BEYOND HOT SPOTS: using space syntax to understand dispersed patterns of crime risk in the built environment.

1
BEYOND HOT SPOTS using space syntax to
understand dispersed patterns of crime risk in
the built environment.

• Bill Hillier
• Ozlem Sahbaz
• Bartlett School of Graduate Studies
• University College London
• b.hillier_at_ucl.ac.uk
• o.sahbaz_at_ucl.ac.uk

2

• The spatial analysis of crime patterns has
  usually been conceived of in terms of the
  analysis of clusters of crime occurrence, or hot
  spots. However, as Newman observed (Newman 1972
  p 109), occurrence is not the same as risk. For
  example, a busy street may have a higher number
  of street crimes that a quiet street, and so will
  appear as a hot spot, even though much higher
  movement rates may mean that the risk to
  individuals is lower than in locations with less
  crime (Alford 1999). We may say that the police
  need to know about occurrence, in order to deploy
  resources, but potential victims need to know
  about risk to work out strategies of avoidance.
• The spatial pattern of risk will often take the
  form of a dispersed pattern of types of location,
  where local conditions facilitate one kind of
  crime or another, rather than a set of spatial
  clusters. Different kinds of crime are easier in
  different spatial conditions residential
  burglary benefits from unsurveyed access, street
  robbery prefers a supply of victims one at a time
  and so on. Understanding the spatial pattern of
  risk often means understanding how and where such
  location types are distributed in the network of
  public space.
• A possible tool for the analysis of such
  dispersed patterns is space syntax, a network
  approach to city spatial form which characterises
  spatial locations numerically, in ways which have
  been shown to reflect other kinds of patterns in
  the built environment such as movement flows
  (Hillier Iida 2005) and land use types and
  mixes (Hillier 2000). So using this approach
  might also allow us to relate patterns of crime
  more precisely to these other aspects of city
  form and life (Hillier 2004. Hillier Shu 2000,
  Hillier Sahbaz 2005).
• Here we show how the use of space syntax in the
  study of a large data base of residential
  burglary and street crime, against a background
  of demographic, socio-economic and physical data,
  can bring to light some unexpected relations
  between the physical, spatial and social
  characteristics of the built environment on the
  one hand, and the spatial patterns of these crime
  types on the other.

3

• What then is space syntax ? It works by first
  reducing the street plan of the city to a least
  line map made up of the fewest lines that cover
  the system. (Turner et al 2004) recently showed
  that if this map is converted into a graph,
  treating lines as nodes and intersections and
  links (the inverse of the usual practice) there
  is probably a unique least line graph for any
  reasonable urban system.
• We then convert the Ieast line map into a graph
  of the segments of lines between intersections,
  assigning three different weights to the
  relations between adjacent segments the distance
  between the centres of segments, whether or not
  there is a directional change to another line,
  and the degree of angular change.
• We then calculate closeness and betweenness
  values for each segment with all three weightings
  separately.This allows us to describe the network
  in terms of shortest paths, fewest turns paths
  and least angle change paths from each segment to
  all others, or if you prefer in terms of its
  metric, topological and geometric structure.
• We also vary the radius from each segment up to
  which the measures are calculated, up to the
  whole system, defining radius also in terms of
  metric distance, numbers of turns and degree of
  angular change.

4

• We then colour the segments from red for
  strongest through to blue for weakest on whatever
  measure we are using, so as to make the pattern
  visible. Right above is a least angle betweenness
  analysis of part of central London at a radius of
  2 kilometres.
• The relevance of this is that these two measures
  reflect the two components of human movement the
  selection of destinations, and the choice of
  routes. The closeness of a segment to all others
  indicates potential as a destination, since,
  assuming distance decay, human trips are less
  probable with increasing distance so a segment
  which is closer to others within a given radius
  offers more potential as a accessible destination
  so you would locate your shop there if you
  could. The betweenness of a segment indicates the
  degree to which its lies on routes between all
  pairs of other segments in the system at a
  certain radius from that segment, so again this
  would guide you in locating your shop. In this
  way, space syntax assesses the movement potential
  of each segment both as destination and as route,
  using different definitions of distance though
  the weightings of shortest path, fewest turns,
  and least angle change, all of which have been
  canvassed by cognitive science as critical for
  human navigation in complex space patterns such
  as cities. With this matrix of measures we can
  test hypotheses about movement and land use
  patterns simply by correlating flows on segments
  with their spatial values.

5

• Extensive studies have shown three things
• - that in most circumstances these spatially
  defined movement potentials account for around
  60 of the differences in pedestrian movement
  flows on segments and about 70 of vehicular
  flows (most recently see Hillier Iida 2005).
• - that both pedestrian and vehicular movement
  patterns are best approximated by the least angle
  analysis, nearly as well by the fewest turns
  analysis, and least well by the shortest path
  analysis.
• - that land use patterns are shaped by the way
  the urban grid affects movement flows, and this
  sets in motion a self-organising process through
  which the city evolves into its more or less
  universal form of a network of linked centres and
  sub-centres at all scales set into a background
  of mainly residential space.
• Since movement and land use patterns are
  suspected of being critical variables in the
  spatial patterning of crime, either beneficially
  or negatively, depending on your paradigm, this
  analysis of the movement potentials created by
  the urban grid provides a powerful and perhaps
  a necessary - tool for exploring spatial patterns
  of crime and their relation to other aspects of
  the life of the city.

6
Five years of street robbery (left) and
residential burglary (right) in a London borough.
While the robbery pattern is highly linear and
follows the network of linked centres, burglary
seems to have no pattern.

