Вы находитесь на странице: 1из 6

BREEDING AND GENETICS

Estimation of Growth Parameters Using a Nonlinear Mixed Gompertz Model

Z. Wang and M. J. Zuidhof1

Livestock Development Division, Alberta Agriculture, Food and Rural Development,


7000-113 St. Edmonton, Alberta, Canada, T6H 5T6

ABSTRACT In order to maximize the utility of simula- ance structure assumed with the random effect accounted
tion models for decision making, accurate estimation of for part of the BW correlation across ages in the same
growth parameters and associated variances is crucial. A individual. The amount of residual variance decreased
mixed Gompertz growth model was used to account for by over 55% with the mixed model. The mixed model
between-bird variation and heterogeneous variance. The reduced estimation biases that resulted from selective
mixed model had several advantages over the fixed ef- sampling. For analysis of longitudinal growth data, the
fects model. The mixed model partitioned BW variation mixed effects growth model is recommended.
into between- and within-bird variation, and the covari-
(Key words: broiler, nonlinear mixed model, Gompertz, growth parameter, inference)
2004 Poultry Science 83:847–852

INTRODUCTION mature weight (Wm) is estimated for each genotype, and


a random error (eit) is associated with each individual
Comparisons of the characteristics of different broiler bird i at age t in the model, which is assumed to be an
strains can be used to identify and improve supply chain independent normally distributed random variable with
efficiencies. In bioeconomic models, genotype-specific mean of zero and a constant variance σe2. This model is
growth curves are used to dynamically estimate daily considered a fixed effect model because there is no ran-
nutrient requirements at different ages (Hancock et al., dom effect associated with Wit. The results obtained with
1995; Gous, 1998; Gous et al., 1999, 2002). In order to fixed model analysis can sometimes be misleading, espe-
maximize the utility of simulation models for decision cially in the presence of competitive interaction and selec-
making, accurate estimation of growth parameters and tive sampling (Le Dividich, 1999). Furthermore, repeated
related variances and covariances across age is crucial measurements of BW from the same bird are likely to be
(Hancock et al., 1995). Many broiler growth data analyses more closely correlated than measures made on different
(Emmans, 1995; Hancock et al., 1995; Gous et al., 1999) birds, and measures made close in time on the same
have been conducted using the well-known Gompertz individual are likely to be more highly correlated than
(Gompertz, 1825; Emmans, 1981; Emmans and Fisher, measures made further apart in time (Littell et al., 2000).
1986) growth function, which describes a single sigmoidal This type of underlying relationship, often inherent in
growth phase. Typically, this growth curve is used to fit repeatedly measured growth data, contradicts the as-
longitudinal BW (Wt) data as a function of age (d) for a sumption that eits are independent of each other with a
particular genetic strain from age t = 0 to maturity. In constant variance over ages. For this type of longitudinal
the current analysis of broiler chicken growth, maturity growth data, there are at least 2 sources of variation
was defined as the mature lean carcass, including bones, (within and between individuals) in Wit. The fixed effect
muscles, and essential metabolic organs (Kwakkel et al., model described above does not account for variability
1993). More complex multiphasic models must be used between individuals (Craig and Schinckel, 2001), and it
to account for other growth phases such as the develop- also ignores the underlying relationships inherent in the
ment of the reproductive tract, or the deposition of fat in longitudinal growth data. The objectives of the current
mature animals. study were 1) to account for at least a portion of increasing
The Gompertz growth function is usually estimated BW variance as age increases, and greater serial correla-
once per genotype (sex × strain), such that a common tions at later ages; 2) to determine whether the mixed
model can reduce the effect of selective sampling and
competitive advantage on estimates of growth parame-
2004 Poultry Science Association, Inc.
Received for publication September 19, 2003.
Accepted for publication December 29, 2003.
1
To whom correspondence should be addressed: martin.zuidhof@
gov.ab.ca. Abbreviation Key: BIC = Bayesian information criterion.

