Bevan Et Al 2013
Bevan Et Al 2013
Bevan Et Al 2013
net/publication/260984079
CITATIONS READS
47 587
4 authors:
Some of the authors of this publication are also working on these related projects:
Sino-British cooperative project; Imperial Logistics; Multidisciplinary approaches to Qin Terracotta Warriors; Bronze weapons; View project
All content following this page was uploaded by Alessio Palmisano on 22 March 2014.
1
In
certain
contexts,
there
are
differences
between
these
two
terms,
but
for
economy
in
the
discussion
below,
they
are
used
interchangeably,
as
are
their
opposites
(inhomogeneity,
heterogeneity
and
non-‐stationarity,
see
below).
One
related
aspect
of
spatial
data
that
does
not
receive
any
attention
here,
but
which
implies
a
limited
form
of
non-‐stationarity,
is
anisotropy
(i.e.
situations
in
which
points
are
found
more
frequently
in
certain
prevailing
directions;
see
Markofsky
and
Bevan
2011
for
archaeological
discussion).
Figure
1.
Hypothetical
examples
of
random,
regular
and
clustered
point
patterns.
In
contrast,
Figures
1b-‐c
depict
two
alternative
patterns
in
which
the
point
distribution
is
(b)
more
‘regular’
(also
often
described
as
‘dispersed’)
or
(c)
more
‘clustered’
(also
often
described
as
‘clumped’
or
‘aggregated’).
Sometimes
such
patterns
are
intuitively
obvious
and
we
could
get
away
without
using
statistics
to
consider
them,
but
often
our
spatial
intuition
is
misleading:
for
example,
some
people
would
suspect
slight
clustering
in
the
figure
1a
(in
fact,
it
is
purely
random),
while
others
might
not
suspect
regularity
in
figure
1b
(in
fact,
there
is
an
arbitrarily
imposed
minimum
distance
between
points).
It
has
therefore
long
been
acknowledged
that
the
role
of
quantitative
spatial
analysis
is
partly:
(a)
to
arbitrate
in
situations
where
spatial
patterns
are
uncertain,
(b)
to
characterise
such
relationships
in
ways
that
are
useful
for
explicit
comparison,
and
(c)
to
offer
a
formal
platform
for
suggesting
possible
processes
and
behaviours
behind
such
spatial
patterns.
In
particular,
we
often
assume
that,
behind
any
patterns
of
regularity
or
clustering
are
also
some
interesting
alternative
processes,
beyond
one
that
is
purely
stochastic
and
Poisson.
Regular
patterns
are
often
thought
to
be
the
result
of
‘inhibition’
processes.
For
example,
for
human
settlements
we
might
think
of
the
way
in
which
the
existence
of
one
settlement
might
inhibit
the
creation
of
another
one
immediately
next
to
it
(e.g.
because
of
competition
over
resources,
see
below).
For
artefact
distributions,
regularity
can
be
generated
by
various
kinds
of
post-‐depositional,
taphonomic
sorting
or
due
to
very
deliberate
human
decisions
about
artefact
placement.
Clustered
patterns,
in
contrast,
are
often
the
result
of
‘attraction’
processes.
We
might
think
of
the
movement
of
people
towards
larger
settlements
because
of
a
variety
of
the
advantages
such
aggregated
locations
might
offer.
For
artefacts,
we
can
think
of
processes
of
discard
and
subsequent
breakage
in
situ
that
encourage
very
clumped
scatters
of
such
finds
in
the
archaeological
record.
2.3
Spatial
Inhomogeneity
So
any
point
pattern
documented
across
a
given
study
area
(of
whatever
archaeological
size,
from
one
observed
under
the
microscope,
to
one
found
on
a
house
floor,
to
one
seen
across
a
whole
landscape)
can
be
thought
of
as
a
realisation
or
one
or
more
underlying
processes
(see
also
O’Sullivan
and
Unwin
2003:
51-‐75).
In
the
simplest
null
case,
a
single
random
Poisson
process
is
involved.
In
other,
still
simple
cases,
a
non-‐random
process
is
at
work,
but
only
one,
with
effects
that
are
homogeneous
across
the
entire
study
area
(even
if
the
pattern
manifests
differently
at
different
spatial
scales,
see
below).
However,
in
many
real
world
examples,
it
is
likely
that
multiple
processes
are
at
work
and/or
that
they
behave
differently
in
different
parts
of
the
study
area
(i.e.
they
can
be
described
as
‘inhomogeneous’,
‘heterogeneous’
and/or
‘non-‐stationary’;
for
an
archaeological
example
with
aggregated
count
data
rather
than
point
patterns,
see
Bevan
and
Conolly
2009).
Given
the
prevalence
of
inhomogeneous
distributions
in
real
life,
it
is
both
theoretically
and
practically
useful
to
distinguish
between
the
‘first-‐order’
and
‘second-‐order’
characteristics
of
a
given
point
pattern
(e.g.
Bailey
and
Gatrell
1995:
32-‐5).
First-‐order
characteristics
are
those
that
describe
the
average
intensity
of
points
across
a
given
region
(if
this
average
intensity
varies
spatially
then
the
point
pattern
can
be
called
inhomogeneous),
and
first-‐order
effects
refer
to
one
or
more
external
processes
or
phenomena
that
encourage
the
intensity
of
points
in
the
study
region
to
vary
at
different
locations.
In
contrast,
the
second-‐order
characteristics
of
a
point
pattern
describe
the
relative
intensity
of
points
as
influenced
by
the
spatial
configuration
of
other
points
in
the
study
area
(i.e.
the
pattern’s
covariance
structure),
and
reflecting
different
kinds
of
internal
interaction
effects
among
points,
such
as
propensities
for
attraction
or
inhibition.
A
basic
lesson
from
many
practical
analyses
is
that
it
is
difficult,
and
often
entirely
misleading,
to
consider
second-‐order
effects
before
properly
accounting
for
first-‐order
effects.
We
return
to
these
issues
in
the
first
two
case
studies
below.
For
now
it
is
simply
worth
noting
that,
we
can
easily
build
simulations
in
which
to
observe
what
kind
of
spatial
pattern
is
produced
by
any
single
realisation
of
a
particular
point
process
of
know
design.
In
real
world
contexts
however,
the
process
that
generated
the
points
is
typically
unknown,
and
the
challenge
becomes
the
degree
to
which
we
can
learn
about
what
the
first-‐
and
second-‐order
effects
might
be
solely
via
analysis
of
the
resulting
pattern.
Ironically,
while
it
is
fair
to
say
that
many
archaeologists
would
be
loosely
and
informally
aware
of
such
complex
spatial
considerations
when
it
comes
to
their
interpretation
of
the
archaeological
record,
the
formal
quantitative
tools
they
have
so
far
used
have
been
stuck
in
some
rather
idealised
and
methodologically-‐quarantined
boxes.
For
example,
at
the
scale
of
landscapes
and
archaeological
sites,
‘predictive
modelling’
(e.g.
Mehrer
and
Wescott
2006;
Verhagen
and
Whitley
2011) 2
has
been
a
commonplace
way
of
assessing
first-‐order
properties,
demonstrating,
for
example,
correlations
between
the
probability
of
discovering
sites
in
a
particular
study
area
and
the
distribution
of
one
or
more
environmental
variables
(e.g.
soils,
slope
steepness,
access
to
water,
etc.).
Conversely,
nearest
neighbour
tests
and
quadrat
counts
have
typically
been
used
to
assess
site
spacings,
with
the
implicit
assumption
that
second-‐order
interaction
effects
are
often
at
work.
However,
rarely
if
ever,
are
these
two
methods
brought
together
to
treat
the
issue
as
an
analytically-‐related
whole.
One
final,
complicating
factor
for
archaeologists
is
the
fact
that
archaeological
observations
are
very
partial,
imperfect
records
of
past
activity.
Much
of
the
variability
in
our
observed
spatial
patterns
in
archaeology
is
due
to
patchy
levels
of
archaeological
preservation
and
investigation.
For
example,
most
site
distribution
maps
are
the
result
of
historically-‐complex
sites
and
monuments
records
or
unsystematic
surveys
–
many
of
the
perceived
clusters
of
observations
are
to
do
with
where
people
have
recently
looked,
where
modern
development
has
recently
exposed
new
archaeology,
etc.
