Quantifying socio-temporal effects of loan delinquency drivers in microfinance
We develop and evaluate a family of discrete-time logit-link (LLink) models, including fixed-effects and frailty extensions, to quantify associations between socio-temporal factors and loan delinquency transitions while accounting for latent borrower heterogeneity. Using monthly records for 1,716 borrowers from a Ghanaian microfinance institution, we model transitions among good, intermediate, and poor repayment states, distinguishing deterioration, persistence, and recovery. The Eid season and long vacation indicators are associated with repayment behaviour, although the magnitude and direction of their estimated effects vary across transitions. Bootstrap tests indicate residual borrower heterogeneity in recovery and persistent poor repayment but provide little support for time-dependent frailty beyond a random intercept. Random Forest and KTBoost perform better for several binary transitions, while the LLink models provide a transparent framework for estimating covariate effects and borrower heterogeneity. For multistate prediction, we propose an Optimised Matthews Correlation Coefficient (OMCC) rule and compare it with the Djeundje and Crook (D\&C) rule. The estimated socio-temporal effects are interpreted as conditional associations rather than causal effects.