The damage levels of the maize spotted stem borers (Chilo partellus Swinhoe) are estimated at 400,000 metric tons, which is equivalent to 13.5% of farmers' annual maize harvest accounting for US$80 million. Despite the economic importance of the pest, information on the incidence under long-term organic and conventional farming systems is lacking. This study evaluated three different link functions [logit, probit, and complementary log-log – (clog-log)] to reduce prediction errors in overdispersed stem borer incidence data for 12 years in four farming systems. The clog-log link function had the lowest Akaike information criterion (AIC) and Bayesian information criterion (BIC) indexes for the pest incidence model in Thika. Contrarily, probit showed the lowest AIC and BIC in the Chuka incidence data model. The residual diagnostic plots with clog-log demonstrated no patterns against the predicted values. Our findings revealed that clog-log link function provided the best fit in beta-binomial mixed models compared to others. We advocate for the use of clog-log for long-term pest incidence data modelling to obtain biologically realistic projections. Users of mixed models must incorporate explicit consideration of suitable link function discrimination, model fit and model complexity into their decision-making processes if they build biologically realistic models.