Background: The Global Burden of Diseases, Injuries, and Risk Factors Study (GBD) 2016 provided comprehensive estimates of health loss globally. Decision makers in Kenya can use GBD subnational data to target health interventions and address county-level variation in the burden of disease. Methods: We used GBD 2016 estimates of life expectancy at birth, healthy life expectancy, all-cause and cause-specific mortality, years of life lost, years lived with disability, disability-adjusted life-years, and risk factors to analyse health by age and sex at the national and county levels in Kenya from 1990 to 2016. Findings: The national all-cause mortality rate decreased from 850·3 (95% uncertainty interval [UI] 829·8–871·1) deaths per 100 000 in 1990 to 579·0 (562·1–596·0) deaths per 100 000 in 2016. Under-5 mortality declined from 95·4 (95% UI 90·1–101·3) deaths per 1000 livebirths in 1990 to 43·4 (36·9–51·2) deaths per 1000 livebirths in 2016, and maternal mortality fell from 315·7 (242·9–399·4) deaths per 100 000 in 1990 to 257·6 (195·1–335·3) deaths per 100 000 in 2016, with steeper declines after 2006 and heterogeneously across counties. Life expectancy at birth increased by 5·4 (95% UI 3·7–7·2) years, with higher gains in females than males in all but ten counties. Unsafe water, sanitation, and handwashing, unsafe sex, and malnutrition were the leading national risk factors in 2016. Interpretation: Health outcomes have improved in Kenya since 2006. The burden of communicable diseases decreased but continues to predominate the total disease burden in 2016, whereas the non-communicable disease burden increased. Health gains varied strikingly across counties, indicating targeted approaches for health policy are necessary. Funding: Bill & Melinda Gates Foundation.
GBD is the most comprehensive epidemiological study to systematically gather, analyse, and produce comparable estimates of health loss and related risk factors across locations, age groups, and sex categories. The methods used in this study have been described extensively elsewhere.7, 26, 27, 28, 29, 31 Briefly, GBD identifies data sources for different aspects of the estimation process. The quality of these data is assessed systematically and corrected for known biases. Subsequently, data are subject to standardised statistical estimation and cross-validation analyses to assess model performance. All estimates produced for GBD report 95% uncertainty intervals (UIs) that account for sampling and non-sampling error associated with data and various assumptions of the modelling process and are derived from the 2·5th and 97·5th percentiles of 1000 draws. UIs incorporate sample size variability within data sources, different availability of data by age, sex, year, or location, and cause-specific model specifications, but do not currently propagate uncertainty from covariates or so-called garbage code correction.7, 26, 27, 28, 29 Unless otherwise specified, we report all estimates in terms of age-standardised rates, standardised using GBD world population standards. Our Article complies with the Guidelines for Accurate and Transparent Health Estimates Reporting (GATHER).32 Complete information on data sources used in GBD is available from the Global Health Data Exchange. Related statistical code for Python versions 2.5.4 and 2.7.3, Stata version 13.1, or R version 3.1.2 is available at GitHub. Results for all locations and years are available to explore using data visualisations from the Institute for Health Metrics and Evaluation (IHME). The GBD framework classifies causes of health loss into mutually exclusive and collectively exhaustive categories organised in a four-level hierarchy. GBD 2016 included 333 causes of disease and disability. Causes of health loss are first organised into three primary categories: communicable, maternal, neonatal, and nutritional (CMNN) disorders; NCDs; and injuries. These broad categories are divided further into increasingly more detailed categories in a consistent and comprehensive manner. Standard estimates for different causes of health loss are produced for different sexes, age groups, and locations, enabling useful comparisons. GBD uses various inter-related metrics to measure population health loss, including number of deaths and mortality rates, years of life lost due to premature death (YLLs), years of life lived with disability (YLDs), disability-adjusted life-years (DALYs), life expectancy, and healthy life expectancy (HALE). GBD 2016 synthesised many Kenyan input data sources for mortality (appendix pp 45–76), morbidity (appendix pp 77–120), and risk factors (appendix pp 121–26). These sources included population censuses, surveys, sample registration systems, and registries, in addition to published research from various sources. We used localised covariates to customise the estimation for each location for each custom cause or sex model. We estimated all metrics nationally and individually for each county of Kenya from 1990 to 2016. We chose the years 1990, 2006, and 2016 as primary comparators for this Article, to provide both a historical perspective and a recent assessment that would be more pertinent to policy-making efforts and incorporate recent public health efforts.33, 34 To account for the reorganisation of Kenyan districts and provinces (the 2010 constitution split provinces into counties), all previous boundaries from 1990 onwards were translated onto the 47 county boundaries from 2016. We then recalculated previous data to conform to this rebuilt map. Briefly, we estimated all-cause under-5 and adult mortality using updated demographic methods that have been developed for GBD. These