Jukuri, open repository of the Natural Resources Institute Finland (Luke) All material supplied via Jukuri is protected by copyright and other intellectual property rights. Duplication or sale, in electronic or print form, of any part of the repository collections is prohibited. Making electronic or print copies of the material is permitted only for your own personal use or for educational purposes. For other purposes, this article may be used in accordance with the publisher’s terms. There may be differences between this version and the publisher’s version. You are advised to cite the publisher’s version. This is an electronic reprint of the original article. This reprint may differ from the original in pagination and typographic detail. Author(s): M. H Kjeldsen, M. Johansen, M.R. Weisbjerg, A.L.F. Hellwing, A. Bannink, S. Colombini, L. Crompton, J. Dijkstra, M. Eugène, A. Guinguina, A.N. Hristov, P. Huhtanen, A. Jonker, M. Kreuzer, B. Kuhla, C. Martin, P.J. Moate, P. Niu, N. Peiren, C. Reynolds, S.R.O. Williams, P. Lund Title: Predicting CO2 production of lactating dairy cows from animal, dietary, and production traits using an international dataset Year: 2024 Version: Published version Copyright: The Author(s) 2024 Rights: CC BY 4.0 Rights url: https://creativecommons.org/licenses/by/4.0/ Please cite the original version: Kjeldsen, M. H et al. (2024). Predicting CO2 production of lactating dairy cows from animal, dietary, and production traits using an international dataset. Journal of Dairy Science, Volume 107, Issue 9, 6771 - 6784. https://doi.org/10.3168/jds.2023-24414. https://creativecommons.org/licenses/by/4.0/ https://doi.org/10.3168/jds.2023-24414 6771 ABSTRACT Automated measurements of the ratio of concentra- tions of methane and carbon dioxide, [CH4]:[CO2], in breath from individual animals (the so-called “sniffer technique”) and estimated CO2 production can be used to estimate CH4 production, provided that CO2 produc- tion can be reliably calculated. This would allow CH4 production from individual cows to be estimated in large cohorts of cows, whereby ranking of cows according to their CH4 production might become possible and their values could be used for breeding of low CH4-emitting animals. Estimates of CO2 production are typically based on predictions of heat production, which can be calculated from body weight (BW), energy-corrected milk yield, and days of pregnancy. The objectives of the present study were to develop predictions of CO2 produc- tion directly from milk production, dietary, and animal variables, and furthermore to develop different models to be used for different scenarios, depending on available data. An international dataset with 2,244 records from individual lactating cows including CO2 production and associated traits, as dry matter intake (DMI), diet com- position, BW, milk production and composition, days in milk, and days pregnant, was compiled to constitute the training dataset. Research location and experiment nested within research location were included as random intercepts. The method of CO2 production measurement (respiration chamber [RC] or GreenFeed [GF]) was con- founded with research location, and therefore excluded from the model. In total, 3 models were developed based on the current training dataset: model 1 (“best model”), where all significant traits were included; model 2 (“on- farm model”), where DMI was excluded; and model 3 (“reduced on-farm model”), where both DMI and BW were excluded. Evaluation on test dat sets with either RC data (n = 103), GF data without additives (n = 478), or GF data only including observations where nitrate, 3-nitrooxypropanol (3-NOP), or a combination of nitrate and 3-NOP were fed to the cows (GF+: n = 295), showed good precision of the 3 models, illustrated by low slope bias both in absolute values (−0.22 to 0.097) and in percentage (0.049 to 4.89) of mean square error (MSE). However, the mean bias (MB) indicated systematic over- prediction and underprediction of CO2 production when the models were evaluated on the GF and the RC test datasets, respectively. To address this bias, the 3 mod- els were evaluated on a modified test dataset, where the CO2 production (g/d) was adjusted by subtracting (where measurements were obtained by RC) or adding absolute MB (where measurements were obtained by GF) from evaluation of the specific model on RC, GF, and GF+ test datasets. With this modification, the absolute values of MB and MB as percentage of MSE became negligible. In Predicting CO2 production of lactating dairy cows from animal, dietary, and production traits using an international dataset M. H Kjeldsen,1* M. Johansen,1 M. R. Weisbjerg,1 A. L. F. Hellwing,1 A. Bannink,2 S. Colombini,3 L. Crompton,4 J. Dijkstra,5 M. Eugène,6 A. Guinguina,7,8 A. N. Hristov,9 P. Huhtanen,8 A. Jonker,10 M. Kreuzer,11 B. Kuhla,12 C. Martin,6 P. J. Moate,13 P. Niu,14 N. Peiren,15 C. Reynolds,4 S. R. O. Williams,13 and P. Lund1 1Department of Animal and Veterinary Sciences, AU Viborg–Research Centre Foulum, Aarhus University, 8830 Tjele, Denmark 2Wageningen Livestock Research, Wageningen University and Research, 6700 AH Wageningen, the Netherlands 3Department of Agricultural and Environmental Science, University of Milan, 20133 Milano, Italy 4School of Agriculture, Policy and Development, University of Reading, RG6 GAR Reading, United Kingdom 5Animal Nutrition Group, Wageningen University and Research, 6700 AH Wageningen, the Netherlands 6VetAgro Sup, UMR 1213 Herbivores, INRAE, Université Clermont Auvergne, 63122 Saint-Genès-Champanelle, France 7Department of Applied Animal Science and Welfare, Swedish University of Agricultural Sciences, SE-901 87 Umeå, Sweden 8Production Systems, Natural Resources Institute, Luke, 31600 Jokioinen, Finland 9Department of Animal Science, The Pennsylvania State University, University Park, PA 16802 10Grasslands Research Centre, AgResearch Ltd., Palmerston North 4442, New Zealand 11Institute of Agricultural Science, ETH Zurich, 8092 Zurich, Switzerland 12Research Institute for Farm Animal Biology (FBN), 18196 Dummerstorf, Germany 13Department of Energy, Environment and Climate Action, Agriculture Victoria Research, Victoria 3821, Australia 14Faculty of Biosciences, Norwegian University of Life Sciences, Ås 1432, Norway 15Animal Sciences Unit, Flanders Research Institute for Agriculture, Fisheries and Food, 9090 Melle, Belgium J. Dairy Sci. 107:6771–6784 https://doi.org/10.3168/jds.2023-24414 © 2024, The Authors. Published by Elsevier Inc. on behalf of the American Dairy Science Association®. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/). The list of standard abbreviations for JDS is available at adsa.org/jds-abbreviations-24. Nonstandard abbreviations are available in the Notes. Received November 9, 2023. Accepted March 26, 2024. *Corresponding author: maria.kjeldsen@​anivet​.au​.dk https://orcid.org/0009-0000-2091-4721 https://orcid.org/0000-0002-2274-8939 https://orcid.org/0000-0002-6514-9186 https://orcid.org/0000-0002-2881-399X https://orcid.org/0000-0001-9916-3202 https://orcid.org/0000-0002-4391-3905 https://orcid.org/0000-0003-3752-4804 https://orcid.org/0000-0003-3728-6885 https://orcid.org/0000-0002-9325-512X https://orcid.org/0000-0002-0884-4203 https://orcid.org/0000-0001-7855-7448 https://orcid.org/0000-0002-6756-8616 https://orcid.org/0000-0002-9978-1171 https://orcid.org/0000-0002-2032-5502 https://orcid.org/0000-0002-2265-2048 https://orcid.org/0000-0001-6858-1284 https://orcid.org/0000-0001-5500-1607 https://orcid.org/0000-0002-4152-1190 https://orcid.org/0000-0003-1321-6487 https://orcid.org/0000-0002-9113-4500 https://adsa.org/jds-abbreviations-24 mailto:maria.kjeldsen@anivet.au.dk 6772 Journal of Dairy Science Vol. 107 No. 9, 2024 conclusion, the 3 models were precise in predicting CO2 production from lactating dairy cows. Key words: tracer gas, cattle, heat production, model evaluation INTRODUCTION Quantification of enteric methane (CH4) production is increasingly important, as it is required to evaluate CH4 mitigation strategies in greenhouse gas inventories and for calculating the carbon footprint of the beef and dairy industry. However, large-scale direct measurement of CH4 with respiration chambers (RC), GreenFeed head chambers (GF), or the sulfur hexafluoride method (SF6) is difficult, labor intensive, and costly. To address these challenges, models for predicting CH4 production in cows fed specific diets have been developed (Appuhamy et al., 2016; Niu et al., 2018), although between-animal variation of CH4 emission is ignored. Furthermore, CH4- reducing feed additives are foreseen to be implemented in farm practice in the near future, and therefore prediction of CH4 will require models that account for the effect of different additives, which requires a comprehensive dataset. The sniffer technique is an alternative approach to es- timate individual enteric CH4 production in large-scale settings (Madsen et al., 2010; Lassen et al., 2012), and it offers an economically favorable alternative compared with other methods (RC, GF, and SF6). Installation of the sniffer equipment in combination with a concentrate bin will allow measurements of the ratio between concentra- tion of CH4 and concentration of carbon dioxide (CO2) in breath exhaled by the cows, when they visit the bin (Madsen et al., 2010). Compared with the approach of predicting CH4 production by a model, between-animal variation is accounted for by the sniffer technique, as the CO2 production is calculated and used to estimate the individual CH4 production based on the ratio of [CH4]:[CO2] in exhaled breath. The sniffer technique as such is therefore not a quantitative measure of emissions, like