Balancing health and financial protection in health benefit package design

Policymakers face difficult choices over which health interventions to publicly finance. We developed an approach to health benefits package design that accommodates explicit tradeoffs between improvements in health and provision of financial risk protection (FRP). We designed a mathematical optimization model to balance gains in health and FRP across candidate interventions when publicly financed. The optimal subset of interventions selected for inclusion was determined with bi-criterion integer programming conditional on a budget constraint. The optimal set of interventions to publicly finance in a health benefits package varied according to whether the objective for optimization was population health benefits or FRP. When both objectives were considered jointly, the resulting optimal essential benefits package depended on the weights placed on the two objectives. In the Sustainable Development Goals era, smart spending toward universal health coverage is essential. Mathematical optimization provides a quantitative framework for policymakers to design health policies and select interventions that jointly prioritize multiple objectives with explicit financial constraints. Our analysis included two steps. First, we computed the health and FRP benefits expected for each candidate health intervention contingent on its inclusion in the benefits package (“Intervention‐level analysis”), assessed in comparison to a counterfactual in which that candidate intervention was excluded. Next, we applied constrained optimization to determine the optimal mix of candidate interventions given an objective (e.g., maximize health gains; maximize financial protection or both) and budget constraint (“Mathematical optimization analysis”). As an illustration, we apply our optimization model to the case study of Ethiopia, a low‐income country in sub‐Saharan Africa with diverse epidemiologic needs. Ethiopia‐specific parameters were derived from the following sources: population estimates from the United Nations Population Division World Population Prospects (United Nations, 2017); national disease‐related mortality, incidence, and prevalence estimates from the 2017 Global Burden of Disease study (James et al., 2018; Roth et al., 2018), and intervention coverage from the 2016 Ethiopia Demographic and Health Survey (Central Statistical Agency [CSA] Ethiopia, 2017). We selected 20 interventions across 23 disease areas and age groups as an illustrative mix of candidate health interventions for benefits package consideration. The interventions were picked based on their relevance to WHO’s UHC essential health service categories (Hogan et al., 2018). We also added a bundle of essential surgical procedures to our intervention choice set (Meara et al., 2015). Tables S1 and S2 in Supporting Information S1 provide details on both WHO’s UHC categories and the specific interventions included in this analysis. All interventions were mapped to disease categories using the International Classification of Diseases 10. Intervention‐level input parameters were drawn from various sources (see Tables S3–S9 in Supporting Information S1). Labor and delivery interventions were broken down by signal function when possible with treatment effects for both neonatal and maternal burden targets (see Table S10 in Supporting Information S1) and then summed to account for the cumulative total health and financial benefits of labor and delivery intervention packages. Signal functions are evidence‐based practices used to measure if basic emergency obstetric care (BEmOC) or comprehensive emergency obstetric care (CEmOC) was available at delivery. For example, availability of uterotonic drugs (oxytocin) and parenteral anticonvulsants are two signal functions for access to BEmOC. Access to caesarean section and blood transfusion are the two signal functions that distinguish CEmOC from BEmOC. We modeled each intervention’s expected impact on population health and FRP using a simplified health‐state model (Figure 1). We assumed that inclusion of an intervention would eliminate out‐of‐pocket (OOP) payments for the intervention and would increase coverage by no more than 5% (i.e., an incremental coverage increase presumed to be achievable within current health system capacity). Interventions not included in the benefits package were assumed to maintain the baseline coverage and OOP payment levels in Ethiopia. Intervention unit cost estimates were obtained from the Disease Control Priorities 3rd edition (DCP3) for the World Bank’s low‐income country grouping (Watkins et al., 2020). When unit cost data were not available from DCP3, we relied on estimates from the literature. National Health Accounts data complemented by the 2018 Noncommunicable Diseases and Injuries Commission Report informed the assumed percent of an intervention’s unit cost that would be paid OOP by households at baseline (Federal Democratic Republic of Ethiopia, Ministry of Health, 2014, 2018); when unavailable for a specific disease category, the health sector average was used (approximately 34%; Table S5 in Supporting Information S1). To estimate the total cost to the government of including an intervention in the benefits package we assumed the government would cover the OOP household expenditures for the baseline coverage population and the full unit cost for all households newly seeking care (the incremental 5 percentage‐point coverage). All cost numbers were in 2016 USD. Some