Generalized Cost-Effectiveness

Generalized Cost-Effectiveness Analysis and the Null Scenario

Generalized cost-effectiveness analysis, usually shortened to GCEA, is the World Health Organization's version of cost-effectiveness analysis for priority setting across a whole health sector. Each intervention in a group of related options is measured against one shared starting point, the situation if all of them stopped, so services already funded face the same test as new candidates. The method underpins the WHO-CHOICE programme (Choosing Interventions that are Cost-Effective), aimed mainly at low- and middle-income settings designing health benefit packages on the way to universal health coverage.

Why WHO moved the comparator away from current practice

The WHO guide to GCEA, edited by Tan-Torres Edejer and colleagues, calls the usual comparison with current practice intervention mix constrained CEA, because existing interventions are never removed unless something aimed at the same condition replaces them. Such an analysis can show that one treatment beats another without testing whether existing services were worth providing.

Murray and colleagues, setting out the WHO proposal in 2000, argued that such studies lack a sectoral perspective in which all possible interventions are compared to find the health-maximising mix for the available resources. Hutubessy, Chisholm and Tan-Torres Edejer put the contrast in efficiency terms. Most studies answer technical efficiency questions about producing one intervention without waste, while allocative efficiency concerns how resources are spread across programmes and needs a sectoral analysis.

Changing the comparator also helps transferability. Current mixes of interventions differ greatly between countries, so marginal results travel poorly, and the WHO guide argues that the null is more comparable across populations, within limits that led WHO-CHOICE to report results for 14 epidemiological subregions rather than one global figure.

Building the null for a cluster of related interventions

The null cannot mean a world in which no intervention ever existed, since today's epidemiology reflects decades of immunisation and treatment. The WHO guide defines it as what would happen if the current interventions in a group ceased today, so prevalence then drifts towards the null hazard rates.

Nor does the null remove every health service at once. Interventions are grouped into clusters that interact in costs or effects, or are mutually exclusive, and each cluster gets its own partial null. In the WHO guide's example, insecticide-treated nets interact with house spraying but not with cardiovascular care, so malaria interventions are analysed with cardiovascular services left in place.

Null hazard rates can come from natural history models or from untreated groups in trials and cohorts. The alternative the guide calls back-adjusting removes the estimated effect of current interventions from the observed rate, and the 2021 WHO-CHOICE update by Bertram and colleagues gives the single-intervention form as follows.

$$\lambda_N = \frac{\lambda_C}{1 - c \cdot e}$$

where $\lambda_N$ is the null hazard rate (for example incidence, remission or case fatality), $\lambda_C$ is the current hazard rate, $c$ is the current coverage of the intervention and $e$ is its effectiveness, both as proportions. Where several interventions act on the same rate, the update uses a multiplicative form.

$$\lambda_N = \frac{\lambda_C}{\prod_{k=1}^{n} (1 - c_k e_k)}$$

where $c_k$ and $e_k$ are the coverage and effectiveness of intervention $k$ and $n$ is the number of interventions acting on that rate.

Combining interventions along an expansion path

Once the null is fixed, each intervention is evaluated alone and then in every plausible combination with others in its cluster. Evans and colleagues note that joint costs can fall below the sum of separate programmes when vehicles, buildings or staff are shared, and that WHO-CHOICE treated effects as additive for different outcomes and multiplicative for the same outcome.

The ratio for each option against the null takes the following form.

$$R_i = \frac{C_i}{D_i}$$

where $R_i$ is the cost per disability-adjusted life year averted for option $i$, $C_i$ is its cost and $D_i$ is the number of DALYs averted compared with the null, so the null sits at the origin of the cost-effectiveness plane, which Bertram and colleagues call the null position.

The expansion path then starts from the option with the lowest ratio against the null and adds options one step at a time, each judged by its own incremental ratio.

$$\text{ICER}_{j,k} = \frac{C_j - C_k}{D_j - D_k}$$

where $\text{ICER}_{j,k}$ is the extra cost per extra DALY averted when the path advances to option $j$ after the previous point $k$. Evans and colleagues note that the incremental ratio against doing nothing is sometimes called the average cost-effectiveness ratio. Weakly dominated options are dropped, so the path follows the same logic as a fully incremental analysis with extended dominance removed, except that the starting point is the null and not current care.

