September 2006 - CGIL

Participants
Larry Schaeffer, CGIL
Ignacy Misztal, Georgia
Janusz Jamrozik, CGIL
Asheber Sewalem, AAFC/CDN
Jalal Fatehi, CGIL
Flavio Schenkel, CGIL
DCBGC – Summary of Meeting
September 6, 2006
Room 141, Dept Animal & Poultry Science
Dr. Lin, AAFC/CGIL
Gerrit Kistemaker, CDN
Pete Sullivan, CDN
Brian VanDoormaal, CDN
Jacques Chesnais, Semex
Filippo Miglior, AAFC/CDN
1 Production Traits
i.) Genetic Evaluation, Domestic
1.1 Robust procedures for Canadian test day model
Jamrozik, Fatehi and Schaeffer
Bethany Muir, HAC
Jarmila Bohmanova, CGIL
Jose Moro, McGill
Sarah Loker, CGIL
Ted Burnside
Valentina Palucci, CGIL
The objective of the research was to apply several robust estimation procedures to the CTDM
using data on Canadian Jersey cows and to compare these results to regular BLUP analysis.
Outliers were defined as a 24 hour yield (milk, fat or protein) that was significantly beyond the
range of the remaining yields made by cows of similar age, in the same lactation, on the same
test-day in the same herd. This definition did not necessarily correspond to the usual definition of
outliers. The model used in this study was the same as the official multiple trait multiple lactation
RR test-day model as is used in routine genetic evaluation by CDN. Estimation procedures were
compared by: sum of squared residuals, sum of absolute residuals, average and standard deviation
of residuals. Correlations between the BLUP and the 5 robust methods were estimated for all
animals and the first two regression coefficients by trait and parity.
Results were summarized in 19 tables. Two ways of calculating residuals in the model were
applied: using values of corrected observations and using original observations when calculating
residuals. All robust procedures were superior over the BLUP method when corrected
observations were used for residuals. Since outliers defined in this study were model dependent,
different methods (values of k) defined outliers in a different way. Methods with smaller k values
detected a larger number of outliers and some of these records should probably have been treated
as normal data.
DIM 5 – 15 were associated with a larger number of outliers compared to the remainder of the
lactation. Differences in ranking of animals were small. Persistency and SCS evaluations were
more subject to change than yield traits between methods. Cow rankings were more affected than
sire rankings. No clear association was noted between the ranking for yields and change in
methods. Researchers concluded that robust estimation procedures can be applied to the CTDM
for genetic evaluation of production traits and SCS, with a simple k method, k=2.75.
It was noted that observations were modified in each iteration. If each observation were compared
to the expectation, one would expect the SSR to vary. A suggestion was made to investigate the
predictive value of EBV. This procedure would be more complicated, but has not been attempted
yet.

There seemed to be no clear relationship between the percentage of daughters with modified data
and the ranking of sires. A question was posed as to whether it would be possible to apply a
different method for outliers to yield traits than to SCS, since there tended to be more erroneous
yield data compared with SCS. This method would be possible, but one would always have to
keep in mind that there is a built-in correlation structure between the 4 traits in the model.
DHI manipulates and screens data before sending it to CDN for inclusion in genetic evaluation. In
addition, they also place holds on records until the subsequent test so that outlying records from
the previous test can be validated. Record screening is only performed on yield traits, not on SCS.
It was noted that this screening is only performed on a cow-to-cow comparison, not cow-to-herd
basis. It was estimated that DHI screens approximately 3-5% of the records for inclusion in
genetic evaluation. Calculating the percentage of records expected to be outliers would not easy
since outlying records are defined by DIM and are specific to each cow and her predicted
lactation curve. Producers are aware of DHI’s editing process so it is no longer beneficial to
attempt to manipulate yields on one test day. This manipulation must be constantly conducted in
order to go unrecognized. Due to owner-supervised testing, it is no longer beneficial to
manipulate owner tests, since supervised tests will tend not to pass the editing process and be
removed every time.
