The objective of this study was to evaluate the effect of ketosis on milk production in early lactation. Data from 1,156 cows enrolled in a previous study from 5 commercial dairy herds in Southern Ontario were analyzed. Blood samples from all cows were obtained 1 week before calving and once a week for the first 2 weeks following calving. Sera were submitted to the laboratory at the University of Guelph to measure blood analytes including β-hydroxybutyrate (BHBA). Results were categorized from 1.0 to 2.0 mmol/L in 0.2 mmol/L increments. Daily milk production for all cows was recorded until 63 d in milk (DIM). Using weekly milk production as an outcome, for each of the selected thresholds, 3 linear mixed models that accounted for repeated measures within cow were fitted to test the effects of ketosis in wk 1 or wk 2 postpartum, or both. The analyses were performed with farm as a random effect or as a fixed effect. Overall, as the concentration of BHBA increased, milk yield was lower in cows with ketosis compared with cows without ketosis. Ketotic cows produced less total milk to 63 DIM when ketosis persisted for the first 2 weeks after calving than if present in wk 1 or 2 (at BHBA ≥ 1.4 mmol/L: −97, −37, and −62 kg, for first 2 weeks after calving, wk 1, and wk 2 after calving, relative to cows without ketosis, respectively, P < 0.0001). Cows ketotic at wk 2 only produced more milk in wk 1 than cows without ketosis (at BHBA ≥ 1.4 mmol/L: 25.9 ± 1.3 kg/d and 23.9 ± 1.2 kg/d, respectively, P = 0.004). When farm was included as a fixed effect, there was a significant interaction between farm and ketosis in wk 2 alone for BHBA ≥1.2 mmol/L, and in wk 1 and 2 for BHBA ≥1.6 mmol/L. The effect of ketosis on yield varied among farms from no significant difference to 7.3 kg/d (P < 0.0001). The degree and duration of ketosis were negatively associated with milk yield in early lactation, but the timing of onset of ketosis affects the association. The threshold at which ketosis reduces milk production in early lactation may vary among farms and this should be explored in a large data set from many herds.