Worked example with prevention and treatment in one cluster

All numbers in this example are illustrative. A cluster for one disease contains a preventive intervention A and a treatment B. Current practice is B alone at 50% coverage, and B is assumed to cut case fatality by 40% in those treated. Costs are annual totals for the population in millions of international dollars.

1. Back-adjust the null. The observed case fatality hazard is 0.08 and the denominator is 1 − 0.5 × 0.4 = 0.8, so the null hazard is 0.08 / 0.8 = 0.10.

2. Model each option against the null. A population model gives the following costs and DALYs averted, and the last column divides one by the other, for example 7,200,000 / 7,000 = 1,028.6 for the combination.

Run separately, the two would cost 2.0 + 6.0 = 8.0 million, but shared programme costs of 0.8 million bring the joint cost to 7.2 million. Their effects would sum to 9,000 DALYs, but prevention removes cases that treatment would otherwise have handled, so the joint gain is 7,000.

3. Build the expansion path. A has the lowest ratio, so the path starts there. Replacing A with B would cost 4,000,000 / 1,000 = 4,000 per extra DALY averted, while adding A to B would cost 1,200,000 / 2,000 = 600. A ratio into B that is higher than the ratio out of B marks B alone as extendedly dominated. The path therefore passes through A and then the combination, and its second step costs 5,200,000 / 3,000 = 1,733.3 per extra DALY averted.

4. Compare with an analysis against current practice. An analysis starting from B would find that adding A costs 1,200,000 / 2,000 = 600 per extra DALY averted and would look attractive, without asking whether B alone was worth 6.0 million. GCEA shows that A alone averts 80% of the DALYs that B averts for a third of the cost.

5. Look across clusters. Suppose total health spending is fixed, this disease currently receives 6.0 million, and another cluster offers interventions at 900 per DALY averted with constant returns. Choosing B over A in this cluster spends 4.0 million to avert 1,000 extra DALYs, whereas the same money in the other cluster would avert 4,000,000 / 900 = 4,444 DALYs.

The null exposes an allocative inefficiency that analysis against current practice would hide. How fast treatment staff can move into prevention remains a question for context-specific analysis.

Methods choices in WHO-CHOICE and how they changed

Shared methods let results for different diseases sit in one league table. Hutubessy and colleagues report that costs were expressed in international dollars, effects in DALYs averted relative to no intervention, facilities were assumed to run at 80% capacity utilisation, and drugs were priced at the lowest generic prices available internationally.

The 2021 update by Bertram and colleagues changed several of these choices. Costs are now taken from the health system perspective in 2010 international dollars, with no cost offsets from avoided future treatment. The time horizon became the lifetime of those affected instead of a 10-year implementation period. Results are shown with a 3% discount rate on costs and either none or 3% on health, health states are valued with Global Burden of Disease disability weights, and one-way sensitivity analysis replaces probabilistic analysis, which the authors judged infeasible for so many inputs.

Thresholds and the move away from GDP rules

Early WHO-CHOICE work used criteria from the Commission on Macroeconomics and Health, and Hutubessy and colleagues report that an intervention averting a DALY for less than average per capita income was called very cost-effective, one costing less than three times that figure cost-effective, and anything above it not cost-effective. Bertram and colleagues later wrote, in 2016, that none of these thresholds should be used alone as a funding rule or as a measure of affordability. In Peru, a contextualised WHO-CHOICE analysis found trastuzumab cost-effective under the three-times rule, yet adding it would have cost more than the entire breast cancer treatment budget. The 2021 update uses no threshold at all, presenting results in league tables or log-scale graphs by multiples of 10. Evans and colleagues explain coarse grouping by uncertainty, judging that 45 international dollars per DALY averted cannot be called more efficient than 60, though 45 can be trusted to beat 450. A cost-effectiveness threshold and a budget impact analysis answer separate questions, and GCEA supplies neither.

