When crop varieties are bred for use in low-productivity environments (LPE), it must be decided whether to select directly, in LPE, or indirectly, in high-productivity environments (HPE). The relative performance of these strategies depends upon both the genetic correlation (r[subscript] G) for yields between and the heritabilities (H[superscript]2) within environments. It was hypothesized that direct selection in LPE may be more effective than indirect selection in HPE in some cases, and that such cases can be predicted on the basis of estimates of r[subscript] G and H[superscript]2;These hypotheses were tested in a population of 116 random oat lines tested in 36 yield trials. These trials were classified as LPE, MPE (medium-productivity environments), or HPE according to their mean yields. Among the 12 designated as LPE, individual trials were low yielding due to N deficiency, P deficiency, or heat stress caused by late sowing. Estimates of H[superscript]2 for grain yields within and r[subscript] G among productivity levels were used to predict expected responses in LPE to selection in LPE, MPE, and HPE. H[superscript]2 was highest in HPE, but r[subscript] G between yields in LPE and HPE was only 0.59. Estimates of r[subscript] G between nonstress and P-deficient, N-deficient, and heat-stressed environments were 0.5 ± 0.24, 1.08 ± 0.16, and 0.06 ± 0.24, respectively, indicating that P-deficient and heat-stressed environments were responsible for the low r[subscript] G between yields in LPE and HPE. For 10% selection based on line means in 2 or 4 two-replicate trials, the greatest yield gain in LPE was predicated to result from selection in MPE, but for selection in 12 six-replicate trials, direct selection in LPE was superior. These predictions were tested in three empirical selection experiments, wherein comparisons of direct and indirect selection for grain yield were made in two populations of oat lines tested in a total of three sets of environments. In two of these experiments, direct selection of LPE was superior to indirect selection in HPE. In all three, increased replication improved the efficiency of direct selection in LPE. These results confirm that neither HPE nor environments in which H[superscript]2 is greatest necessarily maximize yield gain in LPE.