These
issues
can
also
be
conceived
of
as
kinds
of
first-‐order
variation
in
intensity,
but
ideally
we
would
want
to
distinguish
them
from
taphonomic
and
human
behavioural
effects
in
the
past,
ultimately
so
that
we
can
offer
some
useful
archaeological
interpretation.
3.
Multi-scalar
and
Monte
Carlo
Approaches
2
In
archaeology,
this
term
has
developed
an
unnecessarily
narrow
meaning,
related
to
cultural
resource
management
and
models
of
site
location
probabilities.
However,
beyond
archaeology,
it
is
just
a
general
term
for
any
kind
of
model
that
leads
to
explicit
predictions
of
one
kind
or
another.
This
section
moves
from
general
theoretical
considerations
to
explore
the
relevance
of
a
variety
of
recent
methods
for
characterising
point
patterns
and
processes.
It
begins
with
a
hypothetical
example
to
fix
some
ideas
before
considering
a
real,
intra-‐site
case
study.
3.1.
A
Hypothetical
Example
It
may
sometimes
be
very
difficult
to
wholly
separate
complex
first-‐
and
second-‐
order
effects
(some
operating
in
the
past,
some
in
the
present)
in
many
archaeological
datasets,
but
there
remain
many
advantages
to
conceptualising
point
distributions
this
way.
Alongside
the
theoretical
issues
raised
above,
a
whole
host
of
more
advanced
analytical
methods
have
been
discussed
in
the
spatial
statistical
literature
over
the
last
20-‐30
years
(for
a
recent
overview:
Gelfand
et
al.
2010:
263-‐423).
These
have
been
used
in
certain
applied
fields
such
as
astronomy
or
ecology
for
a
long
time,
but
have
been
slow
to
percolate
into
other
disciplines
such
as
archaeology.
Figure
2.
Multi-‐scalar
Monte
Carlo
methods:
(a)
a
point
pattern,
in
a
in
a
notional
10x10m
study
area,
that
is
both
regular
and
clustered
at
different
distances;
(b)
a
pair
correlation
function
identifying
a
switch
from
significant
regularity
to
significant
clustering
at
c.0.25m
(with
the
latter
slowly
tailing
off
thereafter).
The
grey
are
in
b)
is
a
95%
envelope
based
on
999
random
simulations.
Two
key
methodological
advances
over
the
last
35
years
(outside
of
archaeology)
have
been
methods
that:
(a)
deliberately
seek
to
address
point
pattern
and
process
at
several
different
spatial
scales,
and
(b)
employ
a
family
of
randomisation
tests
know
as
Monte
Carlo
simulation
(Robert
and
Casella
2004),
which
leverage
the
speed
of
modern
computational
platforms
to
provide
a
powerful
and
flexible
way
of
testing
spatial
patterns,
particularly
in
cases
where
the
study
area
is
irregular
or
the
underlying
effects
are
complicated
(for
other
archaeological
applications
of
Monte
Carlo
simulation,
see
Fisher
et
al.
1997;
Drennan
and
Peterson
2004;
Crema
et
al.
2010).
Figure
2a,
for
example,
presents
a
toy
example
of
a
point
pattern
produced
by
a
known
process
(see
Lennard-‐
Jones
1924)
in
a
notional
10x10m
study
area
which
leads
to
(a)
a
strong
tendency
for
regular
spacing
over
very
short
distances
(up
to
0.25m
in
this
case),
but
thereafter
also
(b)
a
further
tendency
for
clustering
at
medium
distances
which
gradually
tails
off
to
a
random
pattern
at
larger
ones.
In
this
case,
no
first-‐
order
effects
are
present
and
the
process
operates
in
a
uniform
way
across
the
whole
hypothetical
study
area.
To
what
extent
however,
can
we
find
methods
that
correctly
identify
these
different
scales
of
second-‐order
effect
based
solely
on
analysis
of
the
resulting
site
distribution?
If,
for
example,
we
calculate
a
traditional
nearest
neighbour
index
(Clark
and
Evans
1954;
Hodder
and
Orton
1976:
38-‐51)
that
has
been,
for
better
or
worse,
the
bread-‐and-‐butter
of
archaeological
point
pattern
analysis
for
many
years,
it
misleadingly
suggests
that
the
pattern
is
random
or
only
very
slightly
clustered
(r=0.91).
There
are
however
more
recent
spatial
statistical
methods
that
consider
multiple
scales
of
second-‐order
patterning
and
explore
how
likely
or
unlikely
they
are
to
have
occurred
by
chance
(for
a
technical
overview,
see
Gelfand
et
al.
2010:
263-‐
423).
Perhaps
the
most
common
of
these
is
the
K
function
and
its
more
readable,
slightly
transformed
version,
the
L
function
(originally
Ripley
1977;
and
for
some
exploratory
archaeological
uses,
Orton
2004;
Bevan
and
Conolly
2006;
Vanzetti
et
al.
2010).
Here
we
emphasise
another
related
method,
the
pair
correlation
function
(PCF),
which
is
less
well
known,
but
arguably
more
useful
in
many
circumstances
(several
similar
functions
go
by
other:
see
Ilian
et
al.
2010:
218-‐23;
Wiegand
and
Maloney
2004,
Perry
et
al.
2006).
A
PCF
measures
the
intensity
of
points
in
donut-‐shaped
rings
(annuli)
around
each
point
and,
as
such,
is
not
a
cumulative
statistic
in
the
same
way
as
a
K
or
L
function
(the
latter
two
effectively
measure
the
intensity
of
points
in
ever
expanding
circles
that
include
all
previous,
smaller
ones).
Figure
2b
shows
PCF
results
for
the
simulated
point
pattern.
The
x-‐axis
measures
the
separation
distance
between
points
and
the
observed
results
are
presented
as
a
black
line.
This
observed
result
begins
well
below
the
theoretically
random
threshold
of
y=1,
indicating
the
possible
regularity
of
this
pattern
at
short
distances,
then
climbs
well
above
this
threshold,
indicating
medium
distance
clustering,
before
dropping
slowly
back
down
towards
y=1.
For
a
variety
of
reasons
however,
this
theoretical
y=1
threshold
is
often
an
unreliable
baseline,
and
it
is
more
useful
to
use
Monte
Carlo
methods
that
offer
an
‘envelope’
of
possible
values
that
we
might
expect
under
a
null
model
in
which
the
point
process
generating
this
pattern
is
assumed
to
be
wholly
random.
This
is
done
by
repeatedly
generating
sets
of
an
equivalent
number
of
random
points,
and
then
plotting
maximal
and
minimal
PCF
values
at
each
distance
range.
In
the
case
of
figure
2b,
the
grey
shaded
area
marks
out,
not
the
full
range
of
random
PCF
values,
but
an
envelope
enclosing
the
middle
95%
of
PCF
values
from
999
simulation
runs.3
Where
the
real,
observed
values
are
larger
than
this
envelope,
the
observed
pattern
can
be
considered
clustered
at
that
distance,
whereas
where
they
fell
below
the
envelope,
they
are
more
likely
to
be
regularly
spaced.
In
this
example,
the
PCF
successfully
and
accurately
documents
the
shift
from
3
For
simplicity
and
consistency
in
each
of
the
analyses
developed
in
this
paper,
we
have
run
999
Monte
Carlo
simulations,
have
combined
these
with
the
observed
values
and
have
then
taken
the
25th
and
975th
ranked
values
to
define
the
borders
of
the
envelope
depicted
in
each
plot.
At
first
glance,
it
might
seem
as
if
these
envelopes
could
be
treated
as
also
defining
a
0.05
significance
level,
but
in
fact
this
is
potentially
misleading
for
tests
that
consider
multiple
critical
values
simultaneously.
Alternative
envelope
calculations
that
do
produce
exact
significance
envelopes
are
feasible
(e.g.
see
?envelope
in
the
R
spatstat
package),
but
are
more
complicated
to
implement
consistently
across
the
different
methods
used
here,
so
they
have
not
been
included.
significantly
regular
to
clustered
effects
(the
critical
feature
in
this
kind
of
plots
is
usually
the
point
of
inflection
at
about
y=1,
e.g.
here
at
ca.