RC and GF, but it relies on calculating CH4 emis- sion by combining the predicted CO2 production with gas concentration ratio measured by use of a given instru- ment. The idea is that CO2 production is more accurately predicted from animal, dietary, and production traits than CH4. Therefore, CO2 production from dairy cows can be estimated as in Pedersen et al. (2008) and Madsen et al. (2010) in Equations [1] and [2], respectively: CO2 (L/d) = HPU/d × 180 L CO2/h/HPU × 24 h, [1] where heat-producing units (HPU) are equal to the heat production (HP) of an animal (when expressed in W/d) divided by 1,000 W; and CO2 (L/d) = HP (kJ/d)/21.75 kJ/L CO2, [2] where 21.75 kJ is an estimate of HP when 1 L of CO2 is exhaled due to the metabolism of nutrients of an aver- age cow diet (Chwalibog, 1991). Furthermore, because the unit of HP in Equation [3] is W/d, and 1 W = 1 J/s, therefore HP (kJ/d) = (HP (W/d) × 60 s × 60 min × 24 h)/1,000. One of the equations currently used to estimate HP is based on metabolic BW (BW0.75), ECM (kg/d; Sjaunja et al., 1990), and days in pregnancy (DIP), and it originates from a report by Commission Internationale du Génie Rural (CIGR, 2002), where the following model was developed to quantify needed barn ventilation on group level of dairy cows: Heat production (W/d) = 5.6 × BW (kg0.75) + 22 × ECM (kg/d) +1.6 × 10−5 × DIP3. [3] Measuring the CH4 production from cows in large-scale settings plays a crucial role in identifying low CH4-emit- ting cows, forming the basis for genetic selection aimed at reducing CH4 emission. This approach was used by Manzanilla-Pech et al. (2022), where sniffer data created the basis for calculating genetic correlations between CH4 traits and other phenotypes. The approach of using CO2 as an internal marker has also been used to predict ammonia emissions at barn level (Kai et al., 2017). Measuring gas emissions by RC is considered “the gold standard,” but although the CH4 and CO2 production as such is not measured with the sniffer method, measured CH4 concentration values by the sniffer method are well correlated (r = 0.75, based on random cow effects) with data obtained in RC (Difford et al., 2019). However, based on a minor Danish dataset and using Equation [3] in combination with Equation [1] on data from RC, Hell- wing et al. (2013) concluded on a dataset, which is now a minor part of the current training dataset from which the models are derived, that the sniffer method under- estimated the actual production of CO2 and thereby the production of CH4 as well. A part of the explanation lays in the use of HP as an intermediate step to calculate CO2 production, as HP is dependent on the energy balance of the cow (Huhtanen et al., 2020) and nutrient composition of the diet (Kirchgessner and Muller, 1998). The sniffer method only measures the concentration of CO2 and CH4, and, to estimate CH4 emission from cattle, it relies on a prediction equation for CO2 production. We hypothe- sized that CO2 production can be predicted directly from dietary variables, milk production, and animal variables. The objectives were to (1) identify variables that explain variance in CO2 production from dairy cows, (2) develop a CO2 prediction model with the most determining vari- Kjeldsen et al.: PREDICTING CO2 PRODUCTION Journal of Dairy Science Vol. 107 No. 9, 2024 6773 ables, ignoring that some variables may be difficult to obtain on farms, and (3) develop models that can be ap- plied on commercial farms to estimate CO2 production from dairy cows. MATERIALS AND METHODS Dataset Members of the Feed and Nutrition Network of the Global Research Alliance on Agriculture Greenhouse Gases (FNN, 2023) provided data for the present study. As some research groups have more than 1 experimen- tal location, the dataset contains data from 12 research groups, covering 15 different locations in North Amer- ica, Europe, and Oceania, derived from 76 experiments conducted from 1989 until 2019 (Tables 1 and 2). Some limitations for inclusion of data were predefined: (1) to ensure data quality, measurement of gas exchange should have been performed either in RC or with GF (C-Lock Inc., Rapid City, SD); (2) only data from lactating dairy cows were included; (3) records of CO2 production should be available; and (4) data had to be on an indi- vidual animal level. The initial dataset contained 3,179 individual animal records. Because BW, ECM yield, and DIP were data used to predict HP (CIGR, 1984), they were considered as being highly important in the present study as well due to their expected correlation to CO2 production, but not all datasets included DIP. Data Pre-Processing Data pre-processing was necessary before model development to cope with incomplete and inconsistent records, or use of different units for a given variable. Re- cords based on a diet containing monensin were excluded (n = 23) given its noncompliance with EU regulations. Despite the feed additives 3-nitrooxypropanol (3-NOP) and nitrate not being used in all countries at present, re- cords were kept in the training dataset if these specific additives were supplied to the cows. Records with miss- ing values for CO2 (n = 192) production were also ex- cluded. Furthermore, records from cows with more than 300 DIM (n = 140) were excluded, as they constitute a small group of animals in the dataset, which is not repre- sentative for a commercial farm. After this selection, the pre-training dataset contained n = 2,824 records. Records were categorized into 4 breed groups, (1) Holstein, (2) Jersey, (3) Ayrshire, or (4) other breeds and crossbreeds, and 3 parity groups: (1) first, (2) second, or (3) third parity and higher. Emissions of CO2 were reported as grams per day (g/d) or liters per day (L/d). If the research locations delivered the measured gas exchange in liters per day, the ideal gas law was used Kjeldsen et al.: PREDICTING CO2 PRODUCTION Ta bl e 1. O ve rv ie w o f e ac h re se ar ch lo ca tio n in th e re fin ed tr ai ni ng d at as et (n = 2 ,2 44 ) a fte r d at a pr e- pr oc es si ng R es ea rc h lo ca tio n n Ex pe rim en ts (n )   G as m ea su rin g m et ho d (n )   B re ed s ( n) Pa rit y of la ct at in g co w s ( n) A ar hu s U ni ve rs ity 31 3 18 R es pi ra tio n ch am be r H ol st ei n (2 71 ), Je rs ey (4 2) Fi rs t ( 11 8) , s ec on d (1 12 ), th ird a nd o ld er (8 3) A gR es ea rc h Li nc ol n 28 1 G re en Fe ed O th er s/ cr os sb re ed s Fi rs t ( 3) , s ec on d (6 ), th ird a nd o ld er (1 9) A gR es ea rc h Pa lm er st on N or th 56 5 R es pi ra tio n ch am be r O th er s/ cr os sb re ed s Se co nd (9 ), th ird a nd o ld er (4 7) A gr ic ul tu re V ic to ria R es ea rc h 19 6 7 R es pi ra tio n ch am be r H ol st ei n Fi rs t ( 23 ), se co nd (5 2) , t hi rd a nd o ld er (1 21 ) ET H Z ur ic h 41 2 R es pi ra tio n ch am be r H ol st ei n (7 ), ot he rs /c ro ss br ee ds (3 4) Fi rs t ( 4) , s ec on d (7 ), th ird a nd o ld er (3 0) Fl an de rs R es ea rc h In st itu te fo r  A gr ic ul tu re , F is he rie s a nd F oo d 14 1 8 R es pi ra tio n ch am be r H ol st ei n Fi rs t ( 27 ), se co nd (5 6) , t hi rd a nd o ld er (5 8) IN R A E 11 8 7 R es pi ra tio n ch am be r H ol st ei n Fi rs t ( 17 ), se co nd (5 6) , t hi rd a nd o ld er (4 5) Pe nn St at e 41 8 8 G re en Fe ed H ol st ei n Fi rs t ( 14 8) , s ec on d (1 47 ), th ird a nd o ld er (1 23 ) R es ea rc h In st itu te fo r F ar m A ni m al  B io lo gy 19 1 R es pi ra tio n ch am be r H ol st ei n Se co nd (1 0) , t hi rd a nd o ld er (9 ) Sw ed is h U ni ve rs ity o f A gr ic ul tu ra l  S ci en ce s 45 5 3 G re en Fe ed A y rs hi re (9 6) , o th er s/ cr os sb re ed s ( 35 9) Fi rs t ( 16 7) , s ec on d (1 28 ), th ird a nd o ld er (1 60 ) TU M un ic h 51 1 R es pi ra tio n ch am be r O th er s/ cr os sb re ed s Fi rs t ( 3) , s ec on d (1 7) , t hi rd a nd o ld er (3 1) U SD A B el ts vi lle A gr ic ul tu ra l R es ea rc h  C en te r 45 1 R es pi ra tio n ch am be r H ol st ei n (2 2) , J er se y (2 3) Se co nd (2 2) , t hi rd a nd o ld er (2 3) U ni ve rs ity o f M ila n 62 3 R es pi ra tio n ch am be r H ol st ei n Fi rs t ( 18 ), se co nd (3 3) , t hi rd a nd o ld er (1 1) U ni ve rs ity o f R ea di ng 10 6 6 R es pi ra tio n ch am be r H ol st ei n Se co nd (2 5) , t hi rd a nd o ld er (8 1) W ag en in ge n U ni ve rs ity a nd R es ea rc h 19 5 5 R es pi ra tio n ch am be r H ol st ei n Fi rs t ( 52 ), se co nd (5 5) , t hi rd a nd o ld er (8 8) A ll 2, 24 4 76 R es pi ra tio n ch am be r ( 1, 34 3) G re en Fe ed (9 01 ) Ay rs hi re (9 6) , H ol st ei n (1 ,5 55 ), Je rs ey (6 5) , o th er s/ cr os sb re ed s ( 52 8) Fi rs t ( 58 0) , s ec on d (7 35 ), th ird a nd o ld er (9 29 ) 6774 Journal of Dairy Science Vol. 107 No. 9, 2024 to convert to grams per day, with the conversion factor depending on the temperature and pressure at the specific research location. The outcome of the models is CO2 in grams per day at standard temperature and pressure (0°C and 101.325 kPa). If needed, the outcome of the model can be converted to liters per day as CO2 (L/d) = CO2 (g/d) × 0.509 (L/g). Yield of ECM (3.14 MJ/kg) was calculated according to the respective ECM equations in Sjaunja et al. (1990) based on fat, protein, and lac- tose concentration, taking into account lactose reported as monohydrate or in anhydrous form (15.71 kJ/g and 16.54 kJ/g, respectively). Conversion from true protein to CP was performed with the factor 1.058 (DePeters and Ferguson, 1992). Model Development Individual DMI is often available at research facilities but not on commercial farms. In addition, only some commercial farms continuously track cows’ BW. Due to the difference in data availability in different settings, 3 models were developed to cover