United Republic of Tanzania - HungerMap data

WFP’s Hunger Monitoring Unit in the Research, Assessment and Monitoring Division conducts real-time food security monitoring to track the latest food security trends. In areas where limited or no data is available, we use machine learning-based predictive models to estimate the food security situation.

WFP conducts continuous food security monitoring via computer assisted telephone interviewing (CATI) through call centers. Data is collected on a rolling basis, spread evenly over a past 28/30 calendar days or over a three-month period. The main advantage of this approach is that data is available more frequently and processed daily through automated statistical engines. Daily updates are then produced showing a snapshot of the current food security situation (with a slight time lag of 2-4 days to ensure data quality) over the past 28/30 calendar days.

Call interviews aim to cover all mobile service providers, and telephone numbers are randomly selected from a database of phone numbers or generated using random-digit dialling (RDD) method. To ensure a more representative sample, WFP uses various types of pre/post-stratification and sampling methods, including by weighting results by population at the first or second administrative level and by a demographic variable such as the level of education or water sources which could impact food security, in order to account for the fact that households with more phones are more likely to be selected (e.g. younger, somewhat better-off households who live in urban areas).

In order to compensate for non-response and attrition, key challenges for high frequency mobile phone surveys, new observations are added in each administrative area following the sample design specific for each of the country.

The questionnaire includes questions on household demographics, households’ food consumption, coping strategies used (food-based and livelihood-based), access to food, market and health services, and other country-specific livelihood-related questions. In addition, at the end of the survey, respondents are given the opportunity to share additional information on the food situation in their communities.

For first-level administrative areas where daily updated survey data is not available, the prevalence of people with poor or borderline food consumption (FCS) and the prevalence of people with reduced coping strategy index (rCSI) ≥19 is estimated with a predictive model.

Predictive model design: Training data: The models were trained using FCS and rCSI data spanning over 70 countries across, aggregated at the level of first-order administrative divisions. The input variables used to make the predictions were built using information about population density, rainfall, vegetation status, conflict, market prices, macroeconomic indicators, and undernourishment. For areas where past FCS/rCSI measurements are available, the last available data point is also included as input variable.

Data sources: Beyond the data sources already credited in the Glossary, additional sources used only for the model but not for display purposes are: Gridded Population of the World, Version 4 (GPWv4): Population Count, Revision 11 WFP’s Alert for Price Spikes (ALPS) indicator

Algorithm: XGBoost – a machine learning technique producing predictive models in the of an ensemble of regression trees. The model parameters were optimized by a cross-validated grid-search.

Model evaluation: The accuracy of the model was evaluated on a test set comprising 20% of the historical data, having trained 100 models on subsamples (with replacement) of the remaining 80% of the historical data.

Predictions: The model produces current estimates of the prevalence of people with poor or borderline FCS and rCSI for areas where no food security data is available; we call this nowcasting. For each first-level administrative boundary we report the median and 95% confidence intervals of a distribution of predictions obtained from 100-bootstrap models trained on subsamples (with replacement) of the training data.