Beruflich Dokumente
Kultur Dokumente
The attached
copy is furnished to the author for internal non-commercial research
and education use, including for instruction at the authors institution
and sharing with colleagues.
Other uses, including reproduction and distribution, or selling or
licensing copies, or posting to personal, institutional or third party
websites are prohibited.
In most cases authors are permitted to post their version of the
article (e.g. in Word or Tex form) to their personal website or
institutional repository. Authors requiring further information
regarding Elseviers archiving and manuscript policies are
encouraged to visit:
http://www.elsevier.com/authorsrights
Veterinary Parasitology
journal homepage: www.elsevier.com/locate/vetpar
M.H. Gluck Equine Research Center, Department of Veterinary Science, University of Kentucky, Lexington, KY, USA
Department of Statistics, George Mason University, Fairfax, VA, USA
Department of Statistics, University of Wisconsin, Madison, WI, USA
EquiLab Laboratory, Slangerup, Denmark
Department of Infectious Diseases, University of Georgia, Athens, GA, USA
a r t i c l e
i n f o
Article history:
Received 19 September 2012
Received in revised form 15 April 2013
Accepted 19 April 2013
Keywords:
Pyrantel
Anthelmintic resistance
Cyathostomin
Horses
Hierarchical models
Outliers
a b s t r a c t
Anthelmintic resistance is an increasing challenge for the control of equine parasites. The
fecal egg count reduction test (FECRT) is the practical gold standard method for evaluating
reduction in anthelmintic efcacy, but the interpretation is complicated due to high levels of variability. A hierarchical statistical model was described for analysis of FECRT data
from multiple farms to evaluate the role of biological factors in determining the strongyle
efcacy of pyrantel pamoate in a study performed in Denmark. The model was then used
to describe two notions of farm efcacy, namely conditional and marginal efcacy. The
median of the lower prediction limits was used to describe a robust classication rule. The
performance of the methodology was evaluated using Monte Carlo simulations. The eld
study was performed on 64 Danish horse farms of different breeds. Of 1644 horses, 614
had egg counts 200 eggs per gram (EPG) and were treated. Individual coprocultures were
performed for identication of Strongylus vulgaris from all horses pre-treatment. Thirty-one
farms (48.4%) were positive for S. vulgaris, but pyrantel efcacy was unaffected by the presence of this parasite in the statistical model. Further, there were no signicant effects of age,
gender, or interactions between these, while the pre-treatment egg count was negatively
associated with the egg count reduction. The statistical model classied 81.3%, 10.9%, and
7.8% of farms as no signs of resistance (NR), suspect resistance (SR), and resistance (RE),
respectively. In comparison, arithmetic calculations classied 68.8%, 17.2%, and 14.1% in
the same categories. Using 10,000 simulated data sets, the methodology provided a classication of farms into different efcacy categories with a false discovery of reduced farm
efcacy rate equaling 8.74%. In addition, model-classication was unaffected by presence
of single outlier horses in a separate simulation study.
2013 Elsevier B.V. All rights reserved.
1. Introduction
Anthelmintic resistance has become a universal problem in gastrointestinal parasites of horses (Kaplan et al.,
2004), and it is considered an inevitable biological
Corresponding author. Tel.: +1 859 218 1103; fax: +1 859 257 8542.
E-mail address: martin.nielsen@uky.edu (M.K. Nielsen).
0304-4017/$ see front matter 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.vetpar.2013.04.036
615
and horses were recorded as either being S. vulgaris positive or negative. Farms were classied as S. vulgaris present
if any of the horses tested positive for the presence of S.
vulgaris on the coproculture.
2.3. Statistical analyses
All statistical analyses were performed using the glimmix procedure with SAS version 9.3 (SAS Institute Inc., Cary,
NC, USA). Six horses with post-treatment egg count larger
than the pre-treatment egg count were removed from the
data set prior to the analysis. Biologically, anthelmintic
treatment is expected to reduce egg counts, even if some
efcacy is lost, and thus it was assumed that these six
horses were either incorrectly treated or accidentally not
treated at all. All remaining horses had egg count reductions of 40% or higher.
Data were rst tted to a hierarchical generalized linear mixed logistic regression model where pre-treatment
egg count, age, gender, and the presence of S. vulgaris on
farms were xed effects and horses, farms and regions
were random effects. The overall variability in observed
efcacy can be described to be due to intra-farm and interfarm variability. The intra-farm variability was captured
using a horse specic random effect and inter-farm variability was captured using a farm specic random effect.
The horse specic random effect was introduced to account
for possible differences in the species composition of the
parasitic burden between horses. To account for common
characteristics between farms within a region, a region specic random effect was incorporated. All the explanatory
variables and random effects were combined additively to
model the efcacy on a logit scale. All possible interaction
terms were evaluated during model construction, as well as
residual analyses and goodness of t tests were performed
to check the adequacy of the model t.