• The numerical values that this analysis generates
  can then become the basis of a street segment
  based data table, to which any number of other
  numerical variables can be added for each street
  segment in this case crime counts of various
  kinds, but also demographic, socio-economic and
  physical data such as house types, numbers of
  storeys, existence of back alleys or basements
  and so on. In effect we integrate all the
  variables on the basis of a spatial descriptions
  of all the street segments and so arrive as a
  description of the system of interest as a system
  of spatial location types plus their functional
  characteristics.
• This will allow us to bring greater precision
  into the relation between crime and where it
  occurs. The figure above left a shows the
  distribution of street crime over a five year
  period in a London borough. The pattern is
  strongly linear hot lines ? and closely
  follows the analysis which reflected the network
  of linked centres and sub-centres which can be
  seen behind it. Right shows the same for
  residential burglary, which seems to follow no
  obvious pattern. These spatial distributions form
  our research question. To see them more clearly I
  will magnify them.

7
(No Transcript)
8
(No Transcript)
9
(No Transcript)
10

• We now present some results from an EPSRC funded
  study at UCL of urban design factors in the
  spatial distribution of crime, part of the
  inter-university Vivacity project under the
  Sustainable Urban Environments programme, We use
  this most recently developed form of space
  syntax analysis in conjunction with GIS to
  explore the crime-design relation at the micro as
  well as macro level. We call this technique high
  resolution analysis of crime patterns in urban
  street networks
• Through the study, we seek to address key
  questions in the current debate between two
  opposed paradigms for designing out crime in
  urban areas, one stressing open and permeable
  environments, mixing uses, and higher densities
  using the traditional city as its model, the
  other advocating more closed, relatively
  impermeable, smaller scale, mono-use low density
  environments, using Oscar Newmans notion of
  defensible space as its model. In the US this
  paradigm was goes under the rubric Safescape
  against Defensible Space
• The first is in effect saying having more people
  around makes you safer, so use spatial design to
  maximize co-presence and natural surveillance,
  and in this way make life more difficult for the
  criminal. The other is saying that people
  include criminals, and increasing numbers
  increase their anonymity, so reduce numbers to
  improve control by the resident community.
• We suggest that each of these views is in a way
  true, but they are referring to different things.
  Part of our aim today is to clarify how
  apparently contradictory propositions drawn from
  past research go together though others on both
  sides will turn out to be just plain wrong The
  key will be bringing much great precision in
  describing the relationship between urban design
  variables and crime risk. This greater precision
  is the prime concern of this paper.

11

• So the specific paradigm questions we aim to
  address are
• - Are some kinds of dwelling inherently safer
  than others ?
• - Does density affect crime rates ?
• - Does mixed use reduce or induce crime ?
• - Do people passing by make you safer or more
  vulnerable ?
• - Should residential areas be permeable or
  impermeable ?
• - Do physical and spatial factors interact with
  social and economic factors in crime risk ?
• Here we show that in no case is there a simple
  yes-no answer to these questions, but only
  if-then answers given this and this and this
  design factor, then that may be safe, but
  otherwise unsafe. The key mesage is that spatial
  factors interact, and interact with social
  factors, to alter the vulnerability of locations
  to different kinds of crime, but do so in
  predictable ways.

12

• The study reported here is of 5 years of all the
  police crime data in a London borough made up of
• A population of 263000
• 101849 dwellings in 65459 residential buildings
• 536 kilometres of road, made up of 7102 street
  segments
• Many centres and sub-centres at different scales
• Over 13000 burglaries
• Over 6000 street robberies
• We are focusing our study on residential burglary
  and street robbery as these are the two crimes
  that people most fear today.

13

• Residential burglary and street robbery data
  tables have been created at several levels
• the 21 Wards (around 12000 people) that make up
  the borough for average residential burglary and
  street robbery rates. At this level, spatial data
  is numerically accurate, but reflects only broad
  spatial characteristics of areas. Social data
  from the 2001 Census is available, including
  deprivation index, but at this level patterns
  are broad and scene-setting at best.
• the 800 Output Areas (around 125 dwellings) from
  the 2001 Census, so social data is rich and
  includes full demographic, occupation, social
  deprivation, unemployment, population and housing
  densities, and ethnic mix, as well as houses
  types and forms of tenure. Unfortunately spatial
  data is fairly meaningless at this level due to
  the arbitrary shape of Output Areas.
• the 7102 street segments (between intersections)
  that make up the borough. Here we have optimal
  spatial data, good physical data and council tax
  band data indicating property values which can
  act as a surrogate for social data
• Finally, the 65459 individual residential
  buildings, comprising 101849 dwellings. Here
  spatial values are taken from the associated
  segment, and again we have good physical data
  with Council Tax band as social surrogate. Street
  robbery cannot of course be assigned here.
• So the richest demographic and socio-economic
  data doesnt quite overlap with the richest
  spatial data, but the usefulness of creating data
  tables at different levels with different
  contents will become clear below as we switch
  between levels to seek answers to questions.

14

• First a health warning. Although the database is
  very large, it is confined to one region of
  London, and the findings must be reproduced in
  other studies for us to be sure that they have
  any generality, even in one country.
• Having said that, the area is highly
  differentiated in terms of social composition and
  urban type, from inner city to suburban. This
  will allow spatial propositions to be tested by
  subdividing the data, for example into the 21
  wards which are each made up of a number of
  people comparable to one British Crime Survey
  to see it they hold for each area taken
  separately. We can do the same with house type or
  council tax band, to see if propositions hold for
  each subdivision separately.

15

• We can begin with some broad brush scene-setting
  at the highest level of aggregation the 21
  wards. Left above, burglary per household for the
  21 wards (each with about 12000 people) is
  plotted against Deprivation Index (a multi factor
  index of social disadvantage widely used in the
  UK) falling from left to right, and right above
  the same for street robbery in the ward.
• Using the data on the 21 wards we can first see
  that the concept of a high crime area may need to
  be qualified, as the patterns of residential
  burglary and street crime do not have the same
  peaks and troughs, although there is a broad
  decrease in both with lower social deprivation.