these different sce- narios. Before the continuous predictor variables were included in the model development, Pearson correlation coefficients (r) were calculated (Supplemental Table S1, see Notes). In case of 2 variables being highly correlated (r ≥ 0.5), only the predictor variable with the highest correlation coefficient to CO2 production was chosen. This, for instance, excluded milk production (kg/d) and ECM (kg/d) from being predictor variables in the same model (r = 0.93). They were equally correlated with CO2 production (both r = 0.50), and ECM was chosen for model development, as inclusion of milk production led to a high number of interactions between milk produc- tion and milk nutrients (data not shown). Also, DMI and ECM were highly correlated (r = 0.75), where DMI had the strongest correlation with CO2 production (r = 0.69); therefore they could not be predictor variables in the same model. Furthermore, milk crude fat (CF; g/kg) and milk CP (g/kg) were highly correlated (r = 0.54). Therefore, only milk CF was included in the model development, since its correlation to CO2 production was stronger (r = −0.19) than it was for milk CP (r = −0.09). Based on the described exclusion of predictor variables, continuous predictor variables were DMI, ECM yield, concentra- tions of fat and lactose in milk, BW0.75, DIM, DIP, and dietary CP and CF concentration. Parity and breed were set as discrete factors. Breed was only to some extent confounded to research location (Table 1); therefore it was per default included in all models as a fixed effect. Also, research location and experiment nested within research location were per default included as random effects (allowing individual intercepts) in all 3 models. Method of measuring CO2 production (RC or GF) was confounded with research location, and none of the re- search locations provided data obtained by both GF and RC. Therefore, measurement method was not included as a predictor variable by itself, as it was indirectly in- cluded through the random effect of research location. As some of the variables by nature have different units (e.g., DIM and ECM), all values were centered (mean = 0) by using the “scale” function in R (R Core Team, 2023). No standardization was performed (original varia- tion was kept) to ease the implementation of the mod- els. All statistical analyses were conducted in R 4.3.0 (R Core Team, 2023). The selection of each model was performed with the buildmer function (Voeten, 2023), us- ing the default criterion likelihood ratio test for selection of predictor variables. Using Akaike’s information crite- rion or Bayesian information criterion as criteria instead of likelihood ratio test resulted in the selection of the same variables, regardless of whether the stepwise inclu- sion or elimination followed a “forward” or “backward” direction order. The 3 basic models derived from the buildmer function were tested for increased complexity by adding interactions and afterward testing for inclu- sion of a random slope of one of the predictor variables. Analysis of variance tests were performed to determine the level of significance of increased model complexity (fitting with either “ML” or “REML,” depending on the 2 models compared). Statistical significance was declared at P ≤ 0.05. Based on the significant predictor variables from the model development, a common training data- set (n = 2,259), without missing records for DMI, DIM, ECM, BW, dietary CP, dietary CF, milk CF, parity, and breed was used to derive all 3 models. Another approach would have been to train different models using different datasets, containing the predictor variables of interest to Kjeldsen et al.: PREDICTING CO2 PRODUCTION Table 2. Summary statistics of the continuous parameters included in the training dataset (n = 2,244) Item n Mean SD Minimum Maximum Dietary composition   OM (g/kg DM) 2,227 924 17.0 833 953   CP (g/kg DM) 2,244 167 22.6 81.0 253   CF (g crude fat/kg DM) 2,244 38.8 11.22 12.1 74.0 DMI (kg/d) 2,244 20.7 4.49 6.80 37.2 Milk (kg/d) 2,244 30.1 9.09 2.65 65.7 ECM (kg/d) 2,244 30.6 8.35 2.91 71.5 Milk composition   CF (g crude fat/kg) 2,244 42.7 8.74 13.2 88.5   CP (g/kg) 2,244 33.3 3.92 23.0 53.9   Lactose (g/kg) 2,148 47.7 2.87 26.0 56.3 Days in pregnancy (d) 562 56 60.0 0 233 DIM (d) 2,244 137 69.2 7 299 BW (kg) 2,244 606 82.0 341 969 CO2 (g/d) 2,244 12,402 2,023.0 4,937 20,950 CH4 (g/d) 2,241 397 85.8 136 729 Journal of Dairy Science Vol. 107 No. 9, 2024 6775 maximize the number of records (Niu et al., 2018). How- ever, to perform unbiased comparisons across models, the same dataset was used for all models in the current study. Outliers, here defined as records with residuals >5 × SD (residuals derived with model 1, n = 15), were removed from the dataset. The residuals of each model were plotted against predicted values of CO2 production and against individual variables. The residuals did not show any nonlinear relationship; therefore data were not transformed. The refined training dataset contained 2,244 individual animal records (Tables 1 and 2), where 60% of the records were obtained by RC. Model Evaluation Model performance was tested on 3 datasets from Aarhus University, including (1) solely RC data (without any CH4-reducing feed additives), (2) solely GF data (without any CH4-reducing feed additives), and (3) GF data with observations where only nitrate, only 3-NOP, or both nitrate and 3-NOP were fed (Table 3). All test datasets consisted of data obtained in studies performed after the current training dataset was collected (from 2020 to 2022), and the training dataset was tested again with different applicable models as shown in Table 4. The RC test dataset (n = 103) consisted of data from 5 studies; 4 Latin square designs and 1 crossover design, and records having DIM >300 d were excluded from the test dataset. All animals were Holstein cows and were 136 ± 64.4 DIM (±SD), with a DMI of 21.4 ± 3.76 kg/d and an ECM of 31.3 ± 6.92 kg/d. The percentages of cows in first, second, and third and higher lactation in the dataset were 39%, 47%, and 15%, respectively. The GF test dataset (n = 478) consisted of data from a part of 4 production trials; 3 Latin square designs and 1 continuous trial were included, and in case of the lat- ter, an average from the last week of measuring was in- cluded. Records having DIM >300 d were excluded. All animals were Holstein cows and were 145 ± 52.0 DIM, with a DMI of 21.3 ± 3.06 kg/d and yielding 33.8 ± 6.56 kg ECM/d. The percentages of first-, second-, and third- lactation and older cows in the dataset were 50%, 27%, and 23%, respectively. The last test dataset consisted only of records where the additives nitrate and 3-NOP were supplemented (GF+, n = 295). The data were from a part of 2 production tri- als, which were both Latin square designs. None of the records had DIM >300 d. All animals were Holstein cows and were 107 ± 43.6 DIM, with a DMI of 20.0 ± 3.21 kg/d and an ECM of 31.2 ± 6.14 kg/d. The percentages of cows in first, second, and third and higher lactation in the dataset were 52%, 25%, and 23%, respectively. In addition, the models were also evaluated on a modi- fied and merged version of the 3 test datasets (RC, GF, and GF+; Table 5) and on the training dataset (Table 6). In the modified dataset, the records of the measured CO2 production (g/d) from the test datasets subtracted (where measurements were obtained by RC) or added (where measurements were obtained by GF) the absolute mean bias (MB, here calculated as observed − predicted), from the model evaluation of the specific model on RC, GF, and GF+. The “opmetrics” function from the R package mod- MetricsR (Giagnoni, 2023) was used to obtain the evalu- ation estimates: root mean square error (RMSE), RMSE as percentage of observed mean, mean absolute error Kjeldsen et al.: PREDICTING CO2 PRODUCTION Table 3. Summary statistics of the continuous parameters and CO2 production (g/d) in the test dataset obtained from respiration chambers (RC, without additives, n = 103), GreenFeed (GF, without additives test dataset, n = 478), or GreenFeed only including diets containing nitrate, 3-nitrooxypropanol, or both nitrate and 3-nitrooxypropanol (GF+, n = 295); all data were obtained at Aarhus University (Viborg, Denmark) from 2020 to 2022 Test dataset Mean   SD   Minimum   Maximum RC1 GF2 GF+3 RC GF GF+ RC GF GF+ RC GF GF+ DMI (kg/d) 21.4 21.3 20.0 3.76 3.06 3.21 11.8 13.9 11.7 27.4 30.0 27.6 Diet CP (g/kg DM) 171 165 171 7.9 8.4 11.3 157 149 148 188 188 186 CF (g crude fat/kg DM) 33.9 44.1 41.7 8.52 13.3 15.6 23.7 27.2 26.3 62.0 70.6 69.1 ECM (kg/d) 31.3 33.8 31.2 6.92 6.56 6.14 17.4 16.1 17.4 47.4 54.3 49.6 Milk CF (g crude fat/kg) 39.3 39.4 40.9 6.75 6.25 5.84 23.4 18.6 23.1 58.9 58.7 57.7 DIM (d) 136 145 107 64.4 52.0 43.6 42 16 24 297 290 231 BW (kg) 640 655 642 53.1 67.6 62.8 550 500 496 747 873 858 CO2 (g/d) 14,369 12,625 12,163 1,663.8 1,531.0 1,468.0 11,086 8,617 8,213 18,215 17,782 15,365 1The percentages of cows in first, second, and third and higher lactation in the dataset were 39%, 47%, and 15%, respectively. All cows were Holstein cows. 2The percentages of first, second, and third and older cows in the dataset were 50%, 27%, and 23%, respectively. All cows were Holstein cows. 3The percentages of cows in first, second, and third and higher lactation in the dataset were 52%, 25%, and 23%, respectively. All cows were Holstein cows. 6776 Journal of Dairy Science Vol. 107 No. 9, 2024 (MAE), concordance correlation coefficient (CCC), ratio of RMSE to standard deviation of measured data (RSR), MB, slope bias (SB), and MB, SB, and dispersion as percentage of mean square error (MSE). Both CCC and RSR are dimensionless