Adapting regional results to a country

Regional estimates rarely fit any one country without change. Hutubessy and colleagues describe adjusting unit prices and resource quantities for costs, and coverage, efficacy and adherence for effects. Adaptation can assume efficient delivery, giving an ideal mix, or add local constraints such as staff shortages, current capacity utilisation and local prices.

Shifting to a better portfolio incurs transaction costs that a production-cost ratio leaves out, and where local capacity and infrastructure limits are ignored the theoretical GCEA ratio may differ markedly from one achievable in a given country. The output is a population-specific set of average and incremental ratios showing whether current strategies are justified and what to add if resources grow. Bertram and colleagues present the CHOICE-Spectrum tool for this work, including design of a benefit package and result databases for health technology assessment.

How GCEA differs from neighbouring methods

The WHO guide treats GCEA and analysis against current practice as complements. GCEA identifies an efficient long-run mix, incremental analysis starting at the current mix shows how to reach it, and appraising a new drug in one country remains suited to the conventional design. A cost-effectiveness league table assembled from separate studies mixes comparators, which one null per cluster avoids.

Extended cost-effectiveness analysis and distributional cost-effectiveness analysis address equity questions that GCEA leaves to the wider decision process. Linear programming can add explicit constraints that a league table ignores, though the WHO guide finds such models information intensive and costly, and often hard to communicate.

Limits and common misreadings

The null is never observed, so its quality depends on epidemiological modelling and on how clusters are drawn. Evans and colleagues note that combinations multiply quickly as interventions are added, so only a subset is examined. The 2021 update also assumes that health system constraints can be bought out in the long run, so the result describes an efficient package for a fully functioning system, which a given system may not be able to deliver now.

Several misreadings recur. A ratio against the null is sometimes quoted as if it were an incremental ratio against current care, which can make an expensive addition look cheap, as the combination's 1,029 against 1,733.3 for its actual step shows. GDP-based bands have been used as automatic funding rules, which the WHO-CHOICE authors reject. Leaving an intervention out of the WHO-CHOICE list does not mean it is cost-ineffective, as Bertram and colleagues state in the 2021 update. Efficiency is also only one criterion, and Hutubessy and colleagues stress that it has to be weighed against poverty, equity, implementation capacity and feasibility.

Sources

• Bertram MY, Lauer JA, De Joncheere K, Edejer T, Hutubessy R, Kieny MP, Hill SR. Cost-effectiveness thresholds: pros and cons. Bulletin of the World Health Organization. 2016;94(12):925-930. https://doi.org/10.2471/BLT.15.164418
• Bertram MY, Lauer JA, Stenberg K, Tan Torres Edejer T. Methods for the economic evaluation of health care interventions for priority setting in the health system: an update from WHO CHOICE. International Journal of Health Policy and Management. 2021;10(11):673-677. https://doi.org/10.34172/ijhpm.2020.244
• Evans DB, Tan-Torres Edejer T, Adam T, Lim SS. Methods to assess the costs and health effects of interventions for improving health in developing countries. BMJ. 2005;331(7525):1137-1140. https://doi.org/10.1136/bmj.331.7525.1137
• Hutubessy R, Chisholm D, Tan-Torres Edejer T. Generalized cost-effectiveness analysis for national-level priority-setting in the health sector. Cost Effectiveness and Resource Allocation. 2003;1:8. https://doi.org/10.1186/1478-7547-1-8
• Murray CJL, Evans DB, Acharya A, Baltussen RMPM. Development of WHO guidelines on generalized cost-effectiveness analysis. Health Economics. 2000;9(3):235-251. https://pubmed.ncbi.nlm.nih.gov/10790702/
• Tan-Torres Edejer T, Baltussen R, Adam T, Hutubessy R, Acharya A, Evans DB, Murray CJL, editors. Making Choices in Health: WHO Guide to Cost-Effectiveness Analysis. Geneva: World Health Organization; 2003. https://iris.who.int/handle/10665/42699