Efficiency of plant breeding programs, particularly in perennial crops, may be improved by reducing the time required for evaluation prior to selection. The objective of this work was to study expected selection responses of several variables based on partial bean yields compared to total accumulated bean yield of coffee (Coffea arabica L.) during the first 48 mo of production. One hundred thirty‐eight F5 families of a cross between C. arabica cv. Caturra and Timor hybrid were randomly distributed among four experiments transplanted in 6 × 6 simple lattice designs. Beans were harvested by hand monthly from each tree. Fresh bean yield was determined for each tree on 48 harvests over a 4‐yr period. Sixty‐eight variables were created from bean yield data collected from the 48 harvests; the variables were based on meaningful subsets of the 48 harvests. Heritability was estimated on a family mean basis from components of variance for each variable. Direct selection response for total yield (R) as well as correlated response for total yield (CR), due to selection for a correlated variable, were also computed. The CR/R ratios exceeded 100% for accumulated bean yield of all harvests after 32 mo of age in each experiment. Confidence interval estimates (95% level) of mean CR/R or mean age at which CR/R exceeded 100% suggested an optimum age of 36 mo. Selection for accumulated bean yield at 36 mo‐of‐age was estimated to be 37% more efficient, per year, as selection at 69 mo‐of‐age. The only other variable with potential as an efficient selection criterion was accumulated bean yield for the 5 mo of typically maximum bean production during the first 2 yr, which still requires trees to be 43 mo old before selections are made.