Dairy farms usually decide when to run fans and sprinklers from a single number, the Temperature-Humidity Index (THI), which combines air temperature and humidity. THI ignores air speed, and air speed is the main reason a fan cools a cow at all. So a THI thermostat keeps signalling "heat stress" in a barn where the fans are already doing their job, and the cooling runs for hours it may not need to. A wind-aware index such as the Equivalent Temperature Index for Cattle (ETIC), which adds air velocity to temperature and humidity, could switch cooling on and off more precisely. The payoff would be lower electricity and water bills with the same cow comfort. It could also save less than hoped, or quietly let heat load build up over several days. Nobody has run the randomized field trial that would settle this, and nobody has priced the answer.
We propose a stepped-wedge cluster randomized trial: barns switch from THI-triggered to ETIC-triggered cooling at randomly assigned, staggered dates. The cooling hardware stays exactly the same in both conditions. Only the rule that turns it on changes. Because heat harm in cows accumulates with lags, we analyze the trial with models whose treatment effect can change with time since switching. We also set the non-inferiority margin in money: the largest milk loss the utility savings can cover. The deliverable is a direct answer to a farmer's question, namely whether adding an anemometer to the barn's controller pays back and how fast.
Core Claim
The claim is that ETIC-triggered cooling cuts fan-plus-spray energy and water use compared with THI triggering. It does so without a cumulative rise in core body temperature or loss of milk larger than the economic margin. It holds when the effect is estimated as a curve over weeks of exposure rather than as a single immediate shift. Three things are new together. We randomize the control policy, not the hardware. The primary outcome is net farm-level value, with physiology as a guardrail. The analysis is built to detect delayed harm that a standard analysis would hide.
Two findings motivate the claim. Evaluation of Thermal Indices as the Indicators of Heat Stress in Dairy Cows in a Temperate Climate studied 826 Holstein cows. It found ETIC tracked respiration rate most closely (r = 0.850) and the Comprehensive Climate Index tracked rectal temperature best (r = 0.672), so wind-aware indices carry physiological signal that THI lacks. Effect of Mixed-Flow Fans with a Newly Shaped Diffuser on Heat Stress of Dairy Cows Based on CFD compared a redesigned fan barn with a standard barn. There, THI flagged moderate stress all day, while ETIC flagged it only around 13:00. That gap is the runtime a THI thermostat may be wasting.
The same fan paper also constrains the design. It lowered average ETIC from 26.1 to 24.8 yet found "no significant difference" in respiration rate or rectal temperature. That result is compatible with two readings: the ETIC change did not matter physiologically, or a two-barn, cross-sectional comparison was too weak and too short to see a lagged response. Our design is meant to tell these apart.
We do not claim ETIC is the right mechanistic model. It is an empirical regression-based index like the others. We claim only that it is a cheap, deployable control signal whose economic value can be tested.
Research Program
Setting and clusters. Clusters are separately ventilated barns or pen groups, drawn from several cooperating farms that already have fans and spray lines. Separate ventilation matters because airflow spills across open pens, which rules out randomizing individual cows. Every cluster starts on the farm's existing THI rule. At randomized steps across the hot season, clusters cross to the ETIC rule. By the end, all clusters are on ETIC. The number of clusters and steps will be fixed by simulation before enrolment, not guessed here.
Power calculation. The simulation uses the variance of core body temperature and milk yield in continuous-logger data from MmCows: A Multimodal Dataset for Dairy Cattle Monitoring. That dataset recorded 16 Holsteins with vaginal temperature loggers, ankle posture sensors, indoor THI and individual milk yields.
Stratification. Step order is randomized in blocks that balance cluster-level mean lactation stage and parity. Critical Temperature-Humidity Index Thresholds Based on Surface Temperature for Lactating Dairy Cows in a Temperate Climate reports that lactation stage shifts surface-temperature breakpoints (p < 0.05), so an unbalanced herd mix could fake or mask a policy effect.
