Livestock Science 153 (2013) 20–26

Contents lists available at SciVerse ScienceDirect

Livestock Science
journal homepage: www.elsevier.com/locate/livsci

Comparison of Poisson, probit and linear models for genetic

analysis of number of inseminations to conception and
success at first insemination in Iranian Holstein cows
R. Abdollahi-Arpanahi a,n, F. Peñagaricano b, H. Aliloo a, H. Ghiasi c, J.I. Urioste d
a
Department of Animal Science, University College of Agriculture and Natural Resources, University of Tehran, 31587-77871 Karaj, Iran
b
Department of Animal Sciences, University of Wisconsin, Madison, WI, USA
c
Department of Animal Science, Faculty of Agriculture, Payme Noor University, Iran
d
Departamento de Producción Animal y Pasturas, Facultad de Agronomı́a, Universidad de la Republica, Montevideo, Uruguay

a r t i c l e in f o abstract

Article history: The goals of this study were to estimate genetic parameters and to assess alternative
Received 18 May 2012 models for genetic evaluation of number of inseminations to conception (INS) and
Received in revised form success at first insemination (SF) in Iranian Holstein cows. Two models were considered
21 December 2012
for each trait: linear and probit models for SF, and linear and Poisson models for INS. Data
Accepted 18 January 2013
consisted of 72,124 records of parities 1 to 6 from 27,113 cows having lactation between
1981 and 2007 and distributed over 15 large Holstein herds. Genetic parameters and
Keywords: goodness of fit statistics were estimated using the whole data set and predictive ability
Dairy cows of models was assessed via a 4-fold cross-validation based on mean squared error of
Genetic parameters
prediction (MSEP) and correlation between observed and fitted values. Estimates
Model comparison
of heritability ranged from 0.039 to 0.062 for SF and 0.040 to 0.165 for INS. The
Nonlinear models
Fertility traits performance of linear and probit models was very similar for SF. Predictions of random
effects from these models were highly correlated, and both models exhibited similar
predictive ability. For INS, the linear model performed better than the Poisson model
according to goodness of fit statistics, but these two models showed the same predictive
ability. Overall, nonlinear models did not outperform linear models for genetic evalua-
tions of SF and INS in Iranian Holstein cows.
& 2013 Elsevier B.V. All rights reserved.

1. Introduction success at first insemination in the selection index, dairy

Reproductive performance is economically important
in livestock. Economic losses due to poor fertility can be
attributed to the cost of prolonged calving interval,
increased insemination costs, reduced returns from calves
born and consequently increasing replacement costs
(Chang et al., 2006). By including some fertility traits
such as number of inseminations to conception and
n
Corresponding author. Tel.: þ 98 2632248082;
fax: þ98 2612246752.
E-mail address: r.abdollahi@ut.ac.ir (R. Abdollahi-Arpanahi).
producers can reduce the cost of production and as a
result increase the profit. Furthermore, reliable estima-
tions of genetic parameters are needed for defining a
breeding strategy for these traits.
According to Gonzalez-Recio and Alenda (2005),
proper female fertility performance could be defined as
showing timely manner heat and becoming pregnant
with a low number of inseminations. For instance, redu-
cing the number of inseminations to conception and
increasing the rate of success at first insemination will
lead to increased fertility, and pregnant cows when dairy
producers desire, with reduced costs of used semen,
hormonal treatments and veterinary fees. Therefore, these

