Changing Attitudes and Demand for Bioengineered Soybean Oil

Abstract

Consumer acceptance of bioengineered foods has stalled in repeated international surveys, and the modal explanation — health concerns and environmental scepticism — does not adequately explain the dispersion. Within demographic strata that share income, education, and rural/urban residence, observed purchase intent for genetically modified (GM) soybean oil products spans almost the full unit interval, and the Theory of Planned Behaviour and Risk-Perception Model jointly under-predict the upper tail.

The food-policy and consumer-behaviour literatures have begun to explore non-cognitive consumer dispositions — most prominently **food-technology neophobia**, the dispositional preference for familiar food technologies over novel ones — as the missing explanatory channel. The behavioural-finance methodological literature has begun to use machine-learning methods, particularly Random Forest classifiers combined with SHAP attribution, to recover non-linear interactions between cognitive and dispositional drivers that the linear logit benchmark suppresses. This paper combines those two literatures.

Using a stratified random consumer survey (n = 1,000) with validated TPB, Risk-Perception, and TAM scales, we apply a two-stage Random Forest framework. Stage 1 classifies binary purchase versus non-purchase; Stage 2 classifies product-variant choice among purchasers. SMOTE oversampling addresses class imbalance, 10-fold cross-validation with grid search optimises hyperparameters, MICE multiple imputation handles missing values, and SHAP attribution interprets feature importance.

Stage-1 results are strong: ROC AUC = 0.96, with precision and recall both above 0.90. SHAP attribution identifies health benefits and environmental benefits as the dominant Stage-1 drivers, with food-technology neophobia and passive resistance contributing negatively — consistent with dual-process accounts of consumer behaviour in which deliberate health and environmental reasoning dominates the binary purchase decision when product information is salient. Stage-2 multiclass results are currently pending a diagnostic review of the labelling structure before publication. The paper is positioned as a documentation-stage exemplar of methods-applied behavioural-finance work and serves as a methodological reference for the wider research programme.

Data & Methods