0.25m,
rather
than
the
top
or
bottom
of
observed
humps
in
the
PCF),
and
thereafter
the
slow
tailing
off
of
this
clustering
until
the
pattern
becomes
wholly
random.
In
fact,
the
non-‐
cumulative
nature
of
the
PCF
offers
certain
advantages
over
K
or
L
functions
for
analysing
patterns
with
these
kinds
of
multiple
scales
and
different
kinds
of
interaction,
although,
in
general,
such
methods
offer
complementary
perspectives.
3.2.
Crossbow
Triggers
and
Qin
Terracotta
Warriors
Of
course,
real
archaeological
distributions
rarely,
if
ever,
manifest
themselves
as
such
completely
recovered,
simply
bounded
datasets.
The
first
of
our
three
archaeological
case
studies
therefore
explores
some
of
these
analytical
issues
as
they
arise
at
the
intra-‐site
scale.
The
tomb
complex
of
the
first
Chinese
emperor,
Qin
Shihuang
(259-‐210
BC),
is
famous,
amongst
other
things,
for
its
pits
of
life-‐
sized,
terracotta
warriors,
buried
in
battle
formation,
with
full
military
equipment.
As
an
example,
we
can
consider
the
distribution
of
bronze
crossbow
triggers
(this
being
the
only
part
of
the
crossbow
that
survives
archaeologically)
that
were
found
alongside
the
warriors
in
the
easternmost
part
of
pit
1.
A
plot
of
these
artefacts
against
the
warriors
(Figure
3)
makes
it
clear
that
the
overall
crossbow
trigger
distribution
itself
is
strongly
clustered
in
space
due
both
to
the
shape
of
the
corridors
and
the
nature
of
terracotta
army’s
battle
formation
(with
crossbowmen
only
in
certain
parts
of
the
army,
particularly
along
the
flanks).
This,
in
and
of
itself,
is
probably
not
something
we
need
to
assess
via
more
complicated
statistical
treatment,
but
in
passing,
it
is
worth
noting
that
the
corridors
represent
a
very
irregular
study
area
that
raises
some
complicated
issues
to
do
with
‘edge
effects’
(e.g.
how
we
handle
the
fact
that
the
annuli
must
be
truncated
to
reflect
the
fact
that
points
cannot
fall
beyond
the
corridor
area,
as
well
as
the
possible
inaccuracies
that
arise
from
our
lack
of
knowledge
about
areas
immediately
west
of
the
excavated
portion
of
pit
1).
For
our
purposes
here,
however,
a
key
interest
is
not
how
to
characterise
the
overall
pattern
of
triggers,
but
how
we
might
modify
the
methods
introduced
above
to
consider
patterning
amongst
different
sub-‐groups
of
crossbow
trigger.
More
precisely,
metrical,
typological
and
materials
analysis
of
the
triggers
has
been
able
to
distinguish
subtle
but
undeniably
different
trigger
sub-‐groups
that
suggest
the
existence
of
different
weapon-‐casting
moulds,
different
metallurgical
workshops
and/or
different
organisational
practices.
The
difficult
question
therefore
becomes:
how
do
we
assess
the
spatial
distribution
of
the
trigger
sub-‐
groups
while
controlling
for
the
overriding
spatial
structure
of
trigger
distribution
in
general?
Figure
3.
Intra-‐site
spatial
distributions:
(a)
Qin
terracotta
warriors
(grey
squares)
and
bronze
crossbow
triggers
(black
circles)
in
the
easternmost
parts
of
pit
1,
(b)
a
photograph
of
a
bronze
cross-‐bow
trigger.
A
good
example
is
the
trigger
sub-‐group
shown
in
figure
4a.
This
is
a
group
that,
when
studied
in
detailed,
exhibits
small
but
distinct
morphological
and
typological
differences
from
other
triggers.
To
recap,
when
we
focus
on
the
possible
spatial
patterning
of
this
sub-‐group,
we
clearly
want
to
control
for
the
spatial
structure
of
the
triggers
as
a
whole
(and
by
extension
the
formations
of
crossbowmen).
To
do
so,
we
run
a
Monte
Carlo
simulation
in
which
the
triggers
attributed
to
this
particular
sub-‐group
are
assigned
at
random
amongst
the
overall
trigger
assemblage.
In
fact,
the
group
2
triggers
in
the
pit
are,
themselves,
visibly
clustered,
beyond
the
pattern
imposed
by
the
battle
formation
(Figure
4a)
and,
again,
there
may
not
be
a
need
for
a
formal
method
to
recognise
it
in
this
case.
However,
it
is
useful
to
consider
this
particularly
clear-‐cut
example
as
a
proof
of
concept,
and
in
the
knowledge
that
such
standardised
evaluation
will
be
far
more
important
in
other
less
obvious
cases.
Figure
4b
shows
a
pair
correlation
function
in
which
this
clustering
is
very
evident
in
the
observed
result
substantial
deviation
above
the
95%
envelope.
More
precisely,
the
plot
indicates
particularly
strong
clustering
of
this
sub-‐group
up
to
distances
of
perhaps
3-‐4m
radius
and
then
up
to
7-‐8m,
with
further
possible
clustering
at
much
larger
distances.
There
are
some
interesting
processes
that
are
likely
to
be
behind
such
clustered
patterns
of
trigger
sub-‐groups
in
the
pit.
For
example,
they
may
reflect
different
workshops
producing
marginally
different
crossbow
triggers
and
procedures
for
the
storage
and
placement
of
the
crossbows
in
the
pit
in
batches
(e.g.
zones
of
the
pit
that
were
equipped
with
crossbows
in
one
go).
Applied
more
broadly
to
other
trigger
sub-‐groups,
other
weapon
types
and
other
artefacts
in
the
pit,
such
analyses
can
begin
to
map
out
coherent
activity
spaces
and
explore
how
consistent
they
were
in
size,
arrangement
etc.
(see
especially,
Li
2012;
Martinón-‐Torres
et
al.
in
press).
Figure
4.
Spatial
analysis
of
trigger
groups:
(a)
group
2
triggers
shown
as
triangles
and
the
others
as
grey
dots;
(b)
a
pair
correlation
function
(observed
values
in
black
and
95%
critical
envelope
in
grey)
4.
Inhomogeneous
Point
Process
Models
The
above
case
study
demonstrates,
via
a
deliberately
straightforward
example,
that
such
methods
can
formalise
our
assessment
of
spatial
patterns
at
multiple
scales,
even
in
the
presence
of
other
confounding
spatial
factors
(such
as
the
shape
of
the
corridors
and
the
clustering
of
crossbow
triggers
as
a
whole).
In
their
original
form,
methods
such
as
K,
L
or
pair
correlation
functions
were
not
easily
applied
to
these
kinds
of
inhomogeneous
and
edge-‐affected
cases,
but
such
problems
are
now
becoming
increasingly
tractable.
The
second
case
study
considered
here,
explores
the
potential
of
such
inhomogeneous
approaches
for
assessments
of
site
location
at
the
landscape
scale.
It
considers
some
Iron
Age
I
(ca.
12th-‐11th
centuries
BC)
settlements
documented
by
fairly
systematic
surface
survey
in
the
central
part
of
the
West
Bank
(modern-‐day
Israel
and
Palestinian
Territories,
for
the
survey,
see
Finkelstein
and
Magen
1993;
Finkelstein
and
Lederman
1997).
In
particular,
we
focus
on
an
area
of
hilly
dolomite
upland
of
some
766
sq.km,
across
which
99
sites
of
this
period
have
been
documented
(Figure
5a).
This
choice
of
area
is
deliberate:
it
reduces
the
range
of
complicating
factors
that
need
to
be
considered
below,
both
because
it
was
investigated
in
a
fairly
even
way
by
a
single
archaeological
project,
and
because
it
covers
an
area
of
generally
consistent
underlying
geology
(with
some
knock-‐on
implications
for
soils,
topography,
etc).
Figure
5.
