Global, regional, and national burden of tuberculosis, 1990–2016: results from the Global Burden of Diseases, Injuries, and Risk Factors 2016 Study

Summary Background Although a preventable and treatable disease, tuberculosis causes more than a million deaths each year. As countries work towards achieving the Sustainable Development Goal (SDG) target to end the tuberculosis epidemic by 2030, robust assessments of the levels and trends of the burden of tuberculosis are crucial to inform policy and programme decision making. We assessed the levels and trends in the fatal and non-fatal burden of tuberculosis by drug resistance and HIV status for 195 countries and territories from 1990 to 2016. Methods We analysed 15 943 site-years of vital registration data, 1710 site-years of verbal autopsy data, 764 site-years of sample-based vital registration data, and 361 site-years of mortality surveillance data to estimate mortality due to tuberculosis using the Cause of Death Ensemble model. We analysed all available data sources, including annual case notifications, prevalence surveys, population-based tuberculin surveys, and estimated tuberculosis cause-specific mortality to generate internally consistent estimates of incidence, prevalence, and mortality using DisMod-MR 2.1, a Bayesian meta-regression tool. We assessed how the burden of tuberculosis differed from the burden predicted by the Socio-demographic Index (SDI), a composite indicator of income per capita, average years of schooling, and total fertility rate. Findings Globally in 2016, among HIV-negative individuals, the number of incident cases of tuberculosis was 9·02 million (95% uncertainty interval [UI] 8·05–10·16) and the number of tuberculosis deaths was 1·21 million (1·16–1·27). Among HIV-positive individuals, the number of incident cases was 1·40 million (1·01–1·89) and the number of tuberculosis deaths was 0·24 million (0·16–0·31). Globally, among HIV-negative individuals the age-standardised incidence of tuberculosis decreased annually at a slower rate (–1·3% [–1·5 to −1·2]) than mortality did (–4·5% [–5·0 to −4·1]) from 2006 to 2016. Among HIV-positive individuals during the same period, the rate of change in annualised age-standardised incidence was −4·0% (–4·5 to −3·7) and mortality was −8·9% (–9·5 to −8·4). Several regions had higher rates of age-standardised incidence and mortality than expected on the basis of their SDI levels in 2016. For drug-susceptible tuberculosis, the highest observed-to-expected ratios were in southern sub-Saharan Africa (13·7 for incidence and 14·9 for mortality), and the lowest ratios were in high-income North America (0·4 for incidence) and Oceania (0·3 for mortality). For multidrug-resistant tuberculosis, eastern Europe had the highest observed-to-expected ratios (67·3 for incidence and 73·0 for mortality), and high-income North America had the lowest ratios (0·4 for incidence and 0·5 for mortality). Interpretation If current trends in tuberculosis incidence continue, few countries are likely to meet the SDG target to end the tuberculosis epidemic by 2030. Progress needs to be accelerated by improving the quality of and access to tuberculosis diagnosis and care, by developing new tools, scaling up interventions to prevent risk factors for tuberculosis, and integrating control programmes for tuberculosis and HIV. Funding Bill & Melinda Gates Foundation.


Introduction
Although tuberculosis is a preventable and treatable disease, it is the cause of more than a million deaths each year. 1,2 Tuberculosis was the leading cause of death from a single infectious pathogen in 2016. 1 The ambitious Sustainable Development Goal (SDG) target 3 aims to end the tuberculosis epidemic by 2030, and numeri cal milestones (eg, annual reduction in global tuber culosis incidence of 10% by 2025) have been set to achieve this target. 3 Robust assessment, monitoring, and evaluation of progress towards this SDG target are therefore crucial to inform policy and programme decision making.
