Modelling Indigenous small-scale agriculture and food systems in Guatemala - Hybrid Bayesian inference for data-poor regions

CONTEXT

Participatory system dynamics modelling is a useful method for characterising agricultural systems and the complex dynamics linking their human and agronomic counterparts that determine their long-term behaviour and sustainability. One challenge facing this use of system dynamics methods, nonetheless, is the scarcity of time-series data for many key variables, which hinders the calibration and validation of these models.

OBJECTIVE

This research proposes a new approach for quantifying difficult-to-quantify relationships within system dynamics models of socio-agricultural systems when temporally scarce but spatially rich data (e.g., survey or census data) is available for many socioeconomic model variables of interest.

METHODS

We propose a methodology to quantify system dynamics models that uses Bayesian inference over spatially-explicit data from different regions to estimate the shape of relationships between socioeconomic variables, where the diversity of values across a country can serve to compensate for the lack of time-series data in regions of interest. The hierarchical component of the approach allows for the automatic weighting of each site's data according to its degree of similarity to the case study region. This approach was applied to a model of agricultural systems and food security developed in Tz'olöj Ya', and K'iche', Guatemala with two different Indigenous farming communities.

RESULTS AND CONCLUSIONS

1) Results indicate that the model performs better in non-study site municipalities that are socioeconomically and environmentally similar to the case study sites than in less similar municipalities (R² 0.81–0.98 in the study sites, but <0.5 in many dissimilar regions).

2) The spatial validation procedure across non-case study municipalities shows that trends in population and child chronic malnutrition are relatively well-represented by the model in similar municipalities (R² 0.81–0.99 in case study regions), while forest cover dynamics are much more difficult to generalise across regions (R² 0.26–0.87 in case study regions, and worse elsewhere).

3) The model showed that agricultural system resiliency was best improved not by technological fixes to improve crop productivity, but rather by structural changes to livelihood diversification.

4) These results were possible due to the hybrid approach used: stakeholder participation was central to the identification of key relationships between agronomic and socioeconomic variables, while Bayesian inference and spatial validation allowed for the assessment of the model's validity and geographical limits.

SIGNIFICANCE

The new methodology allows for quantification and testing of system dynamics models of agricultural systems that could otherwise not be formally calibrated or validated due to a lack of time-series data.