Iron
Age
I
sites
in
the
central
West
Bank
and
four
possible
first-‐order
covariates:
(a)
elevation
(light
to
dark
ranges
from
135-‐1010m
ASL),
(b)
average
annual
rainfall
(dark
to
light
ranges
from
c.335-‐720mm),
(c)
ridge
landforms
(darker
is
more
likely
to
be
geomphorphometrically
classified
as
a
ridge),
(d)
topographic
wetness
index
summed
over
a
local
neighbourhood
(darker
is
wetter),
and
(e)
a
prediction
surface
based
on
the
three
significant
covariates
(darker
is
higher
point
intensity).
A
quick
visual
inspection
of
figure
5a
suggests
informally
both
that
there
might
be
a
first-‐order
trend
towards
slightly
greater
densities
of
settlement
at
higher
elevations,
and
also
that
there
might
conceivably
some
regular
spacing
to
some
of
the
settlements.
We
can
therefore
build
some
formal
point
process
models
to
consider
whether
environmental
affordances
such
as
elevation
are
indeed
significant,
and
above
and
beyond
this,
whether
there
is
yet
a
further
second-‐
order
propensity
for
the
location
of
one
settlement
to
inhibit
the
location
of
another
nearby.
We
begin
by
considering,
as
examples,
four
related
environmental
affordances
–
elevation,
average
annual
rainfall,
ridge-‐top
landforms,
and
topographic
wetness
in
a
local
catchment
(Figures
5a-‐d).4
This
4
The
digital
elevation
model
(DEM)
used
here
is
NASA’s
90m
SRTM
dataset
(Jarvis
et
al.
2008).
The
rainfall
data
has
interpolated
from
50mm
contours
of
average
annual
precipitation
(the
original
contours
are
courtesy
of
the
GIS
Center,
Hebrew
University
of
Jerusalem).
Ridge-‐like
landforms
were
defined
from
the
DEM
via
a
fuzzy
feature
classification
across
focal
filter
scales
from
3x3
to
11x11
cells
(Fisher
et
al.
2004).
Catchment-‐based
topographic
wetness
was
calculated
via
focal
filtering
of
a
standard
topographic
wetness
index
surface
(itself
derived
from
the
DEM)
in
a
way
that
summed
all
values
within
a
circular
neighbourhood
of
2.5km
radius
(about
half
an
hour’s
walk
and
a
common
threshold
for
daily
travel
budgets).
selection
is
prompted
in
part
by
many
commentators’
informal
impressions
that
rugged
topography
and
hydrology
were
important
factors
behind
settlement
locations
in
this
region
and
period,
for
a
variety
of
practical
reasons
(e.g.
Zertal
1988;
Gibson
2001;
in
fact
many
other
possible
covariates
have
been
explored
but
are
not
considered
here).
Univariate
regression
of
binned
versions
of
each
of
these
variables
against
site
intensity
(figures
6a-‐d)
suggests
that
rainfall
is
not
a
particularly
good
predictor
of
the
intensity
of
sites
across
the
landscape,
but
that
the
other
three
variables
have
significant
positive
correlations
(p<0.05
or
better).
In
other
words,
sites
are
more
common
a)
at
higher
elevations,
b)
where
the
landforms
are
more
ridge-‐like
in
shape,
and
c)
where
patches
of
ground
offer
better
access
to
surface
water
or
soil
moisture.
Informally,
it
appears
that
ridge-‐
like
locations
might
be
the
most
influential
of
these.
Figure
6.
Univariate
correlations
between
site
intensity
and
a)
elevation
above
sea-‐level
(r2=0.72),
(b)
average
annual
rainfall
(r2=0.23),
(c)
ridge
landforms
(r2=0.82),
and
(d)
topographic
wetness
index
summed
over
a
local
neighbourhood
(r2=0.54).
The
intensities
for
a,
b
and
d
have
been
summarised
in
decile
bins
of
the
covariate,
while
for
c,
the
x-‐axis
probabilities
are
discontinuous
due
to
the
nature
of
the
fuzzy
geomorphometric
classification
used.
These
univariate
regressions
provide
a
guide
to
likely
relationships
between
site
intensity
and
various
first
order
effects
(note,
as
above,
that
we
have
already
removed
other
possible
first
order
effects
by
choosing
a
study
area
that
is
broadly
one
type
of
geology,
and
that
has
been
surveyed
by
one
field
project
with
fairly
consistent
methods).
If
we
now
run
a
multivariate
regression
and
select
the
best
combination
of
these
four
variables
via
stepwise
comparison
(minimising
an
Akaike
Information
Criterion),
we
find
that
rainfall
is
excluded
as
we
might
expect,
that
the
other
three
variables
are
all
significant
(p<0.05
or
better),
and
that
this
new
model
with
a
first-‐order
trend
is
substantially
more
effective
that
a
null,
random
hypothesis.
We
can
then
create
a
predicted
first-‐
order
intensity
surface
(figure
5e)
that
can
be
used
to
return
to
the
question
of
second-‐order
interactions
in
a
more
complete
way.
First,
as
a
point
of
contrast,
it
is
worth
considering
a
pair
correlation
function
of
the
settlements
sites,
along
with
an
envelope
of
wholly
random
Monte
Carlo
simulations
(figure
7a).
The
observed
PCF
shows
something
very
close
to
regularity
at
shorter
distances
of
up
to
ca.1km,
although
the
Monte
Carlo
envelope
suggests
that
this
might
be
of
marginal
significance.
If
we
then
look
at
a
simple
histogram
of
nearest
neighbour
distances
(figure
7b),
we
can
see
a
spike
at
just
over
1,000-‐1,250m
and
a
Monte
Carlo
95%
envelope
suggests
that
this
pattern
is
unlikely
to
occur
by
chance
(for
this
method
with
histograms
of
nearest
neighbour
distances,
see
also
Wilson
and
Melnick
1990).
In
other
words,
the
nearest
neighbour
histogram
provides
a
slightly
more
discerning
picture
of
very
short
distance
patterning
that
confirms
the
evidence
for
regular
spacing
that
was
initially
visible
in
the
PCF
(the
same
observation
is
valid
for
other
multi-‐
scalar
functions
such
as
K
and
L
as
well).
At
this
stage
however,
we
cannot
be
sure
whether
such
regular
spacing
has
been
induced
by
the
spatial
structure
of
some
important
external
influence
on
site
location
(e.g.
evenly
spaced
ridge-‐
tops)
or
is
due
to
internal
processes
that
inhibited
settlements
being
located
close
to
one
another
(e.g.
competition
over
resources).
Figure
7c
seeks
to
tease
apart
the
relative
contribution
of
first-‐
and
second-‐order
effects
by
showing
the
same
nearest
neighbour
histogram
for
the
observed
values,
but
this
time
with
a
simulated
envelope
conditioned
on
the
spatially
inhomogeneous
intensities
predicted
by
our
first-‐order
covariates.
Put
plainly,
the
Monte
Carlo
sets
of
points
are
now
forced
to
respect
the
spatial
inhomogeneity
modelled
by
the
predicted
intensity
surface
(figure
5e).
As
mentioned
above,
we
might
conceivably
anticipate
that
one
or
more
of
the
first
order
variables,
such
as
evenly-‐spaced
ridgelines,
might
have
accounted
for
some
of
the
short
distance
regularity
in
settlement,
but
this
is
not
the
case.
Of
course,
it
is
possible
that
this
continuing
regularity
simply
means
that
some
important
environmental
covariate
has
not
been
considered,
but
nevertheless,
we
are
now
at
least
moving
closer
towards
ruling
this
possibility
out.
Figure
7.
Point
process
models
and
goodness-‐of-‐fit:
(a)
a
pair
correlation
function
of
the
observed
sites
with
a
95%
envelope
from
wholly
random
Poisson
process,
(b)
a
histogram
of
nearest
neighbour
distances
with
a
95%
envelope
from
wholly
random
Poisson
process,
(c)
a
histogram
of
nearest
neighbour
distances
with
a
95%
envelope
also
conditioned
on
the
first-‐
order
covariates
model,
(d)
a
histogram
of
nearest
neighbour
distances
with
a
95%
envelope
also
conditioned
on
both
the
first-‐order
covariates
and
a
second-‐order,
area-‐interaction
model.