847
848 WANG AND ZUIDHOF
TABLE 1. Bayesian information criterion1 fit statistics for fixed and mixed Gompertz growth models
for females and males of 6 commercial boiler strains

Female Male

Strain Fixed model Mixed model Fixed model Mixed model


1 −427.1 −556.0 −235.9 −396.7
2 −364.5 −573.2 −101.2 −334.3
3 −278.9 −474.2 −41.5 −218.3
4 −215.0 −389.9 −0.4 −181.7
5 −316.3 −521.8 −135.7 −328.1
6 −216.3 −408.2 −108.8 −352.7
Average −303.0 −487.2 −103.9 −302.0
1
A lower value indicates a better fit of the model to the data.

ters; and 3) to compare the results with a corresponding ing, the facility was maintained at approximately 5°C
fixed effects Gompertz growth model. below the desired temperature at floor level in the central
hallway, and temperature gradients were established us-
MATERIALS AND METHODS ing thermostatically controlled gas brooders at the back
of each pen. Bird behavior and temperature data at bird
Experimental Design height in the center of each pen were monitored closely
to verify that birds in every pen had a broad thermoneu-
The data for this analysis was taken from an experiment tral zone. To monitor environmental temperatures, tem-
with a 6 × 2 factorial arrangement of treatments, with 6 perature data loggers were placed on a water line in
commercial broiler strain crosses (1 to 6) and 2 sexes (male each pen and recorded the pen temperature at bird level
and female). Chicks were obtained from commercial automatically every 15 min for the duration of the trial.
broiler breeder flocks. To minimize variation in initial A photoschedule of 23L:1D was used throughout the trial
chick weight, parent flocks of similar age (46 ± 3 wk) so that birds would have almost continual access to feed
were used. At hatch, 960 chicks (80 chicks per strain-by- and water. No vaccinations were provided, as the mount-
sex combination) were individually identified and ing of an immune response would divert nutrients other-
weighed. A total of 120 chicks (10 per treatment) were wise channeled toward growth. Approximate stocking
placed into each of 8 replicate pens and grown together density decreased from 10 birds per m2 at hatch to 4 birds
until 16 wk of age. Individual BW data were recorded per m2 by 6 wk.
for each bird twice a week from 0 to 6 wk, weekly from Starter (0 to 21 d), grower (21 to 35 d), roaster (35 to
6 to 10 wk, and biweekly from 10 to 16 wk. Data were 49 d), and roaster II (49 to 112 d) diets were formulated
recorded more often in the early part of the study because to provide 100 and 105% of NRC (1994) recommended
the relative growth rate (as percentage of BW) was much energy and protein levels (formulated on the first 3 amino
higher early, thus improving the rate of maturing esti- acids that were most likely to be limiting: lysine, total
mate. Because the trial from which these data were col- sulfur amino acids, and threonine), respectively. Ade-
lected involved serial dissections of individual birds quate feeder space (5 cm per bird) and water nipple avail-
weekly to 6 wk, we used a subsample of 14 to 18 birds ability (15 birds per water nipple) was provided to ensure
per strain-by-sex combination having a complete set of that these factors would not limit rates of gain. Ad libitum
BW observations to 12 wk in the current analysis. access to feed and water was provided for the duration
Because of a sexual maturity related growth spurt, data of the trial.
after 12 wk were deleted from the data set. The sexual
maturity related growth spurt is a second phase of growth Broiler Growth Model
(Kwakkel et al., 1993, 1995). Because the Gompertz func-
tion describes a single (prepubertal) growth phase, it was The broiler growth data was fitted to the fixed effect
legitimate and necessary to remove the bias of the sexual Gompertz growth function proposed by Emmans (1981).
maturity related growth spurt from the growth parameter
estimates of the first growth phase. This approach was Wit = Wmexp(−exp(b(t−t*))) + eit, [1]
also taken by Hancock et al. (1995).
and the mean and variance can be expressed as
Bird Management and Nutrition
E(Wit) = Wmexp(−exp(b(t−t*))) [2]
The care of the birds met the guidelines of the Canadian V(Wit) = V(eit) = σe2
Council of Animal Care (1993). Because the Gompertz
growth model describes nonlimited growth, care was where Wit is the expected BW of individual i at age t
taken to ensure that the environment did not limit growth. days; Wm is the average mature BW of all birds in the
To provide a thermoneutral environment during brood- same group; b is a rate of maturing; t* is the time in days
NONLINEAR MIXED GROWTH MODEL 849
TABLE 2. Growth parameters and standard error estimates of females and males of 6 commercial boiler strains determined
by fixed1 and mixed2 effect Gompertz growth models