Breeding Values are accounted for with the robust procedure, animals therefore animals that are
outliers because of their superior genetics would not be discounted. The adjustment for
heterogeneous variance and robust estimation procedures are interconnected. If heterogeneous
adjustments were performed prior to estimation, then fewer outliers would be expected. It is
possible to have a different k value for different traits.
1.2 Phantom Genetic Groups and Genetic Trend
Schaeffer
At the previous technical meeting, the possibility of smoothing the effects of phantom genetic
groups was proposed. In this study, a small (non-species specific) population was simulated to
investigate the effects of different genetic grouping methods. Eight generations of selection (for
two correlated traits) and matings were performed. Missing parents were randomly assigned,
therefore results from this study mat not be as bias as they might be in the actual dairy cattle
population. Four different models were investigated; a simple animal model, a model that
included a year of birth effect, a model including phantom parent groupings, and lastly a model
that included year of birth effects and phantom parent groupings. Three different percentages of
unknown parents (0%, 10% and 25%) were simulated for each model. Variance components,
breeding values and genetic trends were estimated.
Correlations between the true BV and estimated BV were calculated. Only small differences
between models were realized. The highest correlation was found for the simple animal (without
year of birth effects or phantom groups). As the percentage of unknown parents increased, the
correlation decreased. When year of birth effects are added to the model, unknown parents cause
problems in estimating breeding values.
Average inbreeding coefficients for animals born were summarized. As the percentage of
unknown parents increased, the inbreeding coefficient decreased. Therefore when an increasing
amount of the pedigree was missing, the inbreeding rate was significantly underestimated. No
statistical differences between models were found when comparing genetic and residual variances
and heritability. The simple animal model gave unbiased estimation for all pedigree structures.

Therefore, one can assume that phantom groups can be ignored under any pedigree structure.
True and estimated genetic trend were compared. The basic animal model (no year of birth or
phantom group effects included) always underestimated the true genetic change in the population,
and bias increased as the percentage of unknown parents increased. Genetic average by
generation shows that bhat solutions did not match true genetic change. If one investigates the
phantom parent group solutions for one replicate one can see that they are not smooth and are
quite large and do not follow true genetic change. Researchers concluded that phantom group
solutions do not have to follow genetic trend, and that year of birth effects should be included in
the estimation model. If phantom group solutions are smoothed, problems will be encountered.
A concern was raised that since this was a small generic data set that the results and conclusions
would not have a practical application. Since only a single sample was generated, we would not
expect the phantom group solutions to be smooth. If several runs were performed and results were
averaged, then the solutions would become smoother.
The study showed that when year of birth effects were included in the model, phantom groups
were not necessary. In addition, the estimated genetic parameters were not different between
models, suggesting that inclusion of phantom group effects for routine estimation of parameters
was not needed.
Since it has been proven that phantom group solutions do not have to follow genetic trend, a
proper protocol for determining groups needs to be established. The study showed the difference
between approaches was negligible as long as there were enough animals in each group. When
group size increased, the effects tended to smooth out, and parents of cows tended to follow
patterns more than parents of bulls. Dams of cows were more likely to be missing compared to
sires of bulls. If a sire of a bull were missing, it would tend not to be an animal that would be
highly used in the population.
Internationally, pathways (SD, DD, SC, DC) have different levels of missing information. The
results of this study show that the solutions for DC increase by year, but only represent 1/3 of the
true genetic change. Number of animals included in the groups would only be an issue under the
assumption of fixed phantom groups; under a random effect assumption, the size of the group
does not matter.
1.3 Effect of Pregnancy on Milk Yield of Canadian Dairy Cattle
Bohmanova, Miglior, Kelly, Kistemaker and Locker
Preliminary results for updating the TDM were presented. The objectives of the current study
were to evaluate effect of pregnancy on milk production using test-day milk yield and
insemination records. Data were from all seven dairy breeds. Only results from Holstein, Ayrshire
and Jersey were summarized in the report. Two models were compared, one with an effect of
days open (12 classes) and one with an effect of stage of pregnancy (9 classes). In order to
validate third lactation information, a fourth calving had to be confirmed. Detailed insemination
data were only uniformly available after 1996.
