Impact of heat stress on feed intake, milk yield, milk composition, and feed efficiency in dairy cows: a meta-analysis

Heat stress compromises dairy production by decreasing feed intake, and milk yield, and may alter milk composition and feed efficiency. However, little information is available for evaluating such effects across different levels of heat stress and cows enrolled in heat stress studies. The objectives of this study were to evaluate the effects of heat stress on dry matter intake (DMI), energy-corrected milk (ECM), milk composition, and feed efficiency (kg ECM/kg DMI) and to investigate the relationship between such effects and heat stress intervention and animal characteristics by using meta-analytical approaches. Data from 31 studies (34 trials) fulfilled the inclusion criteria and were used foranalysis. Results showed that heat stress decreased DMI, ECM, and milk protein concentration, but did not alter milk fat concentration or feed efficiency. Meta-regression confirmed that such reduction in DMI and ECM was significantly associated with increasing temperature humidity index (THI). Over the period of heat stress, for each unit increase in THI, DMI and ECM decreased by 4.13% and 3.25% in mid-lactation cows, respectively. Regression models further revealed that a strong interaction between THI and lactation stage existed, which partially explained the large heterogeneity in effect size of DMI and ECM. Results indicated a need for more research on the relationship between the effect of heat stress and animal characteristics. This study calls for implementation of mitigation strategies in heat-stressed herds due to the substantial decrease in productivity.


INTRODUCTION
Heat stress is a growing concern in sustainable dairy production given the rising global temperature and more frequent temperature extremes in temperate areas (Henry et al., 2018;Polsky & von Keyserlingk, 2017).Beyond dairy cows naturally accumulating large amounts of internal heat load due to metabolism for milk synthesis (Coppock, 1985;Purwanto et al., 2009), excessive exposure to elevated temperature and humidity may exceed their ability to dissipate the accumulated heat load.Such temperature and humidity conditions are usually quantified by a compound index named temperature-humidity index (THI).THI is commonly used to measure the occurrence and severity of heat stress showing a good correlation with physiological parameters including respiration rate and body temperature in heat-stressed dairy cows (Thom, 1959;Yan et al., 2021).Cows are capable of coping with a certain range of THI while maintaining their productivity.However, when THI exceeds a certain threshold, cows fail to manage their thermal balance, heat stress occurs and milk production, reproductive performance, and overall welfare are compromised (Becker et al., 2020;Polsky & von Keyserlingk, 2017), which may further lead to large economic losses (Gunn et al., 2019).
During periods of heat stress, several physiological and behavioral coping mechanisms are activated to dissipate the excessive heat load.One obvious way is to minimize the generation of internal heat load by decreasing metabolic activities, which happens at the cost of feed intake and milk production (Becker et al., 2020;Tao et al., 2020).The change in metabolic activities and the cows' physiological status often modifies milk composition (Becker et al., 2020).However, reduction in feed intake and milk production do not always happen in parallel, thus decreased feed intake may only partially explain the reduction in milk production (Wheelock et al., 2010), which indicates a potential modification in feed efficiency.Feed efficiency measures how efficient and sustainable dairy cows are at converting feed to dairy products for human consumption.Although several studies have investigated feed efficiency in heatstressed dairy cows, they come to different conclusions.Reduction in feed efficiency has been observed in a few studies (Bouraoui et al., 2002;Karimi et al., 2015) and may be attributed to an increase in maintenance energy requirements under heat stress (NRC, 2001).In contrast, feed efficiency has been reported as stable or even increased in some heat stress studies (Hall et al., 2016;Hill & Wall, 2017;Rungruang et al., 2014).
The size of the heat stress effect is influenced by a combination of intensity and duration of exposure to elevated THI and by cows' physiological status (Tao et al., 2020).Knowing how these factors affect milk production may contribute to the development of stressrelief interventions that help optimize production.Such knowledge is crucial to minimize the environmental impact, for instance, by diluting enteric methane emissions (Løvendahl et al., 2018).Many previous studies on the heat stress effect in dairy cows are location-specific, with outcomes usually measured at single THI levels.These studies differ with respect to enrolled cows, experimental designs, and heat stress interventions.A few studies reviewed the impact of heat stress on milk yield and composition (Tao et al., 2020;West, 2003); however, such studies neither compared the variation of effect size among different studies and the interaction between heat stress effect, animal, and stress levels nor were they conducted systematically (no predefined research question or specific search strategy).Although a recent meta-study quantified the negative correlation between DMI and THI, with parity and lactation stage being investigated, the inclusion criteria included beef breeds and cooling intervention which were too broad to discuss the relationship between dairy cows and their environment (Chang-Fung- Martel et al., 2021).Therefore, the objectives of this study were 1) to perform a meta-analysis to evaluate the effects of heat stress on DMI, ECM, milk composition, and feed efficiency in dairy cows; 2) to evaluate the relationship between the effect size of heat stress and the severity of heat stress; and 3) to identify the variations of effect size of heat stress in different parities and lactation stages of cows.