Female Male

Fixed model Mixed model Fixed model Mixed model


3
Strain Parameter Estimate SE df Estimate SE df Estimate SE df Estimate SE df

1 Wm 4.776 0.043 283 4.753 0.055 16 6.141 0.070 282 6.057 0.083 16
b 0.0416 0.0003 283 0.0415 0.0002 16 0.0396 0.0003 282 0.0398 0.0003 16
σe2 0.0122 0.0010 283 0.0067 0.0006 16 0.0239 0.0020 282 0.0116 0.0010 16
σu2 0.0364 0.0141 16 0.0983 0.0359 16
2 Wm 5.003 0.049 300 4.962 0.063 17 6.514 0.090 302 6.498 0.108 17
b 0.0411 0.0003 300 0.0410 0.0002 17 0.0383 0.0004 302 0.0382 0.0003 17
σe2 0.0164 0.0013 300 0.0069 0.0006 17 0.0396 0.0032 302 0.0153 0.0013 17
σu2 0.0624 0.0223 17 0.1973 0.0687 17
3 Wm 4.955 0.057 300 4.885 0.079 17 6.152 0.093 283 6.159 0.114 16
b 0.0416 0.0004 300 0.0414 0.0003 17 0.0405 0.0005 283 0.0403 0.0003 16
σe2 0.0218 0.0018 300 0.0096 0.0008 17 0.0476 0.0040 283 0.0218 0.0019 16
σu2 0.0859 0.0329 17 0.1841 0.0671 16
4 Wm 4.792 0.061 300 4.664 0.087 17 6.611 0.121 233 6.653 0.144 13
b 0.0422 0.0004 300 0.0421 0.0003 17 0.0386 0.0005 233 0.0383 0.0004 13
σe2 0.0270 0.0022 300 0.0128 0.0011 17 0.0545 0.0050 233 0.0212 0.0020 13
σu2 0.1016 0.0399 17 0.2655 0.1048 13
5 Wm 5.069 0.058 249 5.094 0.070 14 7.076 0.122 249 7.123 0.122 14
b 0.0397 0.0003 249 0.0395 0.0002 14 0.0357 0.0004 249 0.0356 0.0003 14
σe2 0.0154 0.0014 249 0.0057 0.0005 14 0.0318 0.0028 249 0.0124 0.0011 14
σu2 0.0694 0.0265 14 0.2007 0.0764 14
6 Wm 5.504 0.069 234 5.439 0.089 13 7.187 0.102 251 7.205 0.118 14
b 0.0400 0.0004 234 0.0400 0.0003 13 0.0372 0.0004 251 0.0372 0.0002 14
σe2 0.0217 0.0020 234 0.0080 0.0008 13 0.0355 0.0032 251 0.0111 0.0010 14
σu2 0.1001 0.0408 13 0.2128 0.0799 14
1
Wit = Wmexp(−exp(b(t−t*))) + eit.
2
Wit = (Wm + ui)exp(−exp(b(t−t*))) + eit.
Wm = mature body weight (kg); b = rate of maturing (d−1); σe2 = the residual BW variance (kg); σu2 = the individual variance in Wm within a
3

population (kg).

at which growth rate is maximum, which can be obtained E(Wit) = ƒ(t; θ)