Accurately assessing the tuberculosis burden over time is difficult because of the paucity of highquality data from many lowincome and middleincome coun tries. 2 The completeness of vital registration data is gradually improving, but many countries still do not have goodquality vital registration systems. 1 Notification data can be of use as a proxy for tuberculosis incidence in countries with highquality health and surveillance systems where underreporting is minimal; 4 however, in most lowincome and middleincome countries, these data are prone to underreporting and cannot be interpreted without additional information on case detection rate. 4,5 To deal with the lack of highquality data in these countries, various methods have been used to estimate tuberculosis incidence (eg, adjusting for under reporting in notification data by use of expert opinion case detection rates, 4 backcalculating incidence from prevalence survey data by use of different assumptions of the average duration of disease, 4 or using a statistical triangulation approach 2,6 ). For the Global Burden of Diseases, Injury, and Risk Factors Study (GBD) 2015, we used a statistical triangulation approach that modelled tuberculosis incidence, prevalence, and mortality simul taneously to generate consistent estimates for these parameters. 2 The burden of tuberculosis varies by several factors including age, sex, location, HIV status, and drugresistance status. Therefore, these factors should be taken into account when investigating tuberculosis trends. Addi tionally, the burden of disease in many countries has shifted from communicable to noncommunicable dis eases in line with sociodemographic development (the epidemiological transition). 7-9 As such, comparing the observed tuberculosis burden to that expected on the basis of a country's sociodemographic level could be useful for guiding investment in research and interventions. 2 For example, countries with a lower tuberculosis burden than expected relative to their sociodemographic development could provide insight into successful programmatic strategies, and countries with a higher burden than expected might need to investigate the reasons why. GBD 2015 2 examined the difference between the observed and expected burden of tuberculosis but did not provide a detailed assessment by drugresistance type and HIV status. For GBD 2016, we assessed the levels and trends in the fatal and nonfatal burden of tuberculosis by drug resistance type and HIV status from 1990 to 2016, for 195 countries and territories. We also aimed to analyse the association between these burdens and the country or territory's Sociodemographic Index (SDI), 1,10-12 which is a composite indicator of income, education, and fertility rate.

Methods
Overview GBD is a systematic, scientific effort to quantify the comparative magnitude of health loss due to diseases, injuries, and risk factors by age, sex, and location over time. The conceptual and analytical framework for GBD and detailed methods have been published elsewhere. 1, 11,13 We describe here the methods we used for the analysis of the burden of tuberculosis for GBD 2016.

Selection of input data
The input data we used to model mortality due to tuberculosis among HIVnegative individuals included

Research in context
Evidence before this study Tuberculosis causes more than a million deaths each year and was the leading cause of death from a single infectious pathogen in 2016. The global burden of tuberculosis has been estimated by several groups, including the WHO Global Tuberculosis Programme and the Global Burden of Diseases, Injuries, and Risk Factors Study (GBD) 2015. Nevertheless, trends in the burden of drug-resistant tuberculosis by HIV status and how the observed burdens differ from the levels expected on the basis of sociodemographic development have not been comprehensively assessed. We searched PubMed with the search terms ("tuberculosis"[MeSH] AND "drug-sensitive" OR "drug-susceptible") OR "tuberculosis, multidrug-resistant"[MeSH] AND ("burden" OR "estimates") AND "trend", with no language restrictions, for publications up to June 7, 2018. We identified ten studies that provided population-based time trends for the burden of multidrug-resistant tuberculosis (incidence, prevalence, or deaths). Of these studies, the most recent period assessed was 1999-2013 in Lebanon. None of these studies assessed the trends in the burden of drug-susceptible or multidrug-resistant tuberculosis by HIV status and compared these burdens with those expected on the basis of a country's socio-demographic position.

Added value of this study
We found that, although HIV infection and drug-resistant tuberculosis have become the main challenges to tuberculosis control efforts, more than three-quarters of incident cases of tuberculosis and deaths due to tuberculosis in 2016 were estimated to occur in HIV-negative individuals who were susceptible to first-line tuberculosis drugs. During the past decade, the global rate of decline for incidence of both drug-susceptible and multidrug-resistant tuberculosis was slower than the corresponding rate of decline for mortality, for HIV-positive and HIV-negative individuals alike. Many countries had higher burdens of drug-susceptible or multidrug-resistant tuberculosis than expected on the basis of their level of socio-demographic development.

Implications of all the available evidence
If current trends in tuberculosis incidence continue, few countries will meet the Sustainable Development Goal target to end the tuberculosis epidemic by 2030. The pace of progress needs to be increased through interventions including improving the quality of and access to tuberculosis diagnosis and care, and integrating control programmes for tuberculosis and HIV.
