Abstract

Introduction
Resilience has become a central concept in smallholder development, marking a shift from a narrow focus on welfare improvement toward strengthening the capacity of individuals, households, and agricultural systems to absorb, adapt, and transform in the face of shocks and stresses. A substantial body of research has sought to operationalise resilience theory by identifying indicator sets that explain real-world resilience dynamics (e.g. Aguilar et al, 2022; Resilience Alliance, 2010; D'Errico and Smith, 2020; McPeak and Little, 2017; Smith and Frankenberger, 2018).

What has received far less attention, however, is when and how resilience capacities are mobilised during periods of shock. While existing studies have linked resilience capacities to well-being outcomes on an annual basis, the triggers and timing of short-term capacity deployment remain poorly understood. This temporal gap is critical because it has direct implications for consumption behaviour (Kabir, 2023). Even households with substantial resilience capacities may deliberately ration short-term consumption if they expect a prolonged crisis. Evidence from rural Tanzania shows that households unable to stabilise consumption over time experienced significantly worse human capital outcomes -such as child stunting- and that the ability to smooth consumption was as consequential for human development as increasing overall food availability (Christian and Dillon, 2018).

Methodology
For the arid and semi-arid lowlands surrounding the Taita Hills in southern Kenya, this study draws on data from 80 smallholder households surveyed on a fortnightly basis over an 18 month period characterised by drought and recovery. These high-frequency household observations were combined with open-source geospatial datasets -including remote sensing products and OpenStreetMap- to link socio-economic and behavioural data with detailed meteorological and biophysical indicators across the same 36 time periods. In addition, the dataset integrates information on a wide range of structural and contextual capacities, such as infrastructure quality, market access, electrification, and soil characteristics. The result is a uniquely granular and temporally resolved dataset that captures both environmental dynamics and household-level responses.

Using a Bayesian regression framework with Hamilton Monte Carlo sampling, the analysis incorporates random effects to account for both time-varying and time-invariant sources of unobserved heterogeneity. This framework enables a direct assessment of how meteorological and biophysical indicators add explanatory power beyond baseline household characteristics. By progressively introducing time-varying covariates -such as rainfall, soil moisture, and temperature variability- the model isolates the extent to which these environmental factors explain fluctuations in household consumption and dietary diversity. Importantly, including these temporal indicators substantially reduced the standard deviation of the time random effect, demonstrating that much of the observed temporal variability in consumption behaviour is attributable to environmental dynamics rather than unexplained or purely stochastic time effects. This provides a robust empirical link between short-term biophysical change and household well-being outcomes.

Findings
Using the reduced Coping Strategy Index and 24-hour Household Dietary Diversity Score as measures of consumption, three key insights emerge: First, the smoothness of consumption is not driven purely by resilience capacities but also reflects expectations of the future via the interpretation of environmental signals. Second, short- and medium-term consumption depends less on on-farm capacities than on the reliability of off-farm incomes. Third, time in-variant resilience capacities are the strongest determinants of household consumption - particularly, infrastructure quality, soil type and demographic characteristics.

We find meteorological 'signals of change' -specifically wind direction, cloud cover, rainfall and diurnal temperature- are associated with shifts in household consumption. However, these relationships are often indirect and mediated by resilience capacities. Evidence from survey round seven, when the short rains stopped prematurely, illustrates this dynamic: south-to-north winds signalled the northward movement of the Inter-tropical Convergence Zone, indicating a lower probability of rainfall. Instead of a direct deterioration in food security, this period coincided with a peak in remittance income and the beginning of a decline in livestock ownership. These changes had contrasting implications: remittances are associated with greater dietary diversity and increased consumption, whereas in come form livestock sales corresponded to the opposite pattern.

On-farm factors such as adequate soil moisture during the maize growth cycle and greater crop diversity are linked to improved food consumption and dietary diversity, while livestock -especially sheep and goats- provide additional buffering when feed and browser vegetation are available. However, income from selling farm produce tends to coincide with reduced consumption and dietary diversity. Off-farm, unskilled labour offers little protection, as declining agricultural output increases local labour supply and depresses wages. Nonetheless, most financial pressures are managed effectively; only medical expenses and maize price rises significantly undermine food security, suggesting that households can smooth consumption in response to predictable financial stress.

Overall, as drought conditions worsened, households tended to reduce consumption and pivot toward available off-farm activities. Asset sales followed, with those holding more divisible assets (e.g. multiple goats rather than a single cow) better able to stagger sales and smooth consumption. When rains returned and grazing resources