as t* = ln(−ln(W0/Wm))/b if the initial chick weight (W0) V(Wit) = σu2g2 (t; θ) + σe2 [5]
is known; and eit is the residual BW of individual i at age Cov(Wit,Wit + j) = σu2g(t; θ)g(t + j; θ)
t. The eits are assumed to be normally distributed with
mean 0 and a constant variance σe2. The same data set was Because the mean of the random effect ui is 0 for a popula-
also fitted to the mixed effect Gompertz growth function tion, the birds in each genotype follow the function ƒ;
this is the same as in the fixed effect model [1]. However,
Wit = (Wm + ui)exp−exp(b(t−t*))) + eit [3] the advantages of the mixed model over the fixed model
are the ability to separate between- and within-bird varia-
tion, the ability to account for a portion of the increasing
where ui is a random deviation of mature BW of the
variance in BW as age increases, and greater serial correla-
individual i from the average mature BW of its genotype
tions at later ages resulting from greater BW variation.
Wm. The uis are assumed to be normally distributed with
Because the function g increases with t, the covariance
mean 0 and variance σu2 and independent of eit. The rest
structure increases with t such that Cov(Wit, Wit + 1) <
of the terms are the same as in equation [1]. Cov(Wit, Wit + 2). This pattern of the variance and covari-
To fit the ui in the model is to allow the mature BW ance structures usually occurs in repeatedly measured
Wim of each bird to vary among birds (Wim = Wm + ui) in growth data (Craig and Schinckel, 2001).
each sex and strain group. This mixed effect growth
model can also be written as
Culling and Selection
Wit = ƒ(t; θ) + uig(t; θ) + eit [4] The mixed effects model handles selective sampling
problems in animal growth experiments better than a
where θ = {Wm, b}, such that the random effect ui is fixed effects model (Craig and Schinckel, 2001). In broiler
multiplied by g(t; θ) = exp(−exp(b(t−t*))). The mean, variance growth trials, birds that fail to thrive due to sickness or
and covariance of the mixed effect growth model [4] can other reasons may be culled before the experiment is
be expressed as completed. The earlier removal (involuntary culling) of
850 WANG AND ZUIDHOF
TABLE 3. Comparison of mature weight (Wm) parameter estimate using complete data set, and after
selective culling. Estimates were derived using fixed and mixed effect Gompertz growth models

Female Male

Fixed model Mixed model Fixed model Mixed model


1 2 3 1 2 3 1 2 3 1
Strain Cull All Bias Cull All Bias Cull All Bias Cull All2 Bias3

1 4.951 4.776 3.7 4.809 4.753 1.2 6.372 6.141 3.8 6.177 6.057 2.0
2 5.243 5.003 4.8 5.046 4.962 1.7 7.013 6.514 7.7 6.658 6.498 2.5
3 5.182 4.955 4.6 4.957 4.885 1.5 6.835 6.152 11.1 6.479 6.159 5.2
4 4.983 4.792 4.0 4.719 4.664 1.2 7.222 6.611 9.2 6.923 6.653 4.1
5 5.398 5.069 6.5 5.155 5.094 1.2 7.635 7.076 7.9 7.330 7.123 2.9
6 5.827 5.504 5.9 5.530 5.439 1.7 7.588 7.187 5.6 7.432 7.205 3.1
1
Estimates obtained with selectively culled data.
2
Estimates obtained with the complete data set.
3
Bias of estimate expressed as a percentage of inflation of the parameter estimates with selectively culled
data: bias = (Cull1–All2)/All2) × 100.

these slow growers from the trials may result in bias of RESULTS AND DISCUSSION
the growth parameter estimates. In order to show the
ability of the mixed model to handle this selective culling Fit Statistics
problem, the bottom 8% of the slower growing birds at
age 70 d were removed from each strain and sex and The BIC fit statistics for fixed and mixed models are
the remainder of the data were fitted to the fixed and summarized in Table 1 for each strain and sex. The BIC
mixed models. values of the mixed models were all substantially lower
than those from the corresponding fixed model. The aver-
age BIC value over the 6 strains for the mixed model was
Statistical Analysis reduced 184.2 (−487.2 vs. −303.0) for female birds and
198.1 (−302.0 vs −103.9) for male birds. Overall, these fit
A total of 3,266 longitudinal BW observations from 96 statistics indicated that mixed models fit the data better
males and 100 females of 6 commercial broiler strains than those of the fixed models.
were used in this analysis. The same data set was fitted
to the fixed [1] and mixed [3] effects growth models using Parameter Estimates
the NLMIXED procedure in SAS software (SAS Institute,
1999) to compare the results obtained from both models. Estimates of the growth parameters obtained with the
The Bayesian information criteria (BIC) fit statistic pro- fixed and mixed effects models are summarized in Table
vided by NLMIXED procedure for both models was used 2. For specific genotypes, the estimates of Wm and b from
to evaluate the appropriateness of the 2 models. The BIC fixed and mixed growth models were similar because the
values do not have an upper or lower limit, and do not expected means for both models are the same. The SE of
imply significant differences between models. However, the Wm estimates obtained from the mixed model were
a lower BIC value indicates a better fit of the model to higher than corresponding SE estimates from the fixed
the data. Parameters estimated by the model with lower model but lower for the b and σe2 estimates. With the
BIC scores increase the probability of obtaining the ob- mixed model, the average increase in Wm SE over the 6
served data. strains was 31% for females and 15% for males. This