1871-1413/$ - see front matter & 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.livsci.2013.01.009
 R. Abdollahi-Arpanahi et al. / Livestock Science 153 (2013) 20–26
fertility traits should be considered in genetic evaluations
and included in the breeding goals (Gonzalez-Recio and
Alenda, 2005).
Number of services to conception (INS) as an indicator
of reproductive efficiency has been defined as the number
of services required for a successful conception. The INS is
directly related to open days, pregnancy rate, non-return
rate and the reciprocal of services per conception. In
contrast to open days and days to first service which are
interval traits that do not account for elapsed time from
first to last insemination and veterinary costs, the variable
INS takes partly these important issues into consideration.
In this sense, Gonzalez-Recio and Alenda (2005) reported
that by including INS in an index with combination of
production traits, the profit would be increased by 229 h/
cow. Therefore, INS should be registered and considered
in genetic evaluations since it would maximize profit-
ability at balancing economically production and fertility.
Success at first insemination (SF) is a trait that is similar
to pregnancy rate (Bormann, 2006) and which is often
included in dairy female selection indices. SF is an econom-
ically important trait due to the cost of the semen, labors
involved in heat checking and breeding for multiple AI, and
the difference in the quality and value between AI calves
and natural calves. Moreover, heifers that became pregnant
at first breeding are going to calve earlier, and hence they
will have a better chance to breed back on the following
year. Heritability estimates for first-service conception
range from 0.03 to 0.22 (Cammack et al., 2009).
Heritability estimates for fertility traits are typically
low, ranging from 1% to 5% when using linear models for
statistical analysis (e.g. Wall et al., 2003). Nonlinear
models could be better alternatives for more accurate
assessment of genetic parameters and genetic merit of the
animals in these traits. Several probability models can
be used for describing the number of inseminations to
conception (INS) and success at first insemination (SF) in
dairy cattle. For categorical or counted variables such as
INS, several alternatives are available, e.g., Poisson, linear
or ordinal threshold models (Peñagaricano et al., 2011;
Vazquez et al., 2009a, 2009b). In general, analysis of
counts with models based on the Poisson distribution
tends to be more appropriate than models based on
normal distribution. For binary outcomes such as SF, the
Bernoulli distribution parameterization using either a
probit or a logit link is a natural approach. Another
alternative, despite the binary nature of the variable, is
the use of a linear model.
The objectives of the current study were, firstly, to
estimate genetic parameters; and, secondly, to assess
alternative models for genetic analysis of SF (probit and
linear) and INS (Poisson and linear). Models were com-
pared based on goodness of fit statistics and their pre-
dictive ability in a 4-fold cross-validation analysis.
2. Materials and methods
2.1. Data
A total of 72,124 records of parities 1 to 6 from 27,113
cows collected from 1981 to 2007 in 15 large Holstein
21
herds of Iran were used in the analysis. Only artificial
insemination mating records were used. Response vari-
ables studied were number of inseminations to concep-
tion (INS) and success to first insemination (SF). These 15
dairy herds are under registration of the Dairy Herd
Improvement Program of the Animal Breeding Centre of
Iran. Artificial insemination (AI) technicians record all
insemination events and are in charge of maintaining an
accurate dataset. All cows had records for both traits.
Reproductive traits in different parities were treated as
repeated measurements. For INS, all the values greater
than 9 were set to 9 in order to reduce possible recording
errors. SF was a binary trait defined as 1 if the cow
became pregnant at first insemination and 0 otherwise.
2.2. Statistical Models
Preliminary least squares analyses were conducted for
both traits using the GLM procedure of Statistical Analysis
System (SAS) (SAS, 2004) to decide which non-genetic
effects had to be included in the final models. The
following linear predictor was common to both traits:
g ¼ Xb þ Z1 a þ Z2 hys þWpe
where g is a function of the expected value of SF or INS; b
is a vector for fixed effects of parity (6 levels), age at
previous calving (10 levels) and months of first insemina-
tion (12 levels); a is the additive genetic effect; hys is
herd-year-season effect; pe is permanent environmental
effect for cows; e is the residual term; and X, Z1, Z2 and W
are incidence matrices relating data to the corresponding
effects.
Random effects were assumed to follow the multi-
variate normal distribution,
0 1 2 0 13
a As2a 0 0
B hys C 6 B C7
A  N6 B 0 Is2hys 0 C7
@ 40, @ A5
pe 0 0 Is2pe
where a, hys and pe are the vectors of additive genetic
animal, herd-year-season and permanent environmental
effects, respectively; s2a , s2hys and s2pe are the additive
genetic, herd-year-season and permanent environmental
effects variances, respectively; A is the additive relation-
ship matrix (with order 32,447  32,447) and I is the
identity matrix of order 838 for herd-year-season and
27,113 for permanent environmental effects. Animal,
herd-year-season and permanent environmental effects
were assumed to be independent of residual effects in the
models. General structure of the data and the pedigree is
presented in Table 1.
2.2.1. Probit model for SF
The Probit model (Gianola, 1982) describes the obser-
vable outcome (SF) using an underlying linear model
z¼ g þe, where e is independent and identically distrib-
uted standard normal random vector. The scoring rule in
this model is:
(
1 ðsuccessÞ
SF ¼
0 ðfailureÞ
22 R. Abdollahi-Arpanahi et al. / Livestock Science 153 (2013) 20–26