16

• The weak areal relation between burglary and
  robbery can be clarified by plotting residential
  burglary in rank order from low to high as black
  dots, and the corresponding rank order for street
  robbery as circles (left above).
• Right above we do the same for the 800 Output
  areas, showing an even greater pattern of
  discrepancy between the two crimes. This suggests
  that, analytically, we should not talk about high
  and low crime areas. Different crimes select
  different areas.

17

• In fact what we find statistically is that
  robbery responds to social deprivation far more
  linearly than burglary. There is more robbery in
  socially disadvantaged areas.
• But burglary, in terms of socio-economic
  grouping, is U-shaped. If we plot burglary rates
  against Council Tax Band a local tax based on
  the value of property, and so a reasonable guide
  to socio-economic status - , the lowest tax
  households have high rates (even of reported
  burglary) but so do the small number of the
  highest tax households.

18

• What we find at the ward level is that six of the
  wards have markedly higher burglary rates than
  all the others, and that they are all contiguous
  to each other in the south east corner of the
  borough. We can see from this image of all the
  7102 segments coloured up for average tax band,
  from red for high through to green for low (the
  blues are segments with non-residential uses
  only), that this part of the borough has a mixed
  population with strong zones of relative
  affluence.

19
High burglary wards

• Multi-variate analysis shows that the 6 wards
  with markedly high residential burglary rates
  have a distinct social, demographic and spatial
  profile. They have significantly
• smaller households
• lower rates of owner occupation
• the highest rates of converted flats
• the lowest rates of residence at ground level
  (largely due to conversion)
• the greatest incidence of basements
• are in more spatially accessible locations that
  means in the more urban south and east rather
  than the more suburban north and west
• 3 of the 6 are centred on a strong town
  centres, and the other 3 are more up-market
  residential areas adjacent to these areas.
• They are far from socially homogenous. They cover
  most of the range on mean tax band. On the ratio
  of high to low level of employment types, they
  include the 3 highest as well as the third
  lowest. They are not particularly high on
  unemployment or lone parent households. This
  suggest a complex process involving social and
  demographic processes with space featuring as
  where different groups choose to live. Can we
  learn more from the next level down, Output Area
  level ?

20

• At the Output Area level, the pattern of burglary
  rates red for high through to blue for low is
  puzzling. High and low rate areas seem to be
  arbitrarily juxtaposed, barely reflecting even
  the shift from19th century terraced housing
  south-east through to inter-war suburbia
  north-west.

21

• In fact at the Output area level (about 125
  dwellings), we can look at demographic, ethnic,
  socio-economic, occupational and household
  composition data from the 2001 Census,
  deprivation index data, and data on dwelling and
  tenure types, in a much more disaggregated way.
  But stepwise regression of the whole database
  (taking all due precautions on distributions and
  inter-correlations) with respect to residential
  burglary tells a rather unexpected story. Apart
  from the Deprivation Index, the analysis seems to
  background social, economic and demographic
  factors, and highlight simple physical factors of
  dwelling types and densities.
• In fact we find
• an increase in burglary with social deprivation
• a strong effect from housing type, with purpose
  built flats and terrace houses beneficial and
  converted flats and flats in commercial buildings
  vulnerable
• but a decrease with increased housing density
  (and people density, but housing is stronger,
  through the two correlate closely )
• These variables are pervasive and consistent. The
  first and second are known, for example from the
  British Crime Survey (BCS 2001), though in this
  case it is notable that the housing type effect
  is found independent of social factors. Let us
  look first at dwelling type. In what follows the
  data are aggregated from the single building
  database of 65450 residential buildings which we
  have constructed, and so are grounded in the most
  disaggregated data available.

22
Burglary rates for dwelling types

• Type sample burglary single multiple
• 
  rate (5yrs) occupancy occupancy
• Flats gt 15st 676 .084
• Flats 6-15st 228 .066
• Flats 5-6st 4249 .092
• Flats 3-4st 11745 .079
• Low terraces T 13993 .121 .129 .117
• Low terraces t 2469 .109 .115 .096
• Linked2-3st 3570 .093 .073 .100
• Tall terraces 1489 .144 .352 .128
• Linked semis 14350 .117 .117 .116
• Standard semis 22312 .135 .141 .112
• Large semis 5465 .193 .223 .175
• Small detached 3535. .166 .173 .139
• Large detached 2190 .200 .294 .161
• Whole sample 65459 .130 .135 .112
• The samples for the high flats are hundreds
  rather than thousands and are uniformly social
  housing. So lets assume that the lower rates may
  be helped by non-reporting. But the lower flats
  represent the whole range of social classes, as
  do all the categories of housing and cannot be so
  explained. In general risk increases with the
  number of side on which you are exposed a
  simple physical factors. But why are multiple
  occupancy risks systematically lower ? Perhaps
  because a multiple burglary is reported as one ?
  Or because the ground floor protects the upper
  floor by offering the first target ? This could
  be a case of difference between occurrence and
  risk.

23

• Type 1 very tall blocks, point block slabs .084
• B 590 .084
• Type 2 tall flats 6-15 storeys .046
• B 228 .046
• We can use the variable of Council Tax Band, a
  local tax based on an assessment of the value of
  the property, running from A for the lowest value
  through to H for the highest, to see how this
  pattern interacts with socio-economic factors. It
  is a far from perfect indicators of social
  advantage, but it is a pretty good approximation
  given sufficiently large samples. For the two
  samples of taller flats, all dwelling are B
  rated, the second lowest, and are social housing
  left over from the times when we built high rise
  housing for socially disadvantaged people. So
  lets momentarily discount these apparently lower
  rates as being likely to be contaminated by
  non-reporting.

24

• Type 3 medium height flats 5-6 storeys - .109
• A 732 .086
• B 588 .193
• C 1098 .118
• D 1031 .111
• E 431 .105
• F 87 .093
• G 23 .087
• But if we take medium rise flats, the data cover
  the range of house values, so the low mean risk
  of .109 over five years compared to the overall
  average of .130, is very unlikely to be an
  artefact of non-reporting. Let us also worry that
  the samples for the higher tax bands are
  increasingly small. But we do seem to see a
  progressive reduction in risk with increasing
  social advantage.