interventions include both a screening step and treatment conditional on a positive screen. For those interventions we estimated the percentage of households that received treatment conditional on screening and assumed those households incurred a unit cost specific to treatment in addition to a baseline screening cost (Table S6 in Supporting Information S1). In the model, individuals seeking preventive services are at risk of catastrophic health expenditures (CHEs, OOP expenses surpassing a certain threshold of consumption expenditures—a measure of lack of FRP) while in the healthy state in Figure 1. Individuals in the diseased state who access care and are treated are also exposed to a risk of CHE while receiving care based on the OOP expenses incurred. Conceptual disease‐state framework to model the health and financial risk protection benefits from inclusion of an intervention within a benefits package. Cases of catastrophic health expenditure occur when out‐of‐pocket (OOP) health‐related payments surpass a certain threshold of consumption expenditures. Patients are at risk of catastrophic health expenditure in the healthy state for preventive care and in the diseased state for curative care. CFR, case fatality ratio Intervention health impact was modeled in two ways: via reduced disease incidence (directly or through risk factor control), and via reduced disease‐related fatality risks. In both instances, we quantified impact in terms of the number of deaths averted from an intervention coverage increase. Focusing on mortality outcomes undervalues diseases with high morbidity relative to mortality (such as mental health disorders) but was implemented for simplicity and due to data availability. The reduction in expected deaths from a coverage change (C i to Ci′) that acts on disease incidence was calculated as: where E i,x is the effectiveness of intervention i on incidence of disease x, A i,x is the share of disease x addressable (in the population) by intervention i, CFR x is the case fatality ratio (CFR) of disease x, and I x is baseline incidence of disease x. When an intervention directly modified mortality risks, the number of deaths averted was calculated based on the intervention impact on the current mortality burden, D x (rather than incidence I x ): where E i,x is instead the treatment effectiveness of intervention i on deaths attributable to disease x, and D x is baseline disease‐related mortality. Intervention coverage increases today can produce both immediate and future health benefits and financial consequences. For example, when a child is treated with antibiotics for meningitis, both health benefits and OOP expenditures happen immediately. By contrast, when the coverage of screening and control of diabetes increases, health benefits are typically delayed while the financial consequences may occur both today (control) and in the future (through the treatment of diabetes complications). For ease of exposition, we considered a “steady state” system in which coverage increases have immediate health and financial benefits. We also considered a scenario with benefit and cost consequence time lags based on the natural history of each disease for a cohort of present‐day individuals. Where delayed effects were relevant, we binned interventions into either short‐, medium‐, or long‐term categories with time delays for the health and FRP benefits of 5, 10, and 20 years, respectively (and with immediate costs). Rotavirus vaccination in newborns is one example of a short‐term delayed intervention; screening and control of hypertension in 50–69 year‐olds is categorized as a medium‐term delayed intervention; and human papillomavirus vaccination of 12‐year‐old females is the only intervention categorized in the 20‐year delay category (Table S3 in Supporting Information S1). Inclusion of an intervention in the benefits package was assumed to produce FRP (i.e., reduction in estimated CHE cases) through two channels. First, we assumed that inclusion of services in the package eliminated all OOP direct medical costs associated with those services. We considered OOP direct medical costs only, which is consistent with the computations of CHE estimates routinely provided by WHO and the World Bank (Wagstaff et al., 2018; World Health Organization & World Bank, 2017). For example, if the benefits package included rotavirus immunization, households would no longer incur the baseline OOP immunization expenses. We refer to this channel as “primary” FRP. In addition, covered services may also provide “secondary” FRP. With expanded coverage of rotavirus immunization, more children are vaccinated, reducing the likelihood of severe rotavirus diarrhea in those same children. The reduced need for diarrheal treatment (e.g., oral rehydration solution) and associated OOP expenses would be averted if rotavirus were included in the benefits package. We estimated both primary and secondary FRP benefits per intervention by mapping all interventions temporally by disease target, that is, rotavirus is an “earlier” intervention compared to oral rehydration solution for the treatment of diarrheal disease (Table S4 in Supporting Information S1). We quantified the FRP benefits in terms of CHE cases averted when each intervention is publicly financed through a benefits package, compared to a counterfactual in which that intervention is excluded. Health expenditures were deemed “catastrophic” when direct household OOP health expenditures would exceed a pre‐defined threshold (t CHE) of