CHAP TER 8

Migration quantiied: constructing

models and linking them with data
Luca Börger, Jason Matthiopoulos, Ricardo M. Holdo, Juan
M. Morales, Iain Couzin, and Edward McCauley
In many cases, the daily movements of animals represent in miniature movements
similar to migration, and require similar mechanisms of operation.
Woodbury (1941)

(. . .) we now probably see more clearly than ever before the intimate relation existing
between the animals and the conditions which influence their migrations.
Adams (1918)

Table of contents
8.1 Introduction 112
8.2 Identifying potential causes of migration: translating hypotheses
into models 112
8.2.1 Mathematical formulations for migration 112
8.2.2 Ecological mechanisms for migration 114
8.2.3 Migration emerging from individual-environment interactions 114
8.2.4 Mass migrations: the emergence of collective movements 115
8.2.5 Interactions between different types of conspeciics: partial and differential
migration and group leadership 115
8.3 Observing the process of migration: different types of data and their uses 117
8.3.1 Types of movement data 117
8.3.2 Preparing the data for analysis 118
8.3.3 Environmental covariates 118
8.4 Inferential approaches to migration: linking models to data 119
8.4.1 Identifying migratory patterns 119
8.4.2 Estimation 120
8.4.3 Model choice 121
8.5 Case studies 121
8.5.1 An individual-based approach: understanding the mechanisms of mass migration
in the desert locust 121

111
112 M I G R AT I O N Q UA N T I F I E D : C O N S T R U C T I N G M O D E L S A N D L I N K I N G T H E M W I T H D ATA

8.5.2 A population-level approach: ungulate migration in the Serengeti ecosystem 125

8.5.3 A hybrid approach 126
8.6 Conclusions 127

8.1 Introduction seasonal movements of animals tracking changing

resource distributions, Hansson and Hylander
‘Why did the warbler on my summer place in New (2009) interpreted it as periodic vertical diel move-
Hampshire start his southward migration on the ments of invertebrate zooplankton between surface
night of the 25th of August?’ (Mayr 1961). Questions and deep waters, and Holland et al. (2006b) as one-
about the motivation and orientation capabilities of way movements of single insect generations that
migratory animals have fascinated natural histori- may, over multiple generations, lead to a return of
ans since Aristotle and Pliny the Elder. More the entire population to its point of origin. Several
recently, faced with the challenges of conserving authors have debated whether migration should be
migratory species (Bolger et al. 2008; Wilcove 2008), deined solely on the basis of movement patterns or
other problems have gained in importance, such as whether it should also include behavioural and
mapping migratory routes (Sawyer et al. 2009; physiological processes, such as we considered in
Strandberg et al. 2009), understanding the role of Theme 2 of this book (Dingle 1996; Dingle and
individual and environmental drivers of migration Drake 2007; Hobson and Norris 2008). In Subsections
patterns (Alerstam 2006; Bolger et al. 2008), or esti- 8.2.1 and 8.2.2, we argue that, to approach migra-
mating the size of trans-continental migratory pop- tion quantitatively, it is necessary to place it in the
ulations (Hahn et al. 2009). Our overall aim in this wider context of animal movement models (con-
chapter is to discuss how models, combined with ceptual as well as mathematical). In Subsections
modern data sources and statistical methods, can be 8.2.3–8.2.6 we discuss the ecological mechanisms
used to test different hypotheses about the causes of that could spontaneously have given rise to migra-
migration. In Sections 8.2–8.4 we structure our pres- tion-like patterns of space use.
entation around these three essential components
(models, data, and inference) and, in Section 8.5, we
illustrate their linkages by means of several case 8.2.1 Mathematical formulations for migration
studies.
Dingle (1996) remarked on the dificulty of clearly
separating migration from other, often equally
8.2 Identifying potential causes of
poorly deined, types of movement such as nomad-
migration: translating hypotheses into
ism, ranging, dispersal or home-ranging (see
models
Chapter 7). Such classiication dificulties usually
The drivers of migration may be investigated by indicate that the phenomena of interest are part of a
translating contrasting, verbal hypotheses into ani- continuum. This need not preclude a discrete clas-
mal movement models. These models must then be siication of movement as long as we are prepared
itted to movement data (see Section 8.3) and the for the occurrence of non-archetypal movement
quality of it evaluated with formal inferential patterns. For example, there is a continuum between
methods (see Section 8.4). A prerequisite for all of the extremes of sedentarism and nomadism, along
this is a deinition of migration. There is, however, a which lie intermediate types of movement that
striking lack of consensus about what constitutes a share characteristics of both (see also Mueller and
migration and which biological processes give rise Fagan 2008). In fact, many authors have remarked
to it (Newton 2008; Chapter 1). For example, Baker upon the intimate connections between seden-
(1978) used migration as a shorthand for any type tarism, dispersal, and migration (Bruderer and
of animal movement, Fryxell (1991) deined it as the Salewski 2008; Mayr and Meise 1930). Here, as in
 IDE N T IF Y IN G POT E N T IAL C AU S ES O F M I GR AT I O N: T R A NS LAT I N G H Y P OT H ES ES I NTO M O D ELS 113

Chapter 7, we propose that by including nomad- its domain in competition with its neighbours might
ism, a better conceptual understanding of the be modelled by an Ornstein–Uhlenbeck process in
dynamics and patterns of animal movements can which the territorial centroid drifts according to the
be obtained. ebb and low of conspeciic pressure. Or, by using
An additional dificulty is that some movement differential game theory, the behaviour of territorial
patterns (such as ranging) are simpler than others individuals balancing territory expansion against
(such as migration). Often, this is because animal the risk of encountering hostile neighbours, condi-
movement patterns are the expression of several tional on the distance from the centroid, can be
different behaviours corresponding to different life modelled in a spatially explicit way (Hamelin and
history priorities (see also Nathan et al. 2008). It may Lewis 2010). There are statistical dificulties in sepa-
therefore be helpful to think about compound move- rating the effects of internal, environmental and
ment patterns comprising two or more elementary conspeciic inluences from a set of data (e.g.
types of movement (e.g. Morales et al. 2004). Once Benhamou 2006). We discuss these further in
again, this is a constructive conceptualization, as Subsection 8.4.2.
long as we are prepared for the fact that transitions Having described the elementary types of move-
between elementary modes of movement might ment, it is necessary to decide how these should be
often be gradual rather than clear-cut. combined to form a compound movement process.
We can classify elementary types of movement Here, we need to ask: (i) What is the time scale over
from a time series of spatial locations by asking: which an elementary type of movement occurs?
what is the minimum set of features that a move- (ii) What is the probability that a given type of ele-
ment model needs to include in order to replicate mentary movement will succeed another? and (iii)
the observed movement geometry? This broad Is there any long term pattern (e.g. periodicity) in
question can be posed more precisely in three parts: how different elementary types of movement occur
(i) Can the patterns of movement be replicated with in relation to each other?
reference solely to the animal’s present or past The lifetime tracks of migratory animals are prob-
positions? (ii) Do we also need to involve informa- ably as complicated as any movement trajectory
tion on the environment, or on individual behav- can be; for part of the year, migratory animals
ioural state and physical condition? (iii) Do we appear to be nomadic and for other parts they
need to involve information on the position of appear to be home-ranging. Their movement can be
conspeciics? affected by multiple environmental gradients, ixed
The modelling literature offers mathematical for- reference points and the movement of conspeciics.
mulations in response to each of these questions. Different types of movement succeed each other
For example, if movement is purely self-referential with stochastic regularity, which may or may not be
we can use models such as the simple random walk seasonal. The modeller’s task is to construct an
(Turchin 1998), the correlated random walk (Turchin appropriate mathematical formulation that is capa-
1998) or the Lévy walk (Bartumeus et al. 2005; ble of reproducing this richness of pattern thus
Edwards 2008). The effect of the environment can facilitating understanding of (and making testable
be incorporated in different ways depending on predictions regarding) the underlying processes.
whether it acts through local conditions/gradients The task of statistical inference is to identify which
(biased random walk, Turchin 1998) or some global, elementary types of movement are found in the
ixed points of reference (Ornstein–Uhlenbeck proc- data, how each of them is speciied parametrically
ess, see Blackwell 1997, 2003). Finally, if we consider and how they succeed each other through time.
conspeciics as part of an animal’s extended envi- Each of these tasks can usefully be guided by eco-
ronment, the same models can be used if we allow logical irst principles. Finally, the speciic question
the environmental conditions/gradients/points of asked will determine the spatiotemporal scale of
reference to be dynamic and interactive (Haydon analysis—for example, for many questions regard-
et al. 2008; Moorcroft and Lewis 2006). For example, ing migration, small-scale temporal variation in
a territorial animal exploiting the resources within movements might be ignored.
114 M I G R AT I O N Q UA N T I F I E D : C O N S T R U C T I N G M O D E L S A N D L I N K I N G T H E M W I T H D ATA

