Computers, Environment and Urban Systems 31 (2007) 206–227 www.elsevier.com/locate/compenvurbsys
Modelling UK residential dwelling types using OS Mastermap data: A comparison to the 2001 census Scott Orford ¤, Jonathan RadcliVe
1
School of City and Regional Planning, CardiV University, Glamorgan Building, King Edward VII Avenue, CardiV CF10 3WA, United Kingdom Received 16 May 2006; accepted in revised form 8 August 2006
Abstract In the UK dwelling type is a commonly used term to refer to the ‘building style’ of a residential property. The concept is used frequently in urban analysis, social science research and Government policy, and has been used as a measure of homogeneity in the construction of Output Areas (OAs) in the UK 2001 census. However, there are no formal deWnitions of dwelling types in the UK beyond a basic statutory deWnition of dwelling-house and Xat. The British Government’s OYce of National Statistics (ONS) also categorises dwellings into diVerent types but provides no clear guidance to how these are deWned. There is no source of dwelling type data for individual properties at a national scale beyond sample surveys. The main source of information is the 2001 census, which provides counts of ONS deWned dwelling types for OAs. The objective of this research is to use OS Mastermap, a high resolution topographic database of Great Britain, to model and provide dwelling type information for individual residential addresses. The success of the modelling exercise is measured by comparing the modelled dwelling type information to data on dwelling type collected in the 2001 census, using Chi-square as a goodness of Wt measure. Small area diVerences in the two measures are analysed in order to ascertain where OS Mastermap does not provide a very good estimate of dwelling type. The analysis highlights problems with the modelling process and discusses the potential of OS Mastermap as a source of dwelling related data and also possible small area errors in the 2001 census. © 2006 Elsevier Ltd. All rights reserved. Keywords: OS Mastermap; Census; Dwelling type; Disclosure; Small area characterisation
*
1
Corresponding author. Tel.: +44 2920 875272; fax: +44 2920 874845. E-mail addresses:
[email protected] (S. Orford),
[email protected] (J. RadcliVe). Fax: +44 2920 874845.
0198-9715/$ - see front matter © 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.compenvurbsys.2006.08.003
S. Orford, J. RadcliVe / Comput., Environ. and Urban Systems 31 (2007) 206–227
207
1. Introduction The last decade has seen the continuing development of geographic information handling technologies and the emergence of rich new sources of digital data (Longley, 2003). Both GIS and remote sensing data are increasingly more detailed in terms of their spatial and non-spatial information, their resolution and their coverage, with examples including the high resolution imagery from Google Earth and Microsoft Local Live. Commentators such as Longley (2002) have begun to examine how these developments are improving not only the digital data infrastructure of many cities but also the opportunities for undertaking urban modelling and urban analysis. The research agenda in urban GIS is now increasingly dominated by the development of integrated databases, methods for their analysis and models relevant to Wne-scale urban geographies (Longley, 2003). Traditional sources of ancillary information such as Censuses of Population and address and property registers are becoming enhanced by their integration with these data sets. The frequency with which remote sensing data are obtained makes them a potential source of updating ancillary information, while the precision of some of the emerging digital framework data makes them increasingly useful as a source of micro-scale structure in the built environment (Longley & Mesev, 1999). This research is concerned with investigating how some of these emerging sources of digital data can be used to model aspects of the micro-scale structure of built form within a GIS. More speciWcally, it will examine how the development of digital framework data in the UK and the availability of high resolution aerial photographs from providers such as Google Earth can be used to create dwelling information for individual properties, enhancing and supplementing existing dwelling information supplied by the UK Census of Population for small areas. The initial aim is to model dwelling type information for individual residential properties captured by the framework data and thus providing context to the small