A novel approach to estimate the impact of health workforce investments on health outcomes through increased coverage of HIV, TB and malaria services

Background Globally, HIV, TB and malaria account for an estimated three million deaths annually. The Global Fund partnered with the World Health Organization to assist countries with health workforce planning in these areas through the development of an integrated health workforce investment impact tool. Our study illustrates the development of a user-friendly tool (with two MS Excel calculator subcomponents) that computes associations between human resources for health (HRH) investment inputs and reduced morbidity and mortality from HIV, TB, and malaria via increased coverage of effective treatment services. Methods We retrieved from the peer-reviewed literature quantitative estimates of the relation among HRH inputs and HRH employment and productivity. We converted these values to additional full-time-equivalent doctors, nurses and midwives (DNMs). We used log-linear regression to estimate the relation between DNMs and treatment service coverage outcomes for HIV, TB, and malaria. We then retrieved treatment effectiveness parameters from the literature to calculate lives saved due to expanded treatment coverage for HIV, TB, and malaria. After integrating these estimates into the tool, we piloted it in four countries. Results In most countries with a considerable burden of HIV, TB, and malaria, the health workforce investments include a mix of pre-service education, full remuneration of new hires, various forms of incentives and in-service training. These investments were associated with elevated HIV, TB and malaria treatment service coverage and additional lives saved. The country case studies we developed in addition, indicate the feasibility and utility of the tool for a variety of international and local actors interested in HRH planning. Conclusions The modelled estimates developed for illustrative purposes and tested through country case studies suggest that HRH investments result in lives saved across HIV, TB, and malaria. Furthermore, findings show that attainment of high targets of specific treatment coverage indicators would require a substantially greater health workforce than what is currently available in most LMICs. The open access tool can assist with future HRH planning efforts, particularly in LMICs. Supplementary Information The online version contains supplementary material available at 10.1186/s12960-023-00854-0.

below (individual study estimates are provided in Table 3 in the later section of this supplementary file). Given the range of estimates, the median of the productivity ratios was used as the multiplier. 2 For example, for estimates from Okello and colleagues the productivity multiplier was calculated by dividing the TB treatment success for DNMs (74%) by that when CHWs were added to the skills mix (56%), which yields the multiplier of 1.32.
Relationship between human resources for health (HRH) density and increased coverage of selected services related to HIV, TB and malaria,

A. Primary analysis of service coverage indicators
Eleven of the 29 candidate indicators were examined using ordinary least squares (OLS) regression (see ""Z.MD_only" and "Z.NM_only" tab in Lives saved calculator: tool") [1]. This framework assumes a curvilinear, or "diminishing returns", relationship between DNM and HIV, TB, and malaria treatment coverage. In this approach, the greatest increases in treatment service coverage are associated with initial investments in DNM. At a certain level of DNM, however, there is an inflection point at which the rate of increase in treatment service coverage is smaller with each additional DNM investment.
First, the relationship between each of the 11 candidate indicators (dependent variables) and the natural logarithm of the number of DNMs per 1000 population (independent variable) was estimated. In the base specification, the number of observations was the total number of LMICs analysed. This analysis yielded four treatment service indicators with a positive and statistically detectable (i.e., P < 0.05) association with DNM concentration (see Table 3 in this document): • Antiretroviral therapy (ART) coverage (% of people living with HIV); • percentage of pregnant women with HIV who receive antiretroviral medicine for prevention of mother-to-child transmission (PMTCT); • TB treatment coverage: the number of new and relapse TB cases per number of incident cases; and • the percentage of children < 5 years with fever who sought treatment at any facility (a proxy indicator of treatment for malaria).
Each treatment service indicator serves as a separate dependent variable in a separate regression. When each of the four regressions was restricted to LMICs only, all but the first HIV indicator (ART coverage for people living with HIV) showed a positive and statistically detectable association. However, when using the full country data set without any income-based country restrictions, all four indicatorstwo for HIV, one for TB and one for malariashowed a positive and statistically detectable association. The full-country results of each of the four OLS regression analyses appear in Table 3 below. Note that the number of countries analysed differed for each indicator, owing to varying data availability for treatment service indicators across countries. These parameters allowed estimation of treatment service coverage for a particular DNM density. For example, for TB treatment coverage a few levels of DNM per 1,000 population, corresponding to a specific level of treatment service coverage, are provided for purposes of illustration (Table 4 below). The regression equation results illustrated in Tables 3 and 4 above were then used to estimate the gains of treatment service coverage associated with investments in DNMs. Users may also identify lower and upper confidence bounds estimates of "lives saved" for additional coverage of each treatment service indicator. These confidence bounds use the standard errors of coefficient estimates (Table 3 above) to derive 95% confidence intervals of lives saved.