15 943 siteyears of vital registration data, 1710 siteyears of verbal autopsy data, 764 siteyears of samplebased vital registration data, and 361 siteyears of mortality surveillance data. We assessed and improved the quality and comparability of data on cause of death through multiple steps, 12 including redistribution of garbage codes to underlying causes of death using GBD algorithms and adjusting for misclassified HIV deaths (ie, deaths caused by HIV being assigned to other underlying causes of death, such as tuberculosis, because of stigma or misdiagnosis). GBD 2016 1 also assessed the overall quality of data for each country (on the basis of completeness, garbage coding, detail of cause list, and time periods covered), and assigned a quality score from zero stars (lowest) to five stars (highest); a score of four to five is considered high quality (quality scores by country are in the appendix p 19). We removed verbal autopsy data for countries with a high prevalence of HIV (using an arbitrary cutoff value of 5% agestandardised pre valence of HIV), because verbal autopsy studies have a poor ability to distinguish deaths due to HIV from deaths due to tuberculosis among people who are HIV positive (HIVtuberculosis deaths). 2 Our input data for the estimation of mortality due to HIVtuberculosis included 382 siteyears of highquality vitalregistration data from countries where data on cause of death directly coded for HIVtuberculosis and tuber culosis were available, and the number of tuberculosis cases (new and retreatment) recorded as HIVpositive, and the number of tuberculosis cases (new and re treatment) with an HIV test result recorded in the WHO tuberculosis register.
In GBD 2016, we included multidrugresistant tuber culosis (without extensive drug resistance) and ex tensively drugresistant tuberculosis by HIV status as new outcomes (case definitions are in the appendix p 3). Input data included the number of cases of tuberculosis that were multidrug resistant, extensively drug resistant, had a drugsensitivity testing result for isoniazid and rifampicin, and that were multidrug resistant with a drugsensitivity result for secondline drugs from routine surveillance and surveys reported to WHO (for data availability by country see appendix p 16); relative risks of mortality for cases of tuberculosis that were multidrug resistant compared with cases that were drug susceptible, and relative risks for cases that were extensively drug resistant compared with multidrug resistant were extracted from studies identified via a systematic review (for details of systematic review see appendix p 37); and the risk of multidrugresistant tuberculosis associated with HIV infection extracted from a metaanalysis. 14 Our input data for modelling nonfatal tuberculosis included annual case notification data, data from pre valence surveys of tuberculosis, data from population based tuberculin surveys, and estimated causespecific mortality rates of tuberculosis among individuals who were HIV positive and HIV negative. Links to data sources and code we used in analyses are in the appendix (pp [40][41].

Fatal tuberculosis
We modelled tuberculosis mortality among people who are HIV negative using the Cause of Death Ensemble modelling (CODEm) strategy, [15][16][17][18] which evaluates a large number of potential models that apply different functional forms (mixedeffects models and spatiotemporal Gaussian process regression models) to mortality or cause fractions with varying combinations of predictive covariates. These covariates included alcohol (L per capita), diabetes (fasting plasma glucose in mmol/L), education (years per capita), lagdistributed income (LDI), indoor air pollution, outdoor air pollution, population density (people per km²), smoking prevalence, SDI, the summary exposure variable scalar (which indicates exposure to risk factors associated with tuberculosis; appendix p 9), and four new covariates added for GBD 2016 (ie, prevalence of tuberculosis, prevalence of latent tuberculosis infection, proportion of adults who are underweight, and the Healthcare Access and Quality [HAQ] Index 19 ). We then selected the ensemble of CODEm models that performed best on outofsample predict ive validity tests (appendix pp 20-23). We estimated HIVtuberculosis mortality using a populationattributable fraction approach, like in GBD 2015 2 (detailed methods and equations are in the appendix pp [34][35][36].