Identification of eligible studies
Peer-reviewed and English-written studies with original data that directly compared feed intake, milk yield and composition, and feed efficiency of lactating dairy cows under both heat stress and thermoneutral conditions were considered for inclusion.Heat stress was defined as THI exceeds 72 (Armstrong, 1994).Eligible studies had to report THI values or had to report both temperature and humidity to enable the calculation of THI.Included studies had to directly report the outcomes as means plus error measurements of means.The inclusion criteria are listed in Supplementary Table 1 (http: / / doi .org/ 10 .6084/m9 .figshare.23895948.v3;Chen, 2023).
Web of Science (WOS) and PubMed were accessed on 03, March 2023 limited to English-language studies.The following terms and their combination: ("dairy cows" OR "dairy cattle") AND ("heat stress" OR "heatstressed" OR "temperature humidity index" OR "THI") AND ((("milk yield" OR "milk production") AND ("dry matter intake" OR DMI OR "feed intake")) OR "feed efficiency") were searched in the above databases in titles, abstracts, and keywords (only titles and abstracts in PubMed).Originally, 378 studies were found on WOS and 135 studies were found on PubMed.After duplicate removal, in total 384 studies were prepared for screening for their eligibility.Screening for eligible studies was conducted by 2 authors separately and at 3 levels: titles, abstracts, and full-text articles to identify qualified studies for further analysis.The screening process was conducted shown in Supplementary Figure 1 (http: / / doi .org/ 10 .6084/m9 .figshare.23895948.v3;Chen, 2023).

Data Collection and Calculations
An Excel spreadsheet was designed to compile data from eligible studies.The compiled data were doublechecked by co-authors.For studies testing the mitigation effect of feed additives in heat-stressed cows, only control groups (no feed additives) were selected.For studies that reported more than one trial, data were organized separately.Daily mean and maximum THI were collected from included studies.Since different equations for THI were used by these included studies, daily mean and maximum THI values were recalculated based on the following equation when temperature and humidity were provided from NRC (1971): where T was the temperature in degrees Celsius, and RH was the relative humidity.
For studies that did not report THI values, but which had figures showing daily fluctuation of THI, daily mean and maximum THI were estimated from the figures.For studies that only reported daily minimum and maximum THI, mean THI was calculated as the mean of those 2 figures.
The duration of heat stress was counted as the total days of exposure to a heated environment despite that the measurements of outcomes may have only been taken within a shorter period of total exposed days.The outcomes were directly collected from enrolled studies as mean values over thermoneutral or heat-stressed Chen et al.: Impact of heat stress: a meta-analysis periods.The ECM yield for each study was calculated using the equation developed by Sjaunja et al. (1990): ECM = milk yield × (383 × fat % + 242 × protein % + 783.2)/3,140.[2]Feed efficiency was defined as the ratio between kg ECM and kg DMI.If studies reported both ECM and DMI and feed efficiency, their feed efficiency was not used, instead, we calculated feed efficiency by dividing ECM by DMI.In cases of missing Standard Deviation (SD) or Standard Error (SE) values, we calculated them from P-values, confidence intervals, and quantiles (Weir et al., 2018).The SE of raw milk yield was applied for the analysisof ECM.The SE of feed efficiency was calculated by assuming the correlation of milk yield and DMI at 0.75 (Kramer et al., 2008).