Table 1
General structure of the data and the pedigree by class of parity in Iranian Holstein cows. The traits under study are success at first insemination (SF) and
number of inseminations to conception (INS).

Parity 1 Parity 2 Parity 3 Parity 4 Parity 5 Parity 6

Data
No. of records 25,519 18,700 12,855 7587 4656 2453
SF (%) 45.96 41.95 41.65 41.65 39.21 37.73
Mean INS 2.05 2.18 2.16 2.17 2.23 2.35
SD INS 1.35 1.42 1.39 1.40 1.41 1.55

Pedigree
No. animals in pedigree 31,447 24,622 18,438 12,473 8116 4745
No. sires 1406 1196 1019 792 603 482
No. dams 15,575 12,211 9114 6081 3901 2149
No. of progeny per sire 18.67 16.75 14.25 12.01 9.91 6.91
No. of animals with both parents known 22,125 16,524 11,705 7372 4457 2315
No. of phenotyped animals with at least one parent unknown 2178 295 93 53 16 15

Therefore, the conditional probability of the success or 2.3. Inference

failure is PðSF ¼ 19gÞ ¼ FðgÞ, where FðgÞ is the standard
normal cumulative distribution function. Using this, the All models were implemented in a Bayesian frame-
likelihood function becomes work (Kadarmideen et al., 2001). Following Bayes rule, the
Y posterior density of model unknowns is pðh9yÞppðy9hÞpðhÞ
pðSF9gÞ ¼ FðgÞSF ½1FðgÞ1SF
where h ¼ fb0 ,. . .,a,hys,pe, s2a , s2hys , s2pe , s2e g is the collection
where SF and g are the vectors of the binary outcomes of model unknowns, pðhÞ is the joint prior distribution of
and the linear predictors, respectively. h, pðy9hÞ is the conditional distribution of the data (a
likelihood function when viewed as a function of h) and
2.2.2. Poisson model for INS pðh9yÞ is the posterior distribution of model unknowns.
A mixed effects Poisson model was also fitted for INS. The joint prior distribution of model unknown was as
Poisson models, which are appropriate for count data, follows:
impose an important constraint in that the mean and
pðhÞppða9s2a Þpðhys9s2hys Þpðpe9s2pe Þpðs2a Þpðs2hys Þpðs2pe Þpðs2e Þ
variance of the conditional distribution of observations,
given the random effects, are assumed to be equal
(Vazquez et al., 2009a). The introduction of a residual Nða90,As2a ÞNðhys90,Is2hys ÞNðpe90,Is2pe Þx2
allows modelling individual differences in propensity and ðs2a 9df a ,Sa Þx2 ðs2hys 9df hys ,Shys Þx2 ðs2pe 9df pe ,Spe Þ
accommodates overdispersion. Thus, the mixed Poisson
model for INS assumed that the number of insemination
x2 ðs2e 9df e ,Se Þ
to conception, given the fixed and random effects, fol-
lowed a Poisson distribution with parameter l, where log where x2 ð:9df:,S:Þ is a scaled inverse chi-square distribu-
(l) ¼ Z þe, and e is independent and identically distributed tion with degrees of freedom df, and scale parameter S.
as e  Nð0, s2e Þ. Therefore the likelihood function A value of 5 was given to the parameter df, in order to
becomes assign a low degree of belief to prior information. Values
  Y kt expfkg for the scale parameter S were obtained from previous
p INS ¼ t9k ¼
t! REML estimates in the study of Ghiasi et al. (2011).
where t (1,2,y,9) and k are vectors of the counting
outcomes and parameters for the Poisson model, 2.4. Implementation
respectively.
For all models, Gibbs samples were obtained using the
MCMCglmm (Hadfield, 2010) package which is available
2.2.3. Linear models
in the R language/environment (R Development Core Team,
Linear mixed models were fitted to SF and INS. Linear
2010). Inferences for each of the models were based on
models are obtained by adding a Gaussian residual
500,000 samples obtained after discarding 200,000 samples
to the linear predictor, that is, y¼ Z þe where y is either
as burn in. A thinning interval of 200 was used for
SF or INS, and e is independent and identically distri-
computing features of the posterior distribution. Conver-
buted Gaussian random variable with null mean and
gence diagnostics and statistical and graphical analysis of
variance s2e . The likelihood function for each of these
Markov chain Monte Carlo sampling output were carried
responses is
out with the Coda (Plummer et al., 2006) package of the R
language/environment (R Development Core Team, 2010).
Y
pfSF9gg ¼ NðSF9g, s2e Þ and, The method of Raftery and Lewis (1992) and visual inspec-
Y tion of trace plots were used to assess convergence of the
pfINS9gg ¼ NðINS9g, s2e Þ, respectively
MCMC outputs.
 R. Abdollahi-Arpanahi et al. / Livestock Science 153 (2013) 20–26 23