The two control rules. Both rules drive the same fans and nozzles. Only the sensor and threshold differ. The ETIC rule uses anemometers at cow height in the resting and feeding zones, because barn-scale stations miss local air pockets, as the CFD fan work shows. We set the ETIC switching threshold on a pilot barn before randomization: the ETIC value at which logged core temperature starts to rise matches the physiological point implied by the farm's current THI threshold. This keeps the comparison about signal quality, not about one rule simply being more lenient. The threshold is pre-registered.
Endpoints. Every endpoint is automated, because neither farmers nor cows can be blinded to fans. The physiological guardrail is continuous core body temperature from vaginal loggers in a sampled subset of cows per cluster. Supporting endpoints are:
- lying time from ankle loggers; MmCows reports sensor-fusion F1 of 0.997 for lying and 0.938 for standing, so the measure is trustworthy;
- respiration rate from video;
- daily parlor milk yield.
Metered kWh and water volume per cluster-day are the resource endpoints.
Analysis. Following Analysis of stepped wedge cluster randomized trials in the presence of a time‐varying treatment effect, we fit an exposure-time-indicator model and a natural-cubic-spline model. Each estimates the policy effect separately for each week since switching, with calendar-period fixed effects to absorb the heat waves the whole region experiences. That paper shows the standard immediate-treatment estimator is a weighted sum of these time-specific effects, with negative weights at late exposure times. Under lagged effects it can badly miss the truth and even flip sign. That failure mode is exactly the one that would make a slow-building harm look like no harm.
Economic layer. The economic analysis is built into the trial, not added afterwards. The non-inferiority margin for milk is defined, before unblinding, as the daily milk loss per cow whose farm-gate value equals the expected utility saving. The trial therefore tests "savings exceed losses" directly.
Net present value (NPV) and payback period combine four inputs:
- sensor capital cost and controller retrofit;
- maintenance and calibration labour, since the tool-fabrication example in Design and experiment of multi-tooth clip cutter shows that hidden labour can offset material savings;
- measured kWh and water savings valued at each farm's actual tariffs;
- the trial's estimated milk effect, including its full confidence interval, propagated by resampling.
The result is a distribution of NPV rather than a point estimate. The concern that engineering papers report static savings without depreciation or price variation is exactly why we vary tariffs and milk price in sensitivity analysis.
Baselines
Status-quo THI rule. The reference point is each farm's existing THI-thermostat rule. We do not include a true no-cooling arm, because withholding cooling in hot weather is unacceptable welfare-wise and irrelevant to the adoption decision. The THI rule is also the strongest competitor. It is what farms already run, it needs no new sensors, and it errs toward over-cooling, so it is the hardest rule to beat on cow outcomes. It runs on the same clusters, hardware and season as the ETIC rule, and the stepped design gives every cluster time under both.
We found no reported result on Lacuna for the energy or water use of THI-triggered barn cooling. The nearest reported numbers are physiological: ETIC's r = 0.850 with respiration rate, and the "no significant difference" in rectal temperature and respiration rate in the fan study despite lower ETIC. Beating this baseline means a confidence interval for resource savings that excludes zero, together with confidence intervals for milk and core-temperature effects at every exposure week that stay inside the economic margin.
Two-barn parallel comparison. The second baseline is the design of the fan study itself: one barn per policy, compared cross-sectionally. We reproduce it by taking pairs of clusters in the same calendar window from our own data, with one on each rule. We then show how wide its uncertainty is and how often it disagrees with the full stepped-wedge estimate.
Immediate-treatment analysis ablation. The third is an analysis ablation on identical data: the standard stepped-wedge model with a single immediate effect. This isolates the key ingredient, lag-aware estimation. If the immediate-effect model declares ETIC non-inferior while the exposure-time model shows core temperature drifting upward in later weeks, the ablation has shown what the design buys.
Sensor-cost ablation. Finally, we recompute NPV assuming ETIC control needs only one barn-level anemometer instead of cow-height sensors in several zones. This tests whether the spatial detail that the CFD work motivates is worth its cost.
