---
title: R: Chrono
---
What does CHRONO do ?
The chronological clustering proposed by
Legendre, Dallot
& Legendre (1985) is computed by program CHRONO. This clustering
method, which had first been described for multivariate time series,
can also be used to segment spatial series
(Galzin & Legendre,
1987). The non-hierarchical method uses a hierarchical
proportional-link linkage algorithm whose connectedness level
(
Co) is determined by the user as an answer to a question of
the program; it is the test of significance, described in the next
paragraph, that makes the method non-hierarchical. The constraint of
spatial or temporal contiguity imposed to the clustering results means
that only objects or object groups that are
adjacent along the
series may eventually groupements. Notice that it is unlikely that changing
the connectedness would produce a major change in the clustering
results, as can be seen in the examples of the
Legendre, Dallot &
Legendre (1985) paper.
At each step of the agglomeration, a permutation test is performed
to decide whether a fusion should be made between the two groups whose
fusion is proposed by the agglomerative algorithm. The null hypothesis
of that test is explicitly described in the output of versions CMS and
VMS:
H is the probability that the null hypothesis is true. The
null hypothesis says that the two groups being tested are
an artifact and should be fused in a single group. Fusion
occurs if H is larger than the probability level ALPHA set
by the user (above).
Answering a question of the program, the user must determine the
alpha rejection level of the null hypothesis (often chosen
values are 0.01, 0.05 or 0.10; one may choose to use a higher level in
order to identify singletons -- see below, as well as the example). One
must realize, though, that this is not a genuine test of statistical
hypothesis, since the data used during the test are the same as those
from which the hypothesis of division into groups has been generated.
Numerical simulations described in the main reference have shown,
however, that for random data sets, the probability for this test of
producing a significant result is equal to the preselected
alpha value.
The program allows to identify
singletons, which are
aberrant samples found along the data series. Because of the constraint
of contiguity imposed on the algorithm, the presence of a singleton may
prevent the formation of a group that should have included objects from
both sides of the aberrant sample. At least three reasons may produce
such aberrant samples: (1) random events, such as modified strata in
sediment cores, or else movements of water masses during repeated
samplings at the same station in aquatic environment; (2) improper
sampling or inadequate preservation of the samples before they are
analyzed; (3) extreme stochastic variations, which lead to rejecting
the null hypothesis while no break has occurred in the succession (type
II error).
If the user requests to identify the singletons, the clustering will
be interrupted, and started again from the beginning after removing the
singleton (see example); the only exceptions to this rule are the
singletons located at the beginning or the end of the data series,
since no group is interrupted by their presence. It is unlikely that
singletons will be identified if the
alpha level is low (less
than 10%), because it then becomes difficult, when testing a single
object against p, to obtain a value which is smaller than that in the
first column of Table 1. Final rule: if an object has a similarity of
zero with all its immediate neighbors, the agglomerative algorithm does
not go down to level
S = 0 to force such an object to pertain
to a group; these unclustered objects are represented by dashes (-) in
the final solution, or by a white square in the Macintosh output graph.
It is recommended to check the data for any object coming out with that
symbol; if its presence in the series seems to have interrupted a
group, this object may either be removed from the analysis if it is
considered aberrant or exceptional (which may have given it a null
similarity with its neighbors).
Last updated on Saturday, March 30, 2013 by Philippe Casgrain