2.5. Genetic parameter estimates 3. Results and discussion

In linear and Poisson models, heritability (h2) and
repeatability (r) can be evaluated respectively in the
observable and log scale, using standard formulae:
2 s2a þ s2pe
2 s a
h ¼ 2 and r ¼ 2
sa þ s2pe þ s2hys þ s2e sa þ s2pe þ s2hys þ s2e
In the Probit model, h2 and r can be evaluated at the
liability scale using the above formulae with s2e ¼ 1.
2.6. Model comparison
Goodness of fit of each of the models was assessed by
computing the mean-squared error statistic (MSE),
X parous cows. In agreement with Badinga et al. (1985),
MSE ¼ n1 ^ 2,
ðyyÞ
where y^ is the conditional expectation function evaluated
at the posterior mean of model unknowns, y is the
corresponding vector of values of the response (either SF
or INS) and n is the number of observations. The condi-
tional expectation functions of the linear, probit and
Poisson models are y^ ¼ g ^ , y^ ¼ Fðg
^ Þ and y^ ¼ expðg
^þ
ð1=2Þs2e Þ, respectively. In addition, Pearson’s correlation
between y^ and y was also obtained. Model comparison
also involved assessing Spearman rank correlations
between predictions of random effects obtained with
different models. Random effects were predicted using
the estimated posterior mean of those effects.
To assess the predictive ability of each model, the
dataset was partitioned into four parts, with a restriction
that imposed that all levels of fixed effects were repre-
sented in each partition. Then, solutions for all fixed and
random effects of the training set (three folds) were
estimated and used to predict observations in the testing
set (retained partition). The mean-squared predictive
error was defined as
XX f
MSEP ¼ n1 ðyy^ Þ2 ,
f
f
where y^ is a conditional expectation function evaluated
at the posterior mean of model unknowns obtained when
the data in fold f was excluded from the analysis. Finally,
the Pearson correlation between observed and fitted
values in the testing set was also estimated.
3.1. Descriptive statistics
Thirty percent of the animals had one record, 23.47%
had two records, 18.66% had three records and the rest
had more than three records (up to 6 per animal). The
average number of service to conception was 2.14 and the
percentage of success at first insemination was 42.96.
Table 1 shows the percentage of success at first insemina-
tion (SF) and mean and standard deviation (SD) of number
of inseminations to conception (INS) for each parity.
The percentage of SF decreased as the parity number
increased. Other studies (Chebel et al., 2004) argued that
multiparous cows had lower conception rate than primi-
who reported that heifers required 1.5 services per con-
ception compared with 2.3 for lactating cows, our results
also showed that INS increased over parity number.
3.2. Estimates of variance and genetic parameters
Estimates of the mean and quantiles (0.025 and 0.975) of
the posterior distribution of variance components, heritabil-
ities, and repeatabilities calculated by different models for
SF and INS are presented in Table 2. For all models, the herd-
year-season variance showed the highest value, highlighting
the importance of management and hygiene practices on
individual farms (e.g. Pérez-Cabal et al., 2009; Vazquez et al.,
2009b). Importantly, variance components are not directly
comparable between models because they belong to differ-
ent distributions. A more meaningful comparison can be
based on ratios between variances components, i.e., h2 and r.
Heritability estimates were low for SF (0.062 and 0.039 for
probit and linear models, respectively) and low to moderate
for INS (0.040 and 0.165 for linear and Poisson models,
respectively). Additionally, following Dempster and Lerner
(1950), the heritability estimate for SF obtained with the
probit model was scaled back to the observed scale: the
estimate was equal to 0.045 and it was smaller than that on
the liability scale due to the loss of information by grouping
into two categories (Dempster and Lerner, 1950).
Working with the same population, Ghiasi et al. (2011)
reported lower heritability estimates for both SF (0.029) and
INS (0.046) traits. It is important to remark that these

