Breeding for high-yielding and broadly adapted varieties of cassava has been the primary target of the International Institute of Tropical Agriculture (IITA) cassava breeding program based in Nigeria. However, this target has been hindered due to the presence of genotype-by-environment interaction (GEI) and phenotypic recurrent selection as a traditional approach of breeding. This approach generates gains slowly due to long breeding cycle, resulting to low rate of realized genetic gain per unit time for complex traits like fresh root yield. Taking advantage of recent advances in computational resources, this study focused on exploring historical data using advance statistical techniques and stochastic simulation. The main objective was to identify a breeding scheme which optimizes the cost of field operation and rate of genetic gain. First, I used classical linear-bilinear model to dissect existing patterns of GEI of 36 elite cassava clones evaluated in 11 locations over 3 growing seasons. This aims to identify the optimum number of environments from target population of environments for future testing of genetic lines for key traits such as fresh root yield, dry matter content, and top yield. Second, I exploited the complex pattern of GEI from 96 varieties assessed in 48 trials using variance structure models on fresh root yield to identify an optimal model that captures GEI and stable clones, identify mega-environments and key environmental covariables driving GEI. Lastly, I used stochastic simulation to assess different breeding scenarios to identify an optimal breeding scheme which maximized genetic gain for cassava in Nigeria by investing the breeding resources in one breeding program for broad adaptation or splitting the resources into two sets of testing locations for narrow adaptation. Key lessons from these studies include: (1) Regardless the number of environments sampled to represent TPE, prediction accuracy of fresh root yield is lower than that of dry matter content and top yield. (2) The testing locations within the same geographic region were clustered and dissimilar from locations in other regions indicating some locations within each cluster may be dropped for field trial to maximize the budget cost. (2) A factor analytic statistical model with three factors was identified as the parsimonious model whose common latent factors captured 79.0% of total genetic variability. (3) Maximization of covariance between latent factor loadings and weather variables was an effective approach for identifying weather conditions driving genotypic response to testing environments. (4) The rate of genetic gain per unit time from genomic-enabled breeding programs were consistently higher than that of phenotypic-based conventional breeding program.