scale urban structures that the data represents. The basic concept of dwelling type relates to the number of properties a dwelling has as its neighbours and the degree and manner to which they are associated with the dwelling. For instance, a dwelling can have no neighbours and stands alone or it can be part of a larger building containing several dwellings. The concept of dwelling type and its deWnitions vary internationally with diVerent countries using diVerent typologies depending upon national and local contexts. For instance, the US has dwelling types such as ‘duplex’ and ‘condominium’, whilst in the UK dwelling types include ‘maisonette’ and ‘semi-detached’. This concept is examined and discussed more critically later in the paper. Dwelling types have distinct geographies within cities and contribute most to the homogeneity of social structure at small spatial scales in urban areas (Tranmer & Steel, 1998). In terms of urban GIS applications, dwelling type has been used (together with tenure) as a measure of homogeneity in the construction of census Output Areas (OAs) in the UK 2001 Census of Population (Martin, Nolan, & Tranmer, 2001), and has also been an important factor in the measurement of urban density and morphology (Longley & Mesev, 1999). Dwelling type is often used as a proxy measure for residential location and environmental quality, particularly in small area classiWcation schemes such as those used in geodemographics. In housing market research, dwelling type is an important factor in explaining the supply and demand of housing, since diVerent dwelling types tend to embody particular bundles of housing attributes. Hence econometric models often use dwelling type to capture interaction eVects between housing attributes within particular
208
S. Orford, J. RadcliVe / Comput., Environ. and Urban Systems 31 (2007) 206–227
bundles (Orford, 2000) and it is sometimes used as a proxy for housing attributes such as dwelling size. It has also been used to explain diVerences in housing choice with respect to life cycle stages (for instance, in retirement and divorce), and accessibility and disability, vacancy levels by tenure, residential mobility and residential satisfaction (Booth & Amato, 1992; Cho, 1997; Day, 2000; Ostrovsky, 2004). Dwelling type is also related to the trade-oV between the list price of the property and the time it is on the market (Ong & Koh, 2000) and is an inXuential factor in explaining the performance of property investment markets and the link between housing and wealth (Gallo, Lockwood, & Rutherford, 2000; Hamelink, Hoesli, Lizieri, & MacGregor, 2000; Wolverton, Hardin, & Cheng, 1999). The aim of this research is to investigate potential sources of information on dwelling type for individual properties in the UK. In particular, the aim is to evaluate the potential of using digital framework data created and maintained by Ordnance Survey (OS), the national mapping agency for Great Britain, to model dwelling type for individual residential addresses. The research will describe a modelling process that uses a speciWc digital database supplied by the OS called OS Mastermap to classify individual residential addresses in the Welsh capital city of CardiV. The success of the modelling will be measured by comparing the estimated dwelling types with dwelling type recorded in the 2001 census at OA level. The analysis will investigate mis-matches at the OA level with respect to the modelling process and also in accuracies of the diVerent data sources. The following section examines issues relating to deWnitions and data sources of dwellings and dwelling types. This is followed by a discussion of the modelling process and the comparison with dwelling type in the 2001 census. Sections 5 and 6 examine the diVerences in dwelling type classiWcations at the OA level and discuss the principal reasons for the mis-matches. The paper concludes by reXecting on how the modelling process can be improved and also what the results of the classiWcation reveals about the accuracy of the 2001 census for particular small areas. It also provides some applications for the modelling output and suggests areas of future research. 