT control or reduced adherence to antihypertensive medication ,,,,,,, although for one of these seven cohort studies, the association between increased co-payments and reduced medication adherence was only found for low medication co-payments, and at high co-payment levels medication adherence was actually found to increase (OR for medication adherence versus baseline of 1 for US$0 co-payments was 0.72 for US$1–US$9 co-payments (p 0.05), and 1.32 for co-payments >US$30 (p 0.05) for US$10–US$29 co-payments, OR = 1.32 (p US$30 Unclear risk of sampling bias Elhayany and Vinker 2001  Israel. Mixed Arab/Jewish patients from Ramle and Lod (deprived populations)n = 260 Cohort - before and after study of intervention. (2-y follow-up) Systolic BP and diastolic BP reduced by 8 and 3.2 mmHg, respectively, 24 mo following intervention to eliminate prescription co-payments (p 0.05. High risk of participant and personnel blinding. Unclear risk of random sequence generation, allocation concealment, and blinding of outcome assessment. Low risk of selective reporting and incomplete outcome data. Wong et al. 2010  Hong Kong. Chinese patients in primary caren = 83,884 Cohort (unclear length of follow-up) OR for medication adherence fee payers versus fee waivers 1.14 (1.09–1.19) Low risk of bias. Ahluwalia et al. 1997  US. Low-income, African-Americans in an urban ambulatory hospitaln = 733 Case-control OR of control when cost of care not a deterrent versus cost as a deterrent 2.35 (1.19–4.67) Unclear risk of differential misclassification bias. PDC, proportion of days covered by medication. Three studies assessed co-payments or costs of medical care (not simply medications), two of which were conducted in the US (an RCT and a case-control study) and one in Hong Kong (a cohort study). (Table 4) ,,. One of the three studies, the cohort study from Hong-Kong, had a low risk of bias . The RCT reported a higher mean BP level amongst individuals with HT who had cost-sharing insurance plans compared to those with free care, although this was non-significant for systolic BP . The adjusted mean difference in diastolic BP between the two groups was 1.9 mm mercury (mmHg) (95% CI 0.3–3.5 mmHg) and the adjusted mean difference in systolic BP was 1.8 mmHg (95% CI −0.6 to 4.5 mmHg). The case-control study reported that cost of care was a deterrent to BP control (adjusted OR for BP control when cost not a deterrent versus cost as a deterrent: 2.35, 95% CI 1.19–4.67) . The cohort study, which was set in Hong Kong, found, conversely, that being a fee payer was associated with improved adherence to prescribed antihypertensive medications compared to fee waivers (adjusted OR for adherence fee payers versus fee waivers 1.14, 95% CI 1.09–1.19) . Two studies evaluated the association of physician remuneration models with HT control or treatment adherence, one an ecological study set in Canada, and one a US cross-sectional study (Table 5). Neither study had a low risk of bias. The US study reported improved rates of HT control amongst patients treated under a capitation model compared to fee-for service patients (adjusted OR for HT control 1.82, 95% CI 1.02–3.27 for capitation versus fee-for-service patients) . The Canadian study reported highest rates of HT treatment and control among practices using a capitation model, compared to fee-for-service and salary models . HT awareness levels were highest in practices with a fixed salary remuneration model. 10.1371/journal.pmed.1001490.t005 Table 5 Findings of quantitative studies examining the association of physician remuneration models with hypertension outcomes. Study Setting and Sample Size Study Design and Length of Follow-up Where Applicable Findings (95% CIs Given in Brackets Where Available). ORs Are Adjusted For Confounding Unless Stated Otherwise. Risk of Bias Assessment Tu et al. 2009  Canada. Primary care in Ontario.n = 135 Ecological Differences in rates of HT awareness (p = 0.22), treatment (p = 0.01) and control (p 140 mmHg systolic and >90 mmHg diastolic BP). Diastolic BP 3.29 mmHg greater in public versus private sector (p = 0.042). Unclear risk of sample bias. Kotchen et al. 1998  US. Inner-city African American population from Milwaukeen = 583 Cross-sectional Unadjusted OR for HT control: Private provider 1.20 (0.62–2.32) versus non-private provider High risk of confounding. Unclear risk of sample bias. de Santa-Helena et al. 2010  Brazil. Patients from family health units in Blumenaun = 595 Cross-sectional OR for non-adherence: Treated by public health service (SUS) 1.8 (1.1–2.7) versus private medical provider. Unclear risk of non-differential misclassification. Yiannakopoulou et al. 2005  Greece. Patients admitted for elective surgery in Athens.n = 1,000 Cross-sectional Medication adherence with private physician 25.1% versus 10% of those with physician in rural areas and 8.8% of with physician from the National Health System (p US$30) were, surprisingly, associated with improved medication adherence. The study authors do not provide an explanation for this result in the paper, but it could be that the subgroup of patients with co-payments of US$30 or more for medications have shared characteristics that were not analyzed in this study, such as high socioeconomic status, which may confound the association between co-payment levels and medication adherence. The association between reduced medication co-payments and improved HT outcomes was replicated in single studies from China, Finland, Israel and Brazil ,,, but not in a study of Hong Kong Chinese, which found that fee payers had improved medication adherence compared to those with fee waivers . The finding of an association between reduced medication co-payments and improved HT outcomes is intuitive and suggests that costs of medications or health