The effects of pre-flowering drought on millet were studied in two experiments, an on-station, dry-season trial under irrigation and an on-farm trial where drought was induced by natural rainfall regimes (one wet, two dry years). The on-station experimental conditions were adjusted to those of Sahel farms, particularly through low plant density, low soil fertility and an irrigation regime that simulated the natural rainfall pattern. The replanting (5–10%) of missing hills in the on-farm trial increased stand variance significantly, whereas small quantities of applied N reduced it. Grain yield and grain number were equally seriously affected by drought in both experiments; yields were 700–800 kg ha⁻¹ in the control years compared with 200–300 kg ha⁻¹ in drought years. Both experiments showed sub-populations of early-, intermediate- and late-flowering millet plants. Fertile tillers were concentrated in the group of early-flowering plants. Only this group when subjected to moisture stress could secure part of the grain production by delaying, to a maximum of 18 d, the development of fertile tillers beyond the period of stress. Later-flowering plants, mostly uniculm, would be either sterile or unaffected by drought, depending on the exact coincidence of panicle development and drought. Drought did not alter the number of fertile hills per stand, because earlier- or later-flowering plants in the same hill compensated for sterile plants. The period of flowering in a stand was stretched over 40–45 d independently of drought, and this occurred both in the on-station experiment comprising 19 plants per treatment and in the on-farm experiment comprising about 700 hills. Although stand variance was reduced by better crop management and resulted in more fertile hills, more panicles per hill, and taller hills of more uniform height, the spread in flowering period of the millet population tended to be unaffected. Stand variability therefore appeared to be of paramount importance in accounting for drought avoidance.