Effect of days open on lactation curves was consistent across breeds. Overall yield was lowest for
cows that were never pregnant and as cows stayed open longer, milk production increased. Effect
of days open seems to be confounded with the production level of the cow. Pregnancy effects
were almost negligible at beginning of lactation and became more important as lactation
progressed. The magnitude of decline in milk production as pregnancy progressed was larger than

eported in other studies. In addition to the effects of pregnancy this drop in milk production was
confounded by effect of DIM.
It was noted that more than 35% of the cows were found to never become pregnant in first
lactation, which was an expected result. However, 60% of the cows were recorded as not yet
becoming pregnant when the test was taken. This percentage was expected to be closer to 30%.
Cows coded as not becoming pregnant until after 310 days in milk had the highest level of milk
production. Perhaps the owners of these cows chose not to inseminate these cows until later in
lactation because they were high producers. A suggestion was made to graph the decline of milk
production in terms of percentage of yield rather than by kilograms of yield so that the decrease
can be compared across breeds. In addition, estimation of the interaction between stage of
lactation and stage of pregnancy was recommended.
2 CONFORMATION TRAITS, i. Genetic Evaluation (Domestic)
2.1 Table showing if Fixed or Random contemporary group was included in analysis
This table was given to the committee for reference, but was not discussed.
2.2 Approximate Reliability Patch for Conformation
Sullivan and VanDoormaal
CDN started providing Parent Averages (PA) for animals whose index or proof could not be
calculated. Due to some inconsistencies between reliabilities calculated in February and those
calculated in May (specifically unexpected drops in reliabilities), a review of the reliability
approximation procedure used in the new type system was initiated. Researchers found no errors
in the methodology, but rather an issue with how the approximation was designed. Drops in
reliability from one proof run to the next, although theoretically possible, could cause problems
with extension. Therefore, it was decided to modify the procedure to ensure that the PA reliability
never went down.
The modification (or patch) substituted the matrix of Tier and Meyer with one that accounted for
knowing sire and dam were mates in the parent contribution of the approximation algorithm. In
this case, the parent contributions to reliability were corrected to be a maximum of 50% (rather
than 40%).
Changes in reliability between the old and new approximation procedures were summarized.
Changes were small for animals with either low or high reliability. Animals with originally
intermediate reliability (20-60%), which would be mostly based on parent information, changed
by 2 to 3 percentage points. After the patch was applied, reliabilities only increased as desired.
Changes in reliabilities between procedures were smaller if old reliabilities were high.
Researchers concluded that extension problems related to decreases in parent average reliabilities
would have mostly applied to animals that did not have contemporaries for their first official
evaluation, since their own performance record would not contribute to the reliability. Small
contemporaries are less common in production evaluations. Therefore, investigation into applying
a similar patch to the reliability approximation procedure for production traits was not suggested.

There was a general concern that reliabilities tend to be overestimated and that warranted further
investigation. The patch was implemented in August.
2.3 Including Reclassifications after 1 st Lactation in Genetic Evaluation
Sullivan, Kistemaker and Van Doormaal
Prior to May 2006, only first classifications in first lactation were included in calculation of bull
evaluations for type and the most recent classification was included for calculation of cow
indexes. Currently, all first lactation classifications (FIRST) are included for calculation of both
indexes and proofs (one data set). Cow indexes were significantly inflated in the previous system,
because only the most recent classification was included but compared to her contemporaries
when she was first classified. Under the current system (using all first lactation classifications)
some over-inflated cow indexes dropped significantly. In addition, cows that were never
classified in first lactation lost their index. Bull proofs were only minimally affected. This study
was conducted to investigate the inclusion of all classifications (ALL) or the cow’s most recent
classification in addition to all her first lactation classifications (LATEST) for genetic evaluation.
A majority of classification data was recorded in first lactation. The evaluation model included an
effect of age*stage*time, an interaction for type of record (first or re-classification) in first or
later lactations (4 in total), and an effect of herd-round-classifier.
Correlations between bull proofs under the FIRST and LATEST methods were high for all traits.