Finally
therefore,
we
can
explore
the
goodness-‐of-‐fit
(via
informal
visual
comparison
here,
though
more
formal
statistical
treatment
is
also
possible)
of
an
explicit
model
of
what
might
be
causing
these
second-‐order
point
interactions.
Perhaps
the
most
relevant
one
is
Baddeley
and
van
Lieshout’s
“area-‐interaction
model”
(1995)
that
generates
patterns
of
inhibition
and
clustering
with
reference
to
a
defined
circular
neighbourhood
around
each
point.
The
implicit
idea
of
this
model
–
that
points
have
notional
territories
of
influence
around
them
–
is
obviously
attractive
given
our
understanding
of
how
many
human
settlements
work.
We
can
draw
upon
our
knowledge
of
the
observed
spacing
between
settlements
and
set
the
parameters
to
suggest
a
radius
for
the
interaction
neighbourhood
of
just
over
655m
radius
(half
the
median
nearest
neighbour
distance),
and
inhibitive
effects
that
are
very
strong
but
not
absolute
within
this
zone.
These
parameters
lead
to
sites
spacings
that
are
often
twice
the
neighbourhood
radius
and
which
often
suggest
formal
or
informal
village
catchments
of
less
than
135
ha,
with
such
a
scale
being
not
unreasonable
given
evidence
for
fairly
small
Iron
I
community
sizes
of
dozens
to
no
more
that
a
couple
of
hundred
people
in
this
area.
Figures
7d
demonstrates
that
the
histogram
of
observed
nearest
neighbour
distances
now
falls
within
the
Monte
Carlo
envelope.
As
we
discuss
below,
other
explanatory
models
might
conceivably
offer
better
or
equivalent
fits,
but
by
narrowing
down
the
range
of
possibilities
in
this
formal
manner,
we
clarify
our
thinking
about
what
might
be
plausible
kinds
of
causal
phenomena
in
a
very
useful
way.
5.
Models
with
Temporal
Uncertainty
Our
final
case
study
is
again
one
focused
on
settlements
and
landscapes,
but
with
a
greater
emphasis
on
diachronic
comparison
in
the
presence
of
uncertain
dating.
Temporal
uncertainty
is
an
elephant
in
the
room
of
much
archaeological
interpretation.
It
is
a
near
ubiquitous
feature
of
archaeological
datasets,
whether
these
are
radiocarbon
dates,
geoarchaeological
deposits
or
individual
artefacts.
There
is
insufficient
room
to
discuss
this
topic
at
length,
but
for
our
purposes
here,
one
primary
risk
in
the
spatial
analysis
of
point
distributions
is
that
they
might
reflect
a
chronological
palimpsest
that
thwarts
our
ability
to
unpick
single-‐
period,
contemporary
point
patterns.
This
is
also
a
topic
to
which
many
archaeologists
have
discretely
turned
a
blind
eye:
for
example,
it
is
no
more
than
a
convenient
analytical
assumption
that
all
99
of
the
Iron
I
settlements
in
the
central
West
Bank
study
area
discussed
above
were
inhabited
at
exactly
the
same
time
during
that
phase
which
spans
a
couple
of
centuries
(indeed
a
few
of
the
unusually
clumped
sites
in
the
eastern
part
of
the
study
area
might
be
seasonally
occupied).
In
other
cases,
the
chronological
range
of
the
sites
under
investigation
is
even
broader
and
the
risk
of
drawing
misleading
conclusions
is
correspondingly
greater.
This
is
especially
true
with
regard
to
the
assessment
of
second
order
effects
and
the
processes
that
lead
to
them.
Michael
Barton
(this
volume)
nicely
outlines
an
example
of
a
regularly
spaced
pattern
of
small
sites
in
north-‐central
Arizona
that
might
be
due
to
patterns
of
shifting
clearance,
short-‐
term
cultivation
and
abandonment,
in
which
many
of
the
sites
involved
belonged
to
the
same
broad
period,
but
might
not
be
strictly
contemporary.
The
implied
processes
of
interaction
are,
in
this
case,
quite
different
and
the
discussion
at
the
end
returns
to
the
well-‐known
problems
of
equi-‐final
models
that
this
raises.
5.1
Aoristic
Methods
One
way
to
engage
more
effectively
with
temporal
uncertainty
is
for
us
to
make
the
best
of
all
our
available
temporal
information,
however
fuzzy.
Occasionally,
we
can
define
an
explicit
probability
distribution
that
suggests
how
likely
it
is
for
an
event
is
to
have
occurred
at
a
certain
stage
in
time
based
on
a
range
of
sources
of
knowledge
(e.g.
diagnostic
artifacts,
clear
stratigraphic
relationships,
absolute
radiocarbon
dates,
etc.).
Even
so,
such
results
rarely
produce
a
normal
distribution
that
can
be
conveniently
summarised
by
a
single
summary
value
and
confidence
interval.
Instead,
we
are
more
likely
to
have
the
kinds
of
irregular
probability
distribution
often
produced,
for
example,
via
Bayesian
modeling
of
calibrated
radiocarbon
dates
and
associated
soft
information.
More
importantly,
such
information-‐rich
cases
are
rare:
in
most
instances,
we
can
only
suggest
a
very
approximately
bounded
‘time-‐span’
within
which
the
event
is
likely
to
have
occurred
and,
within
this,
assume
a
simple,
uniform
probability
distribution
(i.e.
implying
that
an
event
has
a
similar
chance
of
having
occurred
at
any
stage
within
the
time-‐span).
‘Aoristic’
analysis
is
an
approach
that
provides
a
way
of
quantifying
these
temporal
uncertainties
and
incorporating
them
into
subsequent
analysis
(particularly
in
the
case
of
simple
timespans
where
we
assume
uniform
probablities).
It
was
initially
developed
in
criminology
(Ratcliffe
2000),
and
has
subsequently
also
been
adopted
by
some
archaeologists
for
looking
at
both
individual
artefacts
and
larger
archaeological
sites
(Johnson
2004,
Crema
et
al.
2010;
Pentedeka
et
al.
2010).
Figure
8:
Temporal
uncertainty
in
point
patterns.
The
left
panel
depicts
a
simple
hypothetical
point
pattern
and
(above
this)
our
degree
of
temporal
knowledge
about
each
of
the
six
point
events
(A-‐F)
over
five
time-‐blocks
(t1-‐t5).
The
grey
horizontal
bars
represent
the
time-‐
span
of
each
event,
showing
that
point
event
D
has
low
uncertainty
(the
point
event
belongs
exclusively
to
time-‐block
t1)
and
event
B
has
higher
uncertainty
(the
point
event
belongs
somewhere
between
time-‐blocks
t3
and
t5).
The
three
panels
in
the
the
middle
show
three
possible
realisations
of
the
actual
temporal
(middle
panel)
and
spatio-‐temporal
patterns
(right
panel,
for
time-‐blocks
t1
and
t2).
More
precisely,
given
a
specific
set
of
points
with
their
temporal
probability
distributions,
there
will
be
a
limited
number
of
possible
spatio-‐temporal
patterns
that
might
actually
have
arisen.
Instead
of
ignoring
this
uncertainty
and
producing
a
single,
but
misleading,
spatial
analysis,
we
can
generate
different
possible
spatial
patterns
based
on
these
temporal
probabilities
and
then
obtain
a
distribution
of
the
more
and
less
likely
results.
Figure
8
is
a
schematic
representation
of
both
the
problem
and
the
possible
solution:
take,
in
this
case,
six
point
events
(A-‐F)
that
each
occurred
in
one
of
five
time-‐steps
(t1-‐t5),
but
can
often
only
be
ascribed
archaeologically
to
wider
time-‐spans
(i.e.
figure
8,
left).
Each
of
the
actual
scenarios
in
the
middle
and
right
hand-‐panels
of
this
figure
are
possible
realizations
of
the
pattern,
amongst
many
others,
given
the
state
of
our
temporal
knowledge.
The
only
way
to
explore
what
possible
spatial
patterns
might
really
have
been
present
is
therefore
to
analyse
a
whole
host
of
possible
realisations
and
explore
if
there
are
any
first-‐
or
second-‐order
spatial
properties
that
persistently
crop
up.