8.2.2 Ecological mechanisms for migration from the dynamic interaction of the individual with
its environment (e.g. Adams 1918; Mayr and Meise
Active movement allows animals to exploit spatially
1930). Apart from the historical interest in how
separated resources and reduce the risk of mortality.
some migratory patterns might have evolved, these
Passive movement is in addition shaped by environ-
interactions allow us to model environmentally
mental driving forces. In either case, migration is
driven migration from irst principles. There is a
unlikely to occur independently of an animal’s envi-
certain conservation urgency in understanding such
ronment (Adams 1918; Nathan et al. 2008; Schick
environmentally determined migratory patterns
et al. 2008). A full, mechanistic understanding of
because they are the most likely to be disrupted by
migration requires knowledge of: (i) an animal’s life
environmental change (Chapter 11).
history priorities (survival, growth, fecundity) and
In many cases, the environment may be solely
how these vary with age and season, (ii) the set of
responsible for driving migratory movements. If
behaviours (exploring, foraging, commuting, mat-
animals track desirable conditions and environ-
ing, avoiding risk) that are used to satisfy these pri-
mental conditions vary along smooth geographical
orities, (iii) the spatial distribution of resources or
gradients, then seasonality in temperature, rainfall
conditions (food, risk, cover, cues, conspeciics) that
and the availability of resources may lead to regular
may affect movement and (iv) the cognitive abilities
movement of animals between locales (Chapter 7).
of the animal (memory, perception, communication,
In an environment where two resources are spa-
intention) that enable it to acquire and exploit infor-
tially separated, an animal may be forced to under-
mation from its environment. These processes con-
take long-range movement because its resource
tribute to the state of an animal; its membership of a
requirements are not satisied by the local habitat.
particular life-history stage, its energetic condition
As the need for one resource becomes satisied, the
and spatial position. Although it might not always
animal may return to its original position. This may
be necessary to include all these components in
give rise to regular sorties, as seems to be the case in
order to model migration, it is useful to enquire how
plankton migration (Hansson and Hylander 2009).
they can spontaneously lead to migratory move-
Whether this should be classiied as migration is
ments. However there is a caveat when applying
unclear (but see Woodbury 1941 for an early sug-
these insights to genetically hard-wired movement
gestion): Dingle’s (1996) deinition of migration
behaviours; evolved behaviours may become mala-
requires that during a migratory move, animals
daptive over time, making it hard to understand
should not be responsive to resources that would
their existence from observation of the animals’ cur-
otherwise elicit a reaction. We think that this is per-
rent environment. It might be interesting for future
haps too restrictive. For example, movements of
research to explore if/when, and in which systems,
long-ranging capital breeders such as elephant seals
hard-wired behaviour can be identiied in the data.
(Le Boeuf et al. 2000) are the result of extreme sepa-
For example, if there is a component of movement
ration between breeding sites and seasonal foraging
that remains ixed year after year despite inter-an-
grounds. Their annual movements are therefore
nual environmental variation, or a movement com-
characterized by a regularity and spatiotemporal
ponent that is conserved within genetically-distinct
extent that are more often associated with migra-
populations or individuals. Understanding the lex-
tion than central-place foraging. Yet, it is likely that
ibility of migration behaviour to environmental
they would remain close to their breeding grounds
variation is also of applied interest under climate
throughout the year if they could feed at a suficient
change scenarios.
rate.
Perhaps less obvious are migratory patterns that
arise from two-way interactions between the ani-
8.2.3 Migration emerging from individual-
mals and their environment. A nomadic animal liv-
environment interactions
ing in a fragmented environment may, by chance,
It has been postulated for some time that migration- be captured within a small subset of a habitat patch
like patterns can emerge in several different ways network. This is especially likely if the geometry of
 IDE N T IF Y IN G POT E N T IAL C AU S ES O F M I GR AT I O N: T R A NS LAT I N G H Y P OT H ES ES I NTO M O D ELS 115

the network prevents the animal from perceiving, from environmental drivers. Through a combina-
or makes it unwilling to move within, the wider tion of laboratory experiments and individual-
network. Such movement may even be periodic based modelling, recent research has obtained a
under certain conditions. For example, if the animal detailed understanding of the mechanistic basis for
is solely responsible for depleting the resources in a the onset and maintenance of such mass move-
two-patch system it will regularly alternate between ments in locusts, as we describe in detail in the case
patches. Seasonal variation in environmental condi- study in Section 8.5.1.
tions could cause the same patterns, especially at
the edges of species ranges (Mayr and Meise 1930).
8.2.5 Interactions between different types of
Alternatively, an energy-intake maximizer with
conspeciics: partial and differential migration
good knowledge of the position of food patches but
and group leadership
incomplete knowledge of their current value, may
alternate between exploitation and exploration. Partial migration occurs when only one part of a
Hence, even if the individual is not the primary population migrates while the other remains seden-
cause of patch depletion it could nevertheless tary. Differential migration refers to populations
present regular attendance patterns. This could whose members follow different migration strategies
conceivably give rise to patterns such as the traplin- (thus, partial migration could be deined as a special
ing behaviour observed in some hummingbirds case of differential migration, yet this is counter to
and bumble bees (Ohashi et al. 2008; Ohashi and common usage). Populations are not made of identi-
Thomson 2005). Although the limited spatiotempo- cal individuals and some partial migrations are age-,
ral scale of the above examples appears to be stretch- stage- or sex-dependent. For example, Daphnia of
ing the deinition of migration, these are the same different sizes are able to assess the threat level of
essential mechanisms that underlie mass migra- ultraviolet radiation and predation risk and respond
tions (see also Woodbury 1941) and may also explain by adjusting their depth distribution, resulting in
white shark migration in the eastern Paciic size-dependent differences in diel vertical migration
(Jorgensen et al. 2010). (Hansson and Hylander 2009).
Differences in life history priorities and energetic
requirements between different population compo-
8.2.4 Mass migrations: the emergence
nents may interact in unpredictable ways with local
of collective movements
population density and the spatial distribution of
The interaction between several individuals has the resource abundance. For example, increases in red
potential to amplify the mechanisms discussed deer (Cervus elaphus) density in the North Block of
above. For example, several consumers acting the Isle of Rum after cessation of culling (Clutton-
together are more likely to deplete local resources. Brock et al. 1982) led to an increase of male emigra-
When such groups are found in environments con- tion rates, not so for female emigration rates, with
taining few patches, it is possible that consumers virtually no resident males in high female density
will move in synchrony, particularly in the presence areas by the late 1990s (Coulson et al. 2004); males
of conspeciic attraction, e.g. as a predator-avoid- migrated to these high female density areas only
ance strategy. Conversely, conspeciic avoidance during the rut. The underlying cause is probably
can also play a role—recent research has shown that forage availability mediated through the effect of
coordinated insect mass migration can be generated sex-biased body-size differences; the smaller female
by cannibalistic local interactions between individ- red deer are better able to utilize the short-cropped
uals (escape and pursuit responses, Bazazi et al. high-quality vegetation on the greens, which pre-
2008; Simpson et al. 2006) but that insect density sumably has led the males to search for areas with
must be above a critical density for coordinated lower grazing pressure (Conradt et al. 1999). A simi-
motion to result (Buhl et al. 2006). In general, group lar example of differential migration by sex has
dynamics can yield migratory movement through been shown recently in Great bustards (Otis tarda)
the effects of local interactions, even without forcing in central Spain (Palacín et al. 2009).
116 M I G R AT I O N Q UA N T I F I E D : C O N S T R U C T I N G M O D E L S A N D L I N K I N G T H E M W I T H D ATA