25

• Type 4 lower 3-4 storey and smaller flats -
  .084
• A 1018 .096
• B 2198 .081
• C 5673 .080
• D 1136 .065
• E 256 .142
• The overall mean for low rise flats is lower
  still, and again falls with increasing social
  advantage, though there is a marked rise to above
  average risk for rather small sample in the
  E-band, hinting again at the U-shape.

26

• Type 7 low terraces with small T - .111
• B 133 .132
• C 594 .098
• D 1296 .093
• E 358 .159
• F 24 .391
• Low terraces with small back additions again have
  a lower than average risk, but now the relation
  to social advantage really does begin to look
  U-shaped, though again with a caveat on small
  sample size at the low and high tax band ends of
  the data.

27

• Type 6 low terraces with large T 120
• A 66 .180
• B 1176 .111
• C 5013 .116
• D 4201 .107
• E 2070 .117
• F 847 .165
• G 175 .231
• Low terraces with large back additions are also
  well below average risk, but the data has better
  coverage of the range of tax bands, and now
  really does look U-shaped. It really does begin
  to look as though in this dwelling type, the poor
  and the rich are at higher risk compared with the
  middle classes.

28

• Type 8 linked and step-linked 2-3 storeys and
  mixed - .078
• A 175 .137
• B 444 .136
• C 1070 .129
• D 1403 .059
• E 296 .062
• F 53 .019
• G 41 .073
• Linked and step-linked low rise housing (whose
  precise form we have not so far been able to
  ascertain) shows one of the lowest overall rates,
  and again the lowest rates are in the middle of
  the sample and the higher rates at the ends.

29

• Type 5 tall terraces, 3-4 storeys - .193
• B 237 .063
• C 599 .130
• D 446 .213
• E 37 .159
• F - -
• G 75 .393
• With tall terraces, we find the first above
  average risk, but now a shift towards higher risk
  for higher social advantage is clear, even if we
  aggregate the top three bands for greater
  statistical security. Why are tall terraces so
  much more vulnerable. Almost certainly because
  most higher terraces have basements, and
  basements were shown by the Output area data to
  be a significant factor in increased risk.

30

• Type 11 - semis in multiples of 4,6,8 - .117
• B 859 .177
• C 2349 .102
• D 8076 .113
• E 2570 .138
• F 153 .149
• Another form of grouped housing whose precise
  form we have again not been able to ascertain
  with any security, shows a below average rate
  again with the peaks at either end.

31

• Type 10 - standard sized semis - .138
• B 493 .249
• C 3268 .097
• D 4268 .120
• E 10819 .145
• F 2529 .148
• G 507 .152
• Now we are in the realm of the famous English
  semi-detached house, here at standard size. We
  find a clear above average risk on the large
  sample, and a clear U-shape.

32

• Type 12 - large property semis - .199
• B 307 .268
• C 1581 .169
• D 1322 .153
• E 969 .210
• F 606 .211
• G 489 .260
• With larger semis the risk increases markedly,
  and again with a marked U-shape. We may notes at
  this stage that our B tax band people which in
  high flats we might think were non-reporting, now
  seem to be reporting in what we might now think
  of as the predictable numbers. As we saw, the
  rates for smaller detached houses was .160 and
  for large .200, but with insufficient number to
  permit the breakdown into tax bands.
• It is clear that two factors are involved in the
  shifting pattern of risk with dwelling type the
  simple physical fact of degree of exposure on
  how many sides is your dwelling not contiguous
  with others ? and social advantage, with the poor
  and the rich at higher risk. But houses are more
  at risk than flats, the more so as they become
  more detached, and the better off you are the
  more you are at risk in a house and safe in a
  flat.

33

• These result suggest that density may be a
  factor, but in the opposite to the normally
  expected sense. However, we saw that the density
  component in the Output Area was also strongly in
  this direction. So what is going on ? Other
  studies have found density neutral, and few
  studies have shown such seemingly strong effects.
  But at the Output Area level, the findings may be
  unreliable, due to great differences is the
  amount of open or non-residential space included
  in output areas. However, we can switch levels
  and explore the issue further through the
  individual house data base.
• We take the individual dwelling data on GIS, and
  for every dwelling in the borough, create a
  variable for how many other dwellings are, in
  part or wholly, within 30 metres of each
  dwelling, so measuring the density of dwellings
  in the neighbourhood of all dwellings in the
  sample. We then use logistic regression to see
  how far a lower or higher value of this variable
  affects the likelihood of the dwelling being
  burgled at least once (of course we lose
  information on repeats with this technique).
• We find that for the dataset as a whole higher
  density at ground level substantially reduces the
  risk of being burglarised, while the opposite is
  the case for non-ground level density. Neither
  seems to be affected by the presence of other
  variables. More surprisingly, the presence of
  non-residential uses within the buffer is mildly
  beneficial. The pattern holds if we split the
  data and remember this is over 65000
  residential buildings into single and multiple
  dwelling units, though with the multiple
  non-residential uses switch to a slight
  disbenefit.

34

• To test this, we split the data three ways first
  by the 21 wards (each with about 12000 people and
  3-4000 dwellings) and ask if these relation holds
  in every region regardless of social, economic or
  layout and physical type and mix second by
  dwelling type, regardless of the area in which
  they occur third by Council Tax band
• On the first we find that high density at ground
  level reduces vulnerabiilty substantially in 20
  out of 21 wards, and that in 19 cases the
  relation is statistically highly significant, and
  in the other weakly. In the one case where the
  relation is not found, the result is
  statistically insignificant. So this effect seems
  to hold in spite of the great social, economic,
  demographic and morphoplogical variation in
  physical and area type across the borough.
• Splitting the data by dwelling type, we find that
  the relation holds for all dwelling types taken
  separately except tall terraces, and is in each
  case highly significant. In the case of tall
  terraces, there is no significant relation.
  Splitting by tax band, we again find the relation
  holds in all cases, with high statistical
  significance in all cases bar the small sample of
  top tax band cases.
• Higher ground level density does then seem to
  mean lower burglary. But a worrying pattern is
  beginning to appear. You are safer if you live in
  a flat, but you increase the risk to those
  locally living in houses. There could be a simple
  statistical effect here. If flats are harder
  targets, as they seem to be, then within an area
  that is being target, the higher the proportion
  of harder flat targets the greater the risk to
  the smaller proportions of easier target houses
  on the ground. By Ockhams razor, we would
  suggest this is the case.