If
for
instance,
90%
of
all
the
possible
point
patterns
are
spatially
clustered,
we
will
have
a
relatively
high
confidence
that
the
observed
pattern
was
indeed
clustered.
However,
while
from
a
theoretical
standpoint,
it
might
be
tempting
to
consider
each
and
every
possible
spatio-‐temporal
configuration,
in
practical
terms
this
is
computationally
prohibitive
as
the
number
of
possible
scenarios
is
often
intractable.
The
alternative
however
is
simply
to
sample
a
finite
number
of
possible
realisations
via
Monte
Carlo
simulation
and
calculate
the
frequency
of
certain
spatial
patterns
in
the
results
(see
Crema
et
al.
2010
for
further
details;
and
Izquierdo
et
al.
2009
for
a
similar
perspective).
5.2
Middle
Jomon
Settlement
Our
final
case
study
combines
aoristic
and
Monte
Carlo
methods
to
consider
settlement
patterns
amongst
the
Jomon
hunter-‐gatherers
of
central
Japan.
In
a
sense,
it
offers
an
ideal
case
for
tackling
the
issue
of
temporal
uncertainty
because,
while
careful
pottery
study
and
an
amazingly
dense
number
of
emergency
excavations
(Habu
2004;
Kobayashi
2008)
provides
one
some
of
the
most
detailed
spatial
distributions
and
relative
chronologies
known
for
any
prehistoric
complex
hunter-‐gather
groups
worldwide,
it
remains
true
that
some
Jomon
pithouses
and
broader
settlements
can
be
ascribed
to
only
fairly
broad
chronological
rnages
whilst
others
can
be
dated
far
more
accurately.
Moreover,
Jomon
settlement
patterns
exhibit
some
interesting
possible
patterns
that
may
relate
to
changing
demography,
social
practices
and
subsistence
strategies.
For
example,
several
studies
(e.g.
Imamura
1996)
have
indicated
a
sudden
rise
in
the
overall
number
of
Jomon
residential
units
during
the
first
part
of
the
Middle
Jomon
period
(ca.3530-‐2470
cal.
BC),
followed
by
a
rapid
collapse
after
few
centuries.
Some
authors
explain
such
dynamics
as
due
to
increasingly
intensive
use
of
certain
plant
resources
during
the
early
Middle
Jomon
that
made
it
possible
to
maintain
higher
population
densities,
but
which
became
more
problematic
during
a
subsequent
climatic
cooling
phase
in
late
Middle
Jomon
which
may
have
led
to
a
reduction
in
the
overall
availability
of
these
resources
(Imamura
2002,
Habu
2008).
In
terms
of
spatial
patterning,
most
commentators
agree
that
there
were
larger,
more
nucleated
settlements
prior
to
the
proposed
population
collapse
and
more
dispersed,
smaller
settlements
after
it.
While
such
a
broad
dichotomy
seems
plausible
for
the
Middle
Jomon,
it
remains
difficult
to
consider
tempos
of
change
over
smaller
timescales
or
to
compare
these
processes
with
those
in
earlier
or
later
periods
of
Japanese
prehistory.
Aoristic
analysis
and
Monte
Carlo
simulation
can
however
provide
a
good
analytical
framework
for
such
research.
Figure
9:
a)
Distribution
of
excavation
units
containing
at
least
one
pithouse
attributed
to
Early
to
Late
Jomon
(filled
and
hollow
circles)
and
locations
where
at
least
one
pithouse
with
a
probability
of
existence
higher
zero
at
one
of
the
two
temporal
blocks
examined
is
present
(filled
circles).
b)
Number
of
PCFs
with
significant
clustering
(above
the
horizontal
line)
and
significant
dispersion
(below
the
horizontal
line)
for
time
blocks
2950-‐2850
cal.
BC
(grey
shaded
bars)
and
2750-‐2650
cal.
BC
(hollow
bars)
with
1000
simulated
spatio-‐temporal
patterns.
The
case
study
area
for
the
analysis
that
follows
is
located
on
the
western
side
of
Tokyo
Bay,
near
the
modern
city
of
Chiba.
We
have
chosen
an
arbitrary
square-‐
shaped
area
of
15x15
km
within
which
120
separate
open-‐area
excavations
have
documented
some
1418
Jomon
pithouses
that
can
each
be
attributed
to
somwhere
between
the
Early
and
Late
Jomon
period
(ca.
5050-‐1270
cal.
BC;
figure
8a).
For
each
of
these
pithouses,
we
can
define
a
more
precise
time-‐span
of
existence
from
the
description
of
pottery
and
associated
artefacts
available
from
excavation
reports.
We
can
then
spit
the
whole
5050
to
1250
cal.
BC
timeframe
into
arbitrary
chronological
blocks
of
100
years
each
and
calculate
the
probability
that
a
given
pithouse
actually
exists
during
that
block.
For
our
purposes
here,
we
compare
two
distinct
time
blocks
–
2950-‐2850
cal.
BC
during
a
phase
of
population
increase,
and
2750-‐2650
cal.
BC
at
the
observed
peak
in
population.
For
each
of
these
two,
we
then
generate
1000
Monte-‐Carlo
simulations
of
possible
spatio-‐temporal
patterns.
In
these
simulations,
and
again
for
the
purposes
of
this
example,
we
designate
a
new
settlement
to
exist
at
a
certain
location,
if
an
excavated
area
at
that
spot
is
allocated
at
least
one
pithouse
for
that
chronological
block.
The
resulting
simulations
provide
a
series
of
realized
settlement
patterns:
in
order
to
explore
differences
in
the
respective
spatial
distributions
for
each
period,
we
computed
a
series
of
PCFs.
Just
as
in
our
earlier
case
studies,
each
observed
PCF
can
be
compared
to
an
envelope
generated
from
999
Monte
Carlo
sets,
each
with
an
identical
number
of
points
and
each
only
allowed
to
exist
within
the
120
parent
locations
where
Early
to
Late
Jomon
pithouses
has
been
actually
been
recovered.
This
method
of
constrained
randomisation
is
similar
to
the
one
adopted
for
the
case
study
on
bronze
crossbow
triggers,
and
allows
us
to
account
both
for
the
patchy
nature
of
modern
excavation
and
for
some
general
first-‐order,
locational
choices
the
Jomon
may
have
had
over
the
long
term
(the
latter
being
of
great
general
interest
of
course,
but
not
for
our
analysis
here).
Figure
9b
shows
the
frequency
of
instances
among
the
1000
simulated
spatio-‐
temporal
patterns
where
the
observed
results
depart
from
the
95%
envelope
of
random
values.
During
the
earlier
phase
of
population
increase
(2950-‐2850
cal
BC)
settlements
appear
to
be
aggregated
to
varying
degrees
over
distances
up
to
600
or
even
1,000m
radius
(and
indeed,
whether
we
see
this
as
an
observation
about
inter-‐settlement
clustering
or
simply
large
extensive
settlement
areas
is
partly
just
a
question
of
semantics).
At
much
larger
spatial
scales,
there
are
a
few
instances
of
dispersed
patterns,
but
these
are
fairly
rare,
with
an
exception
at
ca.
4,700-‐4,900m,
where
about
70%
of
the
simulated
patterns
showed
dispersion.
Also
notice
how,
at
some
spatial
scales,
(e.g.
at
ca.1,500m),
the
number
of
instances
of
clustering
and
dispersion
are
roughly
equal,
implying
that
our
levels
of
available
information
are
insufficient
to
draw
any
robust
conclusions.
A
few
hundred
years
later,
when
the
population
size
reached
its
peak
(2750-‐2650
cal
BC),
the
settlement
pattern
is
notably
different.
Firstly,
the
short-‐distance
clustering
is
still
present,
but
now
only
occurs
of
a
much
smaller
range
(<300m
radius).
This
indicates
that
clusters
of
nucleated
settlements
may
have
started
to
decrease
in
their
sizes,
an
idea
that
is
supported
also
by
a
small
peak
of
dispersed
patterns
at
ca.
1,000m.
Secondly
there
is
strong
evidence
for
patterns
of
dispersion
(i.e.
regular-‐spacing)
at
separation
distances
of
c.2,500
meters.
