An algorithm intended to set fair public health insurance contributions in Kenya is routinely increasing costs for the poorest households, an investigation by Africa Uncensored, Lighthouse Reports and the Guardian has found.
The scheme, launched nationally in October 2024 as a flagship reform under President William Ruto to replace the previous national insurance, was promoted as a way to digitize the system and extend cover to Kenya’s large informal workforce — roughly 83% of employees, including day labourers, hawkers and smallholder farmers. Instead, the system’s premium assessments, produced by a predictive machine-learning model using proxy means testing (PMT), have prompted protests and a flood of complaints.
Local volunteers register households by entering dozens of observable characteristics — roof type, toilet, ownership of radios or livestock, whether a family has electricity — into a smartphone questionnaire. The model then estimates a household’s annual ability to pay and sets the SHA (Social Health Authority) premium. Volunteers report many extremely poor families being assigned premiums they cannot afford — sometimes 10–20% of their already meagre earnings — and some critically ill people being denied care because they cannot meet the charged amounts. One volunteer, identified in reports as Grace Amani to protect her identity, said: “People are dying, people are suffering.”
The SHA has been overwhelmed with disputes over misclassification and opaque, unaffordable premiums. Those without private insurance who do not pay SHA charges can be refused treatment or face steep hospital bills. Social media and complaint lines are full of accounts from people who previously paid modest amounts now billed far higher sums; a single mother described her monthly contribution being set at 3,500 Kenyan shillings.
An investigative audit compared the SHA model’s predictions with thousands of real households and found systematic overestimation of incomes for the poorest families while underestimating wealthier households. In one example, two farmers were predicted to earn roughly double their actual income because the model recorded assets such as electricity and homeownership as signs of higher means.
Experts say this is an inherent risk of PMT, which infers income from observable assets rather than measuring earnings directly. Poverty is often fluid and asset ownership does not always match current cash flow. Development economists cite high error rates in similar programs elsewhere — one scheme tested in Indonesia excluded 82% of its intended beneficiaries; another in Rwanda produced error rates near 90%.
David Khaoya, a health economist who advised the health ministry, said the model’s developers faced a trade-off: prioritizing correct identification of wealthier households or of the poorest. He said the choice was made to favour correctly classifying higher-income households, a decision that reduces revenue loss from wealthier people being misclassified as poor but increases the risk of overcharging truly poor families.
An independent review by international consultancy IDinsight, shared with the government before rollout and obtained by reporters, warned the model was “inequitable, particularly for low-income households,” over-represented middle-income profiles and lacked sufficient data from poverty pockets. The consultancy also said the model was out of step with Kenya’s current socioeconomic realities after recent economic shocks. Despite those warnings, SHA proceeded with national deployment.
Rollout problems have been extensive: of more than 20 million people registered for SHA, only about 5 million regularly pay premiums. Hospitals report cash shortfalls as promised reimbursements from SHA remain unpaid, and health-sector leaders describe the program as a failed experiment that shifts the burden of identifying and protecting the poor onto an imperfect technical tool.
PMT-based algorithms are being promoted globally to extend services to informal workers who lack steady documented incomes. But auditors and researchers caution these models are imprecise, opaque and damaging to public trust when they determine access to essential services. For many Kenyan families the result has been painful choices between paying for health insurance or buying food, and in some cases being denied lifesaving treatment because the algorithm judged them able to pay when they could not.
The controversy has turned a major government promise into a source of hardship and uncertainty for vulnerable households, raising questions about whether an opaque, proxy-based approach can equitably allocate healthcare costs at national scale.