TABLE 4. The estimated residual variance,1 covariance,2 and correlation3 across age from d 7 to 49
with fixed growth model for strain 2 of female broilers

Age (d) 7 14 21 28 35 42 49
7 0.00028 0.00050 0.00058 0.00067 0.00054 0.00062 0.00024
14 0.80799 0.00138 0.00139 0.00179 0.00155 0.00156 0.00044
21 0.72330 0.77593 0.00234 0.00300 0.00263 0.00194 0.00098
28 0.56945 0.68105 0.87733 0.00498 0.00450 0.00417 0.00362
35 0.41576 0.53657 0.70003 0.82181 0.00602 0.00506 0.00422
42 0.42539 0.48381 0.46324 0.68238 0.75367 0.00750 0.00766
49 0.12048 0.09785 0.16956 0.42781 0.45402 0.73694 0.01438
1
The variances are on the upper left to lower right diagonal of the table.
2
The covariances are in the upper right triangle of the table.
3
The correlations are in the lower left triangle of the table.
NONLINEAR MIXED GROWTH MODEL 851
increase was likely due to the very large decrease in the be more highly correlated than measures made further
number of degrees of freedom with the mixed model apart in time. Although the present experiment was de-
(Table 2). With the mixed model, the average reduction signed to allow all birds to express their genetic potential
in SE for b was 29% for females and 28% for males. Most for growth, these types of relationships may still exist. A
importantly, due to variance partitioning, residual vari- summary of residual variances, covariances, and correla-
ances (σe2) obtained from the mixed model were reduced tions across age for the females of strain 2 is presented
55% in females and 59% in males in the mixed model in Table 4 to illustrate the violation of fixed model as-
compared with the fixed model. This result represents sumptions about the distribution of residuals. It is clear
an improvement in estimation accuracy using the mixed that the variances along the diagonal increased with age,
model. The variance estimates for the random effect u and the covariances among the adjacent sequential resid-
were all significant (P < 0.05), indicating that the mixed uals also increased as age increased. The pattern was
model accounted for between-bird variation in mature similar to the variance and covariance structures assumed
BW. Fitting this random effect in the model allows Wit for the random effects in the mixed model [5]. The addi-
to vary from the genotype mean, which is appropriate, tion of the random effect in the mixed model accounted
because mature BW of individuals are not identical. From for a part of these heterogeneous variances and covari-
a growth modeling perspective, more accurate parameter ance effects in the data, although not all. The net result
estimation will lead to better simulation to predict per- was improved precision in the estimation of growth pa-
formance. rameters and a decrease in residual variance, ultimately
The BIC and variance estimates inferred that the mixed resulting in more accurate inferences based on these esti-
model was a better way to estimate growth parameters. mates. As modeling becomes a more important tool for
The additional random effect u allowed partitioning of predicting growth and performance, it is paramount that
the Wt variation into between- and within-bird variation those reporting growth parameters report not only geno-
and allowed the Wim of individual birds to vary around type means and standard errors, but also an estimate of
the genotype mean. More importantly, the variance and the variation between birds within genotypes.
covariance structure [5] associated with the random effect As with the fixed model, residual errors from the mixed
u in the mixed model can account for a good portion of growth model were assumed to be randomly indepen-
the increasing Wt variance as age increases. This finding dent normally distributed with mean of zero and variance
explained at least a portion of the age-associated increases of σe2. However, for the growth data in the current study,
in variance and correlations across ages that were com- this assumption was not valid. Although the addition of
mon in longitudinal growth data. the random factor ui into the model removed heteroge-
neous variance and correlation across age, it did not re-
Culling and Selection move all of this variation. Further theoretical studies are
encouraged to develop a mixed model with more appro-
The results of the selective culling are summarized in priate covariance structures for residuals.