Furthermore, differential informational states the scene for frequency-dependence in migratory

among individuals can play an important role in strategy.
migratory processes (for a similar suggestion con- A further aspect of aggregation among individu-
cerning the importance of information in animal als during migration is that individuals may ben-
movements see Clobert et al. 2009). In some cases eit from the directional assessment of others. If
more experienced (typically older) individuals, each individual makes their own, error-prone,
dominant individuals, or those with most to lose by assessment, but then (in addition) tends to align
not migrating may lead groups. Couzin et al. (2005) with the direction of motion of others, environ-
developed an individual-based model to explore mental noise can be dampened and long-range
such leadership behaviour. They assumed that noisy gradients more easily assessed. This has
social individuals have a conlict of interests when been termed the ‘many-wrongs principle’ (Simons
making movement decisions—to remain with 2004; Hancock and Milner-Gulland 2006; Codling
group members or to move in a goal-oriented et al. 2007).
(desired) direction of travel. When conlicts between Animal groups may be capable of even more
these desires arise, individuals must reconcile their sophisticated collective assessment of noisy envi-
behavioural tendencies. In the context of migration, ronments. For example Torney et al. (2009) used an
a desired direction could represent the memory (or individual-based model to demonstrate that groups
a genetically encoded representation) of a migra- can climb environmental gradients with minimal
tory route, or a direction based on an individual’s cognitive or sensing capacity even if, at the level of
assessment of a local environmental gradient. This an individual, an explicit assessment of the gradient
model revealed that leadership can spontaneously cannot be made. If grouping individuals employ a
emerge in groups without requiring signalling or context-dependent strategy such that social interac-
individual recognition and that only a few such tions (alignment with and/or attraction towards
‘informed’ individuals can accurately guide a large others nearby) become increasingly dominant when
number of others. Furthermore as group (or popu- the local environment is decreasing in quality, and
lation) size gets larger, the proportion of ‘informed’ they ignore others when the local environment is
individuals required for effective navigation dimin- increasing in quality, then groups can climb gradi-
ishes rapidly, such that only a very small percentage ents that are impossible for individuals in isolation.
of individuals (<5%) need actually be actively Groups can therefore display an awareness of their
migrating and yet collective performance is very environment not present at the individual level.
similar to the maximal possible migration accuracy Understanding the relative contributions of indi-
(such as if all individuals are actively migrating). vidual interactions, genetically hard-wired behav-
If detecting environmental gradients is costly— iours, and environmental variation in determining
energetically or cognitively, or through time lost in synchronization of migration over large spatial
other activities such as feeding or vigilance—then areas, for example in birds (Nowakowski et al. 2005),
one could envisage a scenario where relatively few as well as temporal variation between years, is an
individuals are actually navigating and a large pro- important topic in migration studies. The ideas out-
portion of others are using a cheaper strategy of lined above will be a useful guide to tackling these
responding to social cues. This is what appears to questions. Information transfer need not be limited
occur in honey bee colony dispersal; in inter-nest to conspeciics. Instead, inter-speciic dynamics
movements of social insects such as honey bees, may be very important for the emergence of migra-
where acquiring information about where to go is tory patterns. Escape-and-pursuit responses are
known to be costly, less than 5% of the colony obtain more prevalent in multispecies interactions. A pred-
information and guide the whole swarm (Seeley ator population tracking a migrating prey is the
1985). Thus, in migrating species, individuals may most trivial way in which inter-speciic interactions
accept such costs if they aid their kin by leading could result in species migration through predator
the group, or even among unrelated organisms if avoidance. Similarly, information transfer can also
the risk of not migrating is suficiently high, setting occur between individuals of different species
 OB S E RV IN G T HE PR O C ES S O F M I GR AT I O N: D I F FER EN T T Y P ES O F D ATA A N D T H EI R U S ES 117

which might also lead to the emergence of, or inlu- ies (especially, absence of individual-based infor-
ence the dynamics of, migratory patterns. mation) are provided in the case studies (Sections
8.5.2 and 8.5.3).
Trajectory data or movement path data are
8.3 Observing the process of migration:
obtained using telemetry methods (Bowlin et al.
different types of data and their uses
2005; Cooke et al. 2004; Fedak et al. 2002; Kenward
Traditionally, migration was the most challenging 2001; Millspaugh and Marzluff 2001), ranging from
movement behaviour to observe because of the visual tracking and radio-telemetry to geolocation
large spatial scales involved. Technical advances and real-time satellite telemetry. Thanks also to
are now revolutionizing our ability to collect technical developments (Reynolds and Riley 2002;
detailed movement data over large spatial scales Rutz and Hays 2009; Wikelski et al. 2007), these are
(Green et al. 2009; Moil et al. 2007; Ropert-Coudert the only data with which to answer more detailed
and Wilson 2005; Rutz and Hays 2009; Wikelski questions such as the location and characteristics of
et al. 2007). Depending on the research objectives migratory routes, the type of movement modes
and the characteristics of the study system, many employed during migration (including stopovers)
different methods can be used to observe migration or the balance between active and passive move-
(Section 8.3.1). After collecting appropriate move- ments (e.g. inluence of ocean currents), as a detailed
ment data, several issues relating to data quality record of the actual movements over large spatial
need to be addressed (Section 8.3.2), such as location scales is required. Two examples of the use of trajec-
error or imputation of missing data. Traditionally, tory data in migration studies (Bunnefeld et al. in
these operations were separated from statistical review; Fieberg and Del Giudice 2008) are discussed
modelling, but modern analysis methods incorpo- in Section 8.4.1.
rate the treatment of sampling issues within model Data obtained through mark-recapture methods
itting. Finally, we briely address the importance of (Amstrup et al. 2005; Turchin 1998), such as trap-
good-quality environmental data for understand- ping, photo-ID, transponder tags and ringing, have
ing the emergence of migratory patterns (Section features from both transects and trajectory data.
8.3.3). Hence, although getting more than one observa-
tion for some animals is possible in some experi-
mental set-ups, the number of repeated observations
8.3.1 Types of movement data
is never as high as with telemetry. The spatial reso-
There are three main types of data on movement; lution of the observations is usually error-free
transect data, trajectory data and mark-recapture because the recapture locations are known, but
data. Transect data (Buckland et al. 2001) can be spatial coverage is generally poor. Representation
collected from various platforms (visual or acous- of the environment at irst capture/sighting is
tic, ship, aerial or shore-based) and by various within the observer’s control but the recapture/
methods (strip-, line-, point-transects, cue-count- resighting probability depends on where individu-
ing). Depending on the type of transect and plat- als go. Furthermore, the number of recapture/
form used, surveys can cover large areas and resighting stations are usually fewer than the obser-
because the regions of observation are subject to vation points in a transect survey. Mark-recapture
survey design they can theoretically achieve repre- methods were originally developed to study bird
sentative spatial coverage. However, this is not migration; as early as 1800, the naturalist John
easy in practice and the challenge for survey James Audubon attached strings to the legs of
design is to reduce bias (Buckland et al. 2001; migratory birds, to successfully demonstrate that
Buckland et al. 2004) caused by, for example, coast- the same individuals tended to return the follow-
line geometry and imprecision due to variation in ing spring (which he indeed could verify). A recent
detection probability (Strindberg and Buckland example of the use of mark-recapture data in migra-
2004). Examples and discussions of the advantages tion studies (Ambrosini et al. 2009) is discussed in
and limitations of transect data for migration stud- Section 8.4.1.
118 M I G R AT I O N Q UA N T I F I E D : C O N S T R U C T I N G M O D E L S A N D L I N K I N G T H E M W I T H D ATA