35

• What about the relation of burglary to movement
  then ? Is it perhaps, like density, a more
  complex relation than we thought ? It is. Most
  studies have concluded that proximity to movement
  increases risk, casting doubt on the Jane Jacobs
  belief that passing strangers are more eyes on
  the street. The key argument is that if you are
  located in an accessible location with strong
  movement, then you are more likely to lie on the
  burglars search path. I am sure this is often the
  case, but syntactic studies have also shown that
  within residential areas the more important roads
  for movement have less risk. Can we reconcile
  these different findings ?
• We can do the simple things first and use the
  burgled or not logistic regression on the
  movement variables. The pattern we find suggests
  that both arguments are right, but at different
  scales of the system. At the level of
  to-movement, or accessibility, at the city scale,
  we find a substantial increase in risk. But
  through-movement at this scale has much less
  effect. To my mind, this supports the search path
  hypothesis, because being accessible at this
  level does not imply though movement at this
  level which has a much lower effect. If we go
  down to the local scale, then the accessibility
  penalty becomes negligibly small, and through
  movement becomes beneficial. This does suggest
  that both parties are right, but that is local
  through movement that is beneficial, global not
  so. Note by the way that the simple spatial
  variable of segment connectivity is slightly
  beneficial in this analysis, but this may be to
  do with the relatively small number of cul de
  sacs in this largely nineteenth and early
  twentieth century area.

36

• This suggests that we might examine the space
  structure more closely to see if we can find out
  more by looking at the data at the level of the
  street segment rather than the individual
  building. For this we must digress from results
  to methodology for a moment and talk about the
  problem of a rate.
• The problem of establishing a rate for a small
  spatial element is this. Suppose there is a
  random process of assigning burglaries to houses,
  saying numbering the houses and selecting the
  burgled ones by a random number selector. As
  the process proceed, we will find that whatever
  spatial unit we choose, then over time the number
  of burglaries on that spatial unit will be
  proportionate to the number of targets on that
  unit. So it will appear that there is more
  burglary on units with more targets, and a random
  result might appear to be a pattern a hot spot,
  say.
• If we then try to get over this problem by
  establishing a rate, that is by dividing the
  number of burglaries into the number of targets,
  then we find the inverse problem that for a
  significant part of the process the rate will
  appear to be lower on units with more targets
  since random events occurring on that unit will
  be dividing into a larger denominator. Of course,
  in the long run, as the process is iterated
  infinitely many times, then the number of
  burglaries will be perfectly proportionate to the
  number of targets but we dont know how close
  we are to getting there, and in any case when we
  get there the information about rates is useless
  since all we need to know is the number of
  targets.

37
We can illustrate this problem by showing the
distribution of crimes against targets as simple
numbers and as rates. Left top is the simple
count of burglaries against dwellings for
segments, and right top the rate of one against
the other. Below left is the simple count of
robberies against segment length, and below right
the rate per unit of length. It is clear that
neither pattern in meaningful. We find two kinds
of illusory result the simple figures, and the
rate obtained by dividing robberies into length.
Shorter segments would appear safer with simple
numbers and more dangerous with a rate, and vice
versa with longer segments.
38

• We need to solve this problem, because the
  segment as a spatial unit is critical to all our
  measures, and there would be great gains from
  being able to link spatial descriptors to crime
  figures at the disaggregated level of the
  segment. How can we do it ? Our reasoning is
  this. From the point of view of the occurrence of
  burglary or robbery on segments, the number of
  targets is the primary risk factor for burglary,
  and the length of segment the primary risk factor
  for robbery. By primary risk factor, we mean a
  distribution that would arise from the random
  assignment process. No analysis of spatial units
  will be relevant unless it builds in a solution
  to the primary risk problem.
• The problem can be solved by aggregation. For
  example, if we take all the segments with a given
  number of targets, count all the burglaries on
  those segments, and divide into the total number
  of targets, we no longer have the denominator
  problem, and we have a true rate, though only at
  the aggregate level. The same applies to length
  of segment for robbery. We take all segments
  within a given length band and aggregate the
  total length and the total number of robberies,
  and we have a true rate.

39

• So let us explore primary risk band analysis.
  There are 436 kilometres of street with at least
  one residence in the borough, made up of 4439
  segments with an average length of 98 metres, and
  an average of 17.1 dwellings per segment, of
  which 15 are on the ground floor and 2 on upper
  floors.
• Primary risk band analysis means taking all the
  segments with a given number of dwellings,
  counting the total number residences and the
  total number of burglaries, and dividing the
  latter into the former. The number of dwellings
  on the segment is now not involved in the
  calculation, so we escape the denominator
  effect. The number of dwellings on the segment
  is now a condition for that aggregate.
• For example, if we take the 328 segments with
  exactly one dwelling, which have on average 3.15
  nonresidential units, then we find a total of 197
  residential burglaries have occurred over 5 years
  in the 328 dwellings, a rate over 60 or 12 a
  year, compared to an overall average in the data
  of 3.37 per year.
• If we take the 34 segments with more than 90
  dwellings per segment we find a total of 3708
  dwellings and 419 burglaries over five years, a
  five year rate of 11.3 or 2.26 per year.