Statistical analysis
Response ratio (outcomes measured under heat stress divided by control measured under thermoneutral conditions) in log-scale was applied to assess the effect of heat stress on all outcomes.The log-transformed ratio of means was then converted back to the proportion of effects measured under heat stress to non-heat stress performance to calculate the reduction percentage of outcomes under heat stress.
All statistical analysis was conducted in R using the "metafor" package (Viechtbauer, 2010).Random-effect models were used to quantify the effect sizes assuming that effect size varies among studies, and weights were assigned to each study based on inversed standard deviation of each measured outcome.Study heterogeneity was estimated by I 2 as the percentage of total variation across all included studies rather than the sampling error; I 2 was calculated as where Q was the χ 2 heterogeneity statistic (Cochran, 1954) and k was the number of studies.Publication bias usually derives from small studies with nonsignificant results.Egger's regression test (Egger et al., 1997) and contour-enhanced funnel plots were applied to assess publication bias.Given our objective to assess the variation between studies, subgroup meta-analysis and meta-regression were performed to analyze between-study heterogeneity.Included studies were categorized by the type of intervention (chamber studies versus seasonal studies), lactation stage (early: 0 to 100 d; mid: 101 to 200 d; late: > 200 d), and parity.However, due to the majority of included studies (30 out of 31 studies) only using multiparous cows in their experiments, parity was not analyzed.During metaregression, covariates considered included duration of heat stress, daily mean and maximum THI values under heat stress, and lactation stage (early; mid; late) and their 2-way interactions.
where Y denoted the overall effect size, β 0 denoted the intercept, β 1 , β 2 , β 3 , and β i were the coefficients of explanatory variables, ω was the between-study error, ε was the within-study error.THI included daily mean THI and daily maximum THI.
Covariates were removed once strong correlations (Pearson correlation coefficient | r | > 0.5) were detected as suggested by (Kebreab et al., 2023).Metaregression was conducted by regressing observed effect size against the above identified covariates.The development of models was performed using a forward stepwise approach.Initially, univariate meta-regression was performed and covariates with P-values smaller than 0.2 were then selected for multivariate meta-regression.All models were fitted with rma() function integrated in "metafor."The model comparison was conducted by ANOVA and only variables and their interactions with P-values smaller than 0.1 remained in the final model.Developed models were assessed by permutation test on their robustness (Higgins & Thompson, 2004).
Since the SE of feed efficiency was calculated, a sensitivity meta-analysis was conducted for feed efficiency by using the number of cows in the heat stress group as an alternative way to weigh included studies.

RESULTS
In total, 31 eligible studies were identified, these studies were carried out in Holstein cows originating from 7countries.The characteristics of these studies are listed in Table 1.Among these studies, 6 investigated heat stress effects in different seasons (seasonal studies), while 25 studies were carried out in climate chambers simulating heat environments (chamber studies).Three chamber studies reported 2trials, enabling 34 trials available for the analysisof all outcomes.Among these trials, 5, 25, and 4trials were conducted with cows in early, mid, and late lactation, respectively.Descriptive statistical analysis for all outcomes was made for both thermoneutral and heat stress groups (Table 2).The range of THI under heat stress indicated that the included studies covered both mild and moderate heat stress levels (Armstrong, 1994).
Chen et al.: Impact of heat stress: a meta-analysis Publication bias was assessed by performing Egger's regression test for all studies and assisted by contourenhanced funnel plots (Table 3; Supplementary Figure 2; http: / / doi .org/ 10 .6084/m9 .figshare.23895948.v3;Chen, 2023).In general, we did not detect obvious asymmetry for any of the outcomes, indicating no evidence of publication bias.
One of the objectives of this meta-study was to quantify the effects of heat stress on dairy production.Overall, heat stress decreased DMI and ECM by 19.3% and 17.9%, respectively (Table 3).Compared with thermoneutral conditions, heat stress reduced milk protein concentration by 3.9% (P < 0.001), but had a small, insignificant effect on milk fat concentration (0.1%; P = 0.96).Feed efficiency was 5.4% higher during heat stress; however, the effect was statistically insignificant (P = 0.17).
The effects of heat stress were associated with large heterogeneity (Table 3; Supplementary Figure 3; http: / / doi .org/ 10 .6084/m9 .figshare.23895948.v3;Chen, 2023).Subgroup analysis showed that ECM decreased more in seasonal studies than in chamber studies (25.8 versus 16.0%, P = 0.04).There was no difference between the chamber studies and the seasonal studies of heat stress effect on DMI (P = 0.33).Heat stress was overall associated with a very small, statistically insignificant effect on milk fat concentration.However, analyzing chamber studies and seasonal studies separately showed that in chamber studies heat stress increased milk fat by 2.3% (P = 0.02), whereas in seasonal studies heat stress reduced milk fat by 8.3% (P = 0.07).Heat stress reduced milk protein concentration in most studies (Supplementary Figure 3; http: / / doi .org/ 10 .6084/m9 .figshare.23895948.v3,Chen, 2023), i.e., mean reduction was 3.9% across all studies (P < 0.001), where no difference was found between the chamber and seasonal studies.We did not observe any differences in milk protein (P = 0.91) or feed efficiency (P = 0.42) between the chamber and seasonal studies.
Meta-regression was then performed to account for the variation of effect sizes in DMI, ECM, and protein concentration.Due to differing ECM between types of studies, only chamber studies were selected for metaregression to account for the variation in effect sizes.Given the strong correlation (r = 0.68) between daily mean THI and daily maximum THI under heat stress conditions, daily maximum THI was dropped for further analysis.
Univariate meta-regression revealed that THI was the key covariate (with the smallest P-values; Table 4) among all covariates for DMI and ECM but not for protein concentration.Protein concentration was associated with duration of heat stress (P < 0.01; Table 4) but was not affected by THI.The lactation stage itself did not contribute to the variation in effect size for any outcomes (Table 4).Strong interactions were detected for all outcomes and contributed to explaining the variation in the effect size of heat stress.
The final models of multivariate meta-regression and their parameters are listed in Table 5. Multivariate models with interactions were better than univariate modes in explaining the variation in effect sizes (Table 4 and Table 5).THI had a strong interaction with lactation stage on DMI and ECM (Table 6).Compared with early lactation, mid and late-lactation cows were more tolerant to heat stress when THI was below 77 (Figure 1).In mid lactation, for every unit increase in THI, the effect size of DMI and ECM dropped by 0.0422 and 0.033, corresponding to a reduction in DMI and ECM by 4.13% and 3.25%.However, when THI was above 77, the effect sizes of mid and late lactation cows became larger than early lactation cows (Figure 1).Late lactation decreased DMI and ECM more than mid-lactation with THI values over 77.THI also interacted with the duration of heat stress.For ECM, such interaction tended to buffer the effect size on ECM (P = 0.09) under prolonged heat stress.On the contrary, the interaction between THI and duration tended to intensify heat stress in protein concentration (P = 0.09).