2. Dwellings and dwelling types: deWnitions and data sources The Wrst thing to note is that the term dwelling type is synonymous with other terms used (although not always consistently) in the literature such as accommodation type, house type and property type. The second thing to note is that there are rarely any formal deWnitions of dwelling type categories and this is true with respect to the UK. Instead the UK has various deWnitions of dwelling and dwelling types. The British Government’s OYce of National Statistics (ONS) deWnes a dwelling as a self-contained unit of accommodation (household space) that consists of one household (an unshared dwelling) or part of a converted or shared house where two or more households share basic facilities (a shared dwelling). A household space is the accommodation occupied by an individual household or, if unoccupied, available for an individual household. A household is deWned as one person living alone, or a group of people living at the same address with common housekeeping- i.e. sharing either a living room or at least one meal a day. Dwellings in the UK can be categorised into two statutory types; a dwelling-house or a Xat (SI 2000 no. 2531). The diVerence is essentially whether a building containing the dwelling is divided horizontally or vertically. A Xat is a “separate and self-contained premises constructed or adapted for use for residential purposes and forming part of a building
S. Orford, J. RadcliVe / Comput., Environ. and Urban Systems 31 (2007) 206–227
209
from some other part of which it is divided horizontally” (SI 2000 no. 2531, p. 3). A dwelling-house covers all other dwelling types and by implication these relate to vertical sub-divisions of buildings. Beyond the ONS and statutory deWnitions of dwelling, dwelling-house and Xat there are no other formal deWnitions of dwelling type. The 2001 census uses the categories houses and bungalows, Xats, maisonettes, apartments and bedsits but no clear deWnition of these is provided but rather they are based on common understanding of the terms. Houses and bungalows are sub-divided into three types based on the number of adjoining neighbours (detached, semi-detached and terraced). The classiWcation of Xats is more complex with diVerent deWnitions in each census (see Table 1). Compared to the 2001 census, the 1981 and 1991 censuses have a more detailed classiWcation for Xats whilst the 1981 census has no classiWcation for house. Another important source of deWnitions is from the English, Welsh, Scottish and Northern Ireland House Condition Surveys (EHCS, WHCS, SHCS and NIHCS respectively – see Table 2). These surveys provide the main source of information on the condition and energy eYciency of the housing stock in each country. Dwellings in these surveys are categorised into types similar to those used in the 2001 census with the EHCS and WHCS having a slightly Wner categorisation with respect to terraced houses and purpose built Xats. In addition, the EHCS and NIHCS have bungalow as a separate category (although rather confusingly, the EHCS also includes ‘chalet bungalows’ that have part of the living space Table 1 Dwelling type deWnitions used in the census since 1981a 1981 census
1991 census
2001 census
Permanent building Total
Purpose built: Detached Semi-detached Terraced
Whole house or bungalow: Detached Semi-detached Terraced (including end-terrace)
Self-contained: Total Purpose built Xat Separate outside entrance Shared outside entrance Flat Flatlet Other
Purpose built Xat in: Residential building Commercial building Converted Xat: Separate entrance Flat Flatlet Shared entrance Flat Flatlet
Flat, maisonette or apartment: Purpose built/tenement Part of a converted/shared house (including bedsits) In a commercial building
Not self-contained Total Bedsits Other
Unshared not self-contained Flat Rooms Bedsit Shared not self-contained Flat Rooms Bedsit
Non-permanent Caravans Other
Non-permanent
Temporary structure
The Wrst inclusion of a dwelling type question was in the 1966 inter-decennial census. A dwelling type question was not included in the 1971 census, however. a
210
S. Orford, J. RadcliVe / Comput., Environ. and Urban Systems 31 (2007) 206–227
Table 2 Current national House Condition Survey dwelling type deWnitions English 2003
Welsh 2004
Scottish 2002
Northern Ireland 2001
Detached house Semi-detached house Terraced house Small Medium Large Bungalow (includes chaletc) Converted Xat Purpose built Xat High risea Low riseb
Detached house Semi-detached house Terraced house End Mid
Detached Semi-detached house Terrace Tenement 4-in-a-block
Detached house Semi-detached house Terraced house
a b c
Converted Xat Purpose built Xat High risea Low riseb
Converted Xat Tower/slab Xat
Bungalow (excludes chaletc) Converted Xat Purpose built Xat
Blocks over Wve storeys. Five storey blocks or less. Chalet bungalows have a small proportion of their Xoor space in attic rooms.
