- Record: found
- Abstract: found
- Article: found
Generating statistics from health facility data: the state of routine health information systems in Eastern and Southern Africa
review-article
Author(s):
Abdoulaye Maïga
1
,
,
Safia S Jiwani
1 ,
Martin Kavao Mutua
2 ,
Tyler Andrew Porth
3 ,
Chelsea Maria Taylor
4 ,
Gershim Asiki
2 ,
Dessalegn Y Melesse
5 ,
Candy Day
6 ,
Kathleen L Strong
7 ,
Cheikh Mbacké Faye
8 ,
Kavitha Viswanathan
9 ,
Kathryn Patricia O’Neill
9 ,
Agbessi Amouzou
1 ,
Bob S Pond
10 ,
Ties Boerma
11
Other contributor(s): (Collab)
Publication date (Electronic):
29 September 2019
Journal:
BMJ Global Health
Publisher:
BMJ Publishing Group
Read this article at
There is no author summary for this article yet. Authors can add summaries to their articles on ScienceOpen to make them more accessible to a non-specialist audience.
Abstract
Health facility data are a critical source of local and continuous health statistics. Countries have introduced web-based information systems that facilitate data management, analysis, use and visualisation of health facility data. Working with teams of Ministry of Health and country public health institutions analysts from 14 countries in Eastern and Southern Africa, we explored data quality using national-level and subnational-level (mostly district) data for the period 2013–2017. The focus was on endline analysis where reported health facility and other data are compiled, assessed and adjusted for data quality, primarily to inform planning and assessments of progress and performance. The analyses showed that although completeness of reporting was generally high, there were persistent data quality issues that were common across the 14 countries, especially at the subnational level. These included the presence of extreme outliers, lack of consistency of the reported data over time and between indicators (such as vaccination and antenatal care), and challenges related to projected target populations, which are used as denominators in the computation of coverage statistics. Continuous efforts to improve recording and reporting of events by health facilities, systematic examination and reporting of data quality issues, feedback and communication mechanisms between programme managers, care providers and data officers, and transparent corrections and adjustments will be critical to improve the quality of health statistics generated from health facility data.
Related collections
Most cited references11
- Record: found
- Abstract: found
- Article: not found
WHO and UNICEF estimates of national infant immunization coverage: methods and processes.
Roeland Monasch, Barbara Lautenbach, M Neill … (2009)
WHO and the United Nations Children's Fund (UNICEF) annually review data on immunization coverage to estimate national coverage with routine service delivery of the following vaccines: bacille Calmette-Guérin; diphtheria-tetanus-pertussis, first and third doses; either oral polio vaccine or inactivated polio vaccine, third dose of either; hepatitis B, third dose; Haemophilus influenzae type b, third dose; and a measles virus-containing vaccine, either for measles alone or in the form of a combination vaccine, one dose. The estimates are based on government reports submitted to WHO and UNICEF and are supplemented by survey results from the published and grey literature. Local experts, primarily national immunization system managers and WHO/UNICEF regional and national staff, are consulted for additional information on the performance of specific immunization systems. Estimates are derived through a country-by-country review of available data informed and constrained by a set of heuristics; no statistical or mathematical models are used. Draft estimates are made, sent to national authorities for review and comment and modified in light of their feedback. While the final estimates may not differ from reported data, they constitute an independent technical assessment by WHO and UNICEF of the performance of national immunization systems. These country-specific estimates, available from 1980 onward, are updated annually.
- Record: found
- Abstract: found
- Article: not found
Neighbourhoods and potential access to health care: the role of spatial and aspatial factors.
Laura Bissonnette, Kathi Wilson, Scott Bell … (2012)
The availability of, and access to, primary health care is one neighbourhood characteristic that has the potential to impact health thus representing an important area of focus for neighbourhood-health research. This research examines neighbourhood access to primary health care in the city of Mississauga, Ontario, Canada. A modification of the Two Step Floating Catchment Area method is used to measure multiple spatial and aspatial (social) dimensions of potential access to primary health care in natural neighbourhoods of Mississauga. The analysis reveals that neighbourhood-level potential access to primary care is dependant on spatial and aspatial dimensions of access selected for examination. The results also show that potential accessibility is reduced for linguistic minorities as well as for recent immigrant populations who appear, on the surface, to have better access to walk-in clinics than dedicated physicians. The research results reinforce the importance of focusing on intra-urban variations in access to care and demonstrate the utility of a new approach for studying neighbourhood impacts that better represents spatial variations in health care access and demand. Copyright © 2012 Elsevier Ltd. All rights reserved.