Sensitivity analysis
Given the fact that the variance of feed efficiency was a function of the variance of milk yield and DMI, the calculation of the variance might introduce extra error.Our data did display the possibility that a few studies with low variance contributed much more to the overall effect size than others (Supplementary Figure 3; http: / / doi .org/ 10 .6084/m9 .figshare.23895948.v3;Chen, 2023).Hence, a sensitivity analysis was performed to synthesize studies by weighing them according to their sample size.The result agreed with the inversedweighing method and showed that heat stress did not change feed efficiency (P > 0.99).

DISCUSSION
By using meta-analytical approaches, our study ensured a robust estimation of heat stress effects and the identification of sources of variation in effect sizes (Gurevitch et al., 2018).Our results demonstrated large negative effects of heat stress on DMI and ECM which emphasized the severity of heat stress impact on dairy production.These results are in line with reviews investigating feed intake and milk yield in heat-stressed cows (Chang-Fung- Martel et al., 2021;Tao et al., 2020;West, 2003).Among these reviews, Tao et al. (2020) summarized that for every unit increase of THI above 72, milk yield reduction ranges from 0.30 to 0.88 kg, meaning that cows do not suffer equally under identical climate conditions.This finding is consistent with evidence from our study showing large heterogeneity for DMI, ECM, and milk composition.Our meta-regression further confirmed that such heat stress effects were correlated to both the duration and severity of heat stress and varied among lactation stages.Heat stress directly depresses cows' appetite and consequently decreases feed intake (Polsky & von Keyserlingk, 2017).Given that the maintenance expenditure already increases in heat-stressed cows (NRC, 2001), insufficient feed intake leaves less energy for milk synthesis, leading to a reduction in feed efficiency.This assumption is supported by earlier studies that heat-stressed cows produce less milk than cows with matched DMI under thermoneutral conditions (Baumgard et al., 2011;Fontoura et al., 2022;Wheelock et al., 2010).In contrast, our data showed that the decreases in DMI and ECM were parallel, resulting in maintained feed efficiency.One explanation is the lag effect of heat stress on milk yield.Several studies support the idea that heat stress affects milk yield with a time lag ranging between 24 to 48 h (Rhoads et al., 2009;Wheelock et al., 2010).On the contrary, decreased feed intake is usually observed instantaneously when heat stress emerges (Rhoads et al., 2009;Wheelock et al., 2010).When heat stress is researched during a short period, the reduction in feed intake seems larger than the reduction in milk yield, which may phenotypically result in a stable or even increased feed efficiency.This phenomenon is agreed by a recent meta-regression study showing that only cows undergoing an extended period (more than 12 d) of heat stress would start to decrease their feed efficiency (Souza et al., 2023).Another explanation is that the increased digestibility derived from reduced feed intake may compensate for energy deficiency to optimize feed efficiency.Lowered feed intake is generally considered to decrease the passage rate of diet which leaves longer rumen retention time and may increase dietary digestibility (Colucci et al., 1982;Shaver et al., 1986).These activities were also observed in heat-stressed dairy cows (Christopherson & Kennedy, 1983;Gao et al., 2017) and may account for the maintained feed efficiency seen in the present study.However, it is worth remembering that our calculation of feed efficiency did not consider any changes in body reserves, such a definition is influenced by body tissue mobilization and varies among lactation stages (Connor, 2015).Nutrient deficiency during heat stress may mobilize body reserves (NASEM, 2021).This concern agrees with several studies included in this meta-analysis showing that body weight significantly reduced over the experimental periods (Baumgard et al., 2011;Yue, Ding, et al., 2020).In other words, not accounting for changes in body reserves during heat stress may have caused us to overestimate feed efficiency under heat stress.