No difference in bull proofs was noted between the LATEST and ALL methods. Average
changes in conformation EBV were essentially zero. Approximately half of the cow indexes and
bull proofs for conformation did not change moving from the FIRST to the LATEST or ALL
methods. Of those proofs or indexes that changed a majority changed by plus or minus 1. Range
in change for cow indexes was slightly larger than for bull proofs. Proofs of bulls who had
daughters that reclassified did not change more than proofs of bulls that did not have reclassified
daughters, moving from the FIRST to the LATEST method. As the percentage of daughters
reclassified increased, the reliability of the bulls proof increased more (~2.5).
Differences between proofs changing from the FIRST to LATEST and FIRST to ALL methods
were not large. When a high percentage of daughters were reclassified reliabilities increased more
under the ALL method (to a maximum of ~5) compared to that under the LATEST method. As
the number of reclassifications increased so did the reliability of the cow indexes switching to the
LATEST or ALL methods. In some cases, average change in reliability of cow indexes was an
increase of 30%. Cows that originally scored higher tended to have greater increases in EBV due
to their reclassification in a later lactation. Cows with higher scores were the only cows
reclassified multiple times in later lactations. EBVs increased the most for cows with the largest
number of classifications.
The average phenotype in reclassified daughters was higher compared to the rest of the
population. Animals that were reclassified 6 times were more likely cows scored multiple times
as being excellent. It would be interesting to compare a bull’s first proof to that once his
daughters have been reclassified. All increases were assumed to be genetic. Permanent
environmental variance should be substantial and were not estimated or accounted for under the
proposed model. However, if a permanent environment and frequency of reclass effect was
included in the model, essentially no genetic change would be realized.
Under the new classification system, animals that are presented for classification but not rescored
and raised can be accounted for. The possibility exists that bias was introduced because

classification scores can only increase may partially be removed if a management effect of
percent of animals presented to for a raise can be included.
Concerns that animals were not receiving indexes due to the requirement that a fist lactation
classification be recorded are avoided if the LATEST method is chosen. However, under the
LATEST method, only the most recent reclassification is used if an animal was classified several
times after first lactation. In this way, animals would be removed from contemporaries from
classifications previous to her latest (in later lactations). The remainder of the contemporary
would be reshuffled and as a result EBVs of the remaining contemporary may increase or
decrease just from that one animal being removed. This situation would not occur under the ALL
method, when all classifications would be used and therefore animals would not have to be
removed from contemporary groups.
Clearly the ALL method should be selected over the LATEST method for this reason. If the ALL
method does not pass the trend validation, then the LATEST method should be investigated
further. The possibility of assuming a non-zero permanent environmental variance for the
calculation of reliabilities was suggested to keep the reliabilities from being inflated with a large
number of reclassifications.
There is a computational benefit to having cow and bull evaluations calculated from the same
dataset. As long as cow reliabilities can be kept from increasing too much this would be a
preferred methodology. Perhaps there should be a limit placed on the amount of information later
classifications can contribute to the reliability calculation.
CDN will perform trend validations for Interbull and calculate stability of proofs under the ALL
method. Results will be presented to the Genetic Evaluation Board at the end of September.
3 AUXILIARY TRAITS
i.) Survival, Longevity
3.1.1 Relationships Between Reproduction Traits and Functional Longevity in Canadian
Dairy Breeds
Sewalem, Kistemaker, Miglior, Van Doormaal
This research was a continuation of previous studies that reported relationships between type
traits and longevity. Relationships between calving and fertility traits and longevity were
investigated in the Ayrshire, Jersey and Holstein breeds. A proportional hazards model and
survival kit were used. Animals that had calvings requiring a hard pull or surgery were more
likely to be culled. Animals having small or large calves were more likely to be culled than
animals having medium sized calves. Since the risk of culling was only minimally higher for
large calves compared to medium calves, one should caution making strong conclusions about
this relationship.
Relative risk for culling was highest for animals giving birth to dead calves. As number of
insemination services increases so did the relative risk of culling. Relationships between calving
traits and longevity were stronger than between fertility traits and longevity. The significant
relationship between stillbirth and culling rates was surprising, however, stillbirth was not
analyzed independently from calving ease. Dead calves can result from infections, calving
difficulties and many other factors.

