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

3.1.3. Contribution of the generalist 
exploratory models
Finally, it appears very difficult to com-
pare in a relevant way the different studies
using generalist exploratory models, and to
carry out meta-analyses. From the analysis
of the available studies, it seems to be
unlikely to more accurately identify CMAST
risk factors through these generalised meth-
ods. Moreover, some highlighted risk fac-
tors using generalist approaches are not
relevant, since they do not completely and
accurately remove the different various
levels of dependence and efficiently control
the first kind error (
a-error). An alternative
to the use of the generalist models is to
develop explanatory designed models based
on a more integrative and causal approach.
3.2. Explanatory designed models
With the generalist exploratory models,
the problem is to adapt the data to an exist-
ing statistical model available in a statisti-
cal software. The authors exclude sample

---

500
P. Gasqui, J. Barnouin
data information (lactation selection, lag-
time of cases, etc.) to respect the hypothesis
of the statistical model used. The aim is
to identify risk factors in an analysis
approach. From another point of view, it is
possible to construct a biological model
based on a lot of knowledge developed in
particular with exploratory model results.
This biological model has then translated in
a statistical model which in general, is not
directly available in a statistical software.
The specific model obtained (with biologi-
cal and statistical parts) is an explanatory
model. It is possible to use all sample data
information in this synthesis approach. 
3.2.1. States models
The knowledge of CMAST epidemiol-
ogy allows to try to develop more explana-
tory models based on biological parame-
ters, as previously recommended [62].
Consequently, a nearly accurate approach
was developed via a simulator based on the
definition of states (uninfected, subclini-
cally infected, clinically infected or recov-
ered susceptibles) and probabilities of state
changes (Fig. 2) through three methodo-
logical issues: Markov processes, discrete-
event simulation and differential equations
[1, 2]. While such an approach allows to
study a germ effect at the quarter or udder
level, it does not allow to test potential risk
factor effects, except through many simu-
lation results.