Prediction of 100-d and 305-d Milk Yields in a Multibreed Dairy Herd in Thailand Using Monthly Test-Day Records[1]^/

Skorn Koonawootrittriron^*, Mauricio A. Elzo^2/, Sornthep Tumwasorn, and Wirot Sintala ^3/

Department of Animal Science, Faculty of Agriculture,

Kasetsart University, Bangkok 10900, Thailand.

* Correspondence: Present address ^3/ Sakon Nakhon Agricultural Research and Training Center, P.O. Box 3,

Pungkone, Sakon Nakhon 47160, Thailand. E-mail: skornk@hotmail.com

[1]^/ This research was supported by the Florida Agricultural Experiment Station and a grant from

the Thailand Research Fund under the Royal Golden Jubilee Project,

and approved for publication as Journal Series No.R-08065.

^2/ Department of Animal Sciences, University of Florida, Gainesville, FL 32611-0910, USA

Abstract

The ability of eight procedures to predict 100-d and 305-d milk yields using monthly test-day records was tested using 28,452 daily yields from 88 cows in a multibreed dairy herd provided by the Sakon Nakhon Agricultural Research and Training Center. The eight procedures were: test interval method, gamma function, mixed log linear, and second, third, fourth, fifth, and sixth degree polynomial models. The breed groups represented in the multibreed herd were HF, 1/2HF 1/2RS, and 3/4HF 1/4RS. Prediction of 100-d and 305-d milk yields by the eight procedures were compared with actual 100-d and 305-d milk yields within breed group x lactation number x calving age and breed group x lactation number x calving season subclasses. Least squares means of individual cow differences predicted and actual 100-d and 305-d milk yields were computed for each subclass. Number of significant least square means of differences and ranking of models within and across subclasses for 100-d and 305-d were used to evaluate the predictive ability of the eight procedures. The highest-ranking model for 100-d was model 4 (third degree polynomial) and for 305-d was model 3 (second degree polynomial). However, no procedure was uniformly better across all subclasses. Thus, perhaps several models might be needed for a genetic evaluation of the animals in this multibreed population. If computational simplicity were the primary goal, then perhaps a single model (model 3) might suffice. However, the results of this study apply only to the data set and the multibreed population used here. To obtain results of national relevance, this study needs to be repeated with a larger multibreed population that more accurately represents the Thai multibreed population.

Key words: dairy cattle, milk yield, test-day yield, prediction, multibreed

Introduction

Recording of milk yields is essential for genetic improvement and herd management in dairy cattle. Under increasing pressure to reduce cost, numerous milk-testing schemes have been developed in many countries. One of the most widely used is monthly recording. In Thailand, monthly test-day records are used to compute cumulative productions of milk and fat to 100-d and 305-d for dairy genetic evaluation purposes. The Dairy Promotion Organization (DPO), and probable other organizations in Thailand, compute monthly milk yields using a single test-day milk yield sample, and then, these monthly estimates are used to compute the accumulated 100-d and 305-d milk yields. This procedure is not appropriate for individual animals because it will, in most cases, either overestimate or underestimate accumulated milk yields. Notice that this procedure is different from the test-interval method (Sargent et al., 1968; Norman et al., 1999), which computes total milk yields of intervals between two consecutive test days using the average milk yield of these two test-days.

Although the test-interval method is the most widely used procedure to compute cumulative milk production traits, prediction of these milk traits could potentially be improved by using a linear or a nonlinear function. Koonawootrittriron et al. (2001) showed that the second degree polynomial was the best out of seven models to predict daily and 305-d milk yields and the sixth degree polynomial model was the best for the prediction of 100-d milk yield within breed group x lactation number x calving age and breed group x lactation number x calving season subclasses in a Holstein Friesian-Red Sindhi herd in the Northeast of Thailand, when using all daily lactation records.

Milk recording organizations in Thailand sample milk production traits on a monthly basis. These are the records used for genetic animal evaluation in dairy cattle. Thus, milk prediction models need to be revalidated under a test-day sampling strategy, and then compared to the test-interval method for their ability to predict 100-d and 305-d milk production yields under Thai conditions.

Thus, the objectives of this study were to assess the predictive ability of the test-interval method and of seven models (gamma, mixed log linear, second to sixth degree polynomial models) to predict 100-d and 305-d milk yields based on monthly test-day records relative to the actual 100-d and 305-d milk yields of individual cows within breed group x lactation number x calving age and breed group x lactation number x calving season subclasses.

Materials and Methods

Animals, Management, and records

This study used the same data set as Koonawootrittriron et al. (2001). Thus, only a reduced description of it will be given here. Daily lactation yields (28,452, 5 to 305 d) from 75 Holstein Friesian (HF), 8 ½ HF ½ Red Sindhi (RS), and 5 ¾ HF ¼ RS dams were collected at the Sakon Nakhon Agricultural Research and Training Center (SARTC) between 1997 and 1999.