upstairs). The diVerences in typologies between the four surveys reXect the nature of the housing stock in each country. Other large national surveys, such as the Survey of English Housing (SEH) and the British Household Panel Survey (BHPS) use similar typologies of dwelling type found in these House Condition Surveys (HCS). There are important diVerences in how a dwelling is classiWed in each of the above surveys. In the 1981 and 1991 censuses, the enumerator classiWed a dwelling based upon an external survey, typically of the front elevation. In the 2001 census the householder determined the dwelling type with the exception of vacant properties where this information was again provided by the enumerator. In comparison, dwelling type in the HCS is determined by professional surveyors. In most other national surveys, dwelling type information is provided by the householder based upon their own judgement. This variation in how dwelling type is determined could have important implications, particularly since dwelling type classiWcation tend to be based upon common sense understanding of the terms as oppose to formally deWned criteria. 3. Overview of OS Mastermap as a data source There is no source of dwelling type data for individual properties at a national scale in the UK beyond sample surveys (such as the BHPS). The (now defunct) rates register used in local property taxation did classify individual dwellings using broad categories but the council tax register that replaced it in 1993 does not. The principal national source of dwelling type information is the 2001 census which provides counts of dwelling types deWned by the ONS aggregated to Output Areas and above. So in order to model dwelling types for individual residential addresses, OS Mastermap was obtained for CardiV, Wales. OS Mastermap is a multi-layered geographical database which provides large-scaled (typically 1:1250 in urban areas and 1:2500 in rural areas) highly detailed framework data for Great Britain (Ordnance Survey, 2006). It was created by the Ordnance Survey and has been continuously maintained since its launch in November 2001. OS Mastermap is available in four separate data layers: a topography layer, an address layer, an integrated transport network layer and an imagery layer. This research is using the topography and
S. Orford, J. RadcliVe / Comput., Environ. and Urban Systems 31 (2007) 206–227
211
address layers only. The topography layer is the building block of OS Mastermap and provides information on nine themes in a seamless intelligent topological database. The nine themes include land area classiWcations, buildings, roads, tracks and paths, rail, water, terrain and height, heritage and antiques, structures and administrative boundaries. The themes contain more than 450 million features represented as points, lines and polygons as well as annotation. The features are all geo-referenced to the OS National Grid. The topography layer was developed from OS Land-Line, an earlier Ordnance Survey data product containing the same nine themes but not the fully structured topological database of OS Mastermap. Instead OS Land-Line represented features only by lines and points and this limited its application potential. The OS Mastermap address layer provides OS National Grid co-ordinates for more than 26 million residential and commercial properties in Great Britain. Each address in the layer is represented by a point that is geo-referenced by a grid co-ordinate to a resolution of less than 1 m. The address layer was derived from an earlier Ordnance Survey product called OS Address-Point which was based on addresses originating from the Royal Mail’s postcode address Wle (PAF). The PAF only contains the addresses of properties that have a delivery point for mail. Addresses that do not have a unique delivery point, such as Xats in a building which has a communal entrance and where the mail is sorted internally, may not be recorded in the PAF and hence these addresses may also be missing in the OS Mastermap address layer. Addresses that share the same two-dimensional geographic location, such as Xats built on top of each other in a block, will be recorded separately in the address layer if they have a unique delivery point but they will all share an identical grid reference (Martin & Higgs, 1997). Recording multiple addresses at a single address-point location is also a common procedure for handling new-build when individual properties have not yet been surveyed by the OS. Positional accuracy is an important issue with points in the address layer and the attribute table has a column which describes the status of the grid-references of an address as either ‘Wnal’ or ‘provisional’ depending upon whether or not it has been surveyed. Addresses may also be described as ‘non-geographic’ in the address layer. These refer to delivery points that are not associated with a physical address but rather mail boxes in shops and oYces, such as Post OYce box (PO Box) addresses. Finally, the address layer does not explicitly diVerentiate between commercial and residential addresses although this may be inferred from the format of the address itself in some cases. These issues have a bearing upon the analysis and will be expanded upon in later sections. Every feature in OS Mastermap has a unique identiWer known as a TOID (Topographic IdentiWer). This is a 16 digit number that provides a method of uniquely referencing each feature but cannot be used to understand the geographic location or nature of a feature (Murray & Shiell, 2004). Instead a TOID is guaranteed to remain unchanged throughout the life of the feature and provides a method of associating user-data to OS Mastermap. TOIDs also provide a means to link corresponding features held in the four diVerent layers. For instance, an address record in the address-layer can be linked to its respective building polygon record in the topography layer by a shared TOID found in both attribute tables. OS Mastermap is updated daily and the updated data are provided in a change-only update (COU) format. This means that only those features that have been added, edited or deleted since OS Mastermap was initially provided to the user are made available. Updates can be provided regularly and theoretically on a daily basis, although normally a monthly COU is the most viable for frequent updates.