An emerging type of data for migration studies error must be considered as a possible cause of 180°
comes from methods involving stable isotopes, turning angles.
strontium isotopes and trace elements contained Various techniques are available to deal with
in feathers, blood, ish otoliths or other tissues these problems. For example, sparsely or unevenly
(Barnett-Johnson et al. 2005; Brattstrom et al. 2008; sampled locations may be interpolated, using sim-
Hobson and Wassenaar 2008; Kaimal et al. 2009; ple linear interpolation or more lexible curvilinear
Norris et al. 2005). These are integrative or cumula- methods depending on the study objectives and
tive data, because information about position is characteristics (Tremblay et al. 2006). Unlikely loca-
compounded on to the organism’s chemical tions may be iltered out prior to analysis based on
make-up, and thus they might be classiied as ‘indi- the expected distribution of movement speeds and
rect data on position’ (genetic markers can also be turn angles, or the habitat of the position (e.g. land
considered as part of this group). The advantage of for a ish).
these techniques is that information like natal ori- Using state-space models (SSMs; Patterson et al.
gin, overwintering location, and even overwinter- 2008, see also Section 8.4) all these operations can be
ing habitat use (Norris et al. 2005), can be quickly carried out as part of model itting. SSMs are hierar-
obtained for each individual without the need for chical time-series models, usually estimated by
any tracking or banding device. However, the spa- Bayesian techniques. They comprise two probabil-
tial resolution is generally coarse and strictly deter- istic components, a process model that refers to the
mined by the available information on the spatial biological mechanism and an observation model
geochemistry of the chosen marker(s), which may that describes how observations of the system are
limit the applicability of the techniques in certain obtained. Thus, SSMs explicitly model the data col-
areas (Lank et al. 2007). A statistical problem is that lection process and use information about data
without the knowledge of an individual’s previous quality, as well as other, ancillary data (e.g. sea-sur-
life history a large part of the natural variation in face temperature or geographic boundaries), to
tissue markers cannot be controlled for (Brattstrom make inferences about an animal’s true position
et al. 2008). This problem may be partly circum- (Jonsen et al. 2005; Patterson et al. 2008; Schick et al.
vented by combining techniques, e.g. trace ele- 2008). However, SSMs are technically and computa-
ments and stable isotopes (see Kaimal et al. 2009 tionally demanding, so prompting the development
and references therein). A more general limitation of various ad hoc methods (e.g. Tremblay et al.
of these techniques for detailed movement studies 2009), that focus primarily on data-iltering.
is that, usually, only one location record is obtained We have so far focused on the effects of location
for a particular individual (but see Cherel et al. errors and missing data on trajectory and mark-re-
2009). capture data. Transect data are affected by other types
of error, such as the imperfect detectability of animals
from the transect. These problems can be dealt with
8.3.2 Preparing the data for analysis
separately using distance methods. Alternatively,
Animal location data are never sampled without their treatment may be embedded in the model itting
positional error (D’Eon and Delparte 2005; process (e.g. Royle and Dorazio 2008).
McKelvey and Noon 2001; Witey et al. 2001) and,
depending on the study objectives, even small loca-
8.3.3 Environmental covariates
tion errors can bias the inferences obtained from the
data. For example, Hurford (2009) showed that GPS Good-quality environmental data are essential for
collar measurement errors give rise to a systematic modelling individual-environment interactions and
bias in the distribution of turning angles for succes- thus for understanding the emergence of migratory
sive positions less than 20 m apart; a stationary ani- patterns. This topic is considered in more detail in
mal is mostly incorrectly measured as turning Chapter 7, but it is important to remember that, just
backwards by 180° or moving forwards towards a like movement data, environmental covariate data
ixed point in space. Consequently, measurement may require lengthy preparation prior to modelling
 IN F E R EN T I A L A P P R O AC H ES TO M I G R AT I O N : LI N K I NG M O D ELS TO D ATA 119

(Aarts et al. 2008; Haining 2003; Hunsaker et al. connectivity were produced by aggregation of indi-
2001). viduals after the migration or as the result of a direct
transference of spatial relationships across migra-
tion areas. Application of these methods to a large
8.4 Inferential approaches to migration:
dataset of ringing recoveries of barn swallows
linking models to data
(Hirundo rustica L.) migrating from Western
There are three main questions we can ask in our Palearctic breeding areas to sub-Saharan winter
analyses: (i) Are the observed patterns of movement locations showed that, for this species, migration
migratory? (detecting migration; Section 8.4.1); connects speciic breeding and wintering areas.
(ii) What are the quantitative properties of the These results were also in accordance with informa-
migratory patterns? (estimation; Section 8.4.2) and tion from studies based on different methods.
(iii) What are the underlying mechanisms (model Consequently, this approach might be used to quan-
choice; Section 8.4.3)? tify migratory connectivity in other animal taxa.
Thanks to the current revolution in biotelemetry
technologies we can start mapping the migration
8.4.1 Identifying migratory patterns
routes of many taxa. Several questions arise from
We address here the challenge of devising methods such data: Why do individuals choose some partic-
for identifying migratory patterns and the underly- ular routes? Why do migratory routes of a species
ing movement modes, as a prelude to modelling. converge in speciic geographical areas (e.g.
A central theme of current movement theory is the Strandberg et al. 2009)? To address these questions it
recognition of the fundamental spatiotemporal scal- is important to map all feasible migration routes.
ing of the movement of organisms (Nathan et al. Pinto & Keitt (2009) proposed an extension of the
2008; Chapter 7). Scales extend from a single move- least-cost approach, based on a GIS implementation
ment step, to entire movement paths (deined as of graph theory methods. Using simulations, the
temporal sequences of location data), to the lifetime authors showed that, in landscapes with randomly
track of an organism, to the multigenerational tracks distributed habitat, multiple movement routes con-
of populations. Based on the assumption that the verged in the same areas. Conversely, once the dis-
structure of a movement path is a relection of the tribution of favourable habitat became more
basic processes that produced it, the irst step for clustered, movement corridors became less redun-
obtaining a mechanistic understanding of move- dant and bottlenecks could be identiied. The
ment is to identify the basic functional units of the authors exempliied the approach by showing the
lifetime track, called movement phases (Nathan effects on movement routes of simulated habitat
et al. 2008). destruction on a real landscape. Hence, this
Most populations have a characteristic spatial approach might be employed to map the distribu-
structure (e.g. colonies, territories, metapopulation tion of alternative migration routes and compare it
structure). An interesting question for migratory with observed movement data.
species is whether these spatial relationships Finally, how can we detect migratory patterns in
between individuals stay consistent between areas. the irst place? Ad hoc criteria are generally used to
In other words, do individuals of one colony or determine migration patterns, such as the onset of
population migrate to the same areas? This has been migration or migration distance. A further complica-
termed migratory connectivity and has been stud- tion is that generally the available data rely on peri-
ied extensively in birds (Webster et al. 2002). Recently odic observations of animals (see Section 8.3.1). Thus
Ambrosini et al. (2009) proposed a new method to the resulting response times, like the onset of migra-
quantify migratory connectivity: Mantel correlation tion, are known to occur only within a certain time
coeficients were used to test if the observed con- interval, e.g. a week (such data are called interval-
nectivity deviated markedly from a random distri- censored data, as opposed to known-event times).
bution of individuals and clustering methods were Fieberg and Del Giudice (2008) recently addressed
used to analyse whether observed patterns of this issue. The authors compared non-parametric
120 M I G R AT I O N Q UA N T I F I E D : C O N S T R U C T I N G M O D E L S A N D L I N K I N G T H E M W I T H D ATA

