Weather-station THI underestimates night-time heat load in free-stall dairy barns: hourly agreement across 172 sensors in Brazil and Uruguay

Abstract The temperature-humidity index (THI) computed from weather-station data underpins much of the heat-stress literature in dairy cattle, although cows live in the barn microenvironment. We assessed agreement between the two sources using 372,633 paired hourly records from 172 sensors installed in free-stall barns across 30 municipalities in Brazil and Uruguay, between September 2024 and September 2025. THI was recomputed for both sources with the National Research Council (1971) equation, because the original database applied different rounding rules to the two sources, which alone created 0.53 THI units of bias. Inference relied on a two-stage bootstrap (cities, then devices within cities), since one city holds 82 sensors. Pooled mean bias was 3.05 THI units (95% CI 2.03 to 3.39) and the across-city mean bias was 1.87 (1.17 to 2.42). The discrepancy is not constant: it reaches 4.69 units when station THI is below 50 and falls to 0.84 between 68 and 72, vanishing above 72. Bias peaked during summer nights and was lowest in the afternoon, driven mainly by air temperature at night and by humidity in the afternoon. Consequently, 11.2% of hours deemed non-stressful by the station were already stressful in the barn, and summer nights averaged 0.73 h below THI 64 inside the barn against 3.54 h at the station. Bias ranged from − 3.0 to + 3.7 units across cities, ruling out a single correction factor; both negative values came from sites whose sensor recorded a wider diurnal cycle than the station, indicating a pairing failure.