Why not just…?
Why not just keep the THI thermostat, since it is cheap and standard? THI may well be good enough, and this is a live possibility. Its known weakness is over-triggering in ventilated barns: the fan study found stress flagged all day by THI but only near midday by ETIC. Over-triggering costs money, not cow welfare. The trial settles the question in the farmer's terms. If utility savings are small relative to sensor and upkeep cost, the NPV distribution will say so, and we will report THI as the rational choice.
Why not just do a before/after test or compare two barns? Before/after comparisons mix the policy effect with the season's heat waves and with herd lactation decline. Effect of heat stress on blood biochemistry and energy metabolite of the Dazu black goats is a sequential-exposure study that is exposed to exactly this confounding. A two-barn comparison has two clusters, so barn-specific quirks cannot be randomized away. The fan study's null physiological result is uninterpretable for this reason. The paired-cluster reanalysis in the Baselines section measures how misleading that design would have been on our own data.
Why not randomize the policy day by day within each barn, which is cheaper and gives many more randomization units? This is a serious alternative, and it would work if the policy's effects were immediate. The danger is cumulative. Less cooling runtime could slowly depress intake and milk over days, and daily switching would dilute that effect toward zero, which biases a non-inferiority trial toward a false "safe." We keep sustained exposure periods for this reason. We also add an embedded check: in two clusters, a short run of randomized day-level switching before their step lets us compare the immediate effect with the week-by-week exposure curve.
Why not simulate the policies on existing sensor data such as MmCows? Recorded data only show what happened under the cooling that actually ran. They cannot show how cows would have responded to less runtime. MmCows also logs indoor THI but not air speed at the cow. We use it only for power calculations and endpoint validation.
Why not report milk yield alone and let farmers decide? A milk-only result tells a farmer nothing about whether the sensors pay for themselves. The ~24% milk drop that EFFECT OF HEAT STRESS ON PRODUCTION PERFORMANCE IN DAIRY ANIMALS observed between comfortable and stressful THI in Sahiwal cows shows the stakes of under-cooling. That paper did not translate the loss into margins. Our economic margin does that translation explicitly.
Proposed Experiments and Decisive Checks
Pilot calibration. The pilot calibrates the ETIC threshold against logged core temperature. It passes if the pre-registered threshold reproduces the THI rule's physiological trigger point within the pilot's measurement noise.
Main trial. The main trial reports three things:
- resource savings per cluster-day;
- the exposure-time curves for core temperature, lying time, respiration and milk;
- the NPV and payback distribution.
Decisive outcomes. The central test is whether the upper confidence bound of the milk-loss curve stays below the economic margin at the last exposure weeks, not just on average. A positive average with a late-week breach would mean the ETIC rule defers harm rather than avoiding it. The immediate-effect ablation will show whether a conventional analysis would have missed that breach.
Secondary analyses. We test whether savings differ by lactation-stage stratum, since early-lactation cows have higher breakpoints. We also test whether the sensor-cost ablation changes the adoption verdict.
Interpreting the results. A clear null on savings is equally publishable. It would tell farms that THI thermostats are not leaving money on the table.
References
- Evaluation of Thermal Indices as the Indicators of Heat Stress in Dairy Cows in a Temperate Climate
- Effect of Mixed-Flow Fans with a Newly Shaped Diffuser on Heat Stress of Dairy Cows Based on CFD
- MmCows: A Multimodal Dataset for Dairy Cattle Monitoring
- Critical Temperature-Humidity Index Thresholds Based on Surface Temperature for Lactating Dairy Cows in a Temperate Climate
- Analysis of stepped wedge cluster randomized trials in the presence of a time‐varying treatment effect
- Design of multi-tooth clip cutter for cyclone milling machine
- Effect of heat stress on blood biochemistry and energy metabolite of the Dazu black goats
- EFFECT OF HEAT STRESS ON PRODUCTION PERFORMANCE IN DAIRY ANIMALS