Farmer characterisation and biofortified maize adoption modelling
From segmenting producers' current behaviour to projecting who would adopt the biofortified seed, in what order and to what degree as market penetration advances.
No survey records whether a farmer will adopt new technology. The task was to estimate two behaviours the data did not contain
How farmers behave today and how they might plausibly behave when offered a new seed are two separate questions. The available data answered neither.
No source directly records a farmer's willingness to adopt new agricultural technology. What the data contain is past behaviour — which seed they buy, how much they spend, what yield they obtain — and that does not tell you how they will respond to a variety that has not yet reached the market. The challenge, then, was not to classify an existing variable better, but to estimate two behaviours that the data did not contain. The first is the present: characterising rigorously what type of producer each farmer is today, without reducing it to seed spending, which conflates investment with sophistication — a farmer who spends a lot may be inefficient rather than an advanced adopter of hybrids — and incorporating yield and controlling for local agro-climatic conditions.
The second is the future: anticipating which profiles will adopt the new variety, in what order as availability grows, and what share of their land they will commit. That cannot be read from current data; it has to be modelled, and modelled so that the result can be calibrated against penetration expectations set by the client and audited by third parties.
4 segments
of farmers, identified by triangulating two complementary analytical frameworks
3 stages
in the adoption model: composite score, risk-aversion penalty and continuous conversion curve
9 scenarios
evaluated: three seed generations by three subsidy regimes
From a portrait of the present to an adoption projection: dual segmentation, risk-aversion scoring and a calibrated saturation curve
Segmentation by triangulating two complementary frameworks
Using agricultural data calibrated against official Ministry of Agriculture statistics, department-level z-scores for yield and seed spending are calculated to control for regional heterogeneity. The classification into four segments (OPV/criollo, low, medium, high) is derived from the convergence of two frameworks: unsupervised clustering with PAM, whose separation between groups was assessed with discriminant analysis, and an interpretable quadrant-based segmentation in the two-dimensional z-score space. On top of these runs a hurdle model that combines a deterministic classification of the OPV segment with probabilistic scoring for the rest. The convergence of two complementary frameworks is what makes the segmentation defensible to auditors and domain experts.
Adoption score with a risk-aversion penalty
To anticipate the propensity to adopt, a composite score is built from five weighted components (segment priority, expected economic benefit, room for improvement, entry scale and producer age), with the relevant variables first filtered by correlation, near-zero variance and Boruta selection. A risk-aversion penalty is then applied to that score, bounded through a logistic transformation, which weighs baseline yield, expected income gain, the size of the plot committed and projected loss. Producers with greater exposure to losses receive a larger penalty in the score, in line with the risk-aversion assumption built into the model.
A continuous adoption curve calibrated against market anchors
Adoption is not an on/off switch: each farmer is assigned to one of three behavioural profiles according to their white maize area (near-binary, sigmoid or gradual adoption), and a conversion curve determines what share of their land they devote to the biofortified seed, depending on their position in the ranking and the level of market penetration. The whole architecture is calibrated against five market composition anchors provided by New Seed (penetration of 5%, 10%, 20% and 30%, plus the structural anchor at 100%), minimising the mean absolute deviation through sensitivity sweeps across each axis of the model.
An actionable portrait of the farmer and an engine that projects their adoption
Actionable producer profiles for the market entry strategy
The segmentation identifies which profiles hold the greatest room for improvement and which show the most resistance to change, so that New Seed can decide which farmers to focus its initial distribution effort on. The characterisation separates the advanced adopter from the producer with high spending but low yield, a distinction that seed spending alone did not capture.
Adoption trajectory quantified by penetration level
For each combination of seed variant and subsidy regime, the scoring engine projects which segments take up adoption and in what order as penetration grows, and what aggregate area is converted to biofortified maize at each step. Adoption is no longer a single assumption but a curve calibrated against penetration expectations set by the client.
Individual economic impact with differential valuation of home consumption
For each farmer, the economic difference between their current situation and a simulated scenario of adopting biofortified seed is estimated, under the assumptions defined in the model. The fine detail lies in how the maize the household does not sell is valued: each household is classified by its actual market position — net seller, net buyer or autarkic — and home consumption is valued at the price that corresponds to that position. A seller values it at the selling price they forgo; a buyer, at the purchase price they save; and in the autarkic household, whose maize neither enters nor leaves the market and for which there is no observed price, the value of home consumption is approximated with a shadow price set at the midpoint of the band between selling and buying prices, used as an operational assumption. That economic figure feeds the benefit component of the score and links the characterisation to the programme's scenario evaluation.
What this case shows about the way we work
Segmentation that stands up to audit
Triangulating an unsupervised method with an interpretable one makes it possible to check whether both produce a consistent segmentation structure, and makes the characterisation of farmers more defensible to auditors and domain experts.
Anchored projection, not speculation
The adoption model is calibrated against market anchors provided by the client. ADI builds and tunes the engine; the penetration expectations are set by those who know the market. The projection is grounded in data, not in a crystal ball.
Risk aversion, modelled explicitly
Incorporating a penalty for potential losses makes it possible to represent explicitly a mechanism documented in the literature on agricultural adoption. The model treats it as one of the determinants of propensity, without assuming that it explains the decision on its own.
One framework, parameterised by scenario
The same characterisation and scoring engine runs on different combinations of seed variant and subsidy regime without being rebuilt, which turns the characterisation of the farmer into a reusable planning tool rather than a one-off analysis.
I've been working with Rubén and Julia on a complex modeling project for over a year now, and they have been exceptional partners throughout. As our project has evolved, they have approached every shift collaboratively, receptively, and with flexibility. They have consistently gone above and beyond to ensure we develop both a robust model and a user interface that translates model predictions into clear insights. What we have now is a fully documented, auditable methodology we can put in front of funders and external reviewers with confidence.
Katie Adams
Director of Impact and Cost Effectiveness, New Seed
Capabilities deployed in this project
Shall we talk?
Tell us what you are working on and we will give you a straight answer on how we can help.