and parametric methods for exploring migration the underlying ecological drivers? Inferential mod-
data, using white-tailed deer autumn migration as elling approaches may be split into Bayesian vs. fre-
their study system, and suggested a new parametric quentist (Cox 2006; Gelman and Hill 2007), dynamic
model that also accounts for facultative migrants. vs. static, probabilistic vs. algorithmic/data mining
Furthermore the authors illustrated methods for (Breiman 2001; Hastie et al. 2001), mechanistic vs.
analysing environmental effects (climate) on the statistical (for a discussion see Bolker 2008; Hilborn
timing of migration and suggested that it may be and Mangel 1997). In practice, the results of differ-
necessary to account for interval-censoring. ent approaches will often be similar (e.g. Spiegel-
A different approach to identifying and quantify- halter and Rice 2009) and the improvements in
ing migration patterns has recently been developed computer hardware and statistical software facili-
by Bunnefeld et al. (in review), based on non-linear tate the application of multiple methods to a given
mixed effects modelling of net squared displacement problem. Importantly, using multiple methods is
(NSD) patterns. The idea is that different movement especially powerful if the choice is guided by bio-
behaviours may generate different space use pat- logical knowledge.
terns and an informative measure of these patterns is Notwithstanding the inferential method
given by Net Squared Displacement statistics employed, however, there are statistical dificulties
(Turchin 1998). NSD can be used to identify distinct in separating the effects of internal, environmental
movement phases, like searching vs. homing behav- and conspeciic inluences from a set of data. For
iour (Börger et al. 2008; Moorcroft and Lewis 2006). example, the state–space modelling approach is
For example, Fryxell et al. (2008) characterized differ- quite appealing for dissecting movement trajecto-
ent movement phases of dispersing elk using gener- ries into different movement strategies (Patterson et
alized additive models to analyse net squared al. 2008; Schick et al. 2008) but parameters on speed
displacement patterns, and Kölzsch and Blasius and sinuosity of movement could play off against
(2008) analysed the characteristics of stork migration each other to give similar overall displacements;
by inspecting mean-squared displacement patterns. problems may arise due to the interplay between
Börger et al. (in review) developed a novel approach, observation error in telemetry data and movement
derived from animal movement theory, to character- parameters; or behavioural and movement param-
ize and quantify the movement patterns of dispers- eters may conlict such that, in a model that allows
ers by means of non-linear hierarchical modelling of multiple behaviours, it might become impossible to
net-squared displacement patterns. Building on this separate between a switch in behaviours and ran-
approach Bunnefeld et al. (in review) developed a dom heterogeneity in movement paths. This is a
model to identify migratory individuals and to quan- general problem in statistical analysis, called the
tify the timing, duration, and distance of migration, inverse or non-identiiability problem, which arises
all within the same approach. The authors exempli- every time there are multiple (or even ininite) com-
ied the approach using data from two populations binations of parameter values that produce equiva-
of GPS-collared migratory moose (Alces alces) in lent model its. Thus, the problem cannot be solved
northern Scandinavia, showing that individual dif- by increasing sample size but additional independ-
ferences accounted for a signiicant part of the varia- ent information is necessary. Consequently, we urge
tion in the distance of migration but not in the timing researchers to consider carefully which data will be
or duration and that the model had good explana- necessary to investigate speciic hypotheses—even
tory and predictive power between individuals and the most detailed GPS data might not be suficient
populations. but will need to be integrated with independent
data on the state or behaviour of individuals.
Thanks to the revolution in bio-logging techniques
8.4.2 Estimation
(Green et al. 2009; Moil et al. 2007; Ropert-Coudert
Once the data have been collected, prepared for and Wilson 2005; Rutz and Hays 2009; Wikelski
analysis and migratory patterns have been detected, et al. 2007) such requirements are increasingly easy
what is the best way to obtain inferences regarding to meet.
 C A S E S T U D I ES 121

the autocorrelation pattern was not modelled

8.4.3 Model choice
directly but instead was an emergent property of
Model criticism, model understanding and coni- the movement process. In this example, Morales
dence building are fundamental components of et al. (2004) could discriminate between models that
modern statistical modelling (Gelman and Hill provided very similar it to the data in terms of DIC
2007) and take considerably more time to complete but that had different performance when account-
than model itting itself. Especially for more com- ing for other interesting properties of the data.
plex models, this may involve various steps such as Research on model selection procedures is currently
debugging, software validation, checking the con- very active (e.g. Kuha 2004; Raffalovich et al. 2008;
vergence of numerical algorithms and checking the Sauerbrei et al. 2008; Ward 2008) and we anticipate
it of a model to data using computer intensive ran- further developments in this area.
domization approaches.
Different candidate models can be itted to the
data and model selection approaches can be used to
8.5 Case studies
choose between them or come up with a conidence In previous sections we have discussed the three
set of models (Burnham and Anderson 2002; essential components for testing hypotheses about
Murtaugh 2009). For example, Morales et al. (2004) the causes of migration behaviour—models, data
itted several movement models to reintroduced elk and inference. Here, we illustrate the links between
in Ontario and used the Deviance Information them using three case studies, each taking a differ-
Criteria (DIC) to weight the merits of alternative ent modelling approach to the problem of migra-
models, concluding that individuals alternated tion (individual-based, population-level and a
between ‘encamped’ and ‘exploratory’ movement combination of the two).
modes. The encamped mode was characterized by
short daily displacements and large turning angles
8.5.1 An individual-based approach:
while the exploratory movements included large
understanding the mechanisms of mass
displacements and small turning angles. For some
migration in the desert locust
of the animals, a very similar it to the data was
obtained when step length in exploratory move- The basis of many of the individual-based
ment had mode zero and a long tail or when the (Lagrangian) techniques that are currently used to
distribution had a large positive mode (the Weibull understand animal motion, such as random walks
distribution used for step length has zero mode and (Berg 1983) and bio-social interaction forces among
a long tail for shape <1, and positive mode for individuals (Couzin and Krause 2003), is in statisti-
shape >1). That is, some very small displacements cal physics (statistical thermodynamics/mechan-
could be included as part of the exploratory move- ics), where the aim is to relate the properties of
ment by having a shape parameter of less than one, individual atoms or molecules to the large-scale
which contradicted the idea of exploratory move- (macroscopic, or bulk) properties of materials, liq-
ments being large. Morales et al. (2004) then decided uids and gases. Recently, there has been growing
to constrain the shape of the step length distribu- interest in applying similar mathematical tech-
tion using a strong prior for this parameter and niques to study biological processes in which a large
looked at the posterior predicted distribution of number of individuals interact to produce collective
autocorrelation in step length as another tool for behaviour. Notwithstanding the increased com-
model comparison. Movement data from collared plexity of organisms compared with atoms and
elk showed strong autocorrelation patterns in step molecules, there is an important difference in living
length as they alternated between movement modes systems—the constituents are not at the mercy of
and only the ‘constrained’ model was able to repro- thermal noise, but rather act to propel themselves.
duce this feature of the data, providing good justii- Buhl et al. (2006) and Yates et al. (2009) have
cation for the use of a strong prior in the shape employed this approach to study the onset and
parameter of the exploratory movements. Note that maintenance of collective motion in locust swarms.
122 M I G R AT I O N Q UA N T I F I E D : C O N S T R U C T I N G M O D E L S A N D L I N K I N G T H E M W I T H D ATA