Using the tted statistical model (1.2), the lower limits
of the 95% predicted intervals for pijk were used to assess
reduction in pyrantel efcacy and classify the farms into
signs of resistance (RE), suspect resistance (SR), and no
resistance (NR) categories. Since the prediction intervals
depend on the covariates, either the mean or the median of
these lower limits could be used as a summary measure of
efcacy for each farm. The median was chosen as this summary measure since it is a robust estimator of the center of
the distribution. Using the proposed methods, farms were
classied with three different methods: The model-based
classication, the same model without random effects and
simple arithmetic calculations without statistical modeling.
2.3.1. Statistical model
Let the indices i, j and k represent the horse, farm, and
region respectively. The farms were grouped into the ve
regions dened in Section 2.2.2 and the possible values for
k were 1, 2, 3, 4, and 5.
The number of farms within a region changed with the
region and number of horses within a farm changed with
the farm. nk denoted the number of farms in region k, while
njk denoted the number of horses on farm j of region k. Let
(Xijk , Yijk ) represent the pre-treatment and post-treatment
egg count of the ith horse, in the jth farm which belonged
to the kth region. It was assumed that,
Yijk |pijk , Xijk Bin(Xijk , 1 pijk ),
(1.1)
log
pijk
1 pijk
3
(1.2)
l=1
(1.3)
617
and 0.92, with LPLs between the two values falling in the SR
category. The principles behind this are outlined in Fig. 1.
Where classication discrepancies occurred between
the model-based classication and the arithmetic calculations, data were revisited by removing the one horse with
the lowest arithmetic FECRT on each farm to evaluate the
possible inuence of single outlier horses on classication
using the two methods.
pij (k) =
1
E(pj ) =
mj 10, 000
k=1
pij (k),
i=1
where
2
term of the statistical model (1.2). However, in the probability metric studied in this manuscript, this is not the case.
These two notions of efcacy are alternatives to the arithmetic mean. When the horses are homogeneous and the
variability in efcacy between horses is negligible, these
two notions reduce to the arithmetic mean and hence they
can be viewed as generalizations of arithmetic mean to the
non-homogeneous case. Due to multiple sources of variability and outliers in various farms, the conditional farm
efcacy was used for the classications made in this study.
10,000 mj
2
.
(2.3)
The true status of the farm was then determined using the
0.95 cutoff: if E(p.j ) 0.95, the farms status is NR, otherwise
the farms status is RE. The parameter values were chosen
to be similar to the values obtained from the analysis of the
equine data set.
Now each simulated data set was analyzed using a variant of model (1.2). This tting facilitated the model-based
classication, and each farm was then classied as RE, SR,
or NR. The FDRFE rate for this study was then computed by
comparing the farms true resistance status and the classied status. Specically, 10,000 independent data sets were
generated, and the FDRFE was computed as the proportion
of non-resistant farms (in these 10,000 data sets) wrongly
classied as resistant.
mj
1
mj
Fixed effects
pij ,
i=1
log
pij
1 pij
= 0 + 1 Xi + 2 Ai +
2
l=1
I(Gi = l) + i
i N(0, 12 ),
where i was the horse specic random effect, Ai is the
age of the ith horse. The values of the covariates were
randomly chosen from a farm with 12 horses. The regression coefcients were set as = (0 , 1 , 2 , 3 , 4 ), where
0 = 3.5 (intercept), 1 = 0.001 (pre-treatment egg count),
2 = 0.01 (age), 3 = 0.01 (gender = stallion), and 4 = 0.01
(gender = mare). An outlier effect was induced by altering
the post-treatment egg count for horse 1. The observed
post-treatment egg count for horse 1 was replaced with
the altered value Y1 , given by
Table 1
Summary of parameter estimates for xed and random effects from the
analysis of the observational farm data.
Y1 = Y1 ,
for different values of in the range [1,X1 /Y1 ]. Increasing
decreased the observed efcacy of horse 1; = 1 corresponded to the true data, where = X1 /Y1 corresponded to
the case of zero efcacy. Note that the symbol x denoted
the integer part of x.
3. Results
3.1. Farm data
Sixty-four farms with at least six horses exceeding
200 EPG were included in the study. On each farm, all
horses exceeding this threshold were included in the data
set. Farm size ranged from six to 68 horses (mean = 25.7,
median = 24), with 1644 horses being included in the
Estimate
Intercept
Pre-treatment egg count
Age
Gender (no information)
Gender (mare)
Gender (gelding)
Gender (stallion)
Presence of Strongylus vulgaris 0
Presence of Strongylus vulgaris 1
5.96880
0.00095
0.01128
0.44720
0.60130
0.78310
0
0.02592
0
Standard error
0.8434
0.0002
0.0214
0.8772
0.7936
0.7916
0.3253
Random effects
Estimate
Standard error
Cluster
Farm
Horse
0.5154
0.1705
6.4137
0.3451
0.2137
0.5244
619
Table 2
Summary of type III tests of xed effects from analyzing the observational
farm data set.