Cows were assigned to three breed groups according to their breed composition (HF, 1/2HF 1/2RS, and 3/4HF 1/4RS). Animals of all breed compositions were milked twice a day, and raised under the same nutritional (grass and concentrate plus minerals) and management conditions. Cows were artificially inseminated up to three times, and if not pregnant at 60 d after last insemination, they were placed with a clean up bull. Pregnant cows were dried off two months prior to calving.

Lactation number was classified as first, second, third, and fourth and later lactations. Calving seasons were defined as winter (November to February), summer (March to June), and rainy (July to October). Two calving ages per lactation were defined. This resulted in eight lactation x calving age subclasses: 1) calving age less than 30 months x lactation 1, 2) calving age equal to or greater than 30 months x lactation 1, 3) calving age less than 44 months x lactation 2, 4) calving age equal to or greater than 44 months x lactation 2, 5) calving age less than 60 months x lactation 3, 6) calving age equal to or greater than 60 months for the third lactation, and 7) calving age greater than 60 months x lactation 4 and greater.

Models and Data Analysis

To accomplish the objectives of this research lactation records from the SARTC data set had to be sampled according to the current milk sampling procedure used in Thailand. Monthly sampling of milk production traits is the prevalent system in Thailand. Because of colostrum, sampling usually begins 5 d postpartum. Thus, daily milk yields from the SARTC data set were sampled on days 5, 35, 65, 95, 125, 155, 185, 215, 245, 275, and 305 of each lactation. These 11 measurements per lactation were used as monthly test-day records to assess the predictive ability of the test-interval method (TIM) and the seven prediction equations to predict 100-d and 305-d milk yields used by Koonawootrittriron et al. (2001).

Firstly, the 11 monthly test-day records were used to predict individual cow lactation daily milk yields (5 to 305d) within breed group x lactation x calving age and breed group x lactation x calving season subclasses using the test-interval method and the seven prediction equations. Secondly, accumulated 100-d and 305-d milk yields for individual lactations were computed using these eight prediction procedures. The predicted values for 100-d and 305-d milk yields from each procedure were deviated from their corresponding actual 100-d and 305-d milk yields. Least squares means of 100-d and 305-d milk yield deviations (predicted minus actual milk yields) for the test-interval method and the seven prediction procedures were obtained separately for each set of breed group x lactation x calving season and breed group x lactation x calving age subclasses. The statistical model used was:

d_ijk = + subclass_i + model_j + e_ijk

where

d_ijk = difference between predicted and actual milk production of cow k at 100-d or at 305-d of lactation, within subclass i and model j,

= overall mean,

subclass_i = i^th breed group x lactation x calving season or breed group x lactation x calving age,

model_j = j^th prediction model and,

e_ijkl = residual.

All effects in the model were assumed to be fixed, except for the residual term that was assumed to be independent, identically distributed with mean zero and a common variance. T-statistics were then used to test if the predicted and actual 100-d and 305-d milk yields differed significantly.

For completeness, the prediction equations used to compute accumulated milk yields at 100 d and 305 d are briefly described below. For further details on prediction models 1 to 7, see Koonawootrittriron et al. (2001).

1) Test-interval method:

[1]

where TMY is total milk yield, P₁ is the milk yield of the first test-day record, D₁ is interval between five days after calving and the first record, P_i is the milk yield on test-day i (i = 2,…, k), D_i is the interval between test-record record i – 1 and i (i = 2,…, k), P_k+1 is the milk yield on the last test-day before drying off, and D_k+1 is the interval between the last test-day and the day a cow was dried off (Sargent et al., 1968).

2) Prediction model 1: Gamma function (Wood, 1967):

[2]

where y_t is the milk yield on day t in each subclass, a is the initial yield of lactation, b represents the increasing slope, and c represents the decreasing slope. Computations were done using the natural logarithm of equation 1, i.e.,

where is the residual. Thus, the predicted yield on day t was, y_t = exp (ln y_t).

3) Prediction model 2: mixed log second-degree polynomial (Ali and Schaeffer, 1987),

[3]

where = milk yield on day t, = t /305, = ln(305/t), t = days since calving or days in milk, b₀, b₁, b₂, b₃, and b₄ are regression coefficients, and e_t is the residual.

4) Prediction models 3, 4, 5, 6, and 7: second, third, fourth, fifth, and sixth polynomial regression models,

Model 3: [4]

Model 4: [5]

Model 5: [6]

Model 6: [7]

Model 7: [8]

where = milk yield on day t, t = days since calving, b₀, b₁, b₂, b₃, b₄, b₅ and b₆ are regression coefficients, and e_t is the residual.