parameters. The CCC is the product of Pearson correlation coefficient (r; ranging from −1 to +1) and the bias correction factor (Cb; ranging from 0 to 1). Perfect fit (precision) is indicated by r = 1, and agreement between predicted and observed values (accuracy) is indicated by Cb = 1 and thus CCC = 1. RESULTS Due to the incorporation of data from different research locations in the training dataset, a high level of variabil- ity was present (Table 2). The DMI ranged from 6.80 to 37.2 kg/d, and CH4 and CO2 production varied from 136 to 729 g/d and 4,937 to 20,950 g/d, respectively. Description of Models 1, 2, and 3 The different models developed to be used in different practical settings, depending on data availability, were as follows. Model 1, intended to be used in a situation where in- dividual DMI data are available (“best model”; e.g., at research locations), and where DMI alone described 58% of the variation in CO2 production (g/d) in the present dataset, reads, Model 1 (“best model”): CO2 (g/d) = b0 + (b1 × DMI) + (b2 × BW0.75) + (b3 × Diet CP) + breed + (bDMI,breed × DMI) + (bDMI,parity × DMI) + (bBW 0.75 ,breed × BW0.75), where, b0 is the intercept; b1, b2, and b3 are the coef- ficients of DMI (kg/d), BW0.75 (kg0.75), and diet CP (g/ kg DM), respectively; and bDMI,breed and bBW 0.75 ,breed are the breed-specific coefficients of DMI and BW0.75. The parity-specific coefficient of DMI is bDMI,parity. All coef- ficients are listed in Table 4. An example of using model 1 to calculate the CO2 pro- duction (g/d) from a second-parity Holstein cow, with a DMI of 25 kg DM/d, with 160 g dietary CP per kilogram DM, weighing 600 kg, is as follows: 956 + [122 × 25 (kg DM/d)] + [60.4 × 600 (kg0.75)] + [3.44 × 160 (g CP/kg DM)] − 777 + [206 × 25 (kg DM/d)] + [7.53 × 25 (kg DM/d)] + [−18.5 × 600 (kg0.75)] = 14,197 g CO2 per day. Model 2 was intended for an on-farm setting, where in- dividual DMI data are not available (“on-farm model”); therefore ECM became a significant predictor variable, as ECM alone described 28% of the variation in CO2 production (g/d) in the present dataset. It reads, Kjeldsen et al.: PREDICTING CO2 PRODUCTION Table 4. Coefficients of the 3 models to predict CO2 (g/d) from lactating dairy cows, where model 1 is “best model,” model 2 is “on-farm model,” and model 3 is “reduced on-farm model”1 Item Model 1 Model 2 Model 3 Intercept 956 −6,134 8,781 DMI (kg/d) 122     ECM (kg/d)   213 80.3 MetaBW (kg) 60.4 126   Diet CP (g/kg DM) 3.44     Milk CF (g/kg)   52.5   DIM (d)   −5.13 −4.66 Breed   Ayrshire 0 0 0   Holstein −777 2,117 −49.0   Jersey 1,103 1,364 −2,321   Others/crossbreeds 1,501 4,083 −1,237 Parity   First     0   Second     511   Third and higher     1,587 DIM × Diet CF   - 0.122 −0.149 ECM × DIM   0.386 0.338 ECM × metaBW   −1.18   Milk CF × metaBW   −0.614   DMI × Ayrshire 0     DMI × Holstein 206     DMI × Jersey 204     DMI × others/crossbreds 225     DMI × first parity 0     DMI × second parity 7.53     DMI × third parity 15.7     MetaBW × Ayrshire 0 0   MetaBW × Holstein −18.5 −5.96   MetaBW × Jersey −37.3 −1.03   MetaBW × others/crossbreds −43.2 −33.4   DIM × Ayrshire   0 0 DIM × Holstein   2.06 6.05 DIM × Jersey   2.49 6.02 DIM × others/crossbreds   8.94 11.3 MetaBW × first parity   0   MetaBW × second parity   3.66   MetaBW × third parity   4.01   First parity × milk CF     −4.18 Second parity × milk CF     −10.5 Third parity × milk CF     −28.8 Ayrshire × first parity     0 Ayrshire × second parity     0 Ayrshire × third parity     0 Holstein × first parity     0 Holstein × second parity     775 Holstein × third parity     803 Jersey × first parity     0 Jersey × second parity     608 Jersey × third parity     1,307 Others/crossbreds × first parity     0 Others/crossbreds × second parity     791 Others/crossbreds × third parity     659 1Diet CF = dietary crude fat (g/kg DM), diet CP = dietary crude protein (g/kg DM), DIM = days in milk (d), DMI = dry matter intake (kg/d), ECM = energy-corrected milk yield (kg/d), milk CF = milk crude fat (g/ kg), metaBW = metabolic body weight = body weight0.75 (kg). Journal of Dairy Science Vol. 107 No. 9, 2024 6777 Model 2 (“on-farm model”): CO2 (g/d) = b0 + (b1 × ECM) + (b2 × BW0.75) + (b3 × Milk CF) + (b4 × DIM) + breed + (bDIM,DietCF × DIM × Diet CF) + (bECM,DIM × ECM × DIM) + (bECM,BW 0.75 × ECM × BW0.75) + (bMilkCF,BW 0.75 × Milk CF × BW0.75) + (bBW 0.75 ,breed × BW0.75) + (bDIM,breed × DIM) + (bBW 0.75 ,parity × BW0.75), where, b0 is the intercept; and b1, b2, b3, and b4 are the coefficients of ECM (kg/d), BW0.75 (kg0.75), milk CF (g/kg milk), and DIM (d), respectively. Furthermore, bDIM,DietCF, bECM,DIM, bECM,BW 0.75, and bMilkCF,BW 0.75 are the coefficients of DIM × diet CF, ECM × DIM, ECM × BW0.75, and milk CF × BW0.75, respectively, and bBW 0.75 ,breed and bDIM,breed are the breed-specific coeffi- cients of BW0.75 and DIM, whereas bBW 0.75 ,parity is the parity-specific coefficient of BW0.75. All coefficients are listed in Table 4. An example of using model 2 to calculate the CO2 production from a second-parity Ayrshire cow, with a yield of 30 kg ECM/d, weighing 650 kg, being 110 DIM, with an average milk CF concentration at 35.0 g/kg milk, eating a TMR with a CF content at 40 g/kg DM is as follows: −6,134 + [213 × 30 (kg ECM/d)] + [126 × 650 (kg0.75)] + [52.5 × 35.0 (g/kg)] + [−5.13 × 110 (d)] + 0 + [−0.122 × 110 (d) × 40 (g CF/kg DM)] + [0.386 × 30 (kg ECM/d) × 110 (d)] + [−1.18 × 30 (kg ECM/d) × 650 (kg0.75)] + [−0.614 × 35.0 (g milk CF/kg) × 6500.75 (kg)] + [0 × 650 (kg0.75)] + [0 × 110 (d)] + [3.66 × 650 (kg0.75)] = 11,634 g CO2 per day. Kjeldsen et al.: PREDICTING CO2 PRODUCTION Table 5. Model evaluation of the 3 models to predict CO2 production (g/d) from lactating dairy cows, where model 1 is “best model,” model 2 is “on- farm model,” and model 3 is “reduced on-farm model”1 Item   Test dataset RMSE RMSE, % mean MAE CCC RSR MB SB MB, % MSE SB, % MSE Dispersion, % MSE Model performance2   Model 1 RC 1,456 10.1 1,281 0.66 0.88 1,134 −0.065 60.6 0.44 38.9 GF 1,046 8.29 856 0.76 0.68 −655 −0.040 39.1 0.27 60.6 GF+ 949 7.81 756 0.79 0.65 −587 −0.052 38.2 0.54 61.3   Model 2 RC 1,416 9.85 1,240 0.66 0.85 1,192 0.097 70.9 0.85 28.2 GF 1,187 9.40 993 0.67 0.78 −740 0.059 38.9 0.32 60.8 GF+ 1,199 9.86 959 0.63 0.82 −639 −0.063 28.4 0.35 71.3   Model 3 RC 1,847 12.9 1,635 0.54 1.11 1,619 −0.028 76.9 0.049 23.1 GF 1,138 9.01 916 0.68 0.74 −138 −0.19 1.47 4.89 93.6 GF+ 1,172 9.64 931 0.61 0.80 −77.4 −0.22 0.44 4.68 94.9 Model performance3   Model 1 Modified 806 6.15 616 0.85 0.52 0.27 −0.046 0.000 0.63 99.4   Model 2 Modified 941 7.15 742 0.77 0.61 −0.11 0.025 0.000 0.10 99.9   Model 3 Modified 1,118 8.88 886 0.69 0.72 0.024 −0.17 0.000 3.87 96.1 1Model performance was evaluated based on either of the following. (1) Observations from 3 test datasets (both from Aarhus University, Viborg, Denmark) obtained from respiration chambers (RC, n = 103) or GreenFeed units (GF, n = 478) without additives, or GF only including diets contain- ing nitrate, 3-nitrooxypropanol, or both nitrate and 3-nitrooxypropanol (GF+, n = 295). Or (2) a modified dat set, with RC, GF, and GF+ test datasets merged together (n = 876). The CO2 production (g/d) in the modified dataset was calculated by subtracting mean bias (if measurements were obtained by RC) or adding mean bias (if measurements were obtained by GF and GF+) from evaluation of the specific model on RC, GF, and GF+ test datasets. RMSE = root mean square error, MAE = mean absolute error, CCC = concordance correlation coefficient, RSR = ratio of RMSE to standard deviation of measured data, MB = mean bias, SB = slope bias, and MSE = mean square error. 2Test dataset based on observed CO2 production. 3Test dataset corrected for mean bias of CO2 production. Table 6. Model evaluation of the 3 models to predict CO2 production (g/d) from lactating dairy cows based on the training dataset itself (n = 2,244), where model 1 is “best model,” model 2 is “on-farm model,” and model 3 is “reduced on-farm model”1 Item2   Test data et RMSE RMSE, % mean MAE CCC RSR MB SB MB, % MSE SB, % MSE Dispersion, % MSE Model 1 Training dataset 1,435 11.6 1,115 0.73 0.71 −298 −0.18 4.32 5.27 90.4 Model 2 Training dataset 1,549 12.5 1,230 0.63 0.77 59.0 −0.14 0.145 1.87 98.0 Model 3 Training dataset 1,573 12.7 1,254 0.59 0.78 68.7 −0.071 0.191 0.382 99.4 1RMSE = root mean square error, MAE = mean absolute error, CCC = concordance correlation coefficient, RSR = ratio of RMSE to standard devia- tion of measured data, MB = mean bias, SB = slope bias, and MSE = mean square error. 2Model performance, training dataset based on observed CO2 production. 