Review article Statistical modelling for clinical mastitis in the dairy cow: problems and solutions

2. STATISTICAL MODELLING 
FOR CLINICAL MASTITIS: 
PROBLEMS AND SOLUTIONS 
2.1. The overdispersion problem
2.1.1. Overdispersion at the lactation level
In the dairy cow, the number of CMAST
cases during a lactation (and herd CMAST
incidence rate) can be modelled according
to the Poisson distribution because of its
random nature [36, 50]. Concerning the
herd incidence rate, a binomial distribution
must be performed for a dichotomous var-
iable (lactation with/without CMAST).
Unfortunately, these theoretical distribu-
tions are rarely observed because of the

---

Statistical modelling for clinical mastitis
495
presence of overdispersion of the studied
data [50, 63], which involves defective
adjustments and incorrect statistical tests.
2.1.2. Overdispersion at the animal level
Successive CMAST reoccurrences can-
not be systematically considered as inde-
pendent, depending on udder pathogen
persistence [36]. Nevertheless other char-
acteristics of the observed lactations may
also generate overdispersion. This is the
case when important factors of variation are
not taken into account depending on herd
managerial conditions (housing and graz-
ing periods including different levels of risk
for CMAST occurrence) [37] or when the
true course of the lactation is not considered
[36]. Finally, the CMAST observed distri-
bution at the lactation level would involve
some overdispersion compared to the
expected theoretical distribution.
At the animal level, it is not possible to
consider that consecutive lactations includ-
ing the CMAST cases are independent.
CMAST occurrence in a selected lactation
depends on udder infection status at the
previous lactation for multiparous females
or at the prepartum period for the primipa-
rous [37]. As at the animal level, the basic
characteristics of lactation (duration, par-
ity, yield level) are also important to con-
sider in the modelling process. Even if
the lactation events are summarised in a
dichotomous way (CMAST vs. CMAST-
free), overdispersion can be observed when
the data are adjusted for a binomial theoret-
ical distribution [21].