To split tuberculosis deaths and HIVtuberculosis deaths by drugresistance type, we first estimated the proportions of tuberculosis cases that were multidrug resistant for all locations and years using a spatiotemporal Gaussian process regression. Second, we estimated the proportions of tuberculosis cases that were multidrug resistant by HIV status on the basis of the risk of multidrugresistant tuberculosis associated with HIV from a metaanalysis by Mesfin and colleagues. 14 Third, we used the estimated proportions of cases of tuberculosis that are multidrug resistant by HIV status and the relative risk of death in multidrugresistant cases compared with drugsusceptible cases to calculate the fraction of tuberculosis deaths among HIVnegative individuals attributable to multidrugresistant tuberculosis, and the fraction of HIVtuberculosis deaths attributable to multidrugresistant tuberculosis (detailed methods and equations are in the appendix pp [23][24][35][36]. Finally, we applied the fraction of tuberculosis deaths attributable to multidrugresistant tuberculosis to the number of tuberculosis deaths we estimated using CODEm, and the fraction of HIVtuberculosis deaths attributable to multidrugresistant tuberculosis to our estimated number of HIVtuberculosis deaths, to generate the number of multidrugresistant tuberculosis deaths by HIV status by location, year, age, and sex. To distinguish extensively drugresistant tuberculosis from multidrugresistant tuberculosis, we aggregated the cases of extensively drugresistant tuberculosis and multidrugresistant tuberculosis (with drugsensitivity testing for secondline drugs) up to the GBD superregion See Online for appendix level (for analytical purposes we grouped 21 GBD regions into seven superregions: 13 central Europe, eastern Europe and central Asia; highincome; Latin America and Caribbean; north Africa and Middle East; south Asia; southeast Asia, east Asia, and Oceania; and subSaharan Africa) and calculated the proportion of cases of extensively drugresistant tuberculosis among the cases of multidrug resistant tuberculosis at the superregion level. We then used these proportions and the relative risk of mortality among people with extensively drugresistant tuberculosis compared with those with multidrugresistant tuberculosis to calculate the fraction of extensively drugresistant tuberculosis deaths among all multidrugresistant tuber culosis deaths at the superregion level (detailed methods and equations are in the appendix p 24). These fractions were then applied to the estimated number of multidrug resistant tuberculosis deaths and multidrugresistant HIVtuberculosis deaths in countries within the super regions to calculate the number of deaths due to extensively drugresistant tuberculosis by HIV status by location, year, age, and sex.
We linearly extrapolated mortality for extensively drug resistant tuberculosis back from 2016 assuming mortality was zero in 1992, 1 year before extensively drugresistant tuberculosis was first recorded in USA surveillance data in 1993. 20 Next, we subtracted the number of deaths due to extensively drugresistant tuberculosis from the number of deaths due to multidrugresistant tuberculosis to generate the number of deaths due to multidrugresistant tuberculosis (without extensive drug resistance) by loca tion, year, age, and sex.

Non-fatal tuberculosis
We made several improvements to the statistical trian gulation approach we used in GBD 2015 2 to model non fatal tuberculosis. First, we estimated the prevalence of latent tuberculosis infection by location, year, age, and sex using data from populationbased tuberculin surveys and cohort studies that reported the risk of developing active tuberculosis disease as a function of indura tion size. 11 Next, we divided the inputs on prevalence (from tuberculosis prevalence surveys in lowincome and middleincome countries), incidence (notification data from countries with a rating of four or five stars and estimated incidence from countries with ratings of zero to three stars), and causespecific mortality rate by the riskweighted prevalence of latent tuberculosis infection to model tuberculosis among individuals at risk in each country. A detailed explanation of how we prepared each of these data sources is in the appendix (pp 6-10).
To generate initial estimates of incidence for countries with a rating of zero to three stars, we did a regres sion analysis using mortalitytoincidence ratios (logit transformed) from locations with a rating of four or five stars as input data, with SDI as a covariate. We calibrated the lowest end of the SDI scale with a datapoint from a communitybased cohort study, 21 which reported that 49·2% of people with untreated pulmonary tuberculosis had died at the end of a 5 year followup period, to predict mortalitytoincidence ratios as a function of SDI for all locations and years. We then used the predicted mortalitytoincidence ratios and estimates of causespecific mortality to calculate the agesex specific incidence input for modelling in DisModMR 2.1, 22 the GBD Bayesian metaregression tool. In locations where our estimated mortalitytoincidence ratios were greater than notificationbased mortalitytoincidence ratios, we used the notificationbased ratios to calculate the incidence input. We then generated a final incidence estimate that is consistent with prevalence data and causespecific mortality estimates using a Bayesian metaregression.