6778 Journal of Dairy Science Vol. 107 No. 9, 2024 Model 3 was intended for an on-farm setting, where BW is not a part of the predictor variables (“reduced on-farm model”). It reads, Model 3 (“reduced on-farm model”): CO2 (g/d) = b0 + (b1 × ECM) + (b2 × DIM) + breed + parity + (bbreed,parity) + (bDIM,DietCF × DIM × Diet CF) + (bECM,DIM × ECM × DIM) + (bDIM,breed × DIM) + (bMilkCF,parity × Milk CF), where b0 is the intercept; b1, b2, and bECM,DIM are the co- efficients of ECM (kg/d), DIM (d), and ECM × DIM, respectively; bDIM,DietCF and bECM,DIM are the coefficients of DIM × diet CF and ECM × DIM; bDIM,breed is the breed-specific coefficient of DIM; bbreed,parity is the breed- specific coefficient for each parity; and bMilkCF,parity is the parity-specific coefficient of milk CF. All coefficients are listed in Table 4. An example of using model 3 to calculate the CO2 pro- duction from a first-parity crossbreed cow, with a yield of 28 kg ECM/d, being 100 DIM, eating a TMR with 35 g CF/kg DM, with 37 g CF/kg milk is as follows: 8,781 + [80.3 × 28 (kg ECM/d)] + [−4.66 × 100 (d)] − 1,237 + 0 + 0 + [−0.149 × 100 (d) × 35 (g CF/kg DM)] + [0.338 × 28 (kg ECM/d) × 100 (d)] + [11.3 × 100 (d)] + [−4.18 × 37 (g milk CF/kg)] = 10,727 g CO2 per day. The models predict the CO2 production in grams per day; to calculate the CO2 production in liters per day, see Ma- terials and Methods section. Evaluation on RC and GF Test Datasets When evaluated on the RC test dataset, model 2 was superior with respect to RMSE, RMSE as percentage of mean, MAE, and RSR (Table 5). However, model 1 was superior with respect to MB, MB as percentage of MSE, and dispersion as percentage of MSE when evaluated on the RC test dataset. Model 1 was superior in most of the evaluation pa- rameters when the models were evaluated on the GF test dataset (RMSE, RMSE as percentage of mean, MAE, CCC, RSR, SB, and SB as percentage of MSE). Model 3 performed better than models 1 and 2 with respect to SB, and SB as percentage of MSE, when evaluated on the RC test dataset. In addition, model 3 was superior to models 1 and 2 with respect to MB, MB as percentage of MSE, and consequently dispersion as percentage of MSE on the GF test dataset. Evaluation on GF+ Test Dataset Model 3 had the highest dispersion as percentage of MSE when the models were evaluated on the GF+ test dataset, as a consequence of low MB and MB as per- centage of MSE (Table 5). Oppositely, RMSE, RMSE as percentage of mean, MAE, CCC, RSR, and SB were bet- ter for model 1 when evaluated on the GF+ test dataset. Evaluation on the Modified Test Dataset The predicted CO2 production underestimated the ac- tual measured CO2 production in RC (MB across models was 1,315) and overestimated the actual measured CO2 production using GF units (MB across models was 511). Bearing in mind that the models were developed with a training dataset containing both GF (40% of the records) and RC data (60% of the records), it was decided to ad- dress this by evaluating the 3 models with a modified dataset (see Materials and Methods section). The evalu- ations obtained with the modified test dataset clearly illustrated that nearly all the variation was related to dispersion error (Table 5). Evaluation on the Training Dataset Due to the risk of some of the models simply matching the properties of test dataset better than other models, it was decided to evaluate the models on the training dataset as well (Table 6). Furthermore, this evaluation il- lustrates the predictability of the models if certain animal parameters were not available. Model 1 was superior to the other models with respect to RMSE, RMSE as per- centage of mean, MAE, CCC, and RSR, when evaluated on the training dataset. However, the actual values of MB or SB, and MB, SB, or dispersion as percentage of MSE for model 1 were not superior to model 2 and 3. This was partly caused by the relatively higher MB for model 1; SB was also slightly higher for model 1, causing the dis- persion as percentage of MSE to be somewhat lower than it was for models 2 and 3. Model 2 (without DMI, with BW as predictor variable) performed slightly better than model 3 (without DMI and BW as predictor variables) with respect to RMSE, RMSE as percentage of mean, MAE, CCC, RSR, MB, and MB as percentage of MSE. However, SB and SB or dispersion as percentage of MSE were slightly better for model 3. Based on the compari- son of model 2 and 3 on the training dataset, predicting CO2 production from dairy cows in settings without data on BW is feasible. Kjeldsen et al.: PREDICTING CO2 PRODUCTION Journal of Dairy Science Vol. 107 No. 9, 2024 6779 DISCUSSION Overall Model Evaluation on the Test Dataset The models were developed with a training dataset where 69.3% of the records were Holstein cows, whereas Ayrshire, Jersey and others or crossbreed cows consti- tuted 4.3%, 2.9%, and 23.5% of the records, respectively. The external validation test datasets consisted of only Holstein cows. Furthermore, the models were developed and evaluated with a dataset of cows having ≤300 DIM. It is important to consider this when applying the models to breeds other than Holstein cows or cows in lactation beyond 300 d. Initially, the RC and the GF test datasets were treated as a unified dataset (data not shown), but a systematic underprediction for the RC data and a simultaneous over- prediction for GF data were observed. Therefore, the models were evaluated separately on the RC and GF parts of the test dataset (Table 5). Additionally, the mod- els were evaluated on the GF+ test dataset to investigate the potential impacts of the use of nitrate, 3-NOP, or a combination of nitrate and 3-NOP on the precision and accuracy of the models. The observed underprediction for RC and overprediction for GF could be caused by inherent model characteristics, but the method of gas measurement was confounded with research location (included as a random effect in all the models). Technical differences between the 2 methods, such as the GF units exclusively measuring gases emitted in exhaled air and not from the rectum of the cow, could partly contribute to the observed discrepancies, despite lack of data related to CO2 released from the rectum of the cow. In addition, the GF relies on repeated short-term measurements, typi- cally lasting 2 to 7 min, and repeated at intervals over subsequent days, whereas RC measurements are gener- ally continuous over successive days (typically 2 to 4 d). The GF system has the advantage of being able to record gas data on a much larger number of animals compared with RC systems. Previous studies have compared RC and GF measurements of CO2 emission, but the conclu- sions drawn were limited by the low number of animals (Doreau et al., 2018) and occasional reductions in DMI when cows entered the chambers (Alemu et al., 2017). For CH4 production, Hristov et al. (2018) showed an unexpectedly weak relationship between DMI and CH4 production measured with the GF (13.9 to 35.4 kg DMI, R2 = 0.05), and a much stronger relationship measured with the RC (3.9 to 33.5 kg DMI, R2 = 0.58), indicating a better capability of RC compared with GF to capture variation in gaseous release. However, the variation in DMI was also greater for the RC than the GF data, which could partly explain the better relationship for RC data in the study by Hristov et al. (2018); even a more restricted range of DMI (15.0 to 33.5 kg/d) with the RC still showed a stronger relationship (R2 = 0.41) than with the GF. The model performance was better when the 3 models were assessed based on the GF and the GF+ test datasets, compared with evaluation on the RC test dataset. This is evident from the higher dispersion in percentage of MSE and CCC (except for model 2 on GF+ test dataset). More- over, the RMSE, RMSE as percentage of mean (except for model 2 on GF+ test dataset), and MAE were consis- tently lower when the models were tested on the GF and the GF+ test datasets. The major reason for lower model performance with the RC test data is the more pronounced MB (inaccuracy) compared with the GF test data. The mean of CO2 production in the RC test dataset (14,369 g/d) is also higher than in the training dataset (12,402 g/d), likely contributing to the high MB observed when evaluating the models on the RC test dataset. The higher dispersion in percentage of MSE indicates that a greater fraction of variation is random variation for GF and GF+ (Table 5). However, it is important to acknowledge that these differences are also attributable to the sizes of the respective test datasets (Doreau et al., 2018), with the GF and GF+ test datasets being larger than the RC test dataset. The DMI in the GF+ dataset (based on a part of 2 production trials) was on average lower than DMI in the GF dataset (based on the same 2 production trials, with all observations in it, plus 2 other production tri- als), likely causing the lower mean CO2 production (g/d) in that specific test dataset (Table 3). Furthermore, the variation of the CO2 production within the GF+ dataset was less (smaller SD: Table 3) than for the RC and GF test dataset, and the cows were earlier in lactation (mean DIM was lower for GF+ than RC and GF: Table 3). The evaluation on the GF+ test dataset should therefore be interpreted bearing in mind that the cows in this test dataset generally produced lower amounts of CO2 with less variation; thus the evaluation on the GF+ test dataset indicates somewhat better model performance than the evaluation on the RC and GF test dataset. All 3 models showed a low SB both in absolute values and as percentage of MSE when tested on the RC, GF, and GF+ test dataset (Table 5). This suggests consistently good prediction abilities for determining whether a given cow emits lower or higher amounts of CO2 as compared with the average cow (Figure 1). Furthermore, it indi- cates that GF units rank cows with comparable preci- sion to RC, contrasting with the results from a previous study (Alemu et al., 2017) but partially in agreement with the findings of Rischewski et al. (2017). Precision is of importance, especially when the models are used within a herd to rank individual cows based on their CO2 production, and subsequently ranking them according to their CH4 production by combining estimated CO2 pro- Kjeldsen et al.: PREDICTING CO2 PRODUCTION 6780 Journal of Dairy Science Vol. 107 No. 9, 2024 duction and measured [CH4]:[CO2] ratio in breath using the sniffer technique. However, model 1 and model 2 had noticeable MB when evaluated on the RC and GF dataset, indicating a lack of accuracy and a disparity in absolute values between RC and GF. Assuming that RC data represents the true production of CO2, and that RC are seen as the the gold standard, it is suggested to add the MB from the RC evaluation