Table 3. Selective culling mainly caused the Wm parameter
estimates to inflate upward. The inflation was very small
with the estimates of parameter b and could therefore be ACKNOWLEDGMENTS
ignored. The magnitude of the inflation of the Wm esti-
The financial assistance of the Alberta Agricultural Re-
mate with the fixed model (3.7 to 6.5% for females and
search Institute and the Alberta Chicken Producers. In-
3.8 to 11.1% for males) was much larger than that esti-
kind assistance of Lilydale Poultry Company is gratefully
mated with the mixed model (1.2 to 1.7% for females and
acknowledged. Thanks to David Eisenbart for excellent
2.0 to 5.1% for males). The smaller bias resulted from
technical assistance.
the variance and covariance structure associated with the
random factor u in the mixed model, which allowed the
mixed model to use information in the variance covari- REFERENCES
ance matrix, u, and information from those birds prior to
Canadian Council on Animal Care. 1993. Guide to the Use of
culling. Therefore, with the mixed model, estimation bias Experimental Animals. Vol. 1. CCAC, Ottawa, ON, Canada.
resulting from selective culling was reduced compared Craig, B. A., and A. P. Schinckel. 2001. Nonlinear mixed effects
with the fixed model. This result was consistent with the model for swine growth. Prof. Anim. Sci. 17:256–260.
work of Craig and Schinckel (2001). Emmans, G. C. 1981. A model of the growth and feed intake
of ad libitum fed animals, particularly poultry. Pages 103–
110 in Computers in Animal Production. G. M. Hillyer, C.
Violation of Residual T. Whittemore, and R. G. Gunn, ed. Occasional Publication
Variance Assumptions No. 5. British Society of Animal Science (formerly British
Society of Animal Production), Penicuik, Midlothian, UK.
As mentioned earlier, growth data often show an inher- Emmans, G. C. 1995. Problems in modelling the growth of poul-
ent underlying relationship. Measures made on the same try. Worlds Poult. Sci. J. 51:77–89.
Emmans, G. C., and C. Fisher. 1986. Problems in nutritional
individual are likely to be more closely correlated than theory. Pages 9–39 in Nutrients Requirements of Poultry and
measures made on different individuals, and measures Nutritional Research. Butterworth and Co. Ltd., Cornwall,
made close in time on the same individual are likely to UK.
852 WANG AND ZUIDHOF
Gompertz, B. 1825. On the nature of the function expressive of Kwakkel, R. P., B. J. Ducro, and W. J. Koops. 1993. Multiphasic
the law of human mortality, and on a new mode of determin- analysis of growth of the body and its chemical components
ing the value of Life Contingencies. Trans. R. Philos. Soc. in White Leghorn pullets. Poult. Sci. 72:1421–1432.
115:513–585. Kwakkel, R. P., J. A. Van Esch, B. J. Ducro, and W. J. Koops.
Gous, R. M. 1998. Making progress in the nutrition of broilers. 1995. Onset of lay related to multiphasic growth and body
Poult. Sci. 77:111–117. composition in White Leghorn pullets provided ad libitum
Gous, R. M., E. T. Moran Jr. , H. R. Stilborn, G. D. Bradford, and and restricted diets. Poult. Sci. 74:821–832.
G. C. Emmans. 1999. Evaluation of the parameters needed to Le Dividich, J. 1999. A review. Neonatal and weaner pig: Man-
describe the overall growth, the chemical growth, and the agement to reduce variation. Page 135 in Manipulating Pig
growth of feathers and breast muscles of broilers. Poult. Sci. Production VII. P. D. Cranwell, ed. Australian Pig Science
78:812–821.
Association, Werribee, Victoria, Australia.
Gous, R. M., E. T. Moran Jr., H. R. Stilborn, G. D. Bradford,
and G. C. Emmans. 2002. Modeling energy and amino acid Littell, R. C., J. Pendergast, and R. Natarajan. 2000. Modelling
requirements in order to optimise the feeding of commercial covariance structure in the analysis of repeated measures
broilers. Pages 82–94 in Simposio Internacional ACAV-Em- data. Stat. Med. 19:1793–1819.
brapa sobre Nutricao de Aves. Embrapa Suinos e Aves, Con- National Research Council. 1994. Nutrient Requirements of
cordia, SC. Poultry. 9th rev. ed. National Academy Press, Washing-
Hancock, C. E., G. D. Bradford, G. C. Emmans, and R. M. Gous. ton, DC.
1995. The evaluation of the growth parameters of six strains SAS Institute. 1999. SAS/STAT User’s Guide. Version 8.2 ed.
of commercial broiler chickens. Br. Poult. Sci. 36:247–264. SAS Institute Inc., Cary, NC.

Вам также может понравиться