Joint Food Security Monitor

Afghanistan, 401 areas, 2010-01-01 - 2025-11-01, version 2025-12-01

The Joint Food Security Monitor datasets consist of four components: raw data, transformed indicators, binary alerts, and population exposure estimates. The transformation and alert framework is developed by Penson et al. (2024). For each indicator, multiple transformation methods and time windows are evaluated, and the best-performing ones are automatically selected. These transformations normalize time-varying risk into a stable range and include methods such as z-scores, moving average divergences, and the Relative Strength Index (RSI).

Alert levels (Heightened or Critical) are determined by applying thresholds to the normalized indicators so as to best reproduce historical food insecurity patterns. Thresholds are optimized by minimizing a loss function that balances the False Positive Rate (FPR) and the False Negative Rate (FNR). The loss formulations follow Andrée et al. (2020). The Heightened threshold minimizes the weighted loss with a two-thirds emphasis on false negatives, while the Critical threshold uses a more conservative calibration that places a two-thirds weight on false positives.

Binary alerts from different indicators are combined into a Generalized Linear Model (GLM) to produce population-at-risk estimates. The GLM is estimated using weighted maximum likelihood estimation (WMLE). The weights are calibrated so that the sum of probability-weighted populations best matches each country’s historical total of populations exceeding IPC cutoff phases.

Across the literature, several approaches exist. Andrée et al. (2020) develop a GLM using lagged continuous indicator data and optimize it for prediction of escalations only, defined as transitions from non-critical IPC phases. Training and validation are done on panel data to ensure sufficient escalation events. Penson et al. (2024) develop a GLM using only binary alerts, trained and validated against escalation events. Gbadegesin, Andrée, and Braimoh (2024) fit a GLM using lagged continuous indicators in Afghanistan against all phase data rather than escalation events only.

Because the Joint Food Security Monitor spans many countries with differing prevalence of escalation events, these methods are unified into a hybrid framework. In the first stage, recursive feature elimination (RFE) is used to select binary alerts following the Penson et al. (2024) approach. In the second stage, lagged continuous indicators from the Andrée and Gbadegesin models are added, while keeping the binary indicators fixed, and RFE is reapplied. The calibration and validation frameworks are also unified. Specifically, the escalation-only validation logic of Andrée and Penson is merged with the full-sample validation of Gbadegesin by combining loss functions as follows: loss1 = 1 - x / log(2), where x = w loss(escalations) + (1 - w) loss(all), and w = log(N_escalations) / log(N_all). When escalation events are scarce, this combined loss approximates that of Gbadegesin et al. (2024), while when escalation events are plentiful, it converges toward Andrée et al. (2020) and Penson et al. (2024).

The final GLM is estimated using WMLE with positive-class weights calibrated to match historical country-level population totals. A second loss function, loss2 = 1 - (pooled_rmse^2) / (pooled_rmse_baseline^2), measures fit between estimated and observed country-level population exposures, pooling values above and below the median. This acts as a balanced pseudo R-squared, sensitive to errors in both high and low exposure regimes.

The overall loss combines loss1 (admin-level balanced loss) and loss2 (country-level balanced loss). The weights are determined by effective sample sizes that account for spatial and temporal correlation: N_eff = N / (1 + (N - 1) rho_s) and T_eff = T / (1 + (T - 1) rho_t). The combined loss is calculated as loss = 1 - sqrt(w (1 - loss1)^2 + (1 - w) (1 - loss2)^2), where w = N_eff / (N_eff + T_eff). This mirrors the weighted error logic of Andrée et al. (2020). When the time series is short, model optimization prioritizes cross-sectional fit (loss1); when the time series is long, it prioritizes temporal consistency (loss2). The correlation adjustments prevent over-weighting in cases where many administrative units are identical (e.g., due to coarse IPC assessment levels) or where country-level totals vary little over time.

References:
Andrée, B. P. J., Chamorro, A., Spencer, P., Kraay, A., & Wang, D. (2020). Predicting food crises (Policy Research Working Paper No. 9412). World Bank. https://hdl.handle.net/10986/34510
Gbadegesin, T. K., Andrée, B. P. J., & Braimoh, A. (2024). Climate shocks and their effects on food security, prices, and agricultural wages in Afghanistan (Policy Research Working Paper No. 10999). World Bank. https://hdl.handle.net/10986/42552
Penson, S., Lomme, M., Carmichael, Z., Manni, A., Shrestha, S., & Andrée, B. P. J. (2024). A data-driven approach for early detection of food insecurity in Yemen’s humanitarian crisis (Policy Research Working Paper No. 10768). World Bank. https://hdl.handle.net/10986/41534