Thirdly
at
ca.4,900
there
is
a
relatively
high
number
of
simulations
(>95%)
that
lead
to
aggregated
patterns,
the
opposite
result
to
the
one
seen
in
the
previous
time-‐block.
Clearly,
the
patterns
for
the
two
different
time
periods
are
not
the
same,
with
the
former
being
characterized
by
a
greater
nucleation
and
the
latter
possibly
by
greater
dispersion
at
medium
distances
and
aggregation
at
higher
distances
(i.e.
broad
clumps
of
settlement
activity
with
intervening
spaces
of
some
5km
between
these).
While
we
still
need
to
treat
such
results
cautiously,
the
analysis
suggest
that
the
beginnings
of
the
dispersed
patterning
know
to
be
present
in
anger
by
the
mid
3rd
millenium
BC
is
already
visible
during
the
period
of
peak
settlement,
reflecting
possible
early
instances
of
group
fission
that
have
been
plasuible
argued
as
being
driven
by
the
diminishing
availability
of
local
food
resources.
6.
Discussion
The
theoretical
discussion
and
three
case
studies
above
should
convey
the
degree
to
which
it
is
worth
reinvesting
in
the
formal
modelling
of
point
patterns
and
processes
in
archaeology.
In
any
case,
from
this
discussion
we
can
draw
out
several
practical
conclusions:
a) Careful
definition
of
a
study
area
is
important.
b) Histograms
of
the
distribution
of
nearest
neighbour
distances
offer
a
fairly
robust
way
of
exploring
short-‐range
regularities
in
point
spacing,
and
can
be
made
more
robust
as
a
confirmatory
method
via
Monte
Carlo
simulation.
Traditional
Clark
and
Evans
tests
are
far
less
discerning.
c) Multi-‐scalar
methods
such
as
K,
L
and
pair
correlation
functions
are
potentially
useful
for
understanding
second-‐order
interactions,
but
are
inappropriate
on
their
own
if
there
are
grounds
for
thinking
the
patterns
exhibit
spatial
inhomogeneity.
d) We
can
use
multivariate
regression
models
(in
a
similar
manner
to
established
practices
in
archaeological
predictive
modelling)
to
provide
first
order
measures
of
the
varying
intensity
of
points
across
a
study
area
as
influenced
by
a
range
of
external
variables.
These
can
then
offer
a
platform
from
which
to
consider
second-‐order
interactions
via
the
above
multi-‐scalar
methods
even
where
spatial
inhomogeneity
in
present.
e) The
kinds
of
temporal
uncertainty
present
in
most
archaeological
datasets
can
be
successfully
addressed
in
spatial
analysis
by
adopting
a
probabilistic
and
Monte
Carlo
framework.
It
is
worth
ending
this
chapter
by
revisiting
one
well-‐known
criticism
of
formal
approaches
to
spatial
analysis
and
modeling
in
archaeology.
A
common
suggestion
is
that
such
efforts
are,
at
best,
frustrating
and,
at
worst,
have
little
interpretative
value
(e.g.
Hodder
1977),
because:
a)
several
different
modelled
processes
can
sometimes
be
shown
to
produce
the
same
or
similar
outcomes
(i.e.
they
are
equi-‐final
or
convergent)
and
b)
the
same
model
can
sometimes
be
shown
to
lead
to
quite
different
outcomes
depending
on
its
exact
starting
conditions
or
the
role
played
by
random
chance
(i.e.
it
is
multi-‐final
or
divergent).
However,
while
these
challenges
should
encourage
us
to
avoid
statements
that
imply
a
cast-‐iron
certainty
about
causal
relationships,
they
should
not
dissuade
us
from
trying
to
model
them
at
all:
a
smaller
set
of
equally
plausible
models
is
still
better
than
a
situation
in
which
anything
goes
(see
also
Premo
2010).
In
this
sense,
the
above
discussion
has
led
us
from
a
traditional
quantitative
emphasis
on
simply
‘rejecting
null
hypotheses’
towards
one
in
which
greater
emphasis
is
placed
on
comparing
the
fit
of
a
series
of
potentially
plausible
explanations
(i.e.
in
general
sympathy
with
a
maximum
likelihood
approach).
In
other
words,
Monte-‐Carlo
methods
allow
us
to
embrace
equifinality
by
weighting
alternative
hypothesis
in
probabilistic
terms.
This
approach
should
be
welcomed
as
more
in
tune
with
interpretative
archaeologies
of
landscape
that
often
look
positively
on
the
co-‐existence
(temporary
or
otherwise)
of
rival
explanations:
there
is
no
reason
that
a
similarly
healthy
lack
of
certainty
should
not
be
welcomed
for
quantitative
approaches
as
well.
Acknowledgements
The
crossbow
trigger
case
study
derives
from
a
much
wider
collaboration
between
UCL
Institute
of
Archaeology
and
the
Museum
of
the
Terracotta
Army,
and
our
particular
thanks
to
Marcos
Martinón-‐Torres
and
Thilo
Rehren
and
the
Terracotta
Army
Museum
staff.
The
Iron
I
settlement
survey
dataset
is
one
provided
by
the
West
Bank
and
East
Jerusalem
Archaeological
Database
(digitallibrary.usc.edu/wbarc/),
with
slight
modifications
by
Palmisano.
We
would
like
to
thank
Adi
Keinan
(UCL
Institute
of
Archaeology)
for
her
help
with
this
and
related
West
Bank
data,
as
well
as
Adi
Ben-‐Nun
(GIS
Center,
Hebrew
University
of
Jerusalem)
for
allowing
us
to
make
use
of
the
rainfall
contour
dataset.
The
dataset
for
the
Jomon
case
study
at
Chiba
has
been
obtained
from
the
excavation
reports
kindly
made
available
by
the
Cultural
Properties
Centre
of
Chiba
Prefecture
Education
Foundation.
The
analysis
discussed
here
was
conducted
primarily
in
R
(R
Core
Development
Team
2011),
with
some
pre-‐
processing
in
GRASS
GIS
(GRASS
Development
Team
2008)
and
Landserf
(Wood
2009).
Our
thanks
to
Adrian
Baddeley
for
advice
on
point
process
models
and
the
use
of
the
spatstat
package
(Baddeley
and
Turner
2005),
as
well
as
to
Mark
Lake
and
two
anonymous
reviewers
for
comments
on
a
chapter
draft.
References
Baddeley,
A.
J.
and
van
Lieshout,
M.
N.
M.
1995.
Area-‐Interaction
Point
Processes,
Annals
of
the
Institute
of
Statistical
Mathematics
47.4:
601-‐619.
Baddeley,
A.
J.
and
Turner,
R.
2005.
spatstat:
An
R
Package
for
Analyzing
Spatial
Point
Patterns,
Journal
of
Statistical
Software
12.6:
1-‐41.
Bailey,
T.
and
Gatrell,
T.,
1995.
Interactive
Spatial
Data
Analysis.
Harlow:
Longman.
Bevan,
A.
n.d.
Archaeological
Sites
and
a
Return
to
the
Humble
Point
Pattern,
(paper
under
review)
for
Deweirdt,
E.
and
J.
Bourgeois
(eds.)
Spatial
Analysis
Applied
to
Archaeological
Sites
from
Protohistory
to
the
Roman
Period.
Ghent.
Bevan,
A.
and
Conolly,
J.
2006.
Multi-‐scalar
Approaches
to
Settlement
Pattern
Analysis,
in
Lock,
G.
and
Molyneaux
B.
(eds.)
Confronting
Scale
in
Archaeology:
Issues
of
Theory
and
Practice:
217-‐234.
New
York:
Springer.
Bevan,
A.
and
Conolly,
J.
2009.
Modelling
Spatial
Heterogeneity
and
Nonstationarity
in
Artifact-‐Rich
Landscapes,
Journal
of
Archaeological
Science
36.4:
956-‐964.
Clarke,
D.
(ed.)
1977.
Spatial
Archaeology.
Boston:
Academic
Press.
Crema,
E.R.,
Bevan,
A.
and
Lake,
M.,
2010.
A
probabilistic
framework
for
assessing
spatio-‐temporal
point
patterns
in
the
archaeological
record,
Journal
of
Archaeological
Science
37.5:
1118-‐1130.