40

• We then divide all segments into bands according
  to their number of dwellings, giving an average
  of 94 segments per band of average total length
  9.3 kilometres with an average of 1600 dwellings
  per band. We then calculate the true rates for
  each band, and plot them on a line chart with
  dwellings per segment on the horizontal axis and
  the burglary rate on the vertical (in fact taking
  the log of each).
• We see that the risk of burglary decreases
  steadily with with increasing numbers of
  neighbours on your street segment. We can also
  express this as a simple regression, and we find
  an r-square of .78. Remember we are dealing with
  over 100,000 dwellings here.

41

• We can see the dwellings per segment pattern
  visually by colouring segments from red to blue
  for high to low numbers of dwellings on the
  segment. We see the high dwelling, safer segments
  are often in quite grid like areas.

42

• Is this the density effect we saw before, or is
  it simply to do with the number of neighbours ?
  By adding the numbers of dwellings on the segment
  on which dwelling lies as a further variable in
  the logistic regression on density on the
  residential building data table (so again we
  switch levels), we show that both hold in the
  presence of the other, and so we can have no
  doubt that the two effects are to some extent
  independent, though this varies from one area to
  another and from one housing type to another. But
  in general the numbers of dwellings on the
  segment and the density of dwelling around the
  dwelling both act to reduce vulnerability to
  burglary.
• There could be two reasons for this. One is of
  course surveillance. Another is perhaps more
  interesting. Recently Kate Bowers of JDI showed
  that a burglary in an area was likely to lead to
  others soon after. We suggest there may also be a
  saturation effect. Once a certain number of
  hits have been made in a target area, then the
  target is deemed saturated and the burglars move
  elsewhere. Statistically this could lead to the
  effect we find if block faces were identified as
  target areas, and show why where seems to be
  safety in numbers living on a larger rather than
  smaller block.

43

• This result is very much in accordance with the
  space syntax theory of cities, in which
  residential areas have larger block sizes, and so
  spread the lower levels of movement over fewer
  spaces. Small block sizes occur not in
  residential areas but in live centre areas
  where small blocks facilitate ease of movement
  from all parts to all others, a vital feature of
  a retail or service providing areas.
• This has important implication for current design
  practice. It is clear that permeability is now
  often being provided for its own sake without
  regard to how well used permeabilities will be.
  It is known from previous studies that unused
  access to residential environments is likely to
  increase crime levels.
• It is clear from these new results that
  permeability should be provided sufficient to
  structure real movement patterns in all
  directions, but should not be provided beyond
  this, especially when it will not be well used.
  This implies a larger block size than is often
  now the case.

44

• So, with some complications, higher densities and
  good number of line neighbours seem beneficial.
  But what about the issue of layout ? The
  generator for layout is segment connectivity. A
  1-connected segment can only be the end of a cul
  de sac, while 6-connected must approximate a grid
  like layout.

45

• If we plot segment connectivity from red for 6
  down to blue for 1, we see that 6 connected
  segments occur in two kinds of place in the high
  street, and in the more grid like residential
  areas. There are 81 1-connected segments with at
  least one dwelling, 451 2-connected, 748
  3-connected, 1868 4-connected, 1050 5-connected
  and 261 6-connected - though the 6-connected
  still account for nearly 4000 dwellings.

46

• If we compare this to the plot for the number of
  dwellings per segment, we will see that there is
  some overlap in the residential areas, but in the
  high street areas we find the opposite a very
  low residential rate is associated with
  6-connectivity. In these areas of course we find
  the highest concentrations of non-residential
  activity.

47

• This leads us to a basic morphological fact about
  city layout at the most elementary level of
  segment connectivity and segment length (and so
  number of targets). For pure residential
  segments, length increases with connectivity. But
  for segments with non-residential activity length
  decreases with connectivity. This has to do with
  grid intensification and the way centres and
  sub-centres are formed. We must take great care
  then in trying to link these layout primitives to
  crime.

48

• For example, if we plot segment connectivity
  against the primary risk dwellings per segment
  variable, we find high connectivity for both very
  low and very high numbers of dwellings, the
  former corresponding to high street areas and the
  latter to grid like residential areas.

49

• If we plot out primary risk bands against spatial
  accessibility, one of the configurational
  variable that best models real movement rates, we
  find that like segment connectivity, integration
  has high values both with low number of dwellings
  per segment and the high numbers.

50

• If we then plot burglary rates for the primary
  risk bands against accessibility, we find a
  bifurcation in the data, with one arm rising and
  the other seemingly falling with integration. If
  we then split the primary risk band about half
  way into those with less than 25 dwellings per
  segment on the left and those with more on the
  right, then it seems that the effect of
  integration and therefore movement on crime
  depends on the amount of residence. If this is
  the case, then it would seem to go some way to
  explaining the divergent findings in earlier
  research.

51

• We find another very striking effect when we
  build non-residential uses into the primary risk
  band analysis. Above we plot with squares the
  fall of burglary rates with increasing numbers of
  dwellings for segment with no non-residential
  units. In circles we plot those with 1-4
  nonresidential units, and with black dots those
  with any number of non-residential units. With
  small numbers of dwellings the penalty of mixed
  use is very large. But as the numbers of dwelling
  increases, the lines converge. Residential
  numbers significantly reduce risk in mixed use
  locations. The benefits of a strong residential
  presence do not seem to be confined to
  residential areas.