We used DisModMR 2.1 to simultaneously model age sex specific tuberculosis incidence, prevalence, and mortality among the population who are latently infected and generate consistent trends in all parameters. We then multiplied the DisModMR 2.1 outputs by the prevalence of latent tuberculosis infection to get populationlevel estimates of incidence and prevalence. To distinguish HIV tuberculosis from all forms of tuberculosis, we applied the proportion of cases of HIVtuberculosis among all cases of tuberculosis (estimated from a mixedeffects regression using the adult HIV mortality rate covariate as in GBD 2015 2 ) to the number of incident and prevalent cases of tuberculosis. We then applied the estimated proportion of cases of tuberculosis that are multidrug resistant to our predicted number of cases of tuberculosis, and the estimated proportion of cases of HIVtuberculosis with multidrugresistant tuberculosis (as described earlier for fatal tuberculosis) to our predicted number of HIV tuberculosis cases, to generate the number of cases of multidrugresistant tuberculosis by HIV status. To distinguish extensively drugresistant tuberculosis from multidrugresistant tuberculosis, we calculated the pro portions of cases of extensively drugresistant tuber culosis among the cases of multidrugresistant tuberculosis at the superregion level and applied these proportions to multidrugresistant tuberculosis cases.
Similar to our estimation for fatal tuberculosis with extensive drug resistance, we linearly extrapolated the prevalence and incidence of extensively drugresistant tuberculosis back from 2016, assuming incidence and prevalence were zero in 1992 and in earlier years. Finally, we subtracted the number of cases of extensively drug resistant tuberculosis from the number of cases of multidrugresistant tuberculosis to generate the number of cases of multidrugresistant tuberculosis (without extensive drug resistance) by location, year, age, and sex. We used the GBD world population age standard to calculate agestandardised rates.

SDI
SDI, initially developed for GBD 2015 9 and updated for GBD 2016, 1,10,11 was calculated on the basis of the geometric mean of three indicators: income per capita, average years of schooling, and total fertility rates. SDI scores were scaled from 0 (lowest income, lowest average years of schooling, highest fertility) to 1 (highest income, highest average years of schooling, lowest fertility), and each location was assigned an SDI score for each year. We estimated the average association between SDI and tuberculosis incidence and mortality using a Gaussian process regression, and we then used these associations to estimate expected values at each SDI level.

Role of the funding source
The sponsor of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.
Agestandardised rates of tuberculosis mortality among HIVnegative individuals decreased at varying rates across countries and territories from 2006 to 2016, with the highest annual decreases seen in Kazakhstan (-13·7%   Trends in the annualised rate of change in age standardised incidence and mortality for multidrug resistant tuberculosis varied largely across countries, with no consistent pattern worldwide for HIVnegative (appendix pp 48-74) or for HIVpositive individuals (pp 75, 88). Among HIVnegative individuals, Kyrgyzstan, Lesotho, Namibia, Somalia, Swaziland (eSwatini), and Turkmenistan had the highest agestandardised inci dence of multidrugresistant tuberculosis (ie, >20 per 100 000 population) in 2016, whereas agestandardised mortality for multidrugresistant tuberculosis was highest (ie, >15 per 100 000 population) in Somalia, Lesotho, eSwatini, and Afghanistan in the same year (appendix pp 108-14). More detailed results for HIVnegative individuals broken down by age, sex, and year, and data for HIV positive individuals, are available online.

Observed versus expected tuberculosis burden
In 2016, among HIVnegative individuals, several regions (eg, eastern Europe, central Asia, southeast Asia, south Asia, and subSaharan Africa) had higher than expected (on the basis of SDI) agestandardised incidence and mortality of drugsusceptible tuberculosis ( figure 4). At the regional level, the highest observedtoexpected ratios were in southern subSaharan Africa (13·7 for incidence and 14·9 for mortality), and the lowest ratios were in highincome North America (0·4 for incidence) and Oceania (0·3 for mortality). For multidrugresistant tuberculosis, eastern Europe had the highest observedto expected ratios for incidence (67·3) and mortality (73·0), and highincome North America had the lowest ratios (0·4 for incidence and 0·5 for mortality). We found no association between SDI and incidence of or mortality due to extensively drugresistant tuberculosis (data not shown).