Table 2
Estimates of mean and quantiles (0.025, 0.975) of the posterior distribution of variance components and genetic parameters by response (success or fail
at first insemination, SF, and number of inseminations to conception, INS) and model (linear, probit and Poisson).

SF INS

Linear Probit Linear Poisson

s2a 0.010 (0.009, 0.011) 0.084 (0.067, 0.105) 0.075 (0.064, 0.092) 0.011 (0.010, 0.013)
s2pe 0.010 (0.009, 0.011) 0.081(0. 061, 0. 103) 0.078 (0.063, 0.094) 0.010 (0.009, 0.011)
s2hys 0.017 (0.015, 0.020) 0. 183 (0.155, 0.208) 0.136 (0.117, 0.152) 0.033 (0.029, 0.038)
s2e 0.217 (0.215, 0.219) 1 1.656 (1.631, 1.673) 0.017 (0.016, 0.019)
h
2 0.039 (0.036, 0.044) 0. 062 (0.050, 0.077) 0.040 (0.033, 0.047) 0.165 (0.140, 0.179)
r 0.081 (0.075, 0.086) 0. 121 (0.109, 0.140) 0.078 (0.072, 0.088) 0.298 (0.270, 0.321)

s2
a, Additive genetic variance; s 2
pe , permanent environmental variance; s 2
hys , heard-year-season variance; s 2
e, residual variance; h2, heritability;
r, repeatability.
24 R. Abdollahi-Arpanahi et al. / Livestock Science 153 (2013) 20–26
authors used bivariate Bayesian linear-threshold models to
estimate genetic parameters. Moreover, using records from
dairy cattle in United Kingdom, Kadarmideen et al. (2003)
using a bivariate linear animal model reported heritability
estimates equal to 0.016 and 0.019 for SF and INS, respec-
tively. In addition, Gonzalez-Recio et al. (2005) reported a
heritability of 0.035 for INS in Spanish dairy cattle using an
ordinal-censored threshold model. Also, for the same trait,
Chang et al. (2006) reported a heritability of 0.04 in
Norwegian Red cattle using a censored threshold-linear
model. Therefore, in general, it is expected that SF and INS
show low heritability. Interestingly, in our study, a moder-
ate heritability (i.e. 0.165) was estimated using a Poisson
model with random residual. This result may suggest that
this model capture more genetic variation (i.e. higher
heritability) in INS. This issue requires further research.
Repeatability estimates were low for SF (0.081 and
0.121 for linear and Probit models, respectively), and low
to moderate for INS (0.078 and 0.298 for linear and Poisson
models, respectively). These estimates were higher than
those reported by Kadarmideen et al. (2003). These authors
reported repeatability estimates equal to 0.039 and 0.055
for SF and INS, respectively. The magnitude of the esti-
mates of repeatability reported in our study suggests that
many records are necessary for making accurate selection
decisions.
3.3. Prediction of random effects
From a breeder’s perspective, a central question is
whether predictions of genetic values from these models
yield different breeding decisions or not (i.e., the extent of
re-ranking). The Spearman rank correlations between
random effect predictions (animal and permanent envir-
onmental effects) across models are displayed in Table 3.
The correlations between predictions from binary
response models were high (0.972 and 0.974 for breeding
values and permanent environmental effects, respec-
tively). Also the predictions for count data from linear
and Poisson models were highly correlated (0.971and
0.977 for breeding values and permanent environmental
effects, respectively). Importantly, high and negative cor-
relations were obtained between estimated breeding
values for SF and INF (between 0.806 and  0.828).
These empirical correlations could be a good indication of
genetic correlations. In this sense, the genetic correlation
between INS and SF was estimated as 0.92 using a
bivariate linear model (Kadarmideen et al., 2003).
Table 3
Spearman rank correlations between predictions of random effects
(additive effects above diagonal, permanent environmental effects
below diagonal) obtained from different models (linear, Poisson, Probit)
fitted to different response variables (success or fail at first insemination,
