III data collected for feeding behaviour analysis, feeding events are gener
ally separated by many very short to very long intervals during which no fe
eding occurs. When feeding is clustered in bouts, a meal criterion (that is
the longest ton-feeding interval accepted as part of a meal) must be estim
ated before events can be grouped into meals. Until recently, three methods
that estimate quantitative meal criteria were available. These methods con
sist of fitting a 'broken-stick' (two straight intersecting lines, both wit
h a negative slope) to the frequency distribution (method 1), the log(e)-tr
ansformed cumulative frequency distribution (the log-survivorship curve; me
thod 2) or the log(e)-transformed frequency distribution (method 3) of inte
rvals between events. Recently, new methods have been proposed that fit eit
her two (method 4) or three (method 5) Gaussians to the frequency distribut
ion of log,transformed interval length (log-normal models). We compare the
estimates obtained with these five methods when applied to a data set of 79
575 intervals between visits to food dispensers. These were recorded with 1
6 lactating cows during an average period of 156.6 (s.d. 51.5) days per cow
. Meal criteria were estimated as 1.9, 6.0, 7.5, 32.4 and 49.1 min by metho
ds I to 5, respectively. Estimated daily number of meals ranged from 5.7 to
12.1 per cow and estimated average meal size from 4.0 to 8.4 kg. The obser
ved probabilities of cows initiating feeding in relation to time since feed
ing last showed best agreement with the predictions of the log-normal model
s. We conclude that the first three methods do not, while log-normal models
do, have an adequate biological basis for a clear interpretation of the es
timated meal criteria. Log-normal models are, therefore, the most promising
for estimating meal criteria in cattle and probably iii other species as w
ell.