Program PROC REG of the SAS statistical system (SAS, 1990) was used to compute the regression coefficients in models 1 through 7 and to obtain the predicted daily milk yields (5 to 305 d) for all models.

Accumulated (100 d and 305 d) individual cow actual and predicted milk yields, and predicted (test-interval method, models 1 to7) minus actual accumulated milk yield deviations were computed using the general SAS program (SAS, 1990). Least squares means of accumulated milk yield deviations per subclass were obtained using the LSMEANS statement of PROC GLM (SAS, 1990). T-tests were used to assess the significance of the accumulated deviations per model within and across subclasses.

Results and Discussion

Predicted 100-d milk yields by eight procedures relative to actual yields

Calving age subclasses. Least squares means of 100-d milk yield deviations (predicted minus actual milk yields) of the test-interval method and seven prediction models by breed group x lactation x calving age subclasses are presented in Table 1. Models 2, 4, 5, 6, and 7 had nonsignificant 100-d milk yield deviations (predicted minus actual milk yield (P > 0.05) for all subclasses. The test-interval method (TIM) had two, and model 1 had three significant differences of LS means for 100-d milk yield deviations (at least P < 0.05). Model 3 had four significant differences (at least P < 0.05). Among the models with nonsignificant deviations, two tiers can be distinguished: models 4 and 5 (mean absolute deviation = 6.3 kg), and models 2, 6, and 7 (mean absolute deviation of about 8.6 kg). Thus, models 4 and 5 appear to be good choices for 100-d milk predictions. However, because it is simpler than model 5, model 4 (third degree polynomial) should be the model of choice. Notice that this choice of model for 100-d milk yields using 11 test-day records here differed from the best model found (model 7) when using all daily records by Koonawootrittriron et al. (2001). The number of test-days considered (and probably the test-day values themselves) would likely affect 100-d predictions, and the model that will have the smallest deviations from actual 100-d milk yields.

Table 1. Least squares means of 100-d milk yield deviations (predicted minus actual milk yields) of the test-interval method and seven prediction models by breed group x lactation x calving age subclasses

Breed group^1^/	Lactation number	Calving age	No. of lactations	Actual Yield (kg)	TIM^2^/	Model 1	Model 2	Model 3	Model 4	Model 5	Model 6	Model 7
HF	1	< 30 mo	8	1,061.5	-52.4	-121.8	119.4	-12.7	-2.4	-4.2	-3.9	-8.9
HF	1	> 30 mo	10	1,210.9	-42.3*	-46.7**	-6.6	-62.4**	-32.6	-15.8	-13.2	-13.5
HF	2	< 44 mo	19	1,072.3	-41.1	-12.3	7.7	-18.0	-8.7	-12.5	-8.9	-6.0
HF	2	> 44 mo	24	1,394.8	-50.4*	-41.5	-46.9	-57.9**	-33.3	-14.7	-13.8	-13.1
HF	3	< 60 mo	6	1,214.7	-12.8	68.0	59.5	11.3	10.5	12.6	16.4	16.7
HF	3	> 60 mo	3	1,401.4	33.7	56.6	107.4	32.0	66.1	70.9	77.3	66.5
HF	> 4	all ages	18	1,261.3	-7.2	-31.0	20.1	26.2	34.1	52.6	56.2	56.2
1/2HF 1/2RS	1	> 30 mo	4	900.7	85.1	108.7*	-19.7	-16.2	-5.1	-9.2	-12.4	-12.0
1/2HF 1/2RS	2	< 44 mo	3	1,228.8	-42.2	36.5	36.6	-4.0	-15.8	23.0	39.9	40.6
1/2HF 1/2RS	2	> 44 mo	3	1,125.1	-4.6	-100.4	146.7	-6.9	-0.3	20.8	19.5	19.4
1/2HF 1/2RS	3	< 60 mo	3	1,481.4	-9.1	92.9**	27.5	-1.7	8.2	23.3	19.9	23.3
3/4HF 1/4RS	1	< 30 mo	2	1,415.9	-47.4	-33.6	-2.2	-114.3**	-60.0	-4.3	-8.4	-6.8
3/4HF 1/4RS	1	> 30 mo	3	1,417.0	-5.8	1.3	11.2	-61.8*	-22.1	18.8	19.6	18.7
All	All	All	106	1,234.3	-27.2*	-17.7	8.4	-22.1	-6.3	6.3	8.6	8.7

^1/ HF = Holstein Friesian, RS = Red Sindhi

^2/ Test-interval method

^3/ Model 1: Model 2:

Model 3: Model 4:

Model 5: Model 6:

Model 7:

Where y_t is milk yield at day in lactation t, t = days in lactation,