- Record: found
- Abstract: found
- Article: not found
Improving Imperfect Data from Health Management Information Systems in Africa Using Space–Time Geostatistics
Introduction Public health decision-makers require accurate and timely information on disease-specific treatment burdens within a health system to monitor and plan resource needs [ 1– 4]. A basic requirement is reliable national and sub-national data detailing the number of treatment events for a given disease or condition occurring at health facilities each month or year. In most African settings, this requirement is addressed with a health management information system (HMIS) that coordinates the routine acquisition of treatment records from health facilities and the transfer, compilation, and analysis of these data through district, regional, and national levels. A perfect HMIS requires all health facilities to report promptly in all months, allowing a comprehensive quantification of treatment events through time and space across the health system. The reality of HMIS in Africa and elsewhere stands in marked contrast to this ideal [ 5– 9]. Typically, many facilities never report, or report only intermittently, resulting in spatially and temporally incomplete national data [ 10– 13]. Following several decades of donor investment in HMIS across Africa, the incomplete nature of routine national reporting has shown little improvement [ 3, 14]. Faced with poor data coverage, national treatment burdens are often estimated using rudimentary methods to account for missing values. The objective of this paper is to present a geostatistical model that predicts missing data in order to provide more reliable estimates of national outpatient treatment burdens with known accuracy. The model has been developed and tested using the example of presumed malaria cases in the Kenyan government's formal health sector. Methods The Kenyan HMIS Dataset Data were obtained from the Department of Health Management Information Systems of the Kenyan Ministry of Health. These data consisted of monthly records of diagnoses made at outpatient departments of health facilities across Kenya over an 84-mo period (January 1996–December 2002). Each record included the total number of all-cause diagnoses made at a given facility during a given month. An additional 11 diagnostic codes were available for each monthly record per facility. We selected malaria as the diagnostic code for model development for a number of reasons: (a) it accounted for over a third of all diagnoses made during the period of observation; (b) malaria is a disease that demands accurate quantification for health system planning in the light of increased donor assistance [ 9], particularly in the era when new expensive therapeutics are being adopted [ 9, 15]; and (c) malaria exhibits considerable spatial [ 16, 17] and temporal [ 18, 19] heterogeneity across Kenya. The records available within the routine HMIS data were not structured by age or sex, nor were they distinguished as initial or follow-up visits, and diagnoses were generally not slide-confirmed. The data, therefore, represent total cases (TC) or presumed malaria cases (MC) seen as outpatients each month at health facilities identified by a unique facility code. Data for each facility were matched to an independent database indicating the longitude and latitude of formal government, mission, and private health facilities nationwide. Details of how this spatial database was constructed are provided elsewhere [ 20] and were updated in 2005 [ 21]. In this paper, we focus on the government providers of routine outpatient care in order to assess treatment burdens within this sector, although the techniques presented can be extended to include georeferenced facilities within any given sector. Government health facilities at the district level are structured according to the levels of service they provide, with the most sophisticated being the general hospitals supporting a network of health centres that in turn act as referral points from dispensaries at the periphery. Space–Time Geostatistics A straightforward technique for predicting national MC totals using incomplete data is to scale up the tally of cases from available records in proportion to the number of missing data. This simplistic approach neglects any heterogeneity in the pattern of MC through space and time across the country. A more sophisticated approach is to predict each missing record individually from existing data. In the presence of spatial and temporal heterogeneity in MC, it is intuitive to allow data that are proximate to the record being predicted to have more influence on its prediction than those that are distant. In a traditional geostatistical approach [ 22, 23], the nature of spatial heterogeneity in the variable of interest is modelled explicitly using a variogram function that relates dissimilarity (quantified using semivariance) to spatial separation (termed lag). This function is then used to determine optimal data weightings in an interpolation exercise such as ordinary kriging, which predicts missing values using a weighted linear average of proximate data. Space–time kriging (STK) is an extension of ordinary kriging that considers simultaneously spatial and temporal heterogeneity and can provide more accurate predictions when the variable of interest is distributed through time as well as a space [ 24– 27]. The one-dimensional spatial variogram function is replaced with a two-dimensional space–time variogram, and the kriging algorithms are adapted to make predictions using spatially and temporally proximate data ( Protocol