a household’s per capita total consumption expenditures (y h ). We used the threshold of 10% of per capita household consumption in our main analysis, and included a sensitivity analysis with a threshold of 25% (as in Wagstaff et al., 2018). CHE cases averted were calculated for both households who sought care at baseline (and incurred OOP expenses) and households newly seeking care conditional on an intervention inclusion in a benefits package. CHE cases averted are a function of the change in OOP expenditures as a percent of household per capita total consumption expenditures (Equations (5), (6), (7) below). where T i,x is the number of individuals treated conditional on an intervention inclusion in a benefits package (and assumed shift to coverage Ci′). δH i are the covered households, and POP i,x is the target population for each intervention (which can be the general population, the incident or prevalent populations, depending on the disease). For health interventions that target multiple disease categories (e.g., the basic surgical package targets road traffic injuries and falls), the ensemble Hi′ included households across all relevant disease targets. Households newly seeking care (due to increased coverage) were counted as CHE cases averted if the baseline OOP expenditures (now waived) for the service constituted expenditures above the 10% threshold of consumption expenditures. Since we assumed inclusion of an intervention eliminates all OOP expenditures, we counted all the associated expected CHE cases as CHE cases averted when the intervention is publicly financed. When relevant, secondary FRP benefits were estimated as decreased demand for downstream care: where i indexes the intervention under consideration for benefit package inclusion and j indexes downstream care relevant to the same disease class x. Per intervention, estimates of deaths averted and CHE cases averted were computed including upper and lower bound estimates based on uncertainty available for the treatment effect and the disease burden (deaths, incidence, and prevalence estimates). We conducted scenario analyses to estimate optimistic and pessimistic outcomes (see section 14 “Scenario cases” in Supporting Information S1). Briefly, we took the set of lower, mean, and upper bounds for disease burden (deaths, incidence, and prevalence) in combination with the lower, mean, and upper bounds of each treatment effect estimate (9 total combinations). Finally, we extracted the highest and lowest deaths averted and CHE cases averted outcomes for the optimistic and pessimistic scenarios. We approximated the distribution of household consumption expenditures per capita using a gamma distribution where scale and shape parameters are a function of the per person total household expenditures and Gini coefficient for Ethiopia (see section 5 “Household consumption distribution” in Supporting Information S1). We assumed that household demand for each intervention is constant across all household income levels: this is a strong assumption which was taken due to lack of appropriate available data. The total cost (denoted TC i ) to fund each intervention was estimated as the intervention unit cost times the new treatment population (i.e., the ensemble of households Hi′). All intervention unit costs and OOP costs are summarized in Tables S5 and S6 of Supporting Information S1. To incorporate uncertainty in unit costs, we considered optimistic and pessimistic scenarios as ±50% the base‐case unit costs as we lacked empirical uncertainty ranges for costs. We used integer programming to identify the optimal subset of interventions to include in the benefits package under a given budgetary constraint, denoted B. First, we optimized on a single objective, Obji1, that was defined as either deaths averted (δD i , per intervention i) or CHE cases averted (δCHE i , per intervention i). Second, we also performed a bi‐criterion optimization by optimizing on one objective, Obji1, and incorporating the second objective, Obji2, as a constraint. For instance, if we maximized deaths averted, we included an additional constraint that CHE cases averted exceeded a given constant K. By varying K, a Pareto efficiency frontier was generated. When K equals 0, the outcome is at least as good as the single objective optimization of Obji1. When K equals the maximum possible solution for Obji2, there is forced consistency with the single objective optimization of Obji2. In this analysis, K ranged from 0 to 115,747 for CHE cases averted and from 0 to 6010 for deaths averted. Our bi‐criterion optimization is expressed as: subject to: where Z i is a binary decision variable for inclusion of intervention i (Z i equals 0 or 1), K is a positive integer that captures the increase in the second objective, B is the budgetary constraint, and TC i is the cost of including intervention i (at no cost to treated individuals). We considered a range of constraint values (K) for the second objective that covered the entire feasible funding space to generate the Pareto efficiency frontier; and K was incrementally changed by single units of either deaths averted or CHE cases averted. The technical appendix of Karsu and Morton (Karsu & Morton, 2021) provides a detailed explanation for implementation of the epsilon constraint method which can efficiently modify the value of K iteratively (see also Ehrgott, 2005). The analysis used R statistical software (version 3.5.0; www.r‐project.org). All data, code, and result files are available on request.