B. Benchmarking alternative
A complementary function of the model and calculator developed entails a possibility for the user to set a predetermined level of treatment service coverage for HIV, TB and malaria and then find the level of aggregate workforce requirements to attain that service coverage.
The benchmarking alternative entails use of empirically derived maximum "caps" as well as minimum "floors" (and the range in between the cap and floor) of general DNM concentrations that correspond with existing treatment service coverage levels. The maximum cap was derived by identifying the median level of DNM observed for countries that meet the service coverage target for that particular indicator (e.g., 10.85 DNM for the 12 countries which exceeded 90% treatment coverage for TB). The minimum floor of DNM, by contrast, was created using results from a data envelopment analysis (DEA), which ranks countries according to their efficiency in delivering treatment service coverage per level of DNM. DEA is a statistical analysis tool often applied to microeconomics and operations management. In this context, the tool has been applied to identify countries that maximize the utility of existing resources to achieve a desired end. Specific details for the DEA approach are covered in Supplementary file 2.
Once the DNM maximum cap and floor levels for each treatment service indicator had been determined, a log-linear function was fitted through these points (given the economic principle of diminishing returns to additional DNM especially at higher levels of DNM).

C. Accounting for comorbidity
In countries with a relatively high incidence of HIV and TB, there is a dual burden from HIV/TB coinfection [16]. Despite the comorbidity associated with HIV/TB coinfection, many coinfected patients remain undetected and undertreated [17]. In light of this challenge, WHO outlined guidelines and policy recommendations for enhancing collaborative HIV/TB activities [18].
Based on empirical estimates available in the studies quoted in this subsection, estimates of HIV/TB coinfection and HIV/TB detection rates were integrated when estimating how HRH investments could augment treatment service coverage across these two conditions. In a study from a setting with a high HIV burden, among HIV-positive persons who had screened positive for TB at the time of the study, only 6.3% were receiving TB therapy [19]. This study underscores the sometimes low level of integration of HIV and TB control activities, but indicates that persons who test positive for HIV may also undergo additional TB screening. It is therefore assumed that, in a high HIV/TB coinfection context, additional treatment of, say, 100 HIV-positive persons would increase TB screening for ~6.3 TB-positive persons (i.e., 6.3%).
Next, the increase in TB service treatment coverage attributable to an increase in HIV treatment coverage was calculated for countries with lower HIV/TB incidence. Estimates of HIV/TB incidence were taken from the Global Burden of Disease project [20]. It was assumed that countries with a lower HIV/TB coinfection rate would necessarily have a slightly lower "crossbenefit" for TB detection among persons undergoing HIV treatment. This cross-benefit in TB detection ranged from 3.3% to 6.3% and was automatically integrated into the final TB service treatment coverage percentage. Table 3 provides estimates of the relation between DNM density and treatment service coverage for four treatment service coverage indicators referring to specific HIV, TB and malaria services. Given the instability of the HIV ART treatment indicator results in alternative approaches tried, which exclude high-income countries and have a relatively low sample size, we recognized that further stratification of our regression approaches by country characteristics (e.g., low-income countries -LICsonly) attenuates the DNM / treatment coverage associations or renders them nondetectable. The functionality of the lives saved calculator also includes options to cater to the specific aspects of a country's situation. For example, countries without endemic malaria may wish to incorporate an assumption that HRH investments are not related to malaria burden in that country. This context-specific aspect, as well as the "high or low HIV burden" button, can thus be clicked to activate these aspects into the modelling estimates.