SF, and number of inseminations to conception, INS).
Model Linear (SF) Probit Linear (INS) Poisson
Linear (SF) 0.972  0.808  0.806
Probit 0.974  0.806  0.828
Linear (INS)  0.707  0.705 0.971
Poisson  0.711  0.710 0.977
3.4. Goodness of fit
Goodness of fit statistics for the different models and
traits are presented in Table 4. For SF, both models
showed similar mean-squared error (0.207 versus 0.201
for probit and linear models, respectively). In addition, the
correlations between observed and fitted values were
similar between these models (0.409 and 0.481 for Probit
and linear models, respectively). For INS, the linear model
fitted the data slightly better than the Poisson model both
in terms of MSE and correlations (1.532 versus 1.626 and
0.503 versus 0.434, respectively). It is important to remark
that the correlations between observed and fitted values
were low (i.e. nearly 0.50), suggesting that our models
could explain only part of the variability of the data.
Similar results were reported for other studies working
also with fertility traits in livestock (Matos et al., 1997;
Olesen et al., 1994; Perez-Enciso et al., 1993).
3.5. Predictive ability in cross-validation
In our study, typical model comparison criteria, such
as the likelihood ratio test, Akaike information criterion,
or Bayesian information criterion, cannot be applied to
compare the alternative models due to that the Probit,
Poisson and Linear models differ in their likelihood func-
tion and response variable (e.g., SF, INS). Instead, the
models were compared by their predictive ability in a
cross-validation study. Results of predictive ability for
different traits and models are shown in Table 5.
In addition to overall mean-squared errors of predic-
tion (MSEP) and correlation between observed and fitted
values (COR), we report MSEP and COR by number of
records per animal available in the training set (N). For the
number of inseminations, linear and Poisson model
showed very similar predictive ability. In addition, for
these models, MSEP decreased as N increased; this is
expected because the amount of information available for
predicting random effects increases as N does. In other
words, the largest MSEP was observed when there were
no records of the animal in the training fold, so prediction
was entirely based on the pedigree structure.
Overall, the linear model fits the data slightly better
than the Poisson model. However, both models showed
similar predictive abilities. Tempelman and Gianola (1999)
compared linear and negative binomial models in the
analysis of the number of artificial inseminations to con-
ception in dairy cattle. Interestingly, the predictive ability
of the negative binomial model was only slightly better
Table 4
Mean-squared error (MSE) and correlation (COR) between observed and
fitted values by response (success or fail at first insemination, SF, and
number of inseminations to conception, INS) and model (linear, Poisson
and probit).
SF INS
Linear Probit Linear Poisson
MSE 0.201 0.207 1.532 1.626
COR 0.481 0.409 0.503 0.434
 R. Abdollahi-Arpanahi et al. / Livestock Science 153 (2013) 20–26 25

Table 5 model in term of goodness of fit. However, both models

Mean-squared error of prediction and correlation (COR) between exhibited similar predictive ability. For SF, the probit and
observed and fitted values for a four-fold cross validation study by response
linear models showed similar performance in terms of
(success or fail at first insemination, SF, and number of inseminations to
conception, INS) and model (linear, Poisson and probit). goodness of fit and predictive ability. Overall, nonlinear
mixed models, which have a theoretical advantage over
SF INS linear mixed models for discrete traits, did not outper-
form linear mixed models in our study.
Linear Probit Linear Poisson

