Household behavior and vulnerability to acute malnutrition in Kenya

Acute malnutrition remains a persistent problem in Sub-Saharan Africa, with Kenya reporting 942,000 cases among children aged 6–59 months in early 2022. Acute malnutrition (measured by MUAC) is used as an early warning indicator of nutrition-related crises, but current early warning mechanisms often miss household-level variation by relying on regional or national aggregates. The study adopts a socio-ecological, coupled natural–human systems lens to examine household food security and emphasizes the complex, emergent dynamics between household characteristics/behaviors and their socio-economic and ecological environments. The authors aim to make explicit the role of heterogeneous household adaptive capacity and coping strategies, and to consider interacting risk factors (e.g., climate and economic shocks) rather than assessing them in isolation. The research questions are: (i) why some households are more vulnerable to acute malnutrition than others within the same geographic context, and (ii) whether given risk factors have uniform effects across households. Using an agent-based, evidence-driven computational model seeded and validated with data from West Pokot (Kenya), the paper evaluates fit against observed malnutrition prevalence, generates short-term forecasts, conducts counterfactual experiments on adaptive capacity under climate and COVID-19-related economic shocks, and tests generalizability in Turkana.

Prior research identifies immediate, underlying, and basic causes of child malnutrition, and multiple studies link malnutrition to climate, conflict, economic conditions, intrahousehold dynamics, and women’s empowerment. However, existing early warning systems and correlational models typically operate at aggregated spatial/temporal scales, masking household-level heterogeneity and coping behavior. Qualitative studies have documented differences in coping strategies by income and access to resources, and recent work examines heterogeneity in latent factors like adaptive capacity. Yet gaps remain: limited forward-looking analytical frameworks, inadequate validation, and weak generalizability. Additionally, most analyses consider single-factor effects rather than interactions (e.g., combined climate and market shocks such as those induced by COVID-19). This study responds by integrating household behavioral heterogeneity and multi-level, interacting stressors within a validated, evidence-driven agent-based modeling framework.

Study area: Five wards in West Pokot County, Kenya (Chepareria, Kapchok, Masool, Riwo, Weiwei) spanning pastoral, agro-pastoral, and mixed farming contexts, all dependent on the long rains and subject to environmental degradation and water management challenges. Data sources and processing: Monthly household nutrition surveillance data from Kenya’s NDMA across sentinel sites (Jan 2016–Mar 2020). Each site tracks 30 households monthly. The study uses N = 14,409 observations of individual children from five sentinel sites in West Pokot. Main measure: MUAC; complemented by qualitative interviews (3–5 per sentinel site) for validation. Stressors include: (i) local market conditions via NDMA monthly maize prices relative to 10-year market-specific baselines; (ii) climate via MODIS NDVI (MOD13A3). Observations are geo-coded at ward level using GADM shapefiles. Continuous acute malnutrition rates per ward are mapped to the five-point IPC Acute Malnutrition scale (IPC III/AMN) for comparability. Computational modeling approach: Evidence-driven computational modeling (EDM), a contextualized agent-based modeling approach integrating GIS and empirical validation. The model is seeded with household characteristics, decisions, behaviors, and contextual covariates to simulate acute malnutrition and household adaptation over time. Actors and space: Households are agents, initialized at the ward (ADM3) level, spatially distributed according to population density. Validation and outputs occur at the ward-month level. Decision cycle: Each model round (mapped linearly to calendar time) consists of (1) strategy selection, (2) household action, (3) nutrition outcome assessment, and (4) strategy adaptation. Households update in random sequential order and can observe and adopt best-performing strategies in their local neighborhood when nutritionally insufficient. Strategies and constraints: Nutritional strategies are vectors over three dimensions: own food production, local networks, and buy/barter. The feasible strategy set S per household is constrained by local livelihood plausibility (e.g., pastoralist areas cannot adopt farming strategies) and exogenous conditions X. Climate shocks reduce effectiveness of own-production strategies by a multiplicative factor (e.g., factor 0.2 = 80% reduction). Economic stressors (e.g., price changes) reduce purchasing capacity. Strategy success depends on matching strategy to the household’s nutritional profile h (abilities to produce, network, or buy), and is further modulated by health-related factors hp and monetary assets ha, both implemented as multiplicative factors in [0,1]. Nutritional sufficiency ns is compared to a global threshold n0; sustained insufficiency yields household acute malnutrition w; ward-level prevalence W is the aggregate over households. Adaptation and learning: If insufficient, a household adapts by adopting better-performing neighbor strategies with probability equal to learning rate λ (0.01–0.1), independent of its own profile hs. New strategies can also spawn with a small rate. Learning follows a hill-climbing heuristic. Household assets (ha) distributions are derived from cross-sectional wealth/food asset data; health factors hp and profiles h from NDMA surveys. Exogenous factors X (NDVI and maize prices) are specified monthly at the ward level. Initialization and calibration: Simulations begin with random draws from S. Calibration estimates unobservable parameters central to decision-making: learning rate λ (0.01–0.1), strategy spawn rate (0.01–0.1; baseline ~0.03), and learning mode (hill-climbing). A parameter sweep (~60,000 combinations) optimizes fit to observed NDMA prevalence. Validation metrics include bias (RMSD) and predictive accuracy (macro F1 across IPC categories; Hamming loss). The model is validated in-sample (West Pokot 2017–2018), out-of-sample by temporal split (train 2017, test 2018), true out-of-sample leading-edge forecasts (4-month windows in 2019 with retraining), and generalizability tests in Turkana (2017–2018) with an alternative specification accounting for purely pastoralist livelihoods and conflict. Counterfactual scenarios (Apr–Jul 2020): Three adaptive capacity scenarios—baseline (calibrated), constrained (learning/adaptation reduced by 50%), enabled (increased by 50%)—are evaluated under (i) climate shock (early lean season; 20% reduction in own production availability), (ii) economic shock (COVID-19-related; 50% reduction in income generation options), and (iii) combined climate + economic shock. Assumptions refined with nutrition experts.