S1). Model Development We used STK to predict MC values at facilities where monthly records were missing. The accuracy of geostatistical predictions is greatly influenced by the amount of spatial correlation present in the variable of interest, that is, the extent to which values vary smoothly through space. The spatial structure of MC values at different facilities is confounded by facility-specific factors such as their type, catchment population size, and utilisation. These factors are not constrained spatially in the same way as malaria risks and may vary widely between facilities, regardless of their spatial proximity. To increase the predictive accuracy of STK it was necessary to increase the spatial correlation of the predicted variable by standardising MC by these facility-specific factors. This standardisation was achieved by dividing each monthly MC value by the mean monthly TC (MMTC) at each facility. MMTC was used as a proxy measure of facility catchment populations, reflecting broad utilization rates driven by the facility type and catchment population densities. The modelling framework therefore consisted of several components ( Figure 1). A completed set of TC values was required for each facility (i.e., 84 continuous months) in order to estimate MMTC. This set was provided by a separate STK procedure that predicted missing TC values, *TC (where the asterisk denotes a prediction), using the existing data. The mean of the combined set of TC data and *TC predictions for each facility, *MMTC, was then calculated. *MMTC was considered a more reliable proxy of catchment population than individual monthly TC values, representing a 7-y average less susceptible to both prediction bias and short-term fluctuations in utilisation. The monthly MC data were then standardised by dividing each by the corresponding *MMTC value to estimate a new variable, standardised MC (SMC). This new variable displayed a greater amount of spatial correlation than the raw MC data. SMC data were then used in a second STK exercise to predict *SMC at all missing points. These predictions were then back-transformed to *MC by multiplication by the relevant *MMTC value. Details of the methodological steps involved in the STK exercises to predict *TC and *SMC are detailed in Protocol S1. The above modelling framework resulted in predictions of MC at all facilities and for all months for which data were missing. In combination with the original data, this set represented a complete picture of the treatment burden for presumed malaria at all facilities for all months. This set could be aggregated to provide treatment burdens at any spatial level from the individual facility through to the district, provincial, and national levels for the 7-y period. Further, averaging could be applied to estimate values for any month or year in the set. Model Testing A validation procedure was carried out to test the performance of the model in terms of the accuracy of predictions of MC at different levels of spatial and temporal aggregation. A test set of 6,349 monthly records (representing a 10% sample) was selected from the full dataset using a stratified random sampling that ensured representative proportions of each facility type. The test set was removed from the database, and the STK modelling procedure was repeated in its entirety using the remaining 90% of data to predict MC values for the test set. The resulting predictions were then compared to the reference values to provide a set of known prediction errors that could be considered a sample of the (unknown) errors of the main prediction exercise. The total prediction error for the test set was calculated, along with the mean and standard deviation error nationwide at the level of individual facility-months. A series of subsets was then created from the test set by aggregating records together over space–time units (district-months, district-years, province-months, province-years, and so on), and the magnitude of errors was compared between subsets. The variance of these errors was found to decrease in inverse proportion to the number of records aggregated together in each subset ( Figure S1). This relationship was then used, along with the sample errors, to estimate the total prediction error and associated variance in each space–time unit. Monte Carlo simulation was used to estimate the combined distribution of total prediction errors for all space–time units in each aggregation level. This procedure resulted in, for example, estimates of the range (expressed as a 95% confidence interval) of percentage errors that could be expected for predictions of total MC for all facilities in a district over a month, all facilities in a province over a year, and so on. Results Data Coverage A total of 2,165 government facilities were identified through consultation with district health management teams and other service providers ([ 20, 21]; A. M. Noor and P. W. Gikandi, unpublished data). It was possible to generate a longitude and latitude from various sources for over 92% of these facilities [ 21]. These included 129 hospitals, 474 health centres, and 1,399 dispensaries ( Table 1). The importance of establishing a comprehensive database was demonstrated by the identification in the above exercise of an additional 400 government facilities that were not included in the central HMIS database. A total of 163 facilities were included in this study that could not be georeferenced. Missing MC values for these facilities were estimated using the local district mean for that month. Reporting Rate Underreporting was found to be widespread, although there was