Insects such as the desert locust (Schistocerca that could track the motion of up to several hundred
gregaria) form some of the largest migrating insects simultaneously and in real time (Fig. 8.1, top
groups found in nature (up to 109 individuals). panel). Thus individual positions and motion were
Approximately one ifth of the Earth’s land surface known, from which macroscopic properties such as
is susceptible to invasion by this species. Locusts, the degree of collective motion exhibited by the pop-
counter to popular expectation, are shy, green cryp- ulation could be derived. This latter ‘instantaneous
tic grasshoppers whose natural tendency is to move alignment’ is analogous to ‘order parameters’ in
little and to avoid one another. It is only when physics and was close to 1 if all individuals were
forced to feed together under conditions of resource moving in the clockwise direction, close to −1 if they
limitation (which is particularly likely when sud- were moving counter-clockwise and close to zero if
den drought follows a period of high food availabil- an equal number were moving in each direction at
ity) that locusts change behaviour and become the same time (see Fig. 8.1, lower panels). While this
attracted to one another. Locusts in the gregarious is a deliberately simpliied version of reality, it
state are then prone to forming mobile aggregations allowed the detailed analysis of the essential features
that can extend over tens of kilometres and have a of the onset of collective motion.
devastating impact on local communities through In conjunction with the experiments, to enable a
their consumption of crops. This process of behav- deeper theoretical understanding for this speciic
ioural ‘gregarization’ occurs within a few hours system, the authors used an individual-based mod-
(Simpson et al. 2001) and results from a combination elling approach in which individuals were consid-
of sight and smell of conspeciics, or tactile contact ered similar to interacting particles in physical
to the hind legs. Recently it was discovered that systems; speciically, a ‘self-propelled particle’ (SPP)
release of the neurochemical serotonin is both nec- model in which particles (i.e. individuals) exhibited
essary and suficient for such a transformation a stochastic propensity to align their direction of
(Anstey et al. 2009). The full process of gregarization travel with their neighbours. This had the advantage
involves a complex cascade of physiological changes of being a mathematically tractable minimal model
including black body patterning and shape changes of the dynamics exhibited by the real system. Using
that occur over a longer period (especially across this approach Buhl et al. (2006) revealed that both
moults). Wingless juvenile insects—termed nymphs gregarious hoppers and the self-propelled particles
or ‘hoppers’—are more prone to such aggregation in the model spontaneously formed mobile bands
since their motion is largely restricted to the plane when the local population density exceeded a criti-
and indeed large aggregates of nymphs (called cal value. Below this density, the insects behaved
‘hopper bands’) invariably precede the lying much like particles in a gas, with interactions
swarms of adults. between individuals in close proximity, but no long-
Locust gregarization is thus an exquisite example range order or collective motion at the group level.
of phenotypic plasticity. But what drives the migra- As insect density increased, however, a local behav-
tion of locusts once they are gregarious? Buhl et al. ioural tendency (over a scale in the order of 5–15 cm)
(2006) built an annulus arena in which locusts could for locusts to align their direction of travel with one
move under controlled laboratory conditions (Fig. another could account for the sudden, and spontane-
8.1). This represented (in reasonable space) a typical ous, formation of a mobile band (without necessitat-
subsection of the much larger swarms that form in ing individuals to change their behaviour as density
nature. Such never-ending ‘periodic boundary’ con- changes). Not only did the model provide insight
ditions are frequently used in simulations and into the type of collective behaviour seen (Fig. 8.1)
experiments for physical systems (as in particle but it was shown to be quantitatively accurate and
accelerators). In this environment the density of revealed that this behaviour does not rely on leader-
insects can be controlled and its relation to migra- ship, or external organizing forces such as attraction
tory behaviour studied. to extrinsic features of the environment, or naviga-
In order to quantify the individual and group-level tion up environmental gradients, or changes in indi-
motion, computer vision software was developed vidual behaviour; the sudden direction changes
 C A S E S T U D I ES 123

a b c
“Gas-like” disordered motion Collective motion with Strong and unidirectional
intermittent changes in direction collective motion
1 1 1
0.8 0.8 0.8
0.6 0.6 0.6
0.4 0.4 0.4
EXPERIMENT

0.2 0.2 0.2

Alignment

0 0
–0.2 –0.2 0
–0.4 –0.4 –0.2
–0.6 –0.6 –0.4
–0.8 –0.8 –0.6
–1 –1 –0.8
0 100 200 300 400 0 100 200 300 400 0 100 200 300 400
Time (min) Time (min) Time (min)

1.5 1.5 1.5

1.0 1.0 1.0

0.5 0.5 0.5

Alignment
THEORY

0 0 0
–0.5 –0.5 –0.5
–1.0 –1.0
–1.0
–1.5 –1.5
–1.5
0 2000 4000 6000 8000 0 2000 4000 6000 8000 0 2000 4000 6000 8000
Time steps

Figure 8.1 Mass migration in the desert locust. The top panel shows an example frame of locusts in the experimental arena (see main text for details) in
which the positions have been detected by the computer vision software (drawn as crosshairs over the insects) and the properties used for tracking recoded
(here, from top to bottom above each insect, the area and orientation of the animal and how accurately that estimate can be relied on). The lower panels
show the collective motion observed (experiment) and predicted (theory) from the self-propelled particle model. At low density (A) both the experiment and
model show uncoordinated motion represented by the erratic alignment between the insects. Alignment values close to 1 and −1 represent strong
coordinated clockwise and counter-clockwise motion, respectively (see main text for more information). At intermediate densities (B) the group exhibits
‘intermittent’ motion with sudden transitions between highly-ordered clockwise and counter-clockwise states. Above that density, (C) the group randomly
selects a direction and continues in that direction for the entire 8 hour duration of the experiment.

seen in natural hopper bands can result from inter- neighbours, individuals changed their intrinsic
nally driven stochastic (random) effects. motion characteristics in response to how well
In a subsequent analysis of these experimental aligned their neighbours were. Modelling of the
data by Yates et al. (2009), it was shown that, process was also critical to this study: The authors
although collective behaviour in locusts can be used experimental data to link the individual-based
attributed to their tendency to align with near (Lagrangian) approach with a continuum-based
124 M I G R AT I O N Q UA N T I F I E D : C O N S T R U C T I N G M O D E L S A N D L I N K I N G T H E M W I T H D ATA