Translating service coverage into health impact
The OLS regression analysis described earlier yielded four treatment service coverage indicators for HIV, TB and malaria that show a positive association with DNM density: • ART coverage (% of people living with HIV); • percentage of pregnant women with HIV who receive antiretroviral medicine for prevention of PMTCT; • TB treatment coverage: the number of new and relapse TB cases per number of incident cases; and • the percentage of children < 5 years with fever who sought treatment at any facility (an indicator of treatment for malaria).
These four indicators were used to model how improvements in service coverage, via HRH investments, may translate into lives saved. Relevant empirical data from the literature were reviewed to derive estimates of lives saved.

A. Lives saved due to increased ART coverage
A study reported trends in HIV incidence, ART coverage and mortality statistics from the 30 countries with the greatest AIDS mortality burden [21]. Based on results using data from the population.
If we take the average number of AIDS-related lives saved per 100,000 population (in the first and second country) for a 1 percentage point increase in ART coverage, which assumes additivity of health benefits per unit increase in treatment coverage given low HIV incidence:  [22]. This estimate is also used in the lives saved calculator for a "low HIV burden" context option. measurement is useful, as it provides a monetary value for how much a person is willing to pay for a given good or experience or to avoid an undesired outcome. In the context of HRH, the WTP value provides a monetary value of how much it is worth for a healthcare worker to stay or leave his/her current position, thus affecting the DNM labour supply. The WTP value is calculated as the percentage of the average salary of each occupational group, which indicates the additional resources needed to retain a healthcare worker.
This body of literature is the most applicable for extracting estimates on how increases in remuneration and training may lead to increased HRH density. The included estimates are presented in Table 6 below. There were three studies (from LMICs that included doctor and/or nurse/midwife occupational groups) that met the inclusion criteria (see Table 7 below for further details) [4,5,6]. However, only one of the three studies actually reported the fraction of study participants willing to accept and stay for the pay raise offereddata required to actually calculate an estimate and make the study usable. 11% for D 8% for NMW -Rockers et al. [5] reported that given a 71% increase in pay, 46% of medical students would be willing to work in a lower-quality facility. Dividing the increase in pay that would incentivize working in a poorer-quality location by the percentage found to accept the incentive converts this association to FTEs: 71% / 46% = 1.54. This quotient is interpreted as a 54% increase in salary being necessary to retain 1 FTE. Given the limited number of studies, the estimates across income categories and across occupational groups relied on one overall estimate (1.54), which can be applied to all groups.

Studies on in-service training and retention
A literature search was conducted to identify DCE studies that include in-service training and outcomes related to retention. Given the few studies that focus on in-service training, the content, duration, cost or frequency of in-service training is not considered further. 3 A review of the available empirical estimates showed that there are few to no studies that report any results relating in-service training to worker retention. Thus, the empirical basis for this investment pathway is not supported by the current state of the literature.

1.
Investments are made on top of existing HRH workforce dynamics. It was assumed that the labour market operates at a steady state in which inflows (newly graduated workers or immigrants) and exits (due to death, departure or retirement) occur as they historically have occurred.

2.
The labour market in LMICs is currently operating at full capacity. Many countries in LMICs do not have enough demand to support a larger HRH workforce [43]. As such, any additional workers to be employed will need to be fully supported for compensation, whether for new graduates from medical education (above and beyond steady-state inflows) or for retaining workers in their current position.

