Incluia

In the context of the COVID-19 emergency, and to respond to the closing of schools and the 50%-reduction in the usual occupation of commercial establishments and its effects on employment, the Costa Rica government implemented actions including the delivery of food packages and a cash transfer program (Bono Proteger) to support the disadvantaged population, particularly affected by the socioeconomic crisis.

The COVID-driven socioeconomic crisis affects poor and vulnerable populations in three ways: (1) makes it difficult to operate and deliver support to beneficiaries of social programs; (2) generates new segments of poor and vulnerable populations that require immediate support; and (3) evidences the need to continuously identify vulnerable populations not covered by social policies.

Problem to be solved

Inefficient allocation of social subsidies to the most disadvantaged: In the face of the COVID crisis, it is crucial to improve the effectiveness of these type of social policies, which largely depend on their appropriate targeting.

Populations affected by the problem

Extreme poverty households in Costa Rica (urban and surrounding areas).

The current response to this problem, considering the related institutions

Sistema Nacional de Información y Registro Único de Beneficiarios del Estado (National Information System and Single Registry of State Beneficiaries, SINIRUBE), an entity attached to Instituto Mixto de Ayuda Social (Joint Institute for Social Assistance, IMAS) that manages a technological tool focused on the programs of Costa Rica's social policy.

Proposal to solve the problem using AI

INCLUIA, the AI solution, generates maps of extreme urban poverty for large territories with high spatial resolution. The result is a geographic recommendation system that increases the effectiveness of field social workers, supporting the proactive search for potential beneficiaries of social programs. SINIRUBE developed this solution.

What security considerations, national laws, or standards should be considered to use each source of information?

Review of regulations related to the handling and use of personal data.

Entering into a Confidentiality Agreement to legally support the exchange, management, and safeguarding of information and data confidentiality.

No personally identifiable information was used, and international servers with worldwide service and security certifications were used to reduce the potential risks of physical and virtual attacks on the data.

Progress/results to March 2022

INCLUIA identified households without coverage in SINIRUBE, otherwise very difficult to locate, precisely because they are few and distributed throughout the large metropolitan area of Costa Rica. Forty-one priority clusters were identified for inclusion in Costa Rica's social registry.

During the assessment and knowledge dissemination stage, two dissemination workshops were held: (i) the virtual workshop "Technological Innovation for Mapping Extreme Poverty, Active Search, and Inclusion of Beneficiaries in Social Protection Systems" to more than a dozen institutions (INA, MEP, CONAPAM, IMAS, CCSS, MIDEPLAN, SINIRUBE, CONAPDIS, MTSS, STATE NATION); and ii) in collaboration with the Social Protection and Health Division of the IADB, the virtual workshop "Use of satellite images to identify gaps in the coverage of social protection systems" was held with the participation of more than 40 social protection specialists from the Bank.

INCLUIA has been implemented in the Dominican Republic and Jalisco (Mexico) and will soon be implemented in Mexico City.

Goals for 2022-II

Assess the possibility of scaling the solution to more programs and institutions of the Costa Rican public social sector.

Main implementation challenges

Handling of highly-sensitive personal information. However, from the beginning, it was determined that no personally identifiable information will be handled, and to mitigate this risk, international servers with worldwide service and security certifications that reduce potential risks of physical and virtual attacks on the data will be used.

Main AI challenges identified

• The quality of the data was critical for the project's development. In this case, this impact was positive since the validation of the georeferencing for active households registered in its database resulted in an average of 90% of valid households at the district level. Thus, the opportunity to use SINIRUBE data for the development of the prototype turned out to be greater than anticipated.
• Labels for training computer vision models. The accuracy of the results of a computer vision model depends, among other factors, on the quality of the labels of the images with which the model is trained. This represents a challenge when classifying subjective conditions such as a household's poverty level, which varies from country to country, from city to city, when viewed from the sky, or at street level. A re-labeling campaign was carried out to address this challenge, which reported high accuracy to the model developed but was money and time-consuming, requiring the participation of three experts who produced 11,250 labels in three months. Based on this experience, accurate models requiring a smaller number of labels for training that reduce time and costs are being developed.