In 2016, among HIVnegative individuals, observedto expected ratios were greater than two for incidence of drugsusceptible tuberculosis in 54 countries, for drugsusceptible tuberculosis mortality in 38 countries, for incidence of multidrugresistant tuberculosis in 83 countries, and for multidrugresistant tuberculosis mortality in 96 countries (figure 4). These countries were located mainly in subSaharan Africa and southeast Asia. For HIVpositive individuals in 2016, observedtoexpected ratios were greater than two in 138 countries for inci dence of drugsusceptible tuberculosis, 82 countries for drugsusceptible tuberculosis mortality, 105 countries for incidence of multidrugresistant tuberculosis, and 95 countries for multidrugresistant tuberculosis mortality. Most of these countries were in subSaharan Africa and eastern Europe. Across tuberculosis drugresistance types and by HIVstatus, the highest observedtoexpected ratios were 17·9 for the incidence of drugsusceptible tuberculosis in South Africa to 3188·6 for the incidence of multidrug resistant HIVtuberculosis in eSwatini.
We estimated that the incidence of tuberculosis among HIVnegative individuals has decreased by only 1·3% (95% UI 1·2-1·5) annually during 2006-16. This rate is much lower than the 10% or more annual reduction needed by 2025 to reach the SDG target to end the tuberculosis epidemic by 2030. 3 We identified the countries with the fastest and slowest improvements in tuberculosis incidence during 2006-16. The fastest annual decrease  Figure 4: Ratio of observed to expected age-standardised incidence and mortality on the basis of SDI by GBD region and country in 2016, for drug-susceptible tuberculosis and multidrug-resistant tuberculosis, by HIV status Ratio of observed age-standardised incidence or mortality to that expected on the basis of a country's SDI for a given year. A ratio of one means that observed and expected values are equal. A ratio higher than one means the observed rate is greater than expected, and a ratio of less than one means the observed rate is lower than expected. GBD=Global Burden of Disease. Multidrug-resistant tuberculosis=multidrug-resistant tuberculosis without extensive drug resistance. SDI=Socio-demographic Index.
in incidence was observed in Kazakhstan, where im provements were attributable to advances in diagnostics and effective treatment of newly diagnosed tuberculosis cases. 23 We saw little to no improvement in some countries, including the Philippines and Uruguay. In the Philippines, a high proportion of smearnegative individuals who are positive for tuberculosis by Xpert MTB/RIF assay (Cepheid, USA) has been documented in highrisk populations, including prison inmates and indigenous populations, suggesting that sputumsmear microscopy alone as a routine diagnostic test is inadequate. 24 In Uruguay, a decrease in treatment success rate for new cases of tuberculosis (from 84% in 2010 to 77% in 2015) 25 probably contributes to the country's the lack of progress. Despite improvements in sociodemographic conditions, several regions have fallen behind in their progress to reduce the burden of tuberculosis. In 2016, most countries in Asia, subSaharan Africa, and eastern Europe had a higher burden of tuberculosis (both drug susceptible and multidrug resistant) than expected given their level of sociodemographic development. In many countries, providing treatment services for multidrugresistant tuberculosis remains a challenge, partly because of the high cost of secondline drugs and poor adherence to regimens. 26,27 Globally in 2016, only 22% of people with newly diagnosed drugresistant tuberculosis were estimated to begin treatment, with a treatment success rate of 54%. 28 Evidence suggests that alcohol abuse and HIV infection are associated with increased risk of unsuccessful outcomes in patients with multidrug resistant tuberculosis, but the association is unclear for other factors and comorbidities such as smoking and chronic kidney disease. 29 A more comprehensive under standing of the key drivers of unsuccessful treatment outcomes in these patients is crucial to improving their treatment outcomes.
