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

2.1.3. Overdispersion at the herd level
At the herd level, when the herd is con-
sidered as a whole, or when time disease
evolution for consecutive years is studied,
CMAST cases are not really independent.
A herd is characterised by its specific envi-
ronmental conditions, by germs remaining
in the environment or in the udder. Such a
Figure 1. An overview of mastitis study levels and factors.

---

496
P. Gasqui, J. Barnouin
heterogeneity could also increase overdis-
persion, whatever is the index summarising
the herd disease information.
2.2. Suggested solutions
2.2.1. Improvement of the statistical 
procedures
In epidemiology, the study level of pre-
dilection is the population. In CMAST epi-
demiological studies, the population both
corresponds to a set of herds, animals, lac-
tations and udder quarters. In order to take
into account such a complexity concerning
CMAST risk levels, several modelling
strategies were performed in the literature
[26, 27, 64, 75]. The degree of complexity
of the multivariate models which are used
to predict the risk is the consequence of the
different study levels which are interwoven
in CMAST epidemiology. Such a complex-
ity is better integrated with statistical pack-
ages. For example, the writers of SAS
software (SAS Software System, SAS
Institute Inc., Cary, NC, USA) finalised the
CATMOD procedure (categorical data
analysis), the LOGISTIC procedure (linear
regression models for binary response as
well as ordinary response data), GENMOD
procedure (which fits the generalised linear
models and allows the response probability
to be any number of an exponential family)
and finally the MIXED procedure (includ-
ing random effect models and a variety of
mixed linear models to fit data). These
improvements allow to rule out losses of
information concerning the samples which
are induced by the statistical selection in
order to ensure the mutual independence
and absence of correlation between the
data. These method evolutions permit to
estimate different levels of dependence or
correlation between some statistical units.
Moreover, the introduction of random
effects in the model allows to take into con-
sideration some sources of overdispersion,
in an empirical method.