Multilevel Beta Regression Model for Potato Post-Harvest Losses along the Farm-Market Value Chain in Kenya

Abstract

Post-harvest losses in potato value chain is one of the major threats to food security and farmers livelihoods in Kenya. This paper uses multilevel beta regression modeling approach to study the determinants and spatial heterogeneity of post-harvest potato losses in Kenya farm-market value chain using data from stratified crosssectional survey of 628 value chain actors within Nakuru and Nyandarua counties, the major potato producing highlands in the country. Five models were estimated in succession; Poisson baseline, single-level Negative Binomial (NB), multilevel NB, single-level Beta regression, and the final multilevel Beta regression model. We found that the multilevel beta regression is the best model (AIC=2711.30, log-likelihood=1375.65, LRT: χ 2 = 27.07, df = 2, p < 0.001) confirming appropriateness of Beta distribution for the bounded proportional loss and need for multilevel model. The storage methods were the most powerful explanatory variable (𝑊𝑎𝑙𝑑 χ 2 = 1182.71, 𝑝 < 0.001); the cold storage reduce proportional loss by 13.07 percentage points loss while dark room storage by 7.44 percent relative to the traditional storage methods. Storage duration, market distance, the usage of chemicals, extension and training and transportation method also proved to be highly significant explanatory variables while the education of actor had insignificant impact on potato post-harvest losses. Random intercepts at county level are close to zero and likelihood ratio test indicated a significant model fit improvement with multilevel model. Residual diagnostics by DHARMa showed validity of Beta distribution assumption. The findings are policy-relevant in investment decisions aimed at reduction of postharvest losses along the Kenyan potato value chain through provision of access to cheaper storage infrastructure, integration of extension-input system and transport infrastructure improvement.