Although the burden of tuberculosis remains far off from the expected level in many countries with a high burden of tuberculosis, the prevalence of diabetes-an important risk factor for both tuberculosis and adverse outcomes from tuberculosis treatment-has increased over time worldwide as a consequence of several factors including population ageing and exposure to lifestyle related risk factors, 11,30 creating additional challenges for tuberculosis care and prevention. Because of the interaction of tuberculosis with diabetes and HIV/AIDS, integrating control programmes for the three diseases could help prevent tuberculosis among people with HIV/AIDS and diabetes and reduce the burden of all three diseases. 31 Additionally, efforts to prevent other risk factors for tuberculosis, including smoking and alcohol misuse, could have a complementary effect on the burden of tuberculosis. 2 In countries with a higher burden of tuberculosis than expected on the basis of SDI, an important first step is to identify the reasons for falling behind so that appropriate measures can be taken. Gaps in case detection and delays in diagnosis and treatment most likely contribute to the burden being higher than expected, and countryspecific reasons should also be investigated. Although national tuberculosis programmes were notified of about 6·3 million new or relapsed cases of tuberculosis globally in 2016, 28 we estimated that the number of incident cases in 2016 was 10·4 million, implying a global case detection rate of 61%. Evidence suggests that a routine passive case finding strategy is insufficient for detecting all tuberculosis cases. 32,33 Active case finding has been recommended as a complementary strategy to passive case finding to increase case detection; nevertheless, the effect of active case finding on treatment outcomes and rate of transmission, and longerterm effects on the epidemiology of tuberculosis, have yet to be determined. 32 Also, despite advances in tuberculosis diagnostics, smear microscopy remains the most commonly used diagnostic test in many countries that are endemic for tuberculosis. 34 The Xpert MTB/RIF assay has higher sensitivity for the detection of tuberculosis than smear microscopy, 35 but few countries use Xpert for general tuberculosis case finding. 36 Policies on the use of Xpert MTB/RIF vary largely between countries, with only a subset of patients with tuberculosis being eligible for the test (eg, patients with suspected drug resistance, HIVpositive individuals). 36 A scaleup of Xpert MTB/RIF could help in detecting additional cases, but it has been impeded by several factors, including high costs, reliance on funding from international donors, and the lack of subsidised pricing in the private sector, which is relied on for most tuberculosis cases in some countries. 26 Overall, even with differences in methods used, both GBD 2016 and WHO estimated 10·4 million incident cases of tuberculosis in 2016, although our estimated number of all tuberculosis deaths (1·45 million) is lower than WHO's estimate (1·7 million) for 2016. 28 The 20 countries with the highest burden, as assessed by the number of incident cases, differ between our estimates and WHO's: WHO includes Angola, Brazil, and Thailand; instead, we include Uganda, Zambia, and Zimbabwe. The most notable difference between our and WHO's estimates is between the estimated numbers of tuberculosis deaths among children. We estimated 39 311 deaths (95% UI 34 415-44 847) among children who are HIV negative and younger than 15 years for 2016, which is substantially lower than the estimates from WHO (201 000 deaths) 28 and Dodd and colleagues (200 000 deaths). 37 The input data and the methods used to generate estimates of deaths in children due to tuberculosis are very different between studies: we used vital registration and verbal autopsy data and the CODEm strategy to estimate tuberculosis deaths in children and WHO used the method of Dodd and colleagues from their 2017 study 37 in which child mortality due to tuberculosis was backcalculated from the incidence and casefatality ratio. WHO estimated the incidence of tuber culosis in children by combining results from two app roaches: the case detection rate adjustment approach (ie, incidence=notifications/estimated case detection rate); and the method of Dodd and colleagues from their 2014 study 38 in which incidence was estimated from the annual risk of infection in children, WHO adult smearpositive tuberculosis prevalence data, and demographic informa tion by use of a mathematical model. Both our method and the method used by WHO and Dodd and colleagues have limitations. Specifically, concerns have been raised about the misclassification of tuberculosis deaths in children as deaths due to pneumonia in countries with a high burden of tuberculosis. 39 In this study, we did not redistribute pneumonia deaths to tuberculosis deaths because of a lack of evidence on whether tuberculosis is a cause or comorbidity of acute severe pneumonia in children. 40 The backcalculation approach used by WHO and Dodd and colleagues most likely has substantial uncertainty due to assumptions in the process of estimating annual risk of infection, the prevalence of adult tuberculosis, and case detection rates.