212
S. Orford, J. RadcliVe / Comput., Environ. and Urban Systems 31 (2007) 206–227
4. Dwelling type typology and building polygon topology: a four stage classiWcation process The aim of the modelling process was to use information contained in the topography and address layers of OS Mastermap to assign a dwelling type category to every residential address in CardiV. First, non-geographic addresses were removed from the analysis as these represent mail boxes rather than the delivery points of buildings. Addresses in the attribute table of the address-layer were then joined to building polygons in the attribute table of the topography layer using the shared TOID in both tables as a join item. In cases where the corresponding building polygon TOID was missing for an address record, pointin-polygon analysis was undertaken to provide a geographic match of an address to a building polygon. This was only undertaken on those addresses which had been surveyed and had a positional quality Xag of ‘Wnal’ indicating that the grid-reference was accurate to less than 1 m. Those addresses which had not been matched by the TOID and had a positional quality Xag of ‘provisional’ were omitted from the analysis. Building polygons which were associated with an address were then extracted from the topography layer and a new layer created. Commercial addresses were included in the analysis at this stage since it is common in areas of mixed-use for residential properties to be joined to a commercial property – for instance, when a terraced house has been converted into a shop or oYce – and the dwelling type classiWcation of the neighbouring terraced houses may be aVected if the commercial building is omitted. The dwelling type of each address was then modelled using a simple four stage process. In stage one a custom written GIS script was used to determine building polygons that were Wrst and second order neighbours. Two polygons were classiWed as being Wrst-order neighbours if they shared a common boundary (known as building divisions in OS Mastermap) and second order neighbours if they shared a boundary with the same Wrst-order neighbour. Table 3 illustrates the diVerent dwelling types that can be identiWed by their OS Mastermap polygon topology. Detached houses are free-standing properties that do not have any Wrst or second order neighbours. Semi-detached houses have one Wrst-order neighbour but no second order neighbours. End-terraces also have one Wrst-order neighbour but they also have one second order neighbour allowing them to be diVerentiated from semi-detached houses. Mid-terraces have two Wrst-order neighbours and, depending upon the length of the terrace, have none, one or two second order neighbours. Stage two of the process attempted to diVerentiate between building divisions that were trivial and non-trivial. A trivial building division was one in which two buildings were joined at ground Xoor level, by an entrance porch or a garage for example, but not joined at the upper stories or by a continuous roof line. This is illustrated in Fig. 1a for the case of an entrance porch and Fig. 1b for a garage. These joins would not have an eVect on the commonly accepted deWnitions of dwelling type (i.e. the two detached houses joined by a garage in Fig. 1b would not be regarded as a pair of semi-detached houses) and therefore this must be accounted for in the modelling process. An analysis was made of a sample of properties that were joined by these ‘trivial’ building divisions compared to properties with non-trivial building divisions. Trivial building divisions tended to be either less than 4 m in length or less than 15% of the length of the entire building outline. Using this criterion, dwellings that had been classiWed as semi-detached in stage one were re-classiWed as detached in stage two if they were joined by a trivial building division. Likewise, dwellings that had been classiWed as terraced in stage one were re-classiWed as semi-detached in stage two.
