Using supervised learning to select audit targets in performance-based financing in health: An example from Zambia

Area of Science:

• Health economics
• Health services research
• Health policy and management

Background:

• Performance-based financing (PBF) in healthcare incentivizes service volume but risks data over-reporting.
• Independent verification is crucial to ensure PBF integrity and accurate performance measurement.
• Targeted sampling strategies are needed to optimize verification efficiency and cost-effectiveness.

Purpose of the Study:

• To evaluate alternative sampling strategies for independent verification in PBF programs.
• To compare the effectiveness of random sampling versus predictive modeling for targeted clinic selection.
• To identify methods that enhance the cost-effectiveness of healthcare verification processes.

Main Methods:

• Empirical comparison of various random sampling and predictive modeling techniques.
• Utilized data from a Zambian PBF pilot, including reported and verified quantity indicators for 140 clinics.
• Applied machine learning algorithms, including Random Forest, for predictive sampling.

Main Results:

• Machine learning methods, particularly Random Forest, demonstrated superior performance over traditional sampling.
• Predictive modeling significantly improved the efficiency of identifying clinics for verification.
• The study identified opportunities to increase the cost-effectiveness of independent verification activities.

Conclusions:

• Machine learning-based targeted sampling is a highly effective strategy for independent verification in PBF.
• Random Forest models offer a promising approach to optimize resource allocation in healthcare verification.
• Implementing advanced sampling methods can strengthen the integrity and efficiency of PBF programs.