52

• What about simple segment connectivity the cul
  de sac versus grid debate ? In previous studies,
  we have found simple linear cul de sacs with good
  numbers of dwellings safe, but complexes of
  linked cul de sacs often unsafe, suggesting that
  the safety of cul de sacs depended on them being
  building into a network of through streets rather
  than being used as a generic design principle.
  Here, where there are relatively few cul de sacs,
  and those that exist are of the simple linear
  kind, we do find that, as expected, there are
  lower rates for the (rather small numbers of) 1-
  and 2-connected segments. On the other hand, the
  rate for 6-connected segments is lower than for
  either 5 or 4-connected, though not as low as the
  1- and 2-.
• However, this hides a more complex story. Again
  we find that the number of dwellings per segment
  is the most powerful variable. Although the five
  year mean rate of burglary for 1 and 2 connected
  segments (which for the most part will be cul de
  sacs or near cul de sacs) in the data is .105,
  against an overall average of .130 for all
  segments, cul de sacs with no more than 10
  dwellings have a rate of .209 (calculated again
  as the total number of burglaries over the total
  number of dwellings in such segments, so the rate
  is a true one), while 6 connected (grid like)
  segments with more than 100 dwellings have a rate
  of .086, and the much larger number with 50 have
  a rate of .142, so still lower than small cul de
  sacs.

53

• Social factors also interact. By dividing the
  data into those with above average council tax
  bands (and so more valuable property),
  6-connected segments with more than 50 dwellings
  per segment and higher tax band and no
  nonresidential, the burglary rate is.101,
  compared to .153 for the lower tax bands, and in
  general 6-connected segments with 50 dwellings
  have a rate of .143, reducing to.124 with higher
  tax bands but rising to .191 for lower tax bands.
  Similar effects are found with increasing
  nonresidential uses. More affluent people seem to
  survive better in a well-scaled mixed use,
  grid-like layout than the less affluent.
• A rather different story is found in cul de sacs.
  The overall burglary rate for 1-2 connected
  segments with 1-10 dwellings is .213 for high
  tax, and .143 for low tax. If we exclude those
  with nonresidential the rate is even higher, at
  .241 for high tax but lower, at .092 for for low
  tax. High tax dwellings are more vulnerable in
  small cul de sacs, either than low tax dwellings,
  or high tax dwellings in grid like layouts. We
  might conjecture that the more attractive the
  target, the more the isolation of the cul de sac
  benefits the burglar, while for less attractive
  targets, cul de sacs tend to be off the search
  path.

54

• So our paradigm questions about how residential
  burglary risk varies with built environment
  conditions seem to have complex, but clear,
  answers or at least with more studies and a
  greater range of data, could be answered clearly.
  Simplistic statements from either side of the
  Safescape-Defensible Space divide about burglary
  are not on. Both are often saying cogent things,
  but we need to learn to say them together.
• But there is one factor above all others that
  seems to come out of this data a kind of safety
  in numbers principle which argues that the
  emphasis in the past on small spatial groupings
  is wrong. The benefits seem to come from larger
  residential co-presence communities since this
  is what is needed to establish the dominance of a
  strong residential culture in whatever part of
  the urban system you are in. A stronger because
  more numerous - residential culture cuts out the
  greater risks that isolated residents have in
  mixed use areas, makes both grid layouts and cul
  de sacs safer, and reduces the risks associated
  with larger scale movement while increasing the
  benefits of smaller scale movement. We need to
  scale up our ideas of what makes a good
  residential area.
• But what about street robbery ? In the following
  sequence of images we plot the patterns of four
  different street crimes first during daylight
  hours and second during the evening and night. We
  see the patterns are systematically different in
  that the pattern disperses during the day and
  contracts at night.

55
600 am 1800 pm
1800 pm 600 am
Snatch
56
600 am 1800 pm
1800 pm 600 am
Threatening behaviour
57
600 am 1800 pm
1800 pm 600 am
Including a weapon
58
600 am 1800 pm
1800 pm 600 am
Attack
59
If we plot our primary risk factor for robbery
the length of the street segment against the
simple count of robbery, we find it increases as
expected. But if we plot the aggregate rate for
length risk bands, we do not find a fall, as with
burglary, but a rise followed by a fall, with a
marked dip on the way.
60
If we separate segments with no non-residence
from those with non-residence, we find that the
residential segments do fall consistently, and it
is only the segment bands with non-residence that
rise and fall. Why ? On the right we plot the
increases in residential numbers with increasing
segment length and see that residential numbers
increase with length far more rapidly than
non-residential. So, as with burglary, the fall
in robbery rates seems to be associated with a
higher residential factor. This is our present
conjecture, but we still need to understand the
fluctuations.
61

• But what of the increase in robbery with
  increased non-residential uses. We know of course
  that these will usually be associated with higher
  movement rates, especially retail and similar
  movement dependent uses. So does the risk
  increase, that is the rate per pedestrian
  movement ? We cannot of course observe pedestrian
  flows in all the relevant segments, but we can
  make use of our extensive London data base on all
  day pedestrian (and vehicular) flows on over 367
  street segments in 5 London areas to ascertain
  the average difference in pedestrian flows on
  segments with and without retail.
• Mean pedestrian movement on all 367 segments is
  224.176 per hour. For segments without retail the
  rate is 158.476 for 317 segments, and for retail
  it is 640.714 for 50 segments. This means that
  the movement rate on segments with retail is
  4.042 times those on segments without. The
  average robbery on segments without
  non-residential uses, as shown above left, is
  .0074, while the rate for segments with non
  residential uses is .0176, or 2.4 times as high.
  The rate of increase in robbery in the
  substantially less than the increase in movement
  rates, and dividing one into the other, so
  2.4/4.042, we get 1.68, so that we can say in
  terms of people risk you are 68 safer on busier
  street segments with non residential uses than on
  those without. This of course is a very
  provisional figure, but it is probably a
  conservative one,

62
But what about the spatial characteristics of
segments where robbery occurs. Here we
plot, left, the mean segment connectivity against
increasing robbery rate for primary risk bands,
and, right. The connectivity of the line of
which the segment forms a part. In both we see
a sharp fall in the highest robbery locations.
63
On the left we now plot robbery rates against
time of day, from early morning to late night. On
the right we plot the mean length of segment on
which robberies occur. We see that the highest
rates in the late afternoon and early and late
evening periods occur on shorter lines (with the
exception of the low early morning rates). The
differences are quite small, but we are talking
about over 6000 events here.
64
We now plot the rates for time periods against
the local accessibility measure which best
predicts pedestrian movement. We see the three
high rate periods show a sharp fall. The spatial
characteristics of robbery in high robbery
periods show a marked shift to segregation. On
the right we show that the increase in rate is
also a concentration in fewer spaces, measured as
events over locations.
65
In fact, if we plot the highest robbery segments,
we tend to find them to be poorly connected
segments making links between parts of the grid
which would be otherwise unlinked. Many are
close to tube stations, and a few close to Post
Offices.

