Tracking total spending on tuberculosis by source and function in 135 low-income and middle-income countries, 2000–17: a financial modelling study

Summary Background Estimates of government spending and development assistance for tuberculosis exist, but less is known about out-of-pocket and prepaid private spending. We aimed to provide comprehensive estimates of total spending on tuberculosis in low-income and middle-income countries for 2000–17. Methods We extracted data on tuberculosis spending, unit costs, and health-care use from the WHO global tuberculosis database, Global Fund proposals and reports, National Health Accounts, the WHO-Choosing Interventions that are Cost-Effective project database, and the Institute for Health Metrics and Evaluation Development Assistance for Health Database. We extracted data from at least one of these sources for all 135 low-income and middle-income countries using the World Bank 2019 definitions. We estimated tuberculosis spending by source and function for notified (officially reported) and non-notified tuberculosis cases separately and combined, using spatiotemporal Gaussian process regression to fill in for missing data and estimate uncertainty. We aggregated estimates of government, out-of-pocket, prepaid private, and development assistance spending on tuberculosis to estimate total spending in 2019 US$. Findings Total spending on tuberculosis in 135 low-income and middle-income countries increased annually by 3·9% (95% CI 3·0 to 4·6), from $5·7 billion (5·2 to 6·5) in 2000 to $10·9 billion (10·3 to 11·8) in 2017. Government spending increased annually by 5·1% (4·4 to 5·7) between 2000 and 2017, and reached $6·9 billion (6·5 to 7·5) or 63·5% (59·2 to 66·8) of all tuberculosis spending in 2017. Of government spending, $5·8 billion (5·6 to 6·1) was spent on notified cases. Out-of-pocket spending decreased annually by 0·8% (−2·9 to 1·3), from $2·4 billion (1·9 to 3·1) in 2000 to $2·1 billion (1·6 to 2·7) in 2017. Development assistance for country-specific spending on tuberculosis increased from $54·6 million in 2000 to $1·1 billion in 2017. Administrative costs and development assistance for global projects related to tuberculosis care increased from $85·3 million in 2000 to $576·2 million in 2017. 30 high tuberculosis burden countries of low and middle income accounted for 73·7% (71·8–75·8) of tuberculosis spending in 2017. Interpretation Despite substantial increases since 2000, funding for tuberculosis is still far short of global financing targets and out-of-pocket spending remains high in resource-constrained countries, posing a barrier to patient's access to care and treatment adherence. Of the 30 countries with a high-burden of tuberculosis, just over half were primarily funded by government, while others, especially lower-middle-income and low-income countries, were still primarily dependent on development assistance for tuberculosis or out-of-pocket health spending. Funding Bill & Melinda Gates Foundation.


S1. Summary statistics
Cost of an outpatient visit Cost of an inpatient bedday Incident cases  -drugsusceptible  TB   135  135  135  135  135  135  135  135  135  135  135  135  135  135  135  135  135      In ST-GPR, there is the option to smooth across countries, which is more applicable for estimating disease burden. In financial estimation, we often do not often smooth across space. Health financing variables vary dramatically across countries because health systems are so distinct. We rely relatively more on the underlying data, the linear fit from the first stage of the model, and smoothing across time.
In addition, to reflect reforms and fluctuations in National Tuberculosis Programme (NTP) investment over time, we set a low GPR scale to lower the correlation between data points through time. Finally, we set a low amplitude to narrow variance for countries with good primary data (e.g., country-year>5). Table S5 below shows the parameters used for ST-GPR models run in this analysis: = - To estimate government health spending, out-of-pocket health spending, and prepaid private health spending on TB, we created estimates using costs and volume of TB health services. The equations for our cost-volume estimates for each source are as follows: and at least one of three injectable second-line drugs (i.e., amikacin, kanamycin, or capreomycin). In this study, we refer all drug resistant tuberculosis as MDR-TB.
We assumed pre-treatment spending was primarily on outpatient care. Pre-treatment visits per patient data were taken from World Health Organization (WHO) patient cost surveys as well as peer-reviewed literature 1,2 . As a function of our assumption that all incident cases sought treatment, we assumed that all incident cases had the same utilisation of pre-treatment visits. We assumed that notified and nonnotified cases had the same utilisation of outpatient visits before receiving treatment. To estimate uncertainty around pre-treatment visits, we created 1,000 draws using the 12 data points for each of drug-susceptible (DS) and MDR TB. As such, the data we used was not country-specific. Per the conceptual framework, we assume the spending on pre-treatment care is sourced similarly to the health sector.
To get pre-treatment spending for government health spending, out-of-pocket health spending, and prepaid private health spending, we used the equations below: To ensure that the medical costs reflected the health system perspective, we adjusted the unit costs with the fractions of outpatient spending sourced by government, out-of-pocket, and prepaid private health spending. We modelled 630, 493, and 486 country-year data points from National Health Accounts (NHAs) respectively, logit-transformed, with ST-GPR. Each model used natural log-transformed lag-distributed income (LDI) per capita from GBD, 3 random effects on country, and the logit-transformed fraction of total health spending per each source ( Collaborator Network 4 as covariates.