care consultations may act as a barrier to optimal HT care in the US, and potentially other settings, including LMICs. A relationship between increased medication co-payments and treatment discontinuation has also been reported for diabetes care in the US . Although lacking longitudinal studies, we found a large positive association between having a routine physician or place of care for HT management and treatment, awareness, control, and adherence to antihypertensive treatment, again in the US ,,,,,,–. This finding is consistent with a recent systematic review of the effect of a usual source of care, showing an association with improved preventive services and chronic disease control . Although it is unclear whether having a routine physician or a place for HT care is more important, this may matter less than the implication that the absence of a consistent source of care reduces awareness, treatment, and control of HT. It is possible, however, that this effect is linked to health system financing arrangements, as those without insurance or facing high co-payments may be least likely to have consistent access to care . There were no longitudinal studies looking at differences in outcomes of HT management provided by the private or public sector, and the four cross-sectional studies considering this question were all at risk of bias, were in different settings, and had different findings, so general inferences were not possible ,,,. All four included studies that evaluated complex multi-component national or regional policy interventions reported some improvement in HT care. These studies had significant methodological flaws including, in some cases, a lack of an adequate control group, precluding attribution of the improvement in HT outcomes to the intervention. However, despite their limitations, these studies may be useful for policymakers seeking to understand ways to strengthen health systems for chronic disease care, particularly in LMICs ,. Labhardt et al., for example, demonstrated the feasibility of task shifting from physicians to non-physician health care workers for HT management in Cameroon, outlining the integrated interventions across multiple health system domains required to deliver improvements in health outcomes . Research on health systems factors influencing HT care is unequally distributed geographically. There is a lack of evidence from LMICs, which bear around three-quarters of the global HT burden . Furthermore, even in high-income countries, health systems barriers to care have been seen mainly as financial, while the understanding of how a complex mix of other factors influence care is relatively new. Intellectual and social resources, such as the production and use of knowledge, social capital, and systems for communication have only recently emerged as distinct areas of research. As a result we found only a small number of studies examining the impact of health system factors relating to human resources or physical resources, and no studies evaluating the impact of intellectual or social resources. This meant we were unable to make firm conclusions about the effects of these factors on HT outcomes. A number of included studies used models to conceptualize the mechanisms by which health systems factors may interact with other key variables to influence HT outcomes. For example Moy et al. and Ahluwalia et al. (1997) used the Anderson-Aday model, which illustrates how three types of population characteristics can influence medical care for HT ,. Factors relating to health systems such as the presence of a usual source of care or health insurance are seen as “enabling” factors for medical care. These “enabling” factors interact with “predisposing” factors such as ethnicity, gender, and socioeconomic status, and “need characteristics” such as health status to determine access and outcomes of medical care. Models such as Anderson-Aday are useful to the extent that they can help demonstrate how health systems factors may interact with other key factors in determining HT outcomes. However, the studies reviewed here lack quantitative and qualitative data on the nature and strength of these interactions, highlighting an important gap to be addressed by future research. Study Limitations and Strengths The majority of included quantitative studies were cross-sectional, and the few longitudinal studies we did find were restricted to either health system arrangements relating to financing or to evaluating the effects of complex multi-component interventions. Inferences about temporal and potentially causal relationships between health systems arrangements and HT outcomes, could, therefore, only be made for a limited number of factors. In addition, included quantitative studies were of variable methodological quality, with only one being randomized and a minority having a low risk of bias for all assessed methodological domains. When considering the findings of this review, the risk of publication bias cannot be ruled out, particularly for the positive findings relating to health insurance status, medication and treatment costs and co-payments, and presence of a routine setting of care, where it is possible that studies with null findings are under-published. It was not possible to produce an Egger funnel plot to formally assess the risk of publication bias, for the same reasons that meta-analysis was not performed, namely the heterogeneity in the study designs, outcome measures, analysis strategies, and populations in the included studies. Reporting bias within individual studies may also be a factor, many of which might have explored the effects of multiple factors on HT outcomes, and may have failed to report results for health system arrangements which did not show significant effects. The lack of published