of the given model to the dependent variable of the model (CO2 g/d). This would cause the outcome of the models to reach a more accurate level, if the observed difference between RC and GF data in the current test dataset is considered universal across research groups. A slight underprediction of CO2 production for the measured low levels of CO2 production was evident from the regression lines of the present models on a reduced version of the training dataset where DIP was given (n = 562, Figure 2). However, the underprediction was even Kjeldsen et al.: PREDICTING CO2 PRODUCTION Figure 1. Observed (red) CO2 production (g/d) in the test dataset, and residual values (blue) of CO2 production of the 3 models plotted against predicted values of CO2 production for (A) model 1 (“best model”), tested on respiration chamber (RC) data; (B) model 2 (“on-farm model”), tested on RC data; (C) model 3 (“reduced on-farm model”), tested on RC data; (D) model 1, tested on GreenFeed (GF) data; (E) model 2, tested on GF data; (F) model 3, tested on GF data; (G) model 1, tested on GF data, only including diets containing nitrate, 3-nitrooxypropanol (3-NOP), or both nitrate and 3-NOP (GF+); (H) model 2 tested on GF+ data; and (I) model 3 tested on GF+ data. The red and blue line represent the linear regression lines of observed and residual values, respectively. Journal of Dairy Science Vol. 107 No. 9, 2024 6781 more pronounced when using the previous equations from Madsen et al. (2010) and Pedersen et al. (2008) on the same reduced dataset (Figure 3). This advocates for using the models from the present study in combination with the sniffer method instead of the equations from Madsen et al. (2010) and Pedersen et al. (2008). Gestation Previously, the requirement of metabolizable energy (ME) for pregnancy in dairy cows was described by an exponential function related to number of days pregnant, with an efficiency of 10.5% of ME for fetal tissue deposi- tion (Moe and Tyrrell, 1972). A recent study estimated efficiency of ME for pregnancy in Holstein × Gyr heif- ers to be 14.1% (Sguizzato et al., 2020). However, this estimation was based on a nonlinear development of net energy (NE) for pregnancy, and possible variation was not taken into account (Sguizzato et al., 2020). Accord- ing to Nielsen and Volden (2011), the NE requirement for gestation is only minor when DIP <150, but significantly increased for cows >150 DIP. Thereby HP increases along gestation, assuming a constant efficiency of ME for gestation as indicated in Moe and Tyrrell (1972) and Sguizzato et al. (2020). The gravid uterus and develop- ment of the mammary gland cause HP to increase (Sguiz- zato et al., 2020), as they, especially the gravid uterus, account for significant metabolism of nutrients. Hence, DIP was initially included in the models (and it is a factor in the equation of HP; CIGR, 2002). However, only a few records in the training dataset (n = 562) could provide such data, likely due to lack of recording, and DIM was expected to be used as a close proxy for DIP. However, in the part of the training dataset with DIP available (n = 562 records), DIM was not a very precise indicator of DIP (Supplemental Table S2, see Notes), likely because the time point of a successful insemination of these experimental animals did not follow the same pattern across research locations. Therefore, the effect of DIP as a predictor variable was only tested on the smaller data- set where DIP was available, and there was no effect of DIP on CO2 production (P = 0.30). Dietary Crude Protein Increasing the dietary CP level has been found to increase the energy content in cattle urine (Ramin and Huhtanen, 2013; Hynes et al., 2016), and the energy content in cattle urine is closely linked with the urinary carbon content (Morris et al., 2021). In addition, dietary CP is an indicator of nutrient composition within a diet. In the current training dataset, data on NDF and starch content were not collected from the research locations, and the correlation between dietary CP and CF was low Kjeldsen et al.: PREDICTING CO2 PRODUCTION Figure 2. Measured CO2 production (g/d) plotted against predicted CO2 production in a reduced version of the training dataset (n = 562, see Supplemental Table S3), where observations having missing values of days in pregnancy were not included. The black line represents y (measured CO2 production, g/d) = x (predicted CO2 production, g/d); the red, green, and blue lines represent linear regressions of CO2 production predicted by models 1, 2, or 3, respectively. This reduced version of the training dataset was a part of the training dataset (n = 2,244), which these 3 models were derived from. Regression lines are given for each model. Figure 3. Measured CO2 production (g/d) plotted against predicted CO2 production in a reduced version of the training dataset (n = 562, see Supplemental Table S3), where observations having missing values of days in pregnancy were not included. The black line represents y (measured CO2 production, g/d) = x (predicted CO2 production, g/d); the red and blue lines represent linear regressions of CO2 production pre- dicted by Pedersen et al. (2008) and Madsen et al. (2010), respectively. Regression lines are given for each model. 6782 Journal of Dairy Science Vol. 107 No. 9, 2024 (−0.07). However, dietary CP was positively correlated with urinary nitrogen (g/d) in van Lingen et al. (2018), and in the same study, urinary nitrogen (expressed in g/ kg DMI) was positively correlated with CH4 yield. A part of the explanation could be a higher DM digest- ibility when cows are sufficiently supplied with dietary CP (Oldham, 1984). Thereby, more nutrients are avail- able for intermediary metabolism without increasing the DMI. Excess of absorbed amino acids also causes altera- tions in oxidation, and less efficient conversion of ME to NE, thereby increasing CO2 production (Oldham, 1984). Not surprisingly, dietary CP concentration was therefore a significant predictor variable of CO2 production in model 1 (“best model”), where DMI was also included in the model, and increased dietary CP intake increased CO2 production. Effects of Different CH4-Mitigating Additives or Feedstuffs on CO2 Production A recent study has shown decreased CO2 production and increased CO2 yield (g/kg DMI; Kjeldsen et al., 2024) when dairy cows were fed 3-NOP, even though 3-NOP, by its mode of action, is not expected to affect CO2 metabolism, except from a small increase due to less reduction of CO2 to CH4. These results are in alignment with another study where increased CO2 yield for 3-NOP concentrations of both 60 mg/kg DM (+3%) and 80 mg/ kg DM (+4%) were observed, whereas it was only the diet with 80 mg/kg DM that negatively affected CO2 produc- tion (van Gastelen et al., 2022). Additionally, Maigaard et al. (2024) observed a reduced CO2 production and increased CO2 yield when cows were provided 80 mg 3-NOP/kg DM. In the 3 studies mentioned above, DMI was negatively affected by 3-NOP supplementation for reasons still unclear, which likely at least partly caused the effect on CO2 yield and production. Melgar et al. (2021) and Van Wesemael et al. (2019) did not observe decreased DMI when dairy cows were supplemented with 3-NOP, nor changes in CO2 production or yield; this indicates that a reduction in CO2 production associated with the use of a given potent CH4-mitigating feed addi- tive seems to be related to a potential reduction in DMI. Nitrate acts as an alternative hydrogen sink and com- petes with methanogens in taking up H2 in the rumen (Leng, 2008). Considering the mode of action, nitrate supplementation does not affect CO2 metabolism of the animal, as also not found in the study by Olijhoek et al. (2016), where 5, 14, and 21 g nitrate/kg DM were fed to the cows. However, Wang et al. (2023) included 10 g nitrate/kg DM and observed decreased CO2 produc- tion, when dairy cows were supplemented with nitrate, although likely due to reduced DMI. Increased dietary fat content has also proven to be an effective CH4 mitigation strategy (Beauchemin et al., 2007). The training dataset reflects very different feed rations, and thereby CF levels also varied, from 12 to 74 g/kg DM (Table 2). Metabolism of fat releases more heat (28 kJ/L CO2) than the metabolism of carbohydrates (21 kJ/L CO2; Madsen et al., 2010). However, increas- ing the fat level from 2% to 5% of the diet reduces CO2 production by ~1 percentage unit (Madsen et al., 2010), as the efficiency of using ME to NE of lactation is rela- tively high (estimated to 0.63 in Moraes et al., 2015, and 0.60–0.64 in Moe, 1981), and thus less heat is lost with feeding higher fat concentrations, as long as the mam- mary gland takes up the fatty acids provided by the feed. The study by Maigaard et al. (2024) is one of few to report CO2 emissions when feeding a high level of fat (60–67 g dietary CF/kg DM). They reported a significant effect of fat supplementation on CO2 production, but an interaction was observed between fat and nitrate supple- mentation, and interpretation of the results are affected by this interaction. In conclusion, high (>60 g/kg DM) or low (<30 g/kg DM) CF concentrations of a given diet are not expected to cause less precise estimation of the CO2 production in the current study. CONCLUSIONS Production of CO2 (g/d) from lactating dairy cows can be predicted directly from dietary, animal, and produc- tion traits, without quantifying HP. The absolute values of SB (−0.22 to 0.097) and SB as percentage of MSE (0.049 to 4.89) were very low, which indicates precision of the models. The absolute value of the dependent vari- able (CO2 g/d) should be interpreted accounting for the