considerable variation spatially and temporally ( Figure 2) and between facility types ( Table 1). No facilities reported in all 84 mo, whilst 546 facilities (25%) did not report in any month. A complete 84-mo dataset for each of the 2,165 facilities would consist of 181,860 facility-months. There were 63,642 records, representing an overall reporting rate of 35%. The overall reporting rate varied both within and between years, with a minimum of 6% in December 1997 (this coincided with a nationwide industrial dispute by nurses) and a maximum of 44% in February 1996. The reporting rate also displayed a seasonal pattern, with generally more facilities reporting during the first three quarters of each year (36%) than in the last quarter (31%). A total of 18.67 million cases of presumed malaria were reported, with a mean of 293.4 cases per facility-month. The totals (means) were 3.36 million (716.9) for hospitals, 6.05 million (323.4) for health centres, and 9.26 million (230.2) for dispensaries. Prediction of Treatment Burdens The mean annual total of presumed malaria cases (i.e., the combined total of data plus predictions) at all government facilities between 1996 and 2002 was 6.79 million cases, with a mean of 261.5 cases per facility-month ( Table 2). The corresponding values for each facility type were 1.11 million for hospitals, 1.74 million for health centres, and 3.95 million for dispensaries, with means of 716.0, 300.3, and 211.8 cases per facility-month, respectively. Mean annual totals for each district ( Figure 3) displayed a pattern of spatial heterogeneity that corresponded broadly to a combination of malaria ecology [ 17, 28], population distribution [ 29], and facility locations [ 20]. Model Testing Comparison of data with predictions for 6,349 randomly selected MC data points in the test set yielded mean prediction errors for hospitals, health centres, and dispensaries of 58.2, −8.8, and −4.7 cases per facility-month. The true and predicted sums of the entire national test set were 1,899,234 and 1,891,136, respectively, representing an overall prediction error of −0.4%. The predictive accuracy of the model increased as predictions were made over larger aggregated space–time units ( Table 3). It was estimated that 95% of MC totals for district-months would be predicted to within 35.3% of the true value and that three-quarters would be predicted to within 15.1%. The equivalent errors for predictions of annual totals at the provincial level were 12.2% and 5.5% and at the national level were −1.3% and −0.9%. Discussion Between 1996 and 2002 the Kenyan HMIS contained only 35% of the expected monthly records from government clinics providing outpatient care nationwide. This seriously limits the direct use of these data for planning health service needs, including staffing and disease-specific commodities such as anti-malarial drugs. Inadequate spatial and temporal coverage of information is compounded by a lack of information on precisely where service providers are located: only 82% of government health facilities were included in the national HMIS database. We have recently upgraded the Ministry of Health's service provider lists and have provided spatial coordinates for each health facility, and in this paper we provide a geostatistical model to improve the interpretation of incomplete data of presumed malaria cases reported to the centralised national HMIS database. Our model accurately predicts national annual treatment burdens for presumed outpatient malaria within the government sector with an estimated margin of error of 1.3% and a predicted average of 6.8 million cases per annum over the period of observation. This demonstrates a tangible improvement over the more traditional approach of simply multiplying nationally available data by a proportion of underreporting, which results in a crude estimate of 7.6 million cases. The incidence of malaria and the proportion of individuals with the illness who seek treatment have large spatial and temporal heterogeneity, and failing to account for this heterogeneity leads inevitably to a distortion in estimates of national treatment burden. STK is a method used in atmospheric [ 30, 31] and earth sciences [ 24, 32] that we have adapted for use in our models. It is likely to provide a more precise estimation of national treatment burdens for presumed malaria at outpatient clinics, consequently allowing a more realistic approximation of treatment requirements, including new expensive anti-malarials, in this sector. One prerequisite for STK that might limit wider application outside Kenya is that a ministry of health must have a spatially referenced map of its service providers. In Kenya, this has been made possible by the development of a geographic information system, which is applied in this paper—to our knowledge for the first time in Africa—to national HMIS data. Rather than thinking of this as a limitation to the generalisability of our approach outside Kenya, we would argue that knowing where service providers are located is a must for any health planning agency and that geographic information system frameworks for health services should be developed everywhere. The predictive power of the proposed model decreases as predictions are required at finer spatial and temporal resolutions. Although under- and overpredictions tended to balance out when areas are aggregated, errors at individual facilities were substantial in places. Thus, different models with additional parameters, including facility drainage, facility characteristics, and competition between facilities, are likely to be required to estimate