Total Conflict of interest statement

MSEP 0.234 0.238 1.796 1.795
COR 0.208 0.204 0.279 0.275
The authors do not have any conflict of interest.
N¼0
MSEP 0.236 0.237 2.090 2.072
COR 0.192 0.191 0.260 0.258
Acknowledgment
N¼1
MSEP 0.232 0.233 2.016 2.028
COR 0.197 0.195 0.258 0.264
The authors wish to acknowledge of Animal Breeding
Centre of Iran for providing data.
N¼2
MSEP 0.231 0.234 1.891 1.900
COR 0.213 0.210 0.272 0.272 References
N 42
MSEP 0.236 0.241 1.611 1.606 Badinga, L., Collier, R., Thatcher, W., Wilcox, C., 1985. Effects of climatic
COR 0.214 0.205 0.293 0.286 and management factors on conception rate of dairy cattle in
subtropical environment. J. Dairy Sci. 68, 78–85.
Bormann, J.M., 2006. Pregnancy rate and first-service conception rate in
N represents the number of records of the animal that appears in the
Angus heifers. J. Anim. Sci. 84, 2022–2025.
training set.
Cammack, K.M., Thomas, M.G., Enns, R.M., 2009. Reproductive traits and
their heritabilities in beef cattle. Prof. Anim. Sci. 25, 517–528.
Chang, Y.M., Andersen-Ranberg, I.M., Heringstad, B., Gianola, D.,
Klemetsdal, G., 2006. Bivariate analysis of number of services to
than that of the linear model. Other studies also compared conception and days open in Norwegian red using a censored
the performance of linear and Poisson models, in terms of threshold-linear model. J. Dairy Sci. 89, 772–778.
goodness of fit and predictive ability. In this sense, Chebel, R.C., Santos, J.E.P., Reynolds, J.P., Cerri, R.L.A., Juchem, S.O.,
Overton, M., 2004. Factors affecting conception rate after artificial
Peñagaricano et al. (2011) reported a better performance insemination and pregnancy loss in lactating dairy cows. Anim.
of a Poisson model with a residual term compared with the Reprod. Sci. 84, 239–255.
linear model. On the other hand, some authors reported no Dempster, E.R., Lerner, I.M., 1950. Heritability of threshold characters.
Genetics 35, 212–336.
difference between these models (Perez-Enciso et al., Ghiasi, H., Pakdel, A., Nejati-Javaremi, A., Mehrabani-Yeganeh, H.,
1993) or a slightly better performance of the linear model Honarvar, M., González-Recio, O., Carabaño, M.J., Alenda, R., 2011.
(Matos et al., 1997; Olesen et al., 1994; Pérez-Cabal et al., Genetic variance components for female fertility in Iranian Holstein
cows. Livest. Sci. 139, 277–280.
2009; Vazquez et al., 2009a).
Gianola, D., 1982. Theory and analysis of threshold characters. J. Anim.
For SF, probit and linear models showed the same Sci. 54, 1079–1096.
predictive ability (Table 5). Furthermore, no differences in Gonzalez-Recio, O., Alenda, R., 2005. Genetic parameters for female
fertility traits and a fertility index in Spanish dairy cattle. J. Dairy Sci
MSEP were observed when different records were used in the
88, 3282–3289.
training set. In this sense, Meijering and Gianola (1985) found Gonzalez-Recio, O., Chang, Y.M., Gianola, D., Weigel, K.A., 2005. Number
that nonlinear models for categorical data performed better of inseminations to conception in Holstein cows using censored
than the linear model only when heritability was moderate records and time-dependent covariates. J. Dairy Sci. 88, 3655–3662.
Hadfield, J.D., 2010. MCMC methods for multi-response generalized
to high and the incidence of the binary trait in the population linear mixed models: the MCMCglmm R package. J. Stat. Software
was r10%. In our study, heritability estimates were lower 33, 1–22.
than 0.062 and the percentage of SF was equal to 42.96%. Kadarmideen, H., Rekaya, R., Gianola, D., 2001. Genetic parameters for