Effect
Num DF
Den DF
F value
P>F
Pre-treatment egg
count
Age
Gender
Presence of Strongylus
vulgaris
515
22.27
<0.0001
1
3
1
515
11
4
0.28
0.51
0.01
0.5981
0.6838
0.9403
Num DF: degrees of freedom of the numerator; Den DF: degrees of freedom of the denominator.
number is not dened, this would make the model undened. Alternatively, we performed the same analyses, with
the entire data set by setting the post-treatment values to
equal the pre-treatment values. This did not change any of
the conclusions described in Tables 13 and the numerical values were very close to the values presented in these
tables. Further, the classications of the 64 farms remained
unchanged (results not included).
3.2. Simulation data: results for false discovery of
reduced farm efcacy rate
The simulations evaluating the marginal efcacy
revealed that amongst 61 farms, 56 farms had an expected
mean efcacy greater than 92%, and 8.74% of them were
wrongly classied as resistant or belonging to the suspect
resistant category. This then determined the FDRFE rate for
this study.
Table 3 presents the results of the simulations study
where efcacy is evaluated using the conditional denition.
The results are displayed for each of the twelve farms and
then aggregated across all of the farms. The table compares
the performance of the model with the arithmetic classication. In this study, of the 12,000 total simulated farms,
3161 were resistant and 8839 non-resistant. Both methods
had low rates of FDRFE; the model-based classication had
an FDRFE of 0 compared to an FDRFE of 0.0547 for the arithmetic classication. The model outperformed arithmetic
classication in terms of power in detecting truly resistant
farms: 0.7062 versus 0.5135.
3.2.1. Outlier effects
The inuence of a single outlier horse on a single
farm is illustrated in Fig. 3. The gure clearly shows that
Table 3
Classication of farms according to their anthelmintic efcacy status based on 1000 simulated data sets, each based on 12 farms. Farms are classied using
the model-based and the arithmetic classications. Classication results are presented both separated out on each farm, and aggregated across farms.
Farm
True status
NR
RE
NR
RE
NR
RE
NR
RE
NR
RE
NR
RE
NR
RE
NR
RE
NR
RE
NR
RE
NR
RE
NR
RE
Model-based
NR
2
3
4
5
6
7
8
9
10
11
12
Arithmetic
SR
RE
NR
SR
RE
25
9
40
7
26
17
41
10
283
69
187
53
406
28
346
51
280
99
265
101
537
75
723
32
0
75
0
69
0
79
0
108
1
204
0
203
0
231
0
225
0
261
0
246
1
225
0
120
0
891
0
884
0
878
0
841
0
443
0
557
0
335
0
378
0
360
0
388
0
162
0
125
24
57
31
51
26
69
41
79
251
257
82
337
404
166
345
205
107
495
155
499
396
304
723
107
1
129
9
135
0
99
0
127
33
175
55
183
2
194
1
206
93
122
91
121
118
87
0
96
0
789
0
774
0
806
0
753
0
284
50
293
0
234
0
243
80
103
19
115
24
71
0
74
3159
551
2
2046
0
6242
2585
2626
403
1674
173
4539
model-based classication using the median was unaffected by the outlier, whereas the arithmetic classication
is highly driven by the FECR value of the outlier.
4. Discussion
This study provides a exible model for combining
horse-specic covariates and random effects on multiple
farms within different regions, and a statistical methodology for evaluating anthelmintic efcacy, which is robust to
the presence of outliers in farms. As a summary statistic
following the model t, we introduce two new notions of
efcacy, namely, marginal and conditional efcacies. These
arise due to the introduction of random effects into the statistical model. Finally, we validated our methodology using
simulation.
The choice of cutoff values for determining anthelmintic
resistance deserves more critical attention. Although
guidelines have been published, these are primarily
focused on sheep (Coles et al., 1992), and no recommendations exist for horses. As a result, there is no consensus
for the choice of any of these cutoff values, and each study
tends to use its own denitions (Ihler, 1995; Craven et al.,
1998; von Samson-Himmelstjerna et al., 2007; Lind et al.,
2007; Traversa et al., 2009). As the performance of a diagnostic test critically depends on the choice of cutoff, this
constitutes an underlying problem in assessing reduction
in efcacy of a drug using FECRT data. Clearly, the cutoffs
chosen in this study were arbitrary as well. However, we
based these on historical data available, where pyrantel
621