S. Orford, J. RadcliVe / Comput., Environ. and Urban Systems 31 (2007) 206–227
213
Table 3 Polygon topology and dwelling type typology using OS Mastermap Detached 1st 0 2nd 0
Semi-detached 1st 1 2nd 0
End-terrace 1st 1 2nd 1
Mid-terrace 1st 2 2nd 0
Mid-terrace 1st 2 2nd 1
Mid-terrace 1st 2 2nd 2
Ordnance Survey © Crown Copyright. All rights reserved. 1st: number of Wrst-order neighbours. 2nd: number of second order neighbours.
Stage three of the process attempted to identify Xats. OS Mastermap is a two-dimensional representation of the built environment that shows vertical but not horizontal subdivisions of buildings. As discussed earlier, the statutory deWnition of a Xat in the UK is a premises forming part of a building from some other part of which it is divided horizontally. This, incidentally, is not the same as the ONS deWnition which also classes any property which is deWned as a ‘shared dwelling’ as a Xat regardless of internal sub-divisions. Therefore it is not possible to identify Xats based on the statutory deWnition using the polygon topology. Instead, Xats were identiWed by three methods using the address-layer. The Wrst method simply searched for the occurrence of the word ‘Flat’ or ‘Apartment’ in the address of the property. Second, address-point locations that shared an identical grid-reference were also assumed to be Xats. The assumption here is that Xats with separate addresses within the same building that share a single delivery point, or that are built directly on top of each other as in a tower block, will share the same grid reference in the address layer. However, as discussed earlier, new build properties are often temporarily
214
S. Orford, J. RadcliVe / Comput., Environ. and Urban Systems 31 (2007) 206–227
Fig. 1. Two examples of ‘trivial’ building divisions: (a) joined by entrance porches; (b) joined by a garage.
geo-referenced by a single address-point and therefore could be mistaken for Xats in this method. But such properties will also tend to have a positional quality Xag of ‘provisional’ indicating that their grid-references have not yet been surveyed and hence would have been omitted from the analysis in the earlier stages. The third method used the multiple occupancy count in the address-layer that records the number of households living at a single address. It was assumed that a multiple occupancy count of two or more indicated that the building had been sub-divided internally into Xats but retained a single delivery point for mail and hence only had a single address. These three methods will not identify all the Xats in the study area. In particular, it is not possible to identify Xats in converted buildings and houses in multiple occupation (HMOs) that share a single delivery point and do not have a separate record in the address layer or a multiple occupancy count of two or more. This will have important implications for under-enumerating dwelling types, especially in innercity neighbourhoods where HMOs and sub-divided properties are more common. In the Wnal stage of the process, commercial addresses (7925) and addresses with ‘unclassiWed’ dwelling types (883) were identiWed. Commercial addresses were identiWed if they had an organisation or department name in the address Weld. This will underestimate the number of commercial addresses especially where they are coincident with residential addresses for small business/self-employment. An address had an unclassiWed dwelling type because either the record in the attribute table in the address layer could not be matched to a corresponding building polygon in the topography layer (320 or 36%) or the building did not conform to the dwelling types discussed above and hence could not be modelled using the four stage process (563 or 64%). Finally, since OS Mastermap does not contain any information on the number of Xoors within a building or the height of the building it is not possible to diVerentiate bungalows. In previous research, Lake, Lovett, Bateman, and Langford (1998) extracted dwelling type information on a sample of properties using OS Land-Line and OS Address-Point – precursors to OS Mastermap. This was achieved through visual inspection of the connectivity between individual properties rather than using a modelling process and did not diVerentiate between trivial and non-trivial building divisions. Each address record was placed into an OA and the total number of residential dwelling types in each OA was calculated. In order to compare the OS Mastermap dwelling type classiWcation with the classiWcation in the 2001 census, only addresses that were in existence at the time of the census (April 2001) were included in the calculation. Date information was obtained from the All Fields Postcode Directory (AFPD) which contains a record
S. Orford, J. RadcliVe / Comput., Environ. and Urban Systems 31 (2007) 206–227
215
