Multicollinearity is a very common problem in studies that employ path analysis in agronomic crops, which generates unrealistic results and erroneous interpretations. This study was aimed at assessing the path analysis in data obtained from guava tree full-sib based on modelling multiple regressions applying latent variables to neutralize the effects of multicollinearity. Seven explanatory variables were measured-fruit mass (FM), fruit length (FL), fruit diameter (FD), mesocarp thickness (MT), peel thickness (PT), pulp mass (PM), total number of fruits (NTF)-, plus the main dependent variable, total yield per plant (YIELD). In accordance with the multicollinearity scenario, eleven values were tested with the addition of the constant K to the diagonal of the correlation matrix X' X. Path analysis was applied in two models: all the explanatory variables with direct effect on the dependent one and another model with multiple regression with more than one chain and the presence of latent variables. The path analysis in the multivariate methodology of structural equation modelling (SEM), which uses latent variable prediction, provided better results than the traditional and ridge path analyses.