Drennan,
R.
D.
and
Peterson,
C.
E.,
2004,
Comparing
archaeological
settlement
systems
with
rank-‐size
graphs:
a
measure
of
shape
and
statistical
confidence,
Journal
of
Archaeological
Science,
31.5:
533-‐549.
Finkelstein,
I.
and
Lederman,
Z.
1997.
Highlands
of
Many
Cultures.
The
Southern
Samaria
Survey.
Jerusalem:
Graphit
Press.
Finkelstein,
I.
and
Magen,
Y.,
1993.
Archaeological
Survey
of
the
Hill
Country
of
Benjamin.
Jerusalem:
Israel
Antiquities
Authority.
Fisher,
P.,
Wood,
J.
and
Cheng,
T.
2004.
Where
is
Helvellyn?
Fuzziness
of
multi-‐
scale
landscape
morphology,
Transactions
of
the
Institute
of
British
Geographers
29.1:
106-‐128.
Gibson,
S.,
2001.
Agricultural
Terraces
and
Settlement
Expansion
in
the
Highlands
of
Early
Iron
Age
Palestine:
Is
There
Any
Correlation
Between
The
Two?
In
Mazar,
A.
(ed.),
Studies
in
the
Archaeology
of
the
Iron
Age
in
Israel
and
Jordan.
Sheffield:
Sheffield
Academic
Press.
Gelfand,
A.
E.
and
Diggle,
P.
J.
and
Fuentes,
M.
and
Guttorp,
P.
2010.
Handbook
of
Spatial
Statistics,
London:
CRC/Taylor
and
Francis.
GRASS
Development
Team
2008.
Geographic
Resources
Analysis
Support
System
(GRASS)
Software.
Open
Source
Geospatial
Foundation
Project.
URL:
http://grass.osgeo.org
Gregory,
D.
and
J.
Urry
(eds.)
1985.
Social
Relations
and
Spatial
Structures.
London:
Macmillan.
Habu,
J.,
2004.
Ancient
Jomon
of
Japan,
Cambridge:
Cambridge
University
Press.
Habu,
J.,
2008.
Growth
and
decline
in
complex
hunter-‐gatherer
societies:
a
case
study
from
the
Jomon
period
Sannai
Maruyama
site,
Japan,
Antiquity
82:
571-‐584.
Hodder,
I.
and
C.
Orton
1976.
Spatial
Analysis
in
Archaeology,
Cambridge:
Cambridge
University
Press.
Hodder,
I.
1977
Spatial
studies
in
archaeology,
Progress
in
Human
Geography
1:
33-‐64.
Imamura,
K.,
1996.
Prehistoric
Japan:
New
Perspectives
on
Insular
East
Asia
London:
UCL
Press.
Imamura,
K.,
2002.
Jomon
no
yutakasa
to
genkai.
Tokyo:
Yamakawa.
Izquierdo,
L.
R.,
Izquierdo,
S.
S.,
Galán,
J.
M.
and
Santos,
J.
I.,
2009.
Techniques
to
Understand
Computer
Simulations:
Markov
Chain
Analysis,
Journal
of
Artificial
Societies
and
Social
Simulations
12.
URL:
http://jasss.soc.surrey.ac.uk/12/1/6.html
Jarvis,
A.,
Reuter,
H.I.
Nelson,
A.
and
E.
Guevara,
2008.
Hole-filled
SRTM
for
the
globe
Version
4,
available
from
the
CGIAR-CSI
SRTM
90m
Database.
URL:
http://srtm.csi.cgiar.org
Johnson,
I.,
2004.
Aoristic
Analysis:
seeds
of
a
new
approach
to
mapping
archaeological
distributions
through
time,
in
Stadtarchäologie
Wien
Magistrat
der
Stadt
Wien
(ed.)
Enter
the
Past.
The
E-way
into
the
Four
Dimensions
of
Cultural
Heritage
(CAA
2003):
448–452.
Oxford:
Archaeopress.
Kobayashi,
K.,
2008.
Jomonjidai
no
rekinendai.
In:
Kosugi,
Y.,
Taniguchi,
Y.,
Nishida,
Y.,
Mizunoe,
W.
and
Yano,
K.
(ed.)
Rekishi
no
monosashi:
:
Jomon
jidai
kenkyu
no
hennen
taikei,
257-‐269.
Tokyo:
Douseisha.
Lennard-‐Jones,
J.E.
1924.
On
the
determination
of
molecular
fields.
Proceedings
of
the
Royal
Society
of
London
A
106:
463–477.
Li,
X.
2012.
Standardisation
and
Labour
Organisation
of
the
Bronze
Weapons
for
the
Qin
Terracotta
Warriors,
China
(PhD
thesis,
University
College
London)
Markofsky,
S.
and
A.
Bevan
2011.
Directional
analysis
of
surface
artefact
distributions.
A
case
study
from
the
Murghab
Delta,
Turkmenistan,
Journal
of
Archaeological
Science
39.2:
428-‐439.
Martinón-‐Torres,
M.,
Li,
X.,
Bevan,
A.
Xia,
Y..
Zhao,
K.
and
T.
Rehren
(in
press)
Forty
thousand
arms
for
a
single
emperor:
from
chemical
data
to
the
labour
organization
behind
the
bronze
arrows
of
the
Terracotta
Army,
Journal
of
Archaeological
Method
and
Theory.
Mehrer,
M.
W.
and
Wescott,
K.
(eds.)
2006.
GIS
and
Archaeological
Predictive
Modeling,
London:
CRC/Taylor
and
Francis
O'Sullivan,
D.
and
D.
Unwin
2003
Geographic
Information
Analysis,
Hoboken:
Wiley.
Pentedeka,
A.,
E.
Kiriatzi,
L.
Spencer,
A.
Bevan,
J.
Conolly
2010.
From
Fabrics
to
Island
Connections:
Macroscopic
and
Microscopic
Approaches
to
the
Prehistoric
Pottery
of
Antikythera,
Annual
of
the
British
School
at
Athens
105:
1-‐81.
Premo,
L.,
2010.
Equifinality
and
Explanation:The
Role
of
Agent-‐Based
Modeling
in
Postpositivist
Archaeology.
In:
Costopoulos,
A.
and
Lake,
M.
(ed.)
Simulating
Change:
Archaeology
into
the
Twenty-First
Century.
Salt
Lake
City:
University
of
Utah
Press,
28-‐37.
R
Development
Core
Team
2011.
R:
A
Language
and
Environment
for
Statistical
Computing,
Vienna:
R
Foundation
for
Statistical
Computing.
URL:
http://www.R-‐project.org/
Ratcliffe,
J.H.
2000.
Aoristic
analysis:
the
spatial
interpretation
of
unspecified
temporal
events,
International
Journal
of
Geographical
Information
Science
14.7:
669-‐679.
Robert,
C.
P.
and
Casella,
G.,
2004.
Monte
Carlo
Statistical
Methods
(2nd
Ed.).
New
York:Springer.
Vanzetti,
A.,
Vidale,
M.,
Gallinaro,
M.,
Frayer,
D.W.
and
Bondioli,
L.
2010.
The
iceman
as
a
burial,
Antiquity
84:
681-‐692.
Verhagen,
P.
and
Whitley,
T.G.
2011
Integrating
Archaeological
Theory
and
Predictive
Modeling:
a
Live
Report
from
the
Scene,
Journal
of
Archaeological
Method
and
Theory
(online
first
copy).
Wilson,
S.
M.
and
Melnick,
D.
J.
1990.
Modelling
Randomness
in
Locational
Archaeology,
Journal
of
Archaeological
Science
17.4:
403-‐412
Wood,
J.
2009
LandSerf:
A
Geographic
Information
System
for
the
Visualization
and
Analysis
of
Surfaces,
London
City
University
URL:
http://www.landserf.org/
Zertal,
A.,
1988.
The
water
factor
during
the
Israelite
Settlement
Process
in
Canaan.
In
Heltzer
M.
and
Lipinski
E.
(eds),
Society
and
Economy
in
the
Eastern
Mediterranean
(c.
1500-1000
BC),
341-‐52.
Leuven:
Peeters.