clinical mastitis in Holstein–Friesians in the United Kingdom: a
Therefore, our results are consistent with those reported by Bayesian analysis. Anim. Sci. 73, 229–240.
these authors. Failure of nonlinear models for categorical data Kadarmideen, H.N., Thompson, R., Coffey, M.P., Kossaibati, M.A., 2003.
to outperform linear models for analysing discrete livestock Genetic parameters and evaluations from single- and multiple-trait
analysis of dairy cow fertility and milk production. Livest. Prod. Sci.
data has been observed in other empirical studies (Matos 81, 183–195.
et al., 1997; Olesen et al., 1994; Peñagaricano et al., 2011; Matos, C., Thomas, D., Gianola, D., Perez-Enciso, M., Young, L., 1997.
Vazquez et al., 2009a). Genetic analysis of discrete reproductive traits in sheep using linear
and nonlinear models: II. Goodness of fit and predictive ability.
J. Anim. Sci. 75, 88–94.
4. Conclusion Meijering, A., Gianola, D., 1985. Linear versus nonlinear methods of sire
evaluation for categorical traits: a simulation study. Genet. Sel. Evol.
17, 115–132.
Four alternative linear and nonlinear models were
Olesen, I., Perez-Enciso, M., Gianola, D., Thomas, D.L., 1994. A compar-
investigated for genetic evaluation of SF and INS in Iranian ison of normal and nonnormal mixed models for number of lambs
Holstein cows. Heritability estimates were low for SF and born in Norwegian sheep. J. Anim. Sci. 72, 1166–1173.
low to moderate for INS. Predictions of random effects Peñagaricano, F., Urioste, J.I., Naya, H., de los Campos, G., Gianola, D.,
2011. Assessment of Poisson, probit and linear models for genetic
obtained from different models were highly correlated. analysis of presence and number of black spots in Corriedale sheep.
For INS, the linear model performed better than Poisson J. Anim. Breed. Genet. 128, 105–113.
26 R. Abdollahi-Arpanahi et al. / Livestock Science 153 (2013) 20–26
Pérez-Cabal, M.A., de los Campos, G., Vazquez, A.I., Gianola, D., Rosa,
G.J.M., Weigel, K.A., Alenda, R., 2009. Genetic evaluation of suscept-
ibility to clinical mastitis in Spanish Holstein cows. J. Dairy Sci. 92,
3472–3480.
Perez-Enciso, M., Tempelman, R., Gianola, D., 1993. A comparison
between linear and Poisson mixed models for litter size in Iberian
pigs. Livest. Prod. Sci. 35, 303–316.
Plummer, M., Best, N., Cowles, K., Vines, K., 2006. CODA: convergence
diagnosis and output analysis for MCMC. R News 6, 7–11.
R Development Core Team, 2010. R: A Language and Environment for
Statistical Computing. R Foundation for Statistical Computing,
Vienna, Austria.
Raftery, A.E., Lewis, S.M., 1992. Comment: one long run with diagnostics:
implementation strategies for Markov Chain Monte Carlo. Stat. Sci. 7,
493–497.
Statistical Analysis System (SAS), 2004. SAS Users’ Guide, Version 9.1.
SAS Institute Inc., Cary, North Carolina, USA.
Tempelman, R.J., Gianola, D., 1999. Genetic analysis of fertility in dairy cattle
using negative binomial mixed models. J. Dairy Sci. 82, 1834–1847.
Vazquez, A.I., Gianola, D., Bates, D., Weigel, K.A., Heringstad, B., 2009a.
Assessment of Poisson, logit, and linear models for genetic analysis
of clinical mastitis in Norwegian Red cows. J. Dairy Sci. 92, 739–748.
Vazquez, A.I., Weigel, K.A., Gianola, D., Bates, D.M., Perez-Cabal, M.A.,
Rosa, G.J.M., Chang, Y.M., 2009b. Poisson versus threshold models for
genetic analysis of clinical mastitis in US Holsteins. J. Dairy Sci. 92,
5239–5247.
Wall, E., Brotherstone, S., Woolliams, J.A., Banos, G., Coffey, M.P., 2003.
Genetic evaluation of fertility using direct and correlated traits.
J. Dairy Sci. 86, 4093–4102.