3.1.2 Analysis of Stillbirth Rates (direct and maternal) in Canadian Holsteins
Van Doormaal and Brand
Interest in rates of stillbirth continues to grow. As a result, CDN performed an analysis to
describe differences between maternal and direct stillbirth rates in first and later lactations. In
most recent years, stillbirth rates in first lactation increase to 12% compared to a steady trend of
10% in past years. Later lactation stillbirth rates have continued to increase steadily from 4% to
6% since 2000.
In most recent birth years of sires, maternal stillbirth rates were higher. Distribution of first
calving direct stillbirth rates have not changed from 1992-1994 compared to 2001-2003.
Distribution of first lactation maternal stillbirth rates has shifted to the right. The correlation
between maternal stillbirth rates in first and later lactations was estimated to be 60%. Bulls with a
high maternal stillbirth rate based on first lactation were expected to have high rates for stillbirth
rates for later calvings. In previous reports, the genetic correlation between first and later calving
stillbirth rates was estimated to be 30%. The correlation between direct and maternal stillbirth
rate in first lactation was estimated to be 41%.
Bulls having high maternal stillbirth rates were investigated too see if this problem ran in certain
bloodlines. Bulls such as Bellwood, Rotate and Marshall were found to be undesirable. Bulls such
as Blackstar, Starbuck and Aerostar were found to have desirable rates for maternal stillbirth.
Hopefully this analysis raises huge interest to have maternal and direct stillbirth breeding values
calculated and will create an awareness of a potential problem in certain bloodlines.
ii. Fertility
3.2.1 Characterization of Body Condition Scores in Canadian Dairy Cattle
Monardes, Cue and Moro-Mendez
A visiting scientist from McGill presented results from ongoing research with Body Condition
Score (BCS) in Québec herds. Objectives of the study were to characterize BCS recorded under
commercial conditions and to assess the importance of certain environmental effects on BCS. In
Québec, BCS is recorded by producers and collected by Valacta technicians on test day.
Approximately half of the cows in each parity had only one record of BCS and over 30% of the
records were recorded before 60 DIM. Observations were analyzed as repeated records by cow
using a classification model and a Wilmink model for Ayrshire and Holstein separately in each of
the first 5 lactations.
Average BCS in Holsteins was lower than 3.00 in all parities, and average BCS in Ayrshires was
generally higher than that in Holsteins. First lactation Holsteins tended to have higher average
BCS if they calved at an older age. BCS curves tended to mirror lactation curves and were of
generally the same shape in Holsteins and Ayrshires. Loss of BCS from DIM 1 to 60 ranged from
0.24 units in first lactation to 0.44 units in third lactation. After 60 DIM, BCS increased in each
lactation.
A suggestion was made to plot Wilmink curves for every lactation (by breed) on the same graph
in order to clearly see differences in shape of curves across lactations. In addition, and interaction
of herd-year-month was suggested for inclusion in the model. Researchers should consider a
variance adjustment by herd in order to account for differences in producer scoring methods

scoring. In addition, investigation of a random regression test day model or adjusting BCS for
level of production was suggested. Data being collected by Holstein Canada and Valacta should
be merged and analyzed jointly. Research on the calculation of genetic parameters continues.
iii. Health Traits
3.3.1 Health Traits: Combining Veterinary Data from DS@HR and CDN/Valacta
Cue and Moro-Mendez
The feasibility of merging information on health events recorded by Québec veterinarians
(DS@HR) with those collected by milk recording (Valacta) is being investigated. Data structures
and file layouts were provided. Attempting to match records required merging calving dates
provided by CDN. Approximately 60% of the records matched. Over 157K Holstein and 8.3K
Ayrshire records were available. Further large-scale combination of data will require significant
data editing.
Concern was raised over the different ways events are recorded between Ontario and Québec. All
terminology will be based on the current DS@HR system. Therefore, links to other system
terminology need to be established. To be efficient, researchers should attempt to fit this data
with specifications for Dr. Schaeffer’s proposed model for genetic evaluation of health traits.