• In fact, if we plot the highest robbery segments,
  we tend to find them to be poorly connected
  segments making links between parts of the grid
  which would be otherwise unlinked. Many are close
  to tube stations, and a few close to Post Offices

66
robbery

• If we then plot robbery against accessibility for
  the time periods (so that the time periods are no
  longer in time order), we find that the top 3
  rates for time periods are in the least
  integrated location, while the four lowest are in
  more integrated locations.
• Only one high rate period is in an integrated
  location the post midnight to 3 pm period, when
  robbery returns to integration locations. The
  lesson is Dont go on the high street after
  midnight - but dont leave if before midnight.

67

• We saw early on that the pattern of street
  robbery was linear, and at a distance seemed
  closely linked to the network of linked centres
  and sub-centres that are the product of the
  self-organising process and probably more than
  anything else makes cities the livable places
  they are. We worried that the same process that
  self-organises the city also self-organises its
  greatest risks ? So not quite. We have already
  seen that pedestrian numbers often make risks
  lower even though occurrence may be high. We have
  also seen that the greatest risk are in places
  close to the high street which do not have
  movement.
• We are currently investigating another
  possibility here, one inherent in the very
  measures we use. It is a key difference between
  the closeness and betweenness measures that close
  to close is close, that is if we are adjacent to
  a space which is close to all others then our
  space is also pretty close to all others. But
  betweenness is not transitive in this sense. The
  fact that a segment is strong on through movement
  potential does not imply that its neighbours are
  so. On the contrary, neighbours of strong through
  movement segments typically include both strong
  and weak neighbours.
• This means that, say, a high street made up of
  segments with high accessibility and strong
  through movement potential, will have some
  neighbours with strong through movement potential
  as well as strong accessibility, but others which
  combine strong accessibility (by virtue of being
  attached to the High Street) with weak through
  movement potential. Visual inspection of several
  areas suggest that it is these where the risk is
  highest, and most likely to optimise the one
  victim at a time formula which benefits the
  potential robber.

68

• In conclusion
• Overall, then, the results of this first study
  using the high resolution methodology show a
  complex but clear story - The safety of
  dwellings is affected by two simple interacting
  factors        - the number of sides on which
  the dwelling is exposed to the public realm - so
  flats have least risk and detached houses
  most        - the social class of the
  inhabitantsIn effect all classes tend to be
  safer in flats, but with increasing wealth the
  advantages of living in a flat rather than a
  house increase - Higher local ground level
  densities of both dwellings and people reduce
  risk, but off the ground density increases it. In
  effect, living in a flat off the ground you are
  safer, but you are not benefiting your ground
  level neighbours. But at the same time, other
  factors can make ground level living secure.
•  - Notably, the larger the numbers of dwellings
  on the street segment (the section of a street
  between intersections, and so one face of an
  urban block) the lower the risk of crime. This
  applies to cul de sacs and to through streets,
  and has a greater effect than either being in a
  cul de sac or being on a through street. The more
  immediate neighbours you have the safer you
  are. - Furthermore, spatially integrated street
  segments (more movement potential) tend to have
  lower crime rates with increasing dwellings per
  segment, but higher crime rates with decreasing
  dwellings per segment - in effect spatial
  integration and numbers of neighbours work
  together to produce a bifurcation in the data.
•  - There is greater crime risk on mixed use
  street segments, but this extra risk is
  diminished with increased residential population.
  So small number of residents in mixed use areas
  are at risk, but larger number of residents
  create a residential culture and eliminate the
  mixed use crime risk penalty. - Finally, social
  and spatial factors interact, for example in that
  small numbers of well-off dwellings in cul de
  sacs are more at high risk than a similar group
  of poor dwellings, while the opposite is the case
  in grid like layouts where better off dwellings
  are less at risk than poor dwellings.

69

• Summary
• So overall, the results of this first study are
  in many cases unexpected, but taken together show
  a complex but clear story - the safety of
  dwellings is affected by two interacting
  factors        - the number of sides on which
  the dwelling is exposed to the public realm - so
  flats have least risk and detached houses
  most        - the social class of the
  inhabitantsIn effect all classes tend to be
  safer in flats, but with increasing wealth the
  advantages of living in a flat rather than a
  house increase - higher local ground level
  densities of both dwellings and people reduce
  risk, but off the ground density increases it. In
  effect, living in a flat off the ground you are
  safer, but you are not benefitting your ground
  level neighbours. - there is greater crime risk
  in mixed use areas, but this extra risk is
  diminished with increased residential population.
  So small number of residents in mixed use areas
  are at risk, but larger number of residents
  create a residential culture and eliminate the
  mixed use crime risk penalty. - the larger the
  numbers of dwellings on the street segment (the
  section of a street between intersections, and so
  one face of an urban block) the lower the risk of
  crime. This applies to cul de sacs and to through
  streets, and has a greater effect than either
  being in a cul de sac or being on a through
  street. The more immediate neighbours you have
  the safer you are. - spatially integrated
  street segments (more movement potential) tend to
  have lower crime rates with increasing dwellings
  per segment, but higher crime rates with
  decreasing dwellings per segment - in effect
  spatial integration and numbers of neighbours
  work together to produce a bifurcation in the
  data. - social and spatial factors interact,
  for example in that small numbers of well-off
  dwellings in cul de sacs are more at high risk
  than a similar group of poor dwellings, while the
  opposite is the case in grid like layouts where
  bet