incomplete data at this level [ 33– 35]. Nevertheless, the model probably performs with a margin of accuracy acceptable for health service planning at provincial and district levels, allowing for sub-national setting of priorities and resources. The model development and results presented in this study raise several important questions that require further attention. The current lag time between data being generated (patients treated at a facility) and nationwide HMIS data being available for analysis is approximately 2 y. If predictions of treatment burden are to be made current, then the modelling framework must be extended to enable predictions at times with no contemporary data. A possible approach is to integrate the nationwide HMIS data with data from a much smaller number of “sentinel” facilities, where systems are put in place to obtain reliable data on a month-by-month basis, and to use these up-to-date data to inform the prediction from the full dataset. A second question is how many of these sentinel facility sites would be needed to achieve this purpose with an acceptable level of accuracy, and how their locations might be chosen so as to optimise their utility. The Kenyan HMIS is typical of those found in many sub-Saharan African countries. Complex national health surveillance systems require substantial financial support and a motivated workforce within the health sector. In many resource-poor countries, ministries of health may be confronted with decisions between, say, buying drugs and printing HMIS forms. The quality of Kenya's HMIS is a symptom of an underfunded government sector. There is an urgent need to upgrade HMIS across Africa to provide reliable and timely data that are absolutely critical to planning and monitoring health service provision for disease-specific priorities [ 3, 14, 36, 37]. In the short term, we believe that the utility of even grossly incomplete HMIS data for planning national and sub-national needs can be greatly enhanced using appropriate statistical models. Supporting Information Figure S1 Empirical Relationship between the Size of Subsets of the Test Dataset and the Standard Deviation of Their Mean Prediction Errors Subsets of different sizes n were created from the test set by aggregating across space (by district, province, and nationally) and through time (by month and year), and the mean prediction error μ ɛ of each subset was calculated. These subsets were then placed in bins according to their size n, and the standard deviation of the mean errors in each bin, σ( μ ɛ), was calculated. The x-axis position of each point represents the mean subset size in that bin. The theoretical relationship is shown (line). The purpose of the exercise was to validate the use of this equation as a model for the effect of aggregation on the variance of prediction error. (142 KB EPS) Click here for additional data file. Protocol S1 Space–Time Kriging (27 KB DOC) Click here for additional data file. Editors' Summary Background. In order to allocate health-care resources (such as doctors, nurses, hospital beds, and drugs), public health officials need to know when and where in their country people are getting sick with which diseases. In most African countries, a country-wide health management information system (HMIS) compiles records about how many patients are being diagnosed with and treated for certain diseases. The actual data are meant to be collected and reported monthly by the individual health-care facilities. The HMIS compiles and analyzes these records, giving a picture of which patients are being treated across districts, regions, and the entire country. Ideally, all facilities report their data promptly and comprehensively every month. This allows the construction of a matrix that shows which treatments are used across the country through space (where) and time (when). However, many of the facilities operate under difficult circumstances, and keeping detailed records and reporting them every month is not always at the top of the priority list. As a result, data from many of the facilities are missing for any given month, and the overall national picture is inevitably incomplete. Why Was This Study Done? Almost any survey has to deal with some missing data, and there are various methods to estimate this missing data. Such estimates get harder the more data are missing. When it comes to reports on using health services in Africa, often more than half of the data are missing for a given month. Using sophisticated statistical methods instead of crude estimates is likely to make a big difference when such a big part of the data is missing. The researchers who did this study have adopted a statistical method called kriging to estimate missing data on health service usage. Kriging was originally developed in the earth sciences (such as geology and soil science) for estimating mineral concentrations at locations where no sampling had been done. This study was done to see whether kriging could be used to estimate the missing data on malaria cases in the Kenyan public health system. A better estimate of the missing data would be helpful for allocating malaria treatments to the right places. What Did the Researchers Do and Find? They obtained the monthly records of diagnoses made at outpatient departments of 2,165 health facilities across Kenya for an 84-month period from January 1996 to December 2002. The records included the number of outpatients and their diagnoses. The researchers chose to focus on malaria, for three reasons: (1) malaria is common (accounting for over one-third of the overall diagnoses in Kenya), (2) there