S2.2.2. Notified cases
We assumed that all spending on notified cases was sourced by the government. While we found evidence that indicated that some patients who were treated through National Tuberculosis Programmes incurred out-of-pocket costs, the evidence was limited and the costs were often closer to 0 than to the pattern of out-of-pocket spending in the health sector as a whole. See Table S5 for studies we reviewed on out-of-pocket expenditures incurred by patients in the NTP setting 1, 2,5-9,10-12, 13-21 .
was sourced from the WHO Global TB Database. 22 We used ST-GPR to fill any missing data points for these utilisation variables and create a complete time series of estimates.
All utilisation variables -  Table 1 in the manuscript, we assume that inpatient and outpatient care for non-notified cases is sourced similarly to the health system as a whole. We do not account for any spending on drugs sourced by the government through social insurance for non-notified cases, as we assume that government spending on drugs for TB is accounted for in NTP spending.
As such, the equations below illustrate how we calculated spending on non-notified cases by source: For out-of-pocket and prepaid private spending, we calculated drug spending as follows: with 537 and 505 country-year data points sourced from NHAs respectively. We used data for HC.5  We sourced all estimates from the GBD 2017 study and the WHO Global TB Database. The first stage linear model of ST-GPR was a mixed-effect model, with the logit-transformed fraction of NTP spending/total government health expenditure, natural log-transformed case notifications and the logittransformed fraction of case notifications that were multi-drug resistant from WHO, 22 and natural logtransformed LDI per capita as covariates and random effects on country. To detect and reduce the influence of outlier data points, we used the selected model to measure Cook's distance for each data point. We excluded each data point if Cook's distance was greater than 15/n where n is the total number of data points.

S2.3.2. Modeling total out-of-pocket spending on tuberculosis
For our final model, we logit-transformed -as a fraction of total out-of-pocket health spending from the GBD Health Financing Collaborator Network. 4 To select covariates, we first conducted a lasso regression to determine which covariates were least correlated, conditional on other covariates, with the logit-transformed -fraction as the dependent variable. Covariates with an estimated coefficient of zero were removed from the set of possible covariates. We then used linear mixed effects regression to estimate all models including all possible combinations of the remaining covariates.
We then selected the 1000 best models with the lowest Akaike information criterion (AIC) and the 1000 best models with the lowest Bayesian information criterion (BIC) values. Finally, we completed a 10-fold cross-validation with out-of-sample predictions on these selected models. We selected the best model based on out-of-sample root mean squared error. We used Cook's distance to detect outliers and excluded any points with a Cook's distance greater than 15/n, where n is the total number of -data points.

S2.3.3. Modeling prepaid private spending on tuberculosis
Prepaid private spending includes spending by voluntary insurance (i.e. private and not mandated) and spending from domestic non-governmental organisations. We modelled as a logittransformed fraction of total prepaid private health spending from the GBD Health Financing Collaborator Network. 4 For the linear first stage of ST-GPR, we used a mixed-effects model with the logit-transformed proportion of prepaid private/total health spending and natural log-transformed LDI per capita as covariates with random effects on country. We used Cook's distance to detect outliers and excluded any points with a Cook's distance greater than 15/n, where n is the total number of data points.

S2.3.4. Modeling spending on tuberculosis by function
The following components are mutually exclusive and exhaustive of domestic spending on tuberculosis: Currently, we do not break down NTP spending into any further functions such as diagnostic spending, nor do we estimate spending on immunisation. These are limitations of our current estimates, but we plan to incorporate them into future work.

S3. Currency Conversion
All

S4. Sensitivity analysis
We conducted sensitivity analyses using different key variables and data sources to test if our models were input dependent.
Model 1 below represents our primary estimates. Model 2 addresses one of our key assumptions, which is that all TB cases seek treatment. In Model 2, we assume that only 75% of non-notified cases sought treatment. Model 3 addresses another of our key assumptions, which is that non-notified cases have similar healthcare utilisation patterns as notified cases. In Model 3, we assume that health care utilisation (health facility visits and hospitalisation percentage) are 50% lower for non-notified cases than notified cases. Finally, Model 4 tests the sensitivity of our estimates to different incidence inputs. In Model 4, we use WHO reported estimates of incidence for TB instead of IHME estimates of TB incidence. Despite differences between the mean results across models, the qualitative conclusions drawn from each model remain the same, as illustrated in the figures below:

S6. How these estimates might be used
We expect that a panel of estimated data regarding total TB spending, consisting of four sources, for each low-and middle-income country (LMIC) and each year, would be used by academics, policy makers, practitioners, and other stakeholders. All the estimates are available to the public in 2019 USD, PPP, and as percentage of GDP. We expect that the estimates with uncertainty intervals, available through Global Health Data Exchange (GHDx) for the public, will enable a wide range use.
For policy makers, we also have an interactive website for visualisation and use cases by policy makers and practitioners. In addition, we put all scientific evidence into lay language in our policy report "Financing Global Health," for the broader audience who may or may not have academic interest to specify the underlying methods.