protocols for the included studies did not allow us to estimate the magnitude of this potential bias. A strength of the review is the addition of forward and backward searching methods to the initial database search for articles. A number of additional studies were identified using these methods, before reaching a saturation point at which the only relevant studies being identified were already included. We included only two qualitative studies, which did not contribute important data about the views of policymakers and health care workers on health systems factors affecting HT care, contrary to what we had initially hoped. The use of a conceptual health systems framework facilitated the conduct of the review, enabling systematic generation of terms for the search strategy and for classification of included studies according to the domains in the conceptual framework. However, the classification and reporting of our findings according to health system domain does not encourage the integrated view of health systems that the framework promotes. For example, classifying the effect of usual source of care into the domain of health systems governance and delivery obscures the fact that the delivery of care from a regular source is very much dependent on human and physical resources inputs to the health system. The difficulties in presenting such complexities are perhaps a reflection of the fact that few of the included studies explored inter-linkages between health system components, with the majority exploring the association between health system arrangements and HT outcomes as simple linear relationships. There were some notable exceptions, however; one study, for example, examined the interaction of insurance status and the presence of a usual source of care on HT outcomes . Implications for Policy We found an association between reduced co-payments for health care, including for medications, and improved outcomes of HT care in multiple US studies, and in single studies set in Finland, Israel, and Brazil. This is consistent with a wealth of other evidence on how co-payments reduce uptake of necessary care and has clear implications for policy makers, particularly as the balance of evidence does not suggest that reducing medication co-payments leads to an increase in overall health care expenditure –. On balance, we found health insurance coverage to be associated with improved outcomes of HT care in US settings, suggesting that expanded insurance coverage through The Patient Protection and Affordable Care Act (also known as Obamacare) may improve HT outcomes. Implications for Research This study indicates a number of possible implications for future research. Ultimately, an increase in the number of high quality, longitudinal and randomized studies identifying and analyzing the effect of health system arrangements on HT care is required, particularly in LMICs where the majority of the global burden of HT lies, and where weaknesses in health systems are thought to play a significant role in deficiencies in chronic disease care . The focus on financing has highlighted important barriers to effective care and control of HT but needs to be supplemented by research examining other domains, such as delivery and governance mechanisms, production of knowledge, and the social function in the health systems. Most existing studies have a focus on independent effects of different health systems arrangements, thereby creating a “laundry list” of isolated components. Recognizing the shortcomings of this approach, it is important that future studies attempt to capture the complexities and interactions between health systems arrangements. In addition, future national or regional health systems strengthening programs that aim to improve care for chronic conditions such as HT should be robustly evaluated, using longitudinal controlled study designs where possible. Moving forward, there is a clear need for more robust designs of studies in a much wider range of settings, especially in LMICs. This will ideally include cluster RCTs and prospective longitudinal studies with detailed data on individual and health system characteristics, complemented by qualitative studies to see inside what is often a health systems black box. Such studies also call for consistency in health systems definitions and outcome measures. A particular challenge will be to take account of the complexity of health systems and all health system domains, as well as interpreting studies by not simply as showing what works, but what works in what circumstances . This review should help inform the design of such studies. In particular, the findings are being combined with multi-method appraisals of health systems to understand the barriers faced by patients with HT and their health workers to design cluster randomized trials in several LMICs . Importantly, given that there are many health systems frameworks, this review has shown the practicality of using the one chosen, a framework that is also being used in the multi-method appraisals and that has been found useful in similar previous studies using diabetes as a probe to analyze health systems ,. Research such as this addresses a crucial gap in understanding of how different models of health systems contribute to health. Supporting Information Table S1 Study designs, settings, findings, and risk of bias of included studies. (DOCX) Click here for additional data file. Text S1 PRISMA checklist. (DOCX) Click here for additional data file. Text S2 Search strategy for Medline. (DOCX) Click here for additional data file. Text S3 Tool for assessing risk of bias for observational studies. (DOCX) Click here for additional data file. Text S4 Quality appraisal tool for qualitative studies. (DOCX) Click here for additional data file.