fact that the models were developed on a dataset contain- ing both RC and GF data, causing a relatively high MB for nearly all models in all evaluations (−740 to 1,619). NOTES The current project “Reduced climate impact at cow- level and herd-level” was funded by the Milk Levy Fund, and the PhD project of M. H. Kjeldsen was funded by iCli- mate (Interdisciplinary Centre for Climate Change, Aar- hus University, Aarhus, Denmark) and Arla (Aarhus, Den- mark). Christian Friis Børsting, Giulio Giagnoni, Morten Maigaard, and Wenji Wang provided very valuable data from their experiments performed at Aarhus University (Viborg, Denmark) to the model evaluation. Supplemen- tal material for this article is available at https:​/​/​www​.erda​ .au​.dk/​archives/​0f1bc3258dbf693109b3d9bb4a94237e/​ published​-archive​.html. Because no human or animal subjects were used, this analysis did not require approval Kjeldsen et al.: PREDICTING CO2 PRODUCTION https://www.erda.au.dk/archives/0f1bc3258dbf693109b3d9bb4a94237e/published-archive.html https://www.erda.au.dk/archives/0f1bc3258dbf693109b3d9bb4a94237e/published-archive.html https://www.erda.au.dk/archives/0f1bc3258dbf693109b3d9bb4a94237e/published-archive.html Journal of Dairy Science Vol. 107 No. 9, 2024 6783 by an Institutional Animal Care and Use Committee or Institutional Review Board. The authors have not stated any conflicts of interest. Nonstandard abbreviations used: 3-NOP = 3-nitro- oxypropanol; BW0.75 = metabolic BW; CCC = concor- dance correlation coefficient; CF = crude fat; DIP = days in pregnancy; GF = GreenFeed head chamber; HP = heat production; HPU = heat-producing unit; MAE = mean absolute error; MB = mean bias; ME = metabolizable energy; MSE = mean square error; NE = net energy; RC = respiration chamber; RMSE = root mean square error; RSR = ratio of RMSE to SD of measured data; SB = slope bias. REFERENCES Alemu, A. W., D. Vyas, G. Manafiazar, J. A. Basarab, and K. A. Beauche- min. 2017. Enteric methane emissions from low- and high-residual feed intake beef heifers measured using GreenFeed and respiration chamber techniques. J. Anim. Sci. 95:3727–3737. https:​/​/​doi​.org/​10​ .2527/​jas2017​.1501. Appuhamy, J., J. France, and E. Kebreab. 2016. Models for predicting en- teric methane emissions from dairy cows in North America, Europe, and Australia and New Zealand. Glob. Chang. Biol. 22:3039–3056. https:​/​/​doi​.org/​10​.1111/​gcb​.13339. Beauchemin, K., S. M. McGinn, and H. Petit. 2007. Methane abatement strategies for cattle: Lipid supplementation of diets. Can. J. Anim. Sci. 87:431–440. https:​/​/​doi​.org/​10​.4141/​CJAS07011. Chwalibog, A. 1991. Husdyrernæring, bestemmelse af næringsværdi og næringsbehov. Faculty of Life Sciences, University of Copenhagen, DSR forlag. CIGR. 1984. Report of working group on climatization of animal houses. Scottish Farm Buildings Investigation Unit, International Commis- sion of Agrocultural Engineering. CIGR. 2002. 4th Report of Working Group on Climatization of Animal Houses. Research Center Bygholm, Danish Institute of Agricultural Sciences, Horsens, Denmark. DePeters, E. J., and J. D. Ferguson. 1992. Nonprotein nitrogen and pro- tein distribution in the milk of cows. J. Dairy Sci. 75:3192–3209. https:​/​/​doi​.org/​10​.3168/​jds​.S0022​-0302(92)78085​-0. Difford, G. F., D. W. Olijhoek, A. L. F. Hellwing, P. Lund, M. A. Bjer- ring, Y. de Haas, J. Lassen, and P. Lovendahl. 2019. Ranking cows’ methane emissions under commercial conditions with sniffers versus respiration chambers. Acta Agric. Scand. A Anim. Sci. 68:25–32. https:​/​/​doi​.org/​10​.1080/​09064702​.2019​.1572784. Doreau, M., M. Arbre, Y. Rochette, C. Lascoux, M. Eugène, and C. Mar- tin. 2018. Comparison of 3 methods for estimating enteric methane and carbon dioxide emission in nonlactating cows. J. Anim. Sci. 96:1559–1569. https:​/​/​doi​.org/​10​.1093/​jas/​sky033. FNN. 2023. Feed and nutrition network. Accessed Nov. 9, 2024. https:​ /​/​globalresearchalliance​.org/​research/​livestock/​networks/​feed​ -nutrition​-network/​. Giagnoni, G. 2023. modRMSE. Accessed Nov. 9, 2024. https:​/​/​github​ .com/​giuliogiagnoni/​modRMSE/​#modrmse. Hellwing, A. L. F., J. Madsen, and M. R. Weisbjerg. 2013. Comparison of enteric methane production predicted from the CH4/CO2 ratio and measured in respiration chambers. Poster session presented at Greenhouse Gases & Animal Agriculture, Dublin, Ireland, Jun. 21, 2013–Jun. 27, 2013. Hristov, A. N., E. Kebreab, M. Niu, J. Oh, A. Bannink, A. R. Bayat, T. M. Boland, A. F. Brito, D. P. Casper, L. A. Crompton, J. Dijkstra, M. Eu- gene, P. C. Garnsworthy, N. Haque, A. L. F. Hellwing, P. Huhtanen, M. Kreuzer, B. Kuhla, P. Lund, J. Madsen, C. Martin, P. J. Moate, S. Muetzel, C. Munoz, N. Peiren, J. M. Powell, C. K. Reynolds, A. Schwarm, K. J. Shingfield, T. M. Storlien, M. R. Weisbjerg, D. R. Yanez-Ruiz, and Z. Yu. 2018. Symposium review: Uncertainties in enteric methane inventories, measurement techniques, and predic- tion models. J. Dairy Sci. 101:6655–6674. https:​/​/​doi​.org/​10​.3168/​ jds​.2017​-13536. Huhtanen, P., A. R. Bayat, P. Lund, A. L. F. Hellwing, and M. R. Weis- bjerg. 2020. Short communication: Variation in feed efficiency hampers use of carbon dioxide as a tracer gas in measuring methane emissions in on-farm conditions. J. Dairy Sci. 103:9090–9095. https:​ /​/​doi​.org/​10​.3168/​jds​.2020​-18559. Hynes, D. N., S. Stergiadis, A. Gordon, and T. Yan. 2016. Effects of concentrate crude protein content on nutrient digestibility, energy utilization, and methane emissions in lactating dairy cows fed fresh- cut perennial grass. J. Dairy Sci. 99:8858–8866. https:​/​/​doi​.org/​10​ .3168/​jds​.2016​-11509. Kai, P., A. P. S. Adamsen, M. L. Jensen, L. P. Kasper, and A. Feilberg. 2017. Ammonia emission from Danish cubicle barns for dairy cows: Effect of floor type and manure scraping. Danish Centre for Food and Agriculture (DCA) report no. 110. Danish Centre for Food and Agriculture. Accessed Jun. 26, 2024. https:​/​/​dcapub​.au​.dk/​ djfpublikation/​djfpdf/​DCArapport110​.pdf. Kirchgessner, M., and H. L. Muller. 1998. Comparison of excessive dietary carbohydrate and fat on energy balance. J. Anim. Physiol. Anim. Nutr. (Berl.) 80:18–30. https:​/​/​doi​.org/​10​.1111/​j​.1439​-0396​ .1998​.tb00496​.x. Kjeldsen, M. H., M. R. Weisbjerg, M. Larsen, O. Højberg, C. Ohlsson, N. Walker, A. L. F. Hellwing, and P. Lund. 2024. Gas exchange, rumen hydrogen sinks, and nutrient digestibility and metabolism in lactating dairy cows fed 3-nitrooxypropanol and cracked rapeseed. J. Dairy Sci. 107:2047–2065. https:​/​/​doi​.org/​10​.3168/​jds​.2023​-23743. Lassen, J., P. Løvendahl, and J. Madsen. 2012. Accuracy of noninva- sive breath methane measurements using Fourier transform infrared methods on individual cows. J. Dairy Sci. 95:890–898. https:​/​/​doi​ .org/​10​.3168/​jds​.2011​-4544. Leng, R. 2008. The potential of feeding nitrate to reduce en- teric methane production in ruminants. A report to the De- partment of Climate Change, Canberra, Australia. Accessed Jun. 26, 2024. https:​/​/​www​.researchgate​.net/​publication/​ 265489637_the ​_po ten t i a l_of ​_ feed ing_n i t ra te ​_ to ​_ reduce​ _enteric_methane​_production​_in​_ruminants​?​_tp​=​eyjjb250zxh0 ijp7imzpcnn0ugfnzsi6inbyb2zpbguilcjwywdlijoichjvzmlszsj9fq. Madsen, J., B. Bjerg, T. Hvelplund, M. Weisbjerg, and P. Lund. 2010. Methane and carbon dioxide ratio in excreted air for quantification of the methane production from ruminants. Livest. Sci. 129:223–227. https:​/​/​doi​.org/​10​.1016/​j​.livsci​.2010​.01​.001. Maigaard, M., M. R. Weisbjerg, M. Johansen, N. Walker, C. Ohlsson, and P. Lund. 2024. Effects of dietary fat, nitrate, and 3-nitrooxypro- panol and their combinations on methane emission, feed intake and milk production in dairy cows. J. Dairy Sci. 107:220–241. https:​/​/​doi​ .org/​10​.3168/​jds​.2023​-23420. Manzanilla-Pech, C. I. V., R. B. Stephansen, G. F. Difford, P. Løvendahl, and J. Lassen. 2022. Selecting for feed efficient cows will help to reduce methane gas emissions. Front. Genet. 13:885932. https:​/​/​doi​ .org/​10​.3389/​fgene​.2022​.885932. Melgar, A., C. F. A. Lage, K. Nedelkov, S. E. Räisänen, H. Stefenoni, M. E. Fetter, X. Chen, J. Oh, S. Duval, M. Kindermann, N. D. Walker, and A. N. Hristov. 2021. Enteric methane emission, milk production, and composition of dairy cows fed 3-nitrooxypropanol. J. Dairy Sci. 104:357–366. https:​/​/​doi​.org/​10​.3168/​jds​.2020​-18908. Moe, P. W. 1981. Energy metabolism of dairy cattle. J. Dairy Sci. 64:1120–1139. https:​/​/​doi​.org/​10​.3168/​jds​.S0022​-0302(81)82692​-6. Moe, P. W., and H. F. Tyrrell. 1972. Metabolizable energy requirements of pregnant dairy cows. J. Dairy Sci. 55:480–483. https:​/​/​doi​.org/​10​ .3168/​jds​.S0022​-0302(72)85519​-X. Moraes, L. E., E. Kebreab, A. B. Strathe, J. Dijkstra, J. France, D. P. Casper, and J. G. Fadel. 2015. Multivariate and univariate analy- sis of energy balance data from lactating dairy cows. J. Dairy Sci. 98:4012–4029. https:​/​/​doi​.org/​10​.3168/​jds​.2014​-8995. Morris, D. L., J. L. Firkins, C. Lee, W. P. Weiss, and P. J. Kononoff. 