Our study has several limitations. First, our assessment of the trends in the burden of multidrugresistant tuberculosis was restricted by a paucity of timeseries data for many countries in Asia and Africa. We assumed that these countries have a similar agesex distribution of multidrugresistant tuberculosis to other countries in the same region and used this common distribution to generate trend estimates for countries and years with little data; the lack of data in a particular country is reflected in wide uncertainty intervals. Second, verbal autopsy studies have modest sensitivity in identifying tuberculosis deaths. [41][42][43] However, at the typical range of the cause fraction of deaths due to tuberculosis in India and subSaharan Africa (3-5%), 43 and at the reported level of sensitivity and specificity of attributing tuberculosis as the cause of death in a large, multicentre, verbal autopsy validation study, 43 we estimate that the false positives and false negatives largely cancel out. Third, as noted in our previous publication, 2 the main challenge in our statistical triangulation approach has been the shortage of data from surveys on cause of death and prevalence, particularly from countries in subSaharan Africa with a high prevalence of HIV. We applied sophisticated modelling methods and covariates, using distributions across geo graphies and time to help predict for those locations. Accordingly, the estimates for a location with sparse data are coupled with wider uncertainty intervals. Fourth, to inform the casefatality ratio among patients with untreated tuberculosis for our mortalitytoincidence ratio regression, we used data from a single communitybased followup study done in Bangalore, India; 21 two other communitybased studies, 44 done in India 45 and the USA, 46,47 were not included because of a lack of information about the treatment of tuberculosis or any systematic followup of cases.
Despite these limitations, we made several improve ments in our methods compared with GBD 2015. First, we no longer used casedetection rates based on expert opinion in the process of estimating the incidence of tuberculosis. Instead, we used a mortalitytoincidence ratio approach to better reflect higher mortality and incidence in lowincome and middleincome countries. Second, we strengthened our statistical triangulation approach by incorporating populationbased surveys of latent tuberculosis infection, and modelling incidence, prevalence, and mortality simultan eously among the population who are latently infected to enhance comparability across countries. Because we used Bayesian metaregression to generate an incidence estimate that is consistent with prevalence data or cause specific mortality estimates, our estimated incidence might differ from countries' official statistics (even from those with a four star or five star quality ratings). Third, we improved our estimates of tuber culosis mortality by including additional covariates that have proximal or strong associations with tuberculosis mortality (ie, prevalence of latent tuber culosis infection, prevalence of active tuberculosis disease, proportion of adults who are underweight, and HAQ Index). These improvements, together with substantial efforts to collate data for the estimation of tuberculosis burden, have resulted in changes in GBD 2016 compared with GBD 2015, especially in estimates of mortality. The global number of deaths in 2016 due to tuberculosis was 11% higher than the GBD 2015 estimate for 2015. The increase mainly occurred in some African countries-notably, Burundi, Central African Republic, Congo, Democratic Republic of the Congo, Gabon, Nigeria, Uganda, and Zambia had more than twice the number of estimated deaths in 2016 compared with in 2015. Fourth, in GBD 2015, we did not separately examine the burden of multidrug resistant tuberculosis. Given their epidemiological and clinical importance, we included estimates of multidrug resistant and extensively drugresistant tuberculosis in GBD 2016. Further estim ation and mapping of the burden of tuberculosis by drugresistance type and HIV status at a finer spatial resolution could better inform surveillance and the targeting of resources for interventions. 48 As countries work towards achieving the SDG target to end the tuberculosis epidemic by 2030, contemporary information on the levels and trends of the burden of tuberculosis is essential to track and monitor the progress of control efforts in individual countries. Locations with the greatest improvements in controlling tuberculosis could provide insight into successful programmatic strategies for countries with stagnant progress. Our findings suggest that, if current trends in tuberculosis incidence continue, few countries will meet the SDG target. Progress needs to be accelerated by improving the quality of and access to tuberculosis diagnosis and care, scaling up of interventions to prevent risk factors for tuberculosis, and integrating control programmes for tuberculosis, HIV, and diabetes.