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

2.2.2. Solutions for within lactation 
dependence
To improve the independence between
animals and consecutive clinical cases,
constraints were imposed by the authors in
the sample selection process. Nevertheless,
it is obvious that such a practice induces a
systematic loss of information to the detri-
ment of the accuracy and finally the ability
of the statistical procedure to highlight the
potential risk factors.
The widely adopted solution consists in
removing within the lactation new CMAST
cases occurring in a pre-established time
interval after the primary case. Since no
precise biological justification allows the
choice of a well-defined censure duration,
there is a significant variability in the cho-
sen durations (lag time for a new case)
according to different authors. Lag time for
a new CMAST case varies from 0 to
90 days, including the value 0 [76], 7 [60],
8 [55], 10 [16, 17, 56], 14 [3, 11, 28, 73],
15 [13, 34], 30 [24, 30, 51, 59, 67, 68] or
90 days [57]. Some authors highlight this
lag time variability while varying this
parameter from 4 to 28 days [79] or from 0
to 30 days [50] starting from the same ini-
tial data. Finally, these authors used a value
of 9 days, close to the value of the 8 days
recommended by the International Dairy
Federation (IDF) [50]. The comparison
between different studies is all the more dif-
ficult since some authors do not specify the
value used in their analyses [9, 10, 45, 47,
69]. Moreover, the knowledge or not of
udder bacteriological status [38], including
various levels of accuracy, still complicates
the between-study comparisons. Some
authors have tried to define synthetic indi-
ces allowing the study of CMAST reoccur-
rences without censoring [65]. In other
studies, more explanatory models integrat-
ing biological concepts have been devel-
oped to take into account the notion of non
independence in a simulation model [1, 2]
or in a statistical model [36, 37]. All the
authors who model CMAST using this pro-
cedure used a model based on a Poisson

---

Statistical modelling for clinical mastitis
497
distribution (Poisson regression or Poisson
model through Generalised Linear Models
(GLM)) [61]. 
In some studies data are more signifi-
cantly censored, since the first CMAST
occurrence within the lactation is only con-
sidered. Within these studies, a binary type
response is sometimes chosen: CMAST
free lactations vs. lactations with at least
one CMAST case [11, 14, 16, 17, 32, 39–
41, 55, 72, 77–79, 81, 83]. Such a choice is
recommended [55] and carried out for
genetic studies [23, 42, 43, 52, 53, 60, 67,
70, 80]. According to the same approach,
the expression of this dichotomous response
is sometimes limited to a subperiod of the
lactation: from 7 days before calving to
7 days after calving [66], the 5 first days
after calving [19], the first 14 days [73, 82],
the first 30 days [7, 20, 74], the first 60 days
[6], the first 100 days [18], or even a calen-
dar year instead of the lactation [79]. The
International dairy Federation recommends
considering the 30 days before calving for
primiparous animals [50]. In other studies,
lactation is divided into different succes-
sive periods: according to the 4 seasons
[60], 2 periods of 2 and 7 months [15], or
4 periods [29]. The periods are even fixed
by some authors in a cumulative way [43,
44].
Fixing the duration of the observational
period enables easier between-study com-
parisons than taking into account the whole
lactation, whatever the duration of the lac-
tation. Among the lactation splitters, some
keep only one occurrence within the lacta-
tion [15, 29], or integrate the minimal time
interval between consecutive CMAST
cases (see above). These authors aim to
obtain over one specific period, and not
on the whole lactation, a dichotomous
response. The analysis models are mainly
based on a binomial distribution (logistic
regression, or binomial model through the
Generalised Linear Models (GLM)) [21,
61]. Other authors consider the date of
occurrence of the first CMAST case within
the lactation, choosing the calving date as
the initial moment [71]. This is the main
analysis framework of survival models,
especially the Cox models [33].