is great variation in where and when it occurs across Kenya, and (3) donors are willing to provide additional support for malaria treatment and prevention but require documentation that such help is needed and reaches patients. The numbers of people diagnosed with malaria at each facility for a given month were matched to an independent database that contains information on where every health-care facility is located. Reporting rates varied from month to month and facility to facility, but the overall reporting rate was only 35%, with 25% of the facilities never reporting. The authors then adopted a version of kriging called space–time kriging to fill in missing data (space–time kriging assumes that for a given month a facility that didn't report is likely to be similar to its neighbors, and likely to be more similar to its own and its neighbors' recent numbers than to those further removed in space or time). The calculations resulted in a number of estimates. To test whether these estimates were accurate, the researchers randomly removed a test set of 10% of the monthly records from the full dataset and repeated the estimates based on the remaining 90% of reports. They found that the real and predicted cases across the country differed by less than 1%. At the district level (which is arguably the most useful for most planning purposes), the researchers found that their method can estimate 95% of the malaria cases within 35% of the true value. For 75% of the districts the estimates would be within 15% of the actual numbers. What Do These Findings Mean? In this case, space–time kriging provided a more precise estimate of missing data on diagnoses at the district and provincial levels than other estimates. This is likely to be true not just for malaria but for other diagnoses for which the number and the proportion of patients who have the disease and seek treatment vary by place and time of year. One caveat is that space–time kriging requires a detailed map of where exactly a country's health-care facilities are located. A database based on such a map existed for Kenya (and was used in this study) but doesn't exist in all countries that might benefit from a method like the one described here. The authors argue that knowledge about where health services are located is a must for any health planning agency, and that databases with that information should be developed everywhere. Additional Information. Please access these Web sites via the online version of this summary at http://dx.doi.org/10.1371/journal.pmed.0030271. • Wikipedia's page on kriging (Wikipedia is a free Internet encyclopedia that anyone can edit) • Description of, and access to, a publication entitled “Developing Health Management Information Systems: A Guide for Developing Countries” from the World Health Organization • The Health Metrics Network, a global collaboration focused on strengthening health information systems • The Roll Back Malaria Global Partnership, a multilateral initiative with the stated goal of halving the burden of malaria by 2010 • The Partnership in Statistics for Development in the 21st Century Description, a global initiative to promote a culture of evidence-based policymaking and monitoring • The Kenyan Ministry of Health
Author and article information
Journal
Journal ID (nlm-ta): BMJ Glob Health
Journal ID (iso-abbrev): BMJ Glob Health
Journal ID (hwp): bmjgh
Journal ID (publisher-id): bmjgh
Title:
BMJ Global Health
Publisher:
BMJ Publishing Group
(BMA House, Tavistock Square, London, WC1H 9JR
)
ISSN
(Electronic):
2059-7908
Publication date Collection: 2019
Publication date
(Electronic):
29
September
2019
Volume: 4
Issue: 5
Electronic Location Identifier: e001849
Affiliations
[1
]
departmentInternational Health , Johns Hopkins University Bloomberg School of Public Health , Baltimore, Maryland, USA
[2
]
departmentDepartment of Research , African Population and Health Research Center , Nairobi, Kenya
[3
]
departmentDivision of Data, Research and Policy, Data and Analytics Section , UNICEF , New York City, New York, USA
[5
]
departmentDepartment of Community Health Sciences , University of Manitoba , Winnipeg, Manitoba, Canada
[7
]
departmentMaternal, Newborn, Child and Adolescent Health Department , World Health Organization , Geneva, Switzerland
Author notes
Author information
Abdoulaye Maïga
http://orcid.org/0000-0002-0475-1106
Martin Kavao Mutua
http://orcid.org/0000-0003-1643-9934
Candy Day
http://orcid.org/0000-0002-7290-0855
Cheikh Mbacké Faye
http://orcid.org/0000-0002-4028-0575
Agbessi Amouzou
http://orcid.org/0000-0002-6262-3866
Ties Boerma
http://orcid.org/0000-0001-9922-495X
Article
Publisher ID: bmjgh-2019-001849DOI: 10.1136/bmjgh-2019-001849
PMC ID: 6768347
PubMed ID: 31637032
SO-VID: f2f3bcf5-70aa-4d8a-b2fb-3d78146c2adc
Copyright © © Author(s) (or their employer(s)) 2019. Re-use permitted under CC BY. Published by
BMJ.
License:
This is an open access article distributed in accordance with the Creative Commons Attribution 4.0 Unported (CC BY 4.0) license, which permits others to copy, redistribute, remix, transform and build upon this work for any purpose, provided the original work is properly cited, a link to the licence is given, and indication of whether changes were made. See: https://creativecommons.org/licenses/by/4.0/.
History
Date
received
: 17
July
2019
Date
revision received
: 28
August
2019
Date
accepted
: 11
September
2019
Funding
Funded by: FundRef http://dx.doi.org/10.13039/100000865, Bill and Melinda Gates Foundation;
Award ID: OPP1148933
Categories
Subject:
Analysis
Subject:
1506
Custom metadata
special-feature unlocked
Data availability: