Bevan Et Al 2013

Download as pdf or txt
Download as pdf or txt
You are on page 1of 24

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/260984079

Intensities, Interactions, and Uncertainties: Some New Approaches to


Archaeological Distributions.

Chapter · January 2013

CITATIONS READS

47 587

4 authors:

Andrew Bevan Enrico Ryunosuke Crema


University College London University of Cambridge
118 PUBLICATIONS   1,734 CITATIONS    64 PUBLICATIONS   975 CITATIONS   

SEE PROFILE SEE PROFILE

Xiuzhen Janice Li Alessio Palmisano


Emperor Qin Shihuang's Mausoleum Site Museum, China Ludwig-Maximilians-University of Munich
17 PUBLICATIONS   193 CITATIONS    45 PUBLICATIONS   372 CITATIONS   

SEE PROFILE SEE PROFILE

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

MicroPasts View project

All content following this page was uploaded by Alessio Palmisano on 22 March 2014.

The user has requested enhancement of the downloaded file.


Intensities,   interactions   and   uncertainties:   some   new   approaches   to  
archaeological  distributions  
 
Andrew  Bevan,  Enrico  Crema,  Xiuzhen  Li,  Alessio  Palmisano  
(UCL  Institute  of  Archaeology)  
 
Postprint  of  a  chapter  in  Bevan,  A.  and  Lake,  M.  (eds.),  Computational  Approaches  
to  Archaeological  Spaces,  Walnut  Creek:  Left  Coast  Press.  
 
1.  Introduction  
While  distribution  maps  are  nearly  as  old  as  the  discipline  of  archaeology  itself,  
most   archaeologists   still   rely   on   personal   intuition   with   regard   to   their  
assessment   both   of   the   spatial   patterns   they   recover   and   the   environmental  
processes  and  human  behaviours  that  might  be  behind  these  patterns.  To  some  
extent,   this   general   preference   for   intuitive   readings   of   space   in   the  
archaeological   record   probably   reflects   several   decades   of   disillusionment   with  
quantitative  spatial  methods,  after  a  flurry  of  early  interest  during  the  1970s  (e.g.  
Hodder   and   Orton   1976;   Clarke   1977),   and   a   continuing   wish   to   prioritise   the  
study   of   cultural   spaces   as   subjectively   experienced   and   meaningfully  
constituted   by   their   human   inhabitants   (e.g.   Gregory   and   Urry   eds.   1985).  
Interestingly,   even   the   enthusiastic   uptake   of   Geographic   Information   Systems  
(GIS)   from   the   1990s   onwards   did   little   to   change   this   situation   with   regard   to  
spatial  pattern  analysis,  as  most  off-­‐the-­‐shelf  GIS  software  was  targeted  at  data  
management   and   querying,   digital   cartography   and   enhanced   visualisation,   as  
well  as  certain  focused  modelling  agendas  (e.g.  terrain,  visibility  and  movement).  
Effectively,   the   study   of   distribution   maps   in   archaeology   merely   carried   on   as   it  
was,   with   a   healthy   dose   of   expert   intuition,   and   perhaps   in   slightly   richer   visual  
form.    
 
However,  while  human  involvement  in  the  act  of  interpretation  is  undeniably  a  
crucial   and   enduring   aspect   of   archaeological   research,   there   remain   good  
reasons   to   characterise   spatial   distributions   in   more   formal,   quantitative   ways.  
This  paper  focuses  on  a  set  of  point  pattern  and  process  models  that,  we  argue,  
now  puts  archaeologists  in  a  position  to  return  to  the  analysis  of  spatial  pattern  
and  process  with  renewed  ambition,  especially  with  regard  to  distribution  maps.  
The   first   section   below   considers   current   theoretical   approaches   to   point  
distributions   and   subsequent   sections   then   address   three   cases   studies   that  
highlight  some  important  conceptual  issues  and  new  analytical  opportunities.  
 
2.  Theoretical  Perspectives  
2.1  Point-­based  Simplifications  
A  dot  on  a  map  is  usually  a  considerable  simplification.  Whether  our  concern  is  
about   the   proper   2-­‐   or   3-­‐dimensional   representation   of   a   real   world   entity,   its  
more   complicated   expression   in   space-­‐time,   or   the   possible   webs   of   cultural  
meaning   that   might   envelope   it,   we   certainly   risk   much   by   this   kind   of   spatial  
abstraction.   There   are   also   further   trade-­‐offs   to   do   with   how   we   record   such  
points,  between  time  and  effort  on  the  one  hand,  and  any  possible  archaeological  
insights  we  might  derive  on  the  other.  Do  we  require  great  spatial  accuracy  (such  
that  measured  coordinates  are  close  to  the  actual  absolute  values),  great  spatial  
precision  (where  measurements  of  a  given  location  are  highly  reproducible,  but  
not  necessarily  accurate)  or  some  combination  of  the  two?  Furthermore,  points  
can   also   be   thought   of   as   highly   simplified   events   in   space-­‐time:     as   such,   they  
often   involve   only   a   fuzzy   or   very   loose   sense   of   duration   (how   long   an   event  
lasts  for)  and  equivalence  (to  what  extent  an  entity  in  one  time  step  can  still  be  
considered   the   same   entity   in   the   next).   Indeed,   if   we   observe   such   events   in  
traditional   blocks   of   archaeological   time   (e.g.   periods   or   phases),   this   lumping  
procedure   is   a   further   abstraction   with   its   own   additional   methodological  
implications  and  risks.  
 
An   important   initial   stage   of   spatial   (and   spatio-­‐temporal)   analysis   therefore  
involves   deciding   what   kinds   of   simplification   and   trade-­‐off   are   acceptable   for  
what  applications,  as  well  as  how  best  to  make  use  of  the  information  we  already  
have.  
 
2.1  Spatial  Randomness,  Regularity  and  Clustering  
Assuming  for  a  moment  that  point-­‐based  abstractions  are  sometimes  justifiable,  
what  do  we  then  want  to  say  about  such  spatial  distributions?  The  main  formal  
question  in  the  past  has  been  the  degree  to  which  a  distribution  departs  or  not  
from  what  we  might  expect  if  we  simply  scattered  points  at  random  across  the  
study   area.   The   latter   random,   purely   ‘stochastic’,   process   establishes   a  
theoretical   baseline   usually   referred   to   as   ‘complete   spatial   randomness’   or   CSR.  
Spatial  statisticians  also  tend  to  assume  that  the  underlying  process  responsible  
for   generating   this   random   point   pattern   operates   in   roughly   the   same   way  
across  the  whole  study  area  (i.e.  it  is  ‘homogeneous’  and  ‘stationary’1),  and  that  if  
we   were   to   consider   the   number   of   points   falling   in   each   of   a   series   of   similar  
sub-­‐units   across   this   area,   we   would   find   that   their   densities   (the   more   common  
technical   term   is   point   ‘intensity’,   as   used   hereafter)   follow   a   Poisson  
distribution  and  are  said  to  be  a  realisation  of  a  spatial  Poisson  process.  Figure  
1a  depicts  an  example  of  a  random  distribution  of  points  in  a  rectangular  area,  
generated  according  to  a  Poisson  process.  
 

                                                                                                               
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.  
 

View publication stats

You might also like