- Heterogeneity in household adaptive capacity significantly shapes vulnerability to acute malnutrition; the most vulnerable households tend to be least adaptive. - Adaptation mitigates the impact of climate shocks more effectively than economic shocks; increasing adaptive capacity yields larger mitigation among more vulnerable households. - In-sample (West Pokot 2017–2018): Macro F1 = 0.72; Hamming loss = 0.25; RMSD = 0.01, indicating low bias and high categorical accuracy at ward level. - Temporal split out-of-sample: Train 2017, validate 2017 (F1 = 0.55; HLS = 0.45; RMSD = 0.01). Test on 2018 (F1 = 0.51; HLS = 0.49; RMSD = 0.02). Performance remains robust though lower than full in-sample. - True out-of-sample, leading-edge 4-month forecasts (West Pokot 2019): Jan–Apr F1 = 0.72; HLS = 0.26; RMSD = 0.02. May–Aug F1 = 0.67; HLS = 0.32; RMSD = 0.04. Sep–Dec F1 = 0.33; HLS = 0.56; RMSD = 0.03. Lower F1 in late 2019 partly due to IPC category discretization of continuous prevalence. - Generalizability (Turkana 2017–2018): F1 = 0.52; HLS = 0.48; RMSD = 0.14. The model captures substantial variation but under-predicts unusually high prevalence in some districts (e.g., Lokichar, Lokiriama/Lorengippi), reflecting unforeseen shocks and unobserved dynamics in purely pastoralist settings. - Scenario-based forecasts (Apr–Jul 2020): Combined climate + economic shocks produce the most severe increases in malnutrition; enabling adaptation reduces prevalence relative to baseline, while constraining adaptation amplifies impacts. Strategy adaptation is less effective under economic constraints due to fewer viable alternative strategies.

The findings directly address why some households are more vulnerable than others and whether risks have uniform effects across households. The evidence shows that household-level heterogeneity in adaptive capacity and strategy diversity critically shapes malnutrition outcomes, even within similar geographic and livelihood contexts. Incorporating these dynamics improves predictive validity over short horizons, informing timelier, ward-level targeting. The Turkana analysis indicates that narrower livelihood portfolios and ongoing crises reduce the diversity of coping strategies and lower effective adaptation, increasing vulnerability. These insights suggest that early warning systems should integrate household behavioral parameters—learning rates, strategy sets, and local network effects—rather than relying solely on aggregate indicators. The model’s complex-systems framing (feedbacks, co-evolution, multilevel interactions) highlights that small exogenous shocks can trigger disproportionate effects in areas with low strategy diversity. Programmatically, enhancing access to diverse strategies and enabling learning pathways can mitigate climate impacts and partially buffer economic shocks.

This study introduces and validates an evidence-driven agent-based model linking household behavior, adaptive capacity, and acute malnutrition in Kenya. It demonstrates robust in-sample fit, credible short-term leading-edge forecasts, and meaningful generalizability to a neighboring pastoralist context. Key contributions include: (i) making household adaptive capacity explicit in forecasting malnutrition risk; (ii) showing that adaptation more effectively mitigates climate than economic shocks; and (iii) revealing that the most vulnerable households benefit most from increased adaptive capacity. Policy implications stress targeted, preventive interventions that expand household strategy sets, improve access to infrastructure, education, and markets, and strengthen local networks. Future research should investigate the determinants and modalities of adaptation (individual vs. collective), integrate additional stressors and behavioral mechanisms, expand spatial coverage, and improve measurement to reduce aggregation-induced classification errors.

- Data limitations: NDMA household data collection ceased after March 2020 due to COVID-19; subsequent Family MUAC changes affect comparability. Privacy constraints required aggregation to ward level. - Sampling: Small number of sentinel wards in West Pokot reflects NDMA strategy and may limit generalizability at finer spatial scales. - Measurement and classification: Reclassification of continuous prevalence to IPC categories can depress F1 when small differences straddle thresholds; difficulty correctly classifying rare high-prevalence categories (e.g., Phase 5). - Model scope and unobserved factors: Lower performance in Turkana suggests unobserved dynamics (e.g., conflict, mobility, pastoralist-specific shocks) not fully captured by NDVI and prices; model trained on past patterns struggles with unprecedented spikes. - Structural assumptions: Strategy sets constrained by livelihood plausibility; learning via hill-climbing with fixed λ; multiplicative health and asset effects; these simplifying assumptions may not capture all real-world decision processes. - Temporal mapping and shocks: Linear time scaling and limited exogenous stressor set may overlook timing nuances and interacting shocks beyond climate and market proxies.