Introduction.

Dose-response trials are used with the objective of selecting the efficient herbicide dose in weed management. Data analysis of these experiments has been criticized for the use of statistical models that do not fit the distribution of the response variable, failure to specify the original structure of the experimental design, and the preference for partial models instead of fitting a unique model. Nonlinear mixed models are presented as a more accurate alternative for analyzing these experiments. Objective. To determine the effective herbicide dose using three modeling strategies in dose-response trials. Materials and methods. Two independent experiments were conducted in greenhouses located in Tambor, Alajuela, Costa Rica, during 2012, where the fresh weight in grams (g) of a biotype of Paspalum paniculatum L. was quantified as a function of grams of acid equivalent (GAE) of an applied herbicide, under a randomized complete block design. A four-parameter logistic regression model was used as a basis, and three model variants were fitted. Using penalized information criteria (Akaike information criterion (AIC) and Bayesian information criterion (BIC)), the best-fitting model was chosen. Results. The strategy that considered the experiment and the block within each experiment as random effects proved to be the most accurate. This model estimated the confidence interval (95 %) for the mean effective dose of GAE between 335.12 and 384.32 g. Conclusions. Integrating information from independent experiments as random effects within a unique model generated more accurate estimates of glyphosate’s effective dose.