population-level (Eulerian) approach in which they individuals are, in effect, on a ‘forced march’—stop,
considered motion and interactions between indi- or move perpendicular to the low, and you risk
viduals as approximating diffusion and advection being cannibalized. Leave the swarm and you risk
processes. Speciically, they assumed that migrating starvation and a high predation risk. Individual-
individuals in a swarm have a tendency to exhibit based models have improved our understanding of
random ‘diffusive’ motion, and simultaneously a how such local interactions scale to mass migration.
tendency to align with, and thus to move in the Romanczuk et al. (2009) simulated the insects’
same direction as, others—a form of ‘advection’ or escape and pursuit responses in the face of canni-
‘drift’. This technique is based on the well-studied balism and showed that, even without an explicit
Fokker–Planck equation from physics, where it is alignment term, such behaviours result in coordi-
used to explore the collective properties of gases nated mass migration under a wide range of
and liquids. conditions.
Using this approach it was revealed that locusts This ‘forced march’ mechanism is also prevalent
do not have a ixed random (stochastic or noisy) in other mass migrating insects such as Mormon
component to their motion. Instead, the degree of crickets, Anabrus simplex (Simpson et al. 2006). The
randomness depends on the degree of order of the ‘autocatalytic’ mutual activation between insects,
local swarm. The insects appear to increase the combined with the risk of being poorly aligned with
degree of random motion as the swarm becomes cannibalistic neighbours, results in migrating bands
disordered. Hence, trajectories become more erratic of both locusts and crickets moving over very long
as the swarm loses coherence of motion. Intuitively, distances in relation to body-length (up to 1 km per
one may think that this would result in the swarm day) and these paths tend to be highly linear, inter-
breaking apart and the locusts’ motion becoming spersed with sudden changes in direction. Such
uncoordinated. Remarkably, it has the opposite motion characteristics probably play an important
effect; it allows the system to ‘forget’ its previous role in allowing the insects to search for unpredict-
state once order begins to be lost (to a suficient ably located food patches. In addition, by travelling
degree) and thus not go through a disruptive and with conspeciics, locusts in bands can persist (by
long transition in which individuals show a conlict feeding on others) through very unfavourable envi-
in desired directions of travel. ronments. Thus multiple selection pressures play a
Even though insects align only with near neigh- role in shaping migratory strategies in such insects
bours, swarms can maintain coherent motion over (Bazazi et al. 2008).
length-scales that are many orders of magnitude Swarming insects also beneit from ready access
greater—such as tens of kilometres. Local align- to mates when breeding and lay eggs in close prox-
ment tendency is suficient to explain the motion imity to one another in so-called ‘egg ields’. Females
characteristics of migrating groups, but why do die shortly after egg laying and one inal adaptation
locusts (and other swarming insects) move together to migratory life is to prepare the offspring for life
in this way? It may, at irst, appear like a coopera- as either a solitary, asocial insect, or to be ready to
tive collective decision to leave a nutrient poor area be part of a swarm. If gregarized herself, a female
and to search more widely for food. However, the secretes a speciic chemical into her egg foam, caus-
underlying mechanism is more selish. Food short- ing the biochemical cascade that results in the onset
age, speciically scarcity of protein and salt, results of gregarization even before birth (Simpson and
in insects turning to each other as the only source of Miller 2007). This transgenerational transfer of
these essential nutrients. Bazazi et al. (2008) demon- information means migratory swarms can persist
strated that locusts are highly cannibalistic and that across many generations and consequently plagues
individuals face a risk of being damaged, and even can persist for months or years.
entirely consumed, by conspeciics in mobile Individual-based models of animal movement
groups. Furthermore, they are speciically sensitive are increasingly used to offer insight into vertebrate
to the approach and contact of others from behind mass migration (Conradt et al. 2009; Couzin and
and begin to move in response to these stimuli. The Krause 2003; Couzin et al. 2005; Couzin and Laidre
 C A S E S T U D I ES 125

2009). For example, the tendency for individuals to or top-down) approaches can also be used to make
align their direction of travel with neighbours natu- inferences about the drivers of movement and
rally allows individuals to access information about migration. Three examples of this approach can be
the environment beyond their own direct interac- found in studies of ungulate movement and migra-
tion range. Just as locusts interacting over the order tion in the Serengeti ecosystem of East Africa (Boone
of centimetres can form coordinated mobile groups et al. 2006; Fryxell et al. 2004; Holdo et al. 2009b). With
that can extend over kilometres, migrating ish, the population-level approach, the entire population
ungulates and birds can create large mobile aggre- is subdivided into smaller groups or units (herds,
gates. Information about a change in direction can individuals, occupants of a particular spatial loca-
percolate through such groups very rapidly. For tion) and these are then redistributed across the
example, the detection of a predator by only a small landscape according to a set of movement rules. The
subset of individuals in animal groups can result in new spatial distribution of the population can be
an amplifying wave of turning that spreads faster compared with observed spatial patterns to assess
than the maximum speed of the individuals them- goodness-of-it, for example using maximum likeli-
selves, and this has been observed in insects hood methods. In two of the Serengeti examples
(Treherne and Foster 1981), ish (Radakov 1973) and (Fryxell et al. 2004; Holdo et al. 2009b), the movement
birds (Potts 1984). In such circumstances individu- ‘rule’ consists of a non-linear emigration function
als can respond to environmental change (in this for the probability of departure from a lattice cell.
example the sudden appearance of a threat) even if The function evaluates the value of a particular
they don’t detect it themselves. resource (e.g. energy intake) in the currently-occu-
From the perspective of mass migration, this gen- pied cell in relation to its mean value in the sur-
eration of an ‘effective’ range of interaction with the rounding landscape. Groups of animals departing
environment, which can vastly exceed an individu- each cell are then redistributed in neighbouring cells
al’s own direct perception of the environment, has in proportion to the abundance of the key resource.
the potential to be very important. Many migrations Although this movement model does not necessar-
depend on individual organisms responding to ily result in an ideal free distribution (IFD; for exam-
noisy, long-range and luctuating gradients, be they ple because movement and perception are assumed
thermal fronts, resources, or diffusive gradients, or to be local rather than global, and therefore less than
the Earth’s magnetic ield. By coordinating motion, ideal), it has in common with IFD theory and the
organisms effectively form a self-organized ‘sensor- related marginal value theorem (MVT) the notion
array’—each individual sensory unit (individual) is that animal distributions and movement decisions
subject to error, but collectively, by integrating their incorporate density-dependence. Foraging animals
own estimates with the local direction choices of exert density-dependent feedbacks on their food
neighbours, aggregates are able to detect and move resources and this effect can be modelled explicitly
up gradients that are dificult, or even impossible, (Fryxell et al. 2004; Holdo et al. 2009a).
to detect from an individual’s perspective A different population-level approach was
(Grunbaum 1998a; Torney et al. 2009). This high- employed by Boone et al. (2006) to infer the drivers
lights the possibility that this type of information of the Serengeti wildebeest migration. They used
transfer may be ubiquitous in mass migrating spe- evolutionary programming (EP) to simulate the
cies and may play an important role in aiding emergence of alternative migration trajectories,
migration. and resource maps were used as selective ilters to
simulate the emergence of the most well-adapted
migration routes. This approach is top-down
8.5.2 A population-level approach: ungulate
because the migration pathways are not observed,
migration in the Serengeti ecosystem
but rather are used to generate population-level
In addition to individual-based (‘Lagrangian’ or distribution maps that can be compared with the
bottom-up) approaches based on relocation data data. Unlike the approach adopted by Fryxell et al.
from individual animals, population-level (‘Eulerian’ (2004) and Holdo et al. (2009a), density-dependence
126 M I G R AT I O N Q UA N T I F I E D : C O N S T R U C T I N G M O D E L S A N D L I N K I N G T H E M W I T H D ATA

