Huy Duc Dang; Au Hai Thi Dam;
Thuyen Thi Pham; Tra My Thi Nguyen

Abstract
Purpose – The purpose of this paper is twofold: to explain access to formal and informal credit in agriculture of Vietnam; and to compare the effectiveness between regular econometrics and machine learning techniques.

Design/methodology/approach – The multinomial logit (MNL) regression model and the random forest (RF) technique are employed for comparison purposes. To avoid heteroskedasticity, the robust covariance matrix is computed to estimate the sandwich estimator which in turn provides an asymptotic covariance matrix for biased estimators. Additionally, multicollinearity is tested among independent variables with variance inflation factors less than 3. Adequacy approach and sensitivity analysis are used to determine relevant levels of predictors. For models comparison, statistical evaluation metrics including Cohen’s κ, mean absolute error, root mean squared error and relative absolute error are employed.

Findings – The discrepancy between sensitivity analysis and adequacy approach revealed that MNL is more compatible for explaining determinants of credit participation. Due to insignificant differences in the evaluation metrics between models, the winner of choice is undetermined. Among other determinants, collateral, farmsize, income, procedure, literacy and all risk variables stand out to be critical factors when deciding borrowing schemes. While financially literate farmers tend to acquire loans from both sources, borrowing decisions against different risk sources depend on risk type and famers’ own desire to borrow.

Originality/value – Results of the MNL model are more consistent with literatures, which reinforce the role of collateral in the local credit scheme. Besides, financial literacy and farmers’ perception on different risk sources also influence how farmers’ borrowing strategies vary among sources.

Keywords: Machine learning, Format credit, Informal credit, Multinomial logit regression