3.3.2 Multiple Trait Test Day Model for Health Traits in Dairy Cattle
Schaeffer
A research proposal was submitted to DairyGen to investigate a multiple trait, linear test day sire
model for the analysis of 24-hr milk yields, somatic cell scores, body condition scores and nine
health traits. A summer student, Shannon Cartwright, provided a literature for the proposal.
Model effects include; stage of lactation, gestation length prior to that lactation, calving ease
score, year and month of calving, age*parity, interactions, herd-test-day-parity, sire and cow
within sire. If an animal experienced a disease after the associated test date, it received a 1, and if
no disease event was recorded the animal received a 0. If a disease occurred more than once
between test days, only the first occurrence was recorded. The model assumed that each animal
had milk yield, SCS and BCS recorded and was observed for all nine health traits on every test
day. Therefore, missing data on a test date were not allowed.
To begin with, the height of lactation curve or mean levels of disease occurrence will be
estimated for sires. Since calving ease and BCS will be included as covariates, reduction of the
genetic variation in disease traits is not expected. Cows with more than one disease can be
included and subsequently correlations between the disease traits can be estimated.
Currently, BCS is only collected in Québec herds on test day by milk recording technicians,
however, in the future, this requirement may become mandatory across Canada. In June, Holstein
Canada began classifying animals for BCS on classification day. BCS recorded on classification
day may not be ideal because the event is best recorded before and after calving.
The collection system for health and disease traits will focus on 8 of the most prevalent traits.
Data will be a combination of farmer-recorded events, events collected by veterinarians or by
DHI or on-farm data management systems. Since the proposed model is a sire model, cow

indexes will not be made available. Female indexes for low heritability traits with low incidence
are not useful. CDN calculates female indexes for every trait, but not all are published.
A suggestion was made to measure the occurrence of disease in the month when it is most likely
to occur. If diseases that are more likely to occur in the first month of lactation, assessing the
occurrence of that disease over the entire lactation may disable detection of differences between
sires during the most frequent times. Attempts to analyze health data using a random regression
model have given strange results in previous studies. Other suggestions were to include
fat/protein ratio together with 24 milk yield and SCS, as indicator traits the model. It has been
shown in literature that fat/protein ratio can be a very good indicator of certain diseases.
Research can begin as soon as data are available.
3.3.3 Genetic Evaluation of Health Traits in Canadian Dairy Cattle: Data Characteristics
Fatehi, Schaeffer and Jamrozik
The purpose of the study was to examine health data recorded by DHI and to create a usable data
set for future analysis of genetic parameters and genetic evaluation. Data were recorded from
1997-2006. The most frequently recorded traits were mastitis (37%), cystic ovary (14%), retained
placenta (11.5%) and lameness (9.5%). Records were available for all breeds, however, 96% of
the data were from Holsteins.
Health events were analyzed for 5 lactations; some traits had different rates of occurrences in
different lactations (i.e., cystic ovary). Parity effects in the genetic model should account for these
differences.
The number of animals and number of records relative to number of herds was low, since the data
only included diseased animals. Subsequently, milk production on herdmates was retrieved, and
the assumption was made that these relatives did not have a disease. Only test day records that
qualified for genetic evaluation were included in the analysis. If the animal’s record were
removed (by DHI) due to the occurrence of a disease, this information would not have been
captured. Perhaps all tests, including those that do not qualify for genetic evaluation, should be
included.
It may not be appropriate to eliminate test day records that were less than 20 days apart. Milk
recording performs surprise retests in very short intervals from scheduled tests. All tests are
included in the production evaluations (TDM). If the retest shows that the first scheduled test was
incorrect, then it is removed.
Analyzing several traits with low incidence may be troublesome. A suggestion was made to
analyze traits in larger groups. Since all traits are highly correlated, a better suggestion may be to
analyze all 8 traits simultaneously. In the end, a selection index will be created using a
combination of only a few traits anyways.
National disease collection is scheduled to begin in April 2007.