methods encompass a multistage process that includes various data sources systematically—eg, surveys, censuses, vital registration systems, birth histories, sibling histories, and household death recall. After correction for biases from different data sources, we generated estimates of under-5 and adult mortality using a combination of spatiotemporal and Gaussian process regressions using covariates (appendix pp 1–44) to help inform data sparse estimation.28 For cause-specific mortality, we standardised various data sources—primarily vital registration, verbal autopsy, surveys, and surveillance—and mapped these to GBD cause of death list. Data attributed to causes that could not be underlying causes of death, so-called garbage codes, were redistributed using standard algorithms. The Cause of Death Ensemble model (CODEm) uses country-level covariates and combines and tests various models to provide the most robust estimates for most causes of death.7 Individual cause models are then combined and corrected to be internally consistent with estimates of all-cause mortality using the cause of death correction process, CoDCorrect, with uncertainty calculated for detailed causes, aggregate levels, and all-cause mortality.7, 28 We obtained YLLs by multiplying the number of deaths from each cause in each age group by the reference life expectancy. The reference life table used in GBD estimation is derived from the highest life expectancy observed, currently 86·6 years at birth in Japan.7, 28 We used various data sources that capture information on non-fatal outcomes—including published studies, surveillance data, and hospital data—to generate consistent estimates for incidence, prevalence, remission (where feasible), and mortality attributable to different causes, using a Bayesian meta-regression tool, DisMod-MR 2.0. We obtained YLDs by multiplying the prevalence of different sequelae attached to causes with the disability weight attributed to the health state for that sequela. These disability weights were derived consistently from large population surveys across different and representative parts of the world and have been updated continuously through internet-based surveys, making them the most valid measure of disease severity available.27, 35 The sum of YLLs and YLDs yields DALYs—a measure of overall health loss.29 DALYs combine the health effects of both fatal and non-fatal conditions, providing a common currency to facilitate useful comparisons across different causes of health loss. We calculated HALE using multiple-decrement life tables and YLDs by age, sex, location, and year.29 Another useful feature of GBD is the systematic quantification of health loss attributable to priority risk factors selected largely based on relevance to policy making and the availability of data to facilitate valid estimation.26 To estimate risk factors, meta-analyses of published research are used in which the relative risk of a specific risk factor for mortality or morbidity was calculated. Subsequently, the distribution of the same risk factor of interest is established across different locations, sexes, and age groups, as defined in GBD systematic hierarchy. We used counterfactual analysis to ascertain the health loss attributable to a specific risk factor by comparing the observed distribution of the risk factor in question with its theoretical minimum risk possible at the population level.26 Similar to the causes of health loss, risk factors are organised into a hierarchical structure. Three broad categories of behavioural, metabolic, and environmental and occupational risk factors are divided further into more detailed categories to facilitate useful comparisons. At all levels of aggregation, different ways of combining the effects of various risk factors are espoused, since their combined effects on health loss could be independent, joint, or mediated through a different factor that has to be taken into account.26 A GBD comparative risk assessment framework is used to estimate exposure, attributable deaths, and attributable DALYs by age, sex, and location-year.36 Data for relative risks and exposure come mainly from published studies, surveys, and censuses that meet quality criteria. We applied the standard GBD subnational estimation process to produce county-level estimates in Kenya. In short, subnational locations undergo the same estimation process as is used for country-level estimates. In locations with scant data, we used covariate estimation to borrow strength from similar subnational, country, region, and super-region models. The same spatiotemporal Gaussian process regression and ensemble modelling processes are used for subnational locations as for country-level modelling. If subnational locations are estimated, national estimations are based on the sum of the subnational values (the subnational values are not derived from a precalculated national value). The Socio-demographic Index (SDI) is a summary indicator calculated as the mean of the scaled values of total population fertility, educational attainment in the population older than 15 years, and per-capita income.7, 28 The SDI for Kenya in 2016 was 0·52 on a theoretical scale of 0–1.7 We used Gaussian regression methods to estimate the relation between SDI and each age, sex, cause, and health measure, then we used these relations to estimate expected values based on SDI alone for each age, sex, location, and year. The funder had no role in study design, data collection, data analysis, data interpretation, or writing of the report. All authors had full access to study data and had final responsibility for the decision to submit for publication.