does not affect migration in this model. In both information about resource availability (Couzin
approaches, however, competing models consist of et al. 2005; Grunbaum 1998b; Hancock and Milner-
alternative linear combinations of resources (e.g., Gulland 2006).
water versus forage, or rainfall vs. NDVI) that are
used to generate alternative population-level dis-
8.5.3 A hybrid approach
tribution maps, which are then compared with
observed distributions to ind the model structures In light of the potential limitations of both individ-
and parameters that maximize the it between ual- and population-based approaches to making
models and data. inferences about movement behaviour, it is clear
A potential shortcoming of the population-level that an ideal approach would combine information
approach is that it cannot distinguish speciic indi- from both levels. In rare cases, researchers are able
viduals or groups of animals. Models that adopt this to tag very large numbers of individuals and follow
approach may be better described as ‘redistribution their movements in real time. We provided an
or lux’ models than ‘movement’ models. With relo- example in Section 8.5.1, in which video capture
cation data, whether based on capture–release or and specialized software enabled the researchers to
radio-tracking data (see Section 8.3.1), true move- gather data on individual movement behaviour in
ment trajectories can be inferred for speciic indi- its social context and to draw striking insights into
viduals. With a population-level approach, the links between rule-based movement decisions
movement rules must be inferred from shifts in spa- at the individual level and collective movement
tial distributions but, as the example in Fig. 8.2 behaviour. Most often, these data are lacking, but
shows, when the identities of individuals are when population-level census data are available in
unknown, alternative movement models can explain conjunction with relocation data for individual ani-
any given population-level shift in spatial distribu- mals, the potential exists for a hybrid approach to
tion. In the absence of any external environmental drawing inferences about the mechanics of move-
information, models A and B in Fig. 8.2 are equally ment and migration (see also Mueller and Fagan
probable, but imply quite different movement 2008).
behaviours. Even when environmental data are Another example of a combined approach is pro-
readily available, some movement parameters may vided in Sibert and Fournier (2001) for application
be unidentiiable in the absence of individual-level in isheries. Dart tags provide large amounts of data
information (see Section 8.4.2 for a discussion of the at the population level, but they do not provide
identiiability problem). tracking data on the movement behaviour of indi-
The individual-based approach has the advan- vidual animals; data are often available from ‘archi-
tage of allowing inferences to be made on a wider val’ tags (Sibert 2001). The two data types can be
range of processes inluencing movement behav- combined into a single likelihood framework by
iour, including the role of memory versus resource exploiting the known relationship between the
tracking (Dalziel et al. 2008; Mueller and Fagan 2008; advection–diffusion equation (used to model move-
van Moorter et al. 2009) and shifts between discrete ment at the population level) and the biased ran-
movement regimes or modes (Jonsen et al. 2006; dom walk (used to model individual tracks). Also,
Morales et al. 2004; Patterson et al. 2008) that would within a likelihood framework, a more general
be dificult or impossible to infer from population- approach might consist of an iterative method that
level data. At the same time, the bottom-up jointly maximizes the likelihood of the modelled
approach, because of the nature of the data, must distributions given the observed distributions
often do without the population-level information (Fryxell et al. 2004) and the likelihood of the mod-
that modulates movement behaviour, not only in elled displacement kernels given the observed tra-
the context of density-dependence and its effects on jectories (Dalziel et al. 2008). Both components of
forage abundance and predation risk (Hancock and the model share environmental data, but they also
Milner-Gulland 2006), but also because it may con- feed off each other. Such an approach is appealing
vey critically important social cues including because it can lead to stronger inference about
 C O NC LU S I O N S 127

t1 t2

(a)

(b)

Figure 8.2 Drawing inferences about individual movement behaviour from population-level distributions. The spatial coniguration shown at t1 shifts to a
new distribution at t2. Because individual animals have not been tagged, we do not know the individual movement trajectories that led to this distributional
shift. Out of many possibilities, we show two candidates, with mean displacement distances for the three individuals in question of 1.2 (case A) and 1.7
(case B) cell widths.

competing movement models than either the wider context of animal movement models (concep-
individual- or population-based methods would in tual as well as mathematical). Hence, we discussed
isolation. the ecological mechanisms that could spontaneously
A rule-based bottom-up model of movement have given rise to migration-like patterns of space
behaviour can result in a hierarchy of predictions at use and showed that there are several potential ways
different spatial scales. For a set of candidate mod- in which migratory patterns can emerge from the
els containing assumptions about how individuals interaction within and between groups of animals
respond both to their immediate physical environ- and their environment. On this basis alone, we may
ment and to population-level variables such as con- argue that observed migratory patterns do not have
speciic density, population distribution patterns a single cause and neither will the original cause
and individual movement tracks could be used to necessarily remain observable. We highlighted that
discard candidate models that it the data poorly at migration is best seen as lying on a continuum from
either or both spatial scales (Grimm et al. 2005). sedentary to nomadic movement patterns and not
as a clearly distinct movement behaviour.
Second, models of migration must be conditioned
8.6 Conclusions
on data about the movement of animals. We dis-
Here we have discussed how models, combined cussed the characteristics of different types of move-
with modern data sources and statistical methods, ment data and highlighted the importance of data
can be used to test different hypotheses about the quality. The processes giving rise to migration are
causes of migratory behaviour. First we argued that multiple and complex and the methods used to
to detect the drivers of migration, it is useful to start observe migration impose several constraints. Both
by translating contrasting hypotheses into models of of these facts may make it dificult to identify unique
movement in order to identify which hypotheses explanations for observed patterns.
can actually generate migratory patterns. We briely Third, we discussed the use of inferential meth-
outlined the controversies regarding the deinition ods to evaluate the support from the data for differ-
of migration and argued that to resolve such disa- ent models. On the technical side, model criticism,
greements it is necessary to place migration in the model understanding and conidence building are
128 M I G R AT I O N Q UA N T I F I E D : C O N S T R U C T I N G M O D E L S A N D L I N K I N G T H E M W I T H D ATA

fundamental components of modern statistical and migration and we discussed three studies
modelling and take considerably more time to com- using this approach to understand ungulate migra-
plete than model itting. More generally, inference tion in the Serengeti ecosystem of East Africa. With
is composed of three main aspects: identiication of a population-level approach, competing movement
migratory patterns, estimation of migration param- rules are used to generate alternative population-
eters and selection of migration models. Novel ana- level distribution maps, or alternative migration
lytical developments may address these three trajectories, which are then compared with
aspects separately but, since they are all non-inde- observed distributions to ind the most appropriate
pendent parts of inference, they are best performed combinations between models and data. However,
as part of the same inferential approach. Further- as inference is based on shifts in spatial distribu-
more, given the complexity of the migration proc- tions of individuals, in the absence of any external
ess, and the dificulties in collecting movement data environmental information or of individual-level
for migrants, in many cases it will be impossible to information, alternative movement models may
identify unique explanations for observed patterns explain any given population-level shift in spatial
based on location data alone. This problem of non- distribution. On the other hand, it is more dificult
identiiability cannot be solved by increasing sam- for individual-based models to use important pop-
ple size but, rather, by collecting additional ulation-level (large scale) information such as pop-
independent information (e.g. individual state or ulation density, forage abundance and predation
behaviour). Thanks to the current revolution in bio- risk. Thus, a combined approach would be desira-
logging techniques this should be increasingly ble and we discuss two case studies where this has
achievable. allowed the drawing of striking insights into the
Finally, we illustrated the links between models, links between rule-based movement decisions at
data and inference using three case studies. Mass the individual level and collective movement
migration in the desert locust provided a clear behaviour.
example of the power of combining individual- In conclusion, the increasing availability of
based models build on a solid theory (in this case, detailed movement data, combined with advances
interacting particle systems developed in thermo- in computing power and statistical software, has
dynamics) with well-designed experiments on a led to a renaissance in migration studies. Making
large number of individuals to identify the causes the best use of these opportunities to understand
of the emergence of migratory patterns. Conversely, the causes of migration involves making choices
population-level approaches can also be used to between a wide variety of different models, data
make inferences about the drivers of movement and types of inference.