Cultivars for use in low‐productivity environments (LPE) can be selected directly, in LPE, or indirectly, in medium‐ or high‐productivity environments (MPE or HPE, respectively). These contrasting strategies were compared by means of predictions and an empirical selection experiment in which random populations of oat (Avena sativa L.) lines were selected for grain yield in LPE, MPE, or HPE; selected lines were then evaluated at each productivity level. The influence of replication (across environments) of the selection unit on the relative effectiveness of direct and indirect selection for yield in LPE was also examined. At low levels of replication, the greatest predicted and realized response in LPE resulted from selection in MPE. However, when replication was increased, direct selection resulted in the greatest predicted and realized yield gains in LPE. A positive correlation between degree of replication and the relative effectiveness of direct selection in LPE is expected when: (i) heritability (h2) is greater in HPE than in LPE and (ii) when the genetic correlation (rg) between yield in HPE and in LPE is less than one. These conditions were shown to exist in the present case, and may occur often in plant breeding programs. The results indicate that separate breeding programs may result in the greatest yield gains for some low‐productivity cropping systems. Estimates of h2 in HPE and LPE, and rg between yields in HPE and LPE, can be used to predict such situations.