In the present study, the effect of heat stress on milk fat concentration was inconsistent.The chamber studies showed increased fat concentration during heat stress.This increase in fat has frequently been observed, yet not been explained (Smith et al., 2013;Tao et al., 2018).Regardless of whether heat stress alters fat concentration or not, the amount of fat yield decreases due to depressed milk yield.Milk protein is more sensitive to heat stress in dairy cows (Chang-Fung- Martel et al., 2021), our data also confirmed the idea that heat stress depresses milk protein concentration.Under heat stress, the depletion of amino acids due to oxidative stress, immune response, and gluconeogenesis decreased the availability of amino acids for milk protein synthesis and may account for low milk protein and milk yield (Sammad et al., 2020).The reduction of protein also Chen et al.: Impact of heat stress: a meta-analysis Days in milk were calculated as the midpoint of their time spans for different lactation stages and for both thermoneutral and heat stress groups.
indicates that the estimation of feed efficiency in heatstressed dairy cows solely based on raw milk yield and feed intake without considering milk composition is inappropriate.
Large heterogeneity observed in this study may relate to the differences in heat tolerance of experimental cows and heat stress interventions.Heat tolerance in terms of decline in milk production has low to moderate heritability and varies among breeds, parity, lactation stage, and morphological traits of cows (Macciotta et al., 2017;Nguyen et al., 2016).Such individual variations are usually quantified on the scale of climate indexes (for instance THI) and suggest the need to identify THI thresholds for individual animals (Sánchez et al., 2009).Accumulation of heat load in cows is determined both by the duration and intensity of exposure to heat stress.Our subgroup analysis revealed that seasonal studies tended to show larger effect sizes than chamber studies.This difference may arise from prolonged exposure and slightly higher THI in seasonal studies than chamber studies, which was further supported by our regression analysis showing that both THI and duration of heat stress contribute to explaining the variation in effect size on ECM and protein concentration.However, our data suggested that duration did not alter DMI under heat stress.This result ties well with recent meta-study showing that DMI decreases only during the first 5 d of heat stress over 20 d (Souza et al., 2023).Our metaregression confirmed that heat stress effects were complex, meaning that a single variable is insufficient to account for the variation of effect size.This is supported by Polsky and von Keyserlingk (2017) in their review arguing that THI can only be used as a rough measure of heat stress.To date, different indices have been developed for quantifying heat stress levels with THI as the most used one (Becker & Stone, 2020).These indices are solely based on environmental parameters, such as temperature, humidity, wind speed, and solar radiation, assuming all cows suffer equally under identical environmental conditions.Our analysis demonstrated an interaction between THI and lactation stage on DMI and ECM in heat-stressed cows, which is in line with a review showing that the effect of heat stress may differ between different lactation stages (Tao et al., 2020).Though it is generally believed that high-yielding cows would suffer more stress than low-yielding cows due to a larger internal heat load accumulated from milk synthesis (Spiers et al., 2004), mid-lactation and multiparous may be more sensitive to heat environment (Moore et al., 2023).Our analysis showed that cows in mid or late-lactation stages were more tolerant of mild heat stress than cows in early lactation.However, this phenomenon was reversed under moderate or severe heat stress.These results suggest that beyond the heat  The log ratio of means between heat stress group and thermoneutral group; the effect sizes were converted to true reduction in percentage of heat stress compared with thermoneutral conditions. 3 The percentage of total variation across studies due to heterogeneity rather than sampling error.

