Methodology

InsightOut seeks to help nonprofits and policymakers to locate and identify potential beneficiaries for targeted social policies. Organizations select outcomes and broad regions of interest, along with other preferences, and this tool delivers:

1. Targeting maps identifying community clusters with high risk or prevalence of the selected outcome, to serve as a starting point for research and service delivery.
2. Optimized surveys developed with machine learning models that are concise but highly accurate in predicting whether the household may be facing the selected outcome.

Demographic and Health Surveys (DHS):

The Demographic and Health Surveys (DHS) Program run by USAID, ICF International, and partner organizations worldwide is the main source of data for this tool. DHS is a global program that conducts large-scale standardized socioeconomic surveys across over 90 countries since 1984. The surveys provide granular data on individual and household conditions spanning health, nutrition, education, economic prosperity, and more. Additional information about the DHS program can be found here: https://dhsprogram.com/.

This tool currently supports analysis in the following outcomes:

1. Poverty (Wealth Index)

This tool currently supports analysis in the following countries based on their corresponding survey years:

1. Nigeria (2024)
2. Democratic Republic of the Congo (2023)

Targeting maps identify community clusters with high risk or prevalence of the selected outcome, to serve as a starting point for research and service delivery. Here's how it works:

1. Users select a country, outcome, and broad region(s) of interest. These regions are generally the broadest administrative or political unit of the country.
2. The household-level DHS data is pulled for the outcome of interest and aggregated at the community cluster level. Community clusters are groupings of households mapped to the same geolocation to preserve household privacy and confidentiality. This geolocation is approximately the geographic center of the households, displaced between 2-5 kilometers. Across surveys, there are usually between 20-50 households per community cluster.
3. The map visualizes community clusters based on these values and returns a list ranking community clusters. Polygons are created around these community cluster points, so every point in the region is associated with a value that corresponds to its closest community cluster.

Optimized surveys are generated with machine learning models to be concise but highly accurate in predicting whether a household in the region may be facing the selected outcome. Here's how it works:

1. Users select a country, outcome, and a broad region of interest. The customized survey will identify questions that produce the best model to predict the outcome in the region of interest.
2. Users select what "best model" means based on their organization's preferences and concerns. This will determine what performance metric is used to optimize the machine learning model. The options include:
   1. Cast a Wide Net: Prioritize identifying most households in need, even if some households are falsely identified as experiencing the need (machine learning model optimizes on recall).
   2. Avoid misallocating resources: Prioritize identifying households most confidently in need, even if some households are falsely identified as not experiencing the need (machine learning model optimizes on precision).
   3. A balance of concerns (a) and (b), (machine learning model optimizes on f1)
3. The tool generates two survey versions: a concise short form (approximately 10 questions) and a more comprehensive long form (approximately 20 questions). Each survey consists of yes/no questions administered to the household head, with each question weighted by its importance in predicting the outcome. The administrator sums the weights of applicable questions to get a final household score; the lower the score, the greater the likelihood of the outcome. The PDF demonstrates an example of survey administration and score interpretation.