3.
The user can input the number of workers targeted for the investment. To calculate the association between HRH investments and treatment service coverage, the starting point for calculating the impact of investments will be the number of health workers targeted by occupational group and, for the case of pre-service education and salary pay raises, the relative size of investment to be made (i.e., 10% increase in salary, increasing the number of graduates to be supported in the labour market by 20%).

4.
The calculator will not estimate the cost implications of the investments. The costs of HRH investments will vary considerably across country and health system contexts. In addition, it is assumed that the salary support given to health workers is discounted over the full lifetime of each worker.

5.
Investment effects will be static. While all investments will take time to implement and to take effect in the health system, the estimates are simplified to reflect the result after all effects have worked through the system.

6.
The main effects of different HRH investment options will be additive. If multiple investment types are chosen, resulting estimates of FTEs will be assumed to be linearly additive in a steady-state labour market.

7.
Estimates will be stratified by occupational group (DNMs and CHWs) and country income level (low, lower-middle and upper-middle) to the extent supported by the empirical literature. Because associations of HRH investments with density and performance may vary substantially by occupational group and the depth of the labour market for health workers across income levels, calculations will be sensitive to occupational group targeted by country income level to the extent possible. Note that given the data requirements of using all countries with available data (including those beyond LMICs) to estimate the relation between HRH density and service coverage levels, identified associations reflect an average across all country income levels.

8.
Estimates for CHWs will be based on a productivity multiplier. Data on the availability of CHWs across countries is not as readily available as it is for DNMs. The peer-reviewed literature will therefore be used to estimate CHW performance using a productivity multiplier (see earlier section). For countries without CHW stock data, the income-level strata median ratio of DNM density to CHW density will be used, calculated from the set of countries for which data for all three health worker categories are available.

9.
Estimates for clinical officers (COs), a default option for an additional category of healthcare workers, will be based on a productivity multiplier. This productivity ratio is estimated from a study which calculates sub-Saharan Africa CO salaries relative to nurse salaries (i.e., US$ 1915 / US$ 1865 = 1.026).
[44] By default, the productivity ratio of COs relative to nurses and midwives is 1.03. The number of COs will be multiplied by the productivity ratio and then folded into the total number for nurses and midwives. The lives saved from additional COs will therefore be calculated in terms of nurses and midwives.

10.
Associations between investment options and HRH availability or productivity will be based on aggregated estimates from the empirical literature.

12.
Service coverage increases resulting from HRH productivity investments will not be occupation specific. The optimal skills mix for producing health services, whether specific to HIV, TB, malaria or more generally, is unknown and likely to be highly specific to health systems structures in different countries. Additional assumptions would need to be made to identify the relative attribution of service coverage increases due to HRH productivity across health workforce categories. Furthermore, no assumptions or limits are made about scale economies, the marginal rates of substitution between capital and labour inputs or the pathway for labour market expansion in the production function for health.

13.
There are diminishing returns to HRH availability investments on service coverage outcomes. This relationship is reflected in the log functional form transformation used in the regression analysis for Step 2 (see earlier section of this supplementary file). The assumption of diminishing returns is not applied to service coverage increases resulting from HRH productivity investments, as there is no empirical basis for this from the existing literature. 14.
The empirical relationship between HRH density and service coverage is based on regression results that aggregate DNMs. While regression estimates were analysed for doctors and nurses/midwives separately, the resulting statistical significance of these relationships did not hold as strongly across service coverage indicators as analyses that combined these groups.

15.
The relationship between HRH inputs and HRH availability and productivity are measured without uncertainty estimates. Given that a key goal of the calculator involves arriving at a point estimate of lives saved due to HRH investments, we used the peer-reviewed literature as well as regression-based parameter estimates to derive the quantitative inputs. Whereas we acknowledge that each of these inputs is measured with error, for ease of interpretation of the calculator tool we did not apply formal uncertainty analyses (e.g., the delta method) to the standard errors of each input. Rather, we used only the standard error estimates from the regressions in Table 3 above) to arrive at lower and upper bounds of total lives saved in the calculator tool.