2021. Relationship between urinary energy and urinary nitrogen or carbon excretion in lactating Jersey cows. J. Dairy Sci. 104:6727–6738. https:​/​/​doi​.org/​10​.3168/​jds​.2020​-19684. Kjeldsen et al.: PREDICTING CO2 PRODUCTION https://doi.org/10.2527/jas2017.1501 https://doi.org/10.2527/jas2017.1501 https://doi.org/10.1111/gcb.13339 https://doi.org/10.4141/CJAS07011 https://doi.org/10.3168/jds.S0022-0302(92)78085-0 https://doi.org/10.1080/09064702.2019.1572784 https://doi.org/10.1093/jas/sky033 https://globalresearchalliance.org/research/livestock/networks/feed-nutrition-network/ https://globalresearchalliance.org/research/livestock/networks/feed-nutrition-network/ https://globalresearchalliance.org/research/livestock/networks/feed-nutrition-network/ https://github.com/giuliogiagnoni/modRMSE/#modrmse https://github.com/giuliogiagnoni/modRMSE/#modrmse https://doi.org/10.3168/jds.2017-13536 https://doi.org/10.3168/jds.2017-13536 https://doi.org/10.3168/jds.2020-18559 https://doi.org/10.3168/jds.2020-18559 https://doi.org/10.3168/jds.2016-11509 https://doi.org/10.3168/jds.2016-11509 https://doi.org/10.1111/j.1439-0396.1998.tb00496.x https://doi.org/10.1111/j.1439-0396.1998.tb00496.x https://doi.org/10.3168/jds.2023-23743 https://doi.org/10.3168/jds.2011-4544 https://doi.org/10.3168/jds.2011-4544 https: / / www .researchgate .net/ publication/ 265489637_the _potential_of _feeding_nitrate _to _reduce _enteric_methane _production _in _ruminants ? _tp = eyjjb250zxh0 ijp7imzpcnn0ugfnzsi6inbyb2zpbguilcjwywdlijoichjvzmlszsj9fq https: / / www .researchgate .net/ publication/ 265489637_the _potential_of _feeding_nitrate _to _reduce _enteric_methane _production _in _ruminants ? _tp = eyjjb250zxh0 ijp7imzpcnn0ugfnzsi6inbyb2zpbguilcjwywdlijoichjvzmlszsj9fq https: / / www .researchgate .net/ publication/ 265489637_the _potential_of _feeding_nitrate _to _reduce _enteric_methane _production _in _ruminants ? _tp = eyjjb250zxh0 ijp7imzpcnn0ugfnzsi6inbyb2zpbguilcjwywdlijoichjvzmlszsj9fq https: / / www .researchgate .net/ publication/ 265489637_the _potential_of _feeding_nitrate _to _reduce _enteric_methane _production _in _ruminants ? _tp = eyjjb250zxh0 ijp7imzpcnn0ugfnzsi6inbyb2zpbguilcjwywdlijoichjvzmlszsj9fq https://doi.org/10.1016/j.livsci.2010.01.001 https://doi.org/10.3168/jds.2023-23420 https://doi.org/10.3168/jds.2023-23420 https://doi.org/10.3389/fgene.2022.885932 https://doi.org/10.3389/fgene.2022.885932 https://doi.org/10.3168/jds.2020-18908 https://doi.org/10.3168/jds.S0022-0302(81)82692-6 https://doi.org/10.3168/jds.S0022-0302(72)85519-X https://doi.org/10.3168/jds.S0022-0302(72)85519-X https://doi.org/10.3168/jds.2014-8995 https://doi.org/10.3168/jds.2020-19684 6784 Journal of Dairy Science Vol. 107 No. 9, 2024 Nielsen, N. I., and H. Volden. 2011. Animal requirements and recom- mendations. Pages 41–54 in NorFor—The Nordic Feed Evaluation System. H. Volden, ed. Wageningen Academic Publishers, Wagenin- gen, the Netherlands. Niu, M., E. Kebreab, A. N. Hristov, J. Oh, C. Arndt, A. Bannink, A. R. Bayat, A. F. Brito, T. Boland, D. Casper, L. A. Crompton, J. Dijkstra, M. A. Eugène, P. C. Garnsworthy, M. N. Haque, A. L. F. Hellwing, P. Huhtanen, M. Kreuzer, B. Kuhla, P. Lund, J. Madsen, C. Martin, S. C. McClelland, M. McGee, P. J. Moate, S. Muetzel, C. Muñoz, P. O’Kiely, N. Peiren, C. K. Reynolds, A. Schwarm, K. J. Shingfield, T. M. Storlien, M. R. Weisbjerg, D. R. Yáñez-Ruiz, and Z. Yu. 2018. Prediction of enteric methane production, yield, and intensity in dairy cattle using an intercontinental database. Glob. Chang. Biol. 24:3368–3389. https:​/​/​doi​.org/​10​.1111/​gcb​.14094. Oldham, J. D. 1984. Protein-energy interrelationships in dairy cows. J. Dairy Sci. 67:1090–1114. https:​/​/​doi​.org/​10​.3168/​jds​.S0022​ -0302(84)81410​-1. Olijhoek, D. W., A. L. F. Hellwing, M. Brask, M. R. Weisbjerg, O. Hojberg, M. K. Larsen, J. Dijkstra, E. J. Erlandsen, and P. Lund. 2016. Effect of dietary nitrate level on enteric methane production, hydrogen emission, rumen fermentation, and nutrient digestibility in dairy cows. J. Dairy Sci. 99:6191–6205. https:​/​/​doi​.org/​10​.3168/​jds​ .2015​-10691. Pedersen, S., V. Blanes-Vidal, H. Jørgensen, A. Chwalibog, A. Haeusser- mann, M. J. W. Heetkamp, and A. Aarnink. 2008. Carbon dioxide production in animal houses: A literature review. Agric. Eng. Int. CIGR J. 10:BC 08 008. R Core Team. 2023. R: A language and environment for statistical com- puting. R Foundation for Statistical Computing, Vienna, Austria. Accessed Nov. 9, 2023. https:​/​/​www​.R​-project​.org/​. Ramin, M., and P. Huhtanen. 2013. Development of equations for predict- ing methane emissions from ruminants. J. Dairy Sci. 96:2476–2493. https:​/​/​doi​.org/​10​.3168/​jds​.2012​-6095. Rischewski, J., A. Bielak, G. Nürnberg, M. Derno, and B. Kuhla. 2017. Rapid Communication: Ranking dairy cows for methane emissions measured using respiration chamber or GreenFeed techniques during early, peak, and late lactation. J. Anim. Sci. 95:3154–3159. https:​/​/​ doi​.org/​10​.2527/​jas​.2017​.1530. Sguizzato, A. L. L., M. I. Marcondes, J. Dijkstra, S. C. Valadares Filho, M. M. Campos, F. S. Machado, B. C. Silva, and P. P. Rotta. 2020. En- ergy requirements for pregnant dairy cows. PLoS One 15:e0235619. https:​/​/​doi​.org/​10​.1371/​journal​.pone​.0235619. Sjaunja, L. O., L. Baevre, L. Junkkarinen, J. Pedersen, and J. Setälä. 1990. A Nordic proposal for an energy corrected milk (ECM) for- mula. 27th Session of the International Committee for Recording the Productivity of Milk Animals. EAAP Public 50:156. van Gastelen, S., J. Dijkstra, J. M. L. Heck, M. Kindermann, A. Klop, R. de Mol, D. Rijnders, N. Walker, and A. Bannink. 2022. Methane mitigation potential of 3-nitrooxypropanl in lactating cows is influ- enced by basal diet composition. J. Dairy Sci. 105:4064–4082. https:​ /​/​doi​.org/​10​.3168/​jds​.2021​-20782. van Lingen, H. J., J. G. Fadel, A. Bannink, J. Dijkstra, J. M. Tricarico, D. Pacheco, D. P. Casper, and E. Kebreab. 2018. Multi-criteria evalu- ation of dairy cattle feed resources and animal characteristics for nutritive and environmental impacts. Animal 12:s310–s320. https:​/​/​ doi​.org/​10​.1017/​S1751731118001313. Van Wesemael, D., L. Vandaele, B. Ampe, H. Cattrysse, S. Duval, M. Kindermann, V. Fievez, S. De Campeneere, and N. Peiren. 2019. Reducing enteric methane emissions from dairy cattle: Two ways to supplement 3-nitrooxypropanol. J. Dairy Sci. 102:1780–1787. https:​ /​/​doi​.org/​10​.3168/​jds​.2018​-14534. Voeten, C. C. 2023. buildmer: Stepwise elimination and term reordering for mixed-effects regression. Accessed Nov. 3, 2023. https:​/​/​cran​.r​ -project​.org/​web/​packages/​buildmer/​. Wang, W., P. Lund, M. Larsen, and M. R. Weisbjerg. 2023. Effect of nitrate supplementation, dietary protein supply, and genetic yield index performance, methane emission, and nitrogen efficiency in dairy cows. J. Dairy Sci. 106:5433–5451. https:​/​/​doi​.org/​10​.3168/​ jds​.2022​-22906. ORCIDS M. H Kjeldsen https:​/​/​orcid​.org/​0009​-0000​-2091​-4721 M. Johansen https:​/​/​orcid​.org/​0000​-0002​-2274​-8939 M. R. Weisbjerg https:​/​/​orcid​.org/​0000​-0002​-6514​-9186 A. L. F. Hellwing https:​/​/​orcid​.org/​0000​-0002​-2881​-399X A. Bannink https:​/​/​orcid​.org/​0000​-0001​-9916​-3202 S. Colombini https:​/​/​orcid​.org/​0000​-0002​-4391​-3905 L. Crompton https:​/​/​orcid​.org/​0000​-0003​-3752​-4804 J. Dijkstra https:​/​/​orcid​.org/​0000​-0003​-3728​-6885 A. Guinguina https:​/​/​orcid​.org/​0000​-0002​-9325​-512X A. N. Hristov https:​/​/​orcid​.org/​0000​-0002​-0884​-4203 P. Huhtanen https:​/​/​orcid​.org/​0000​-0001​-7855​-7448 A. Jonker https:​/​/​orcid​.org/​0000​-0002​-6756​-8616 M. Kreuzer https:​/​/​orcid​.org/​0000​-0002​-9978​-1171 B. Kuhla https:​/​/​orcid​.org/​0000​-0002​-2032​-5502 C. Martin https:​/​/​orcid​.org/​0000​-0002​-2265​-2048 P. Niu https:​/​/​orcid​.org/​0000​-0001​-6858​-1284 N. Peiren https:​/​/​orcid​.org/​0000​-0001​-5500​-1607 C. Reynolds https:​/​/​orcid​.org/​0000​-0002​-4152​-1190 S. R. O. Williams https:​/​/​orcid​.org/​0000​-0003​-1321​-6487 P. Lund https:​/​/​orcid​.org/​0000​-0002​-9113​-4500 Kjeldsen et al.: PREDICTING CO2 PRODUCTION https://doi.org/10.1111/gcb.14094 https://doi.org/10.3168/jds.S0022-0302(84)81410-1 https://doi.org/10.3168/jds.S0022-0302(84)81410-1 https://doi.org/10.3168/jds.2015-10691 https://doi.org/10.3168/jds.2015-10691 https://www.R-project.org/ https://doi.org/10.3168/jds.2012-6095 https://doi.org/10.2527/jas.2017.1530 https://doi.org/10.2527/jas.2017.1530 https://doi.org/10.1371/journal.pone.0235619 https://doi.org/10.3168/jds.2021-20782 https://doi.org/10.3168/jds.2021-20782 https://doi.org/10.1017/S1751731118001313 https://doi.org/10.1017/S1751731118001313 https://doi.org/10.3168/jds.2018-14534 https://doi.org/10.3168/jds.2018-14534 https://doi.org/10.3168/jds.2022-22906 https://doi.org/10.3168/jds.2022-22906 https://orcid.org/0009-0000-2091-4721 https://orcid.org/0000-0002-2274-8939 https://orcid.org/0000-0002-6514-9186 https://orcid.org/0000-0002-2881-399X https://orcid.org/0000-0001-9916-3202 https://orcid.org/0000-0002-4391-3905 https://orcid.org/0000-0003-3752-4804 https://orcid.org/0000-0003-3728-6885 https://orcid.org/0000-0002-9325-512X https://orcid.org/0000-0002-0884-4203 https://orcid.org/0000-0001-7855-7448 https://orcid.org/0000-0002-6756-8616 https://orcid.org/0000-0002-9978-1171 https://orcid.org/0000-0002-2032-5502 https://orcid.org/0000-0002-2265-2048 https://orcid.org/0000-0001-6858-1284 https://orcid.org/0000-0001-5500-1607 https://orcid.org/0000-0002-4152-1190 https://orcid.org/0000-0003-1321-6487 https://orcid.org/0000-0002-9113-4500 Kansi_Kjeldsen-2024-Predicting_CO2_production_of_lactating 1-s2.0-S0022030224007847-main Predicting CO2 production of lactating dairy cows from animal, dietary, and production traits using an international dataset INTRODUCTION MATERIALS AND METHODS Dataset Data Pre-Processing Model Development Model Evaluation RESULTS Description of Models 1, 2, and 3 Evaluation on RC and GF Test Datasets Evaluation on GF+ Test Dataset Evaluation on the Modified Test Dataset Evaluation on the Training Dataset DISCUSSION Overall Model Evaluation on the Test Dataset Gestation Dietary Crude Protein Effects of Different CH4-Mitigating Additives or Feedstuffs on CO2 Production CONCLUSIONS NOTES REFERENCES