4
Egger's test for detecting publication bias in meta-analysis.P-values larger than 0.05 mean that no publication bias was detected.The percentage of total variation across studies due to heterogeneity rather sampling error.
load generated from milk production, the distinct physiological status of the cow, for instance, the negative energy balance and immune status in early lactation and the fetus growth in late lactation, may also have impacts on the effect size of heat stress which needs further investigation.Unfortunately, due to the biased ratio between primiparous and multiparous cows used by the studies we identified, we were unable to analyze the interaction between the effect sizes of heat stress and the parity of stressed cows.
Our study set strict criteria for selecting eligible studies during the literature search which may have caused us to lose some potential for quantifying outcomes individually, because many studies were excluded.Though no obvious publication bias was detected in the present study, only 5 and 4 trials were performed during early and late lactation, respectively, and there are probably important nuances to be explored by extending the number of studies at these 2 lactation stages.For metaanalysis, the risk of bias is usually assessed by applying the Cochrane risk of bias tool (Sterne et al., 2019).This tool is specially developed for clinical trials and is sensitive to experimental design.Considering the dissimilarities between clinical trials and trials identified in the present study, for instance, groups of cows were balanced by milk yield or daysin milk during random-ization, the risk of bias for individual studies was not assessed in the present meta-analysis.Traditionally, weighting studies based on inversed variance can incorporate information on the quality of included studies and decrease the influence of small studies (Gurevitch et al., 2018).For calculated variables with unavailable variance, the sample size may be used to weight studies as supported by our sensitivity analysis showing reasonable estimation of effect size.
Heat stress management is usually performed at the herd or group level by appropriate housing, cooling, or dietary modification (Becker & Stone, 2020).The interaction detected in our study between cows and their surrounding environment may suggest future care at the cow or small group level to sustainably use resources for cooling stressed cows.This work can be combined with the development of a compound heat stress index that incorporates animal characteristics (Becker & Stone, 2020).To implement such strategies in dairy cows, the present study suggests that further research is needed to investigate the effect in early and late lactating cows and maybe for different parities.
To accurately estimate such effects, a prolonged experimental period is necessary to dilute the lag effect in heat-stressed cows.
Chen et al.: Impact of heat stress: a meta-analysis The percentage of total variation across studies due to heterogeneity rather than sampling error.
Chen et al.: Impact of heat stress: a meta-analysisTable 1. Characteristics of studies included for meta-analysisNo.cow; multi = multiparous cow; na = not mentioned.
Chen et al.: Impact of heat stress: a meta-analysis

3P
-values associated with the permutation test of model robustness.4 Daily mean temperature-humidity index under heat stress.5 Lactation stage was categorized as early, mid, and late lactation.6 Duration of heat stress (days).

Table 2 .
Descriptive summary of daily mean temperature-humidity index (THI), duration of heat stress, Days in milk, DMI, milk yield, and milk composition in dairy cows across identified studies (n = 34) on heat stress 1 Standard deviation.2

Table 3 .
Effects of heat stress on daily DMI, ECM, milk composition, and feed efficiency in dairy cows from identified studies (n = 34) on heat stress Item 2

Table 4 .
Univariate analysis of the effect sizes of heat stress on daily DMI, ECM, and milk protein concentration Item 1Daily mean temperature-humidity index (THI) values under heat stress; 2 Duration of heat stress (days); 3 Lactation stage was categorized as early, mid, and late lactation, estimates in mid and late lactation denote the difference compared with early lactation; 4

Table 5 .
Multivariate regression models explaining the effect of heat stress on daily DMI, ECM, and milk protein concentration P-values indicate the relationship between outcomes and the fitted models.

Table 6 .
Parameters of multivariate models for explaining the effect of heat stress on daily DMI, ECM, and milk protein concentration 1Daily mean temperature-humidity index under heat stress.2Duration of heat stress (days).