of when a postcode was introduced or terminated. Addresses containing postcodes that were introduced after June 2001 (the closest date to the census for which data is available in the AFPD) were excluded from the calculation. This method, however, will not identify dwellings that were built after June 2001 but were assigned postcodes in existence prior to this date e.g. housing sub-divisions and housing in-Wll. Data on the total number of dwellings and dwelling type for ‘unshared’ and ‘shared’ household spaces was obtained for the 2001 census. ‘Shared’ households spaces, by the ONS deWnition, relate to two or more households living in part of a converted or shared house (including bedsits) and sharing basic facilities. Hence, shared dwellings will always be classiWed as Xats in the 2001 census, regardless of whether the building is actually sub-divided into Xats. The number of shared dwellings was calculated by subtracting the number of unshared dwellings from the total number of dwellings. Since dwelling type typology varies slightly between the OS Mastermap classiWcation and the 2001 census, dwelling types were aggregated into four key categories: detached, semi-detached, terrace and Xats. Chi-square analysis was used to measure the goodness of Wt between the OS Mastermap estimations and the 2001 census counts using a 5% level of signiWcance as an indication of a poor match between dwelling types estimated at OA level. 5. A comparison of OS Mastermap and census classiWcations of dwelling type A perfect match is not expected due to diVerences with the two data sources and problems relating to the modelling process. Crucially there are several diVerences in the census and OS Mastermap deWnitions of dwelling and dwelling types and in the methods they use to arrive at the total numbers of dwelling types in a particular area. It has already been discussed how the two data sources adopt diVerent deWnitions of dwelling and dwelling type based upon household spaces and physical spaces so this section will focus on some of the problems with estimating dwelling type counts in each data source. First, there is the problem of accuracy in both data sets. The positional accuracy of some of the features in the OS Mastermap topography and address layers may cause misclassiWcation (or un-classiWcation) of dwelling types with 320 addresses in the address layer being unclassiWed due to the problem of not being able to assign them to a building polygon. The census relies upon the householder to be able to classify the dwelling type of their property and, as was discussed, not all dwellings conform to commonly understood dwelling type conventions. At the OA level, census data can be suppressed or manipulated due to issues of disclosure altering the number of dwelling types in small areas (Martin, 2006). Manipulation includes several methods such as record swapping with similar records within broad geographical areas and adjusting small counts. CardiV also suVered from an under-enumeration of around 3000 people in the 2001 census, although the under-count was concentrated within three wards (ONS, 2004). These were typical of ‘hard to count’ wards, having characteristics likely to be associated with under-enumeration, such as a large number of HMOs. In order to provide a fully adjusted set of census statistics covering 100% of the population, the ONS undertook a One Number Census (ONC) project. The aim of the project was to estimate under-enumeration and adjust the census database accordingly so that all statistics added up to ‘one number’. The ONC methodology took into account the uneven distribution of under-enumeration by targeting ‘hard to count’ locations. Households and persons estimated to have been missed by the Census were imputed and the characteristics of these households and individuals were modelled, including
216
S. Orford, J. RadcliVe / Comput., Environ. and Urban Systems 31 (2007) 206–227
dwelling type. Thus the ONC imputation process may have a signiWcant impact on the accuracy of dwelling type recorded in the census in areas of low response in the initial census enumeration. OA boundaries are synthetic in nature, having been generated automatically using OS Address-Points rather than from aggregations of building polygons and land parcels, and so it is common for individual building outlines to cut across OA boundaries. This often makes it diYcult to determine the OA into which a particular building should be placed and so the allocation is based upon the location of the address-point instead. The AFPD will not identify all the properties built after the census and so OS Mastermap will overestimate the number of dwellings in some areas. With respect to the modelling process, there may be mis-matches due to the failure to diVerentiate between trivial and non-trivial boundaries and also the fact that OS Mastermap will underestimate the number of Xats, particularly in HMOs. 6. Goodness-of-Wt of dwelling type estimates in OS Mastermap and the 2001 census Table 4 shows the diVerence in dwelling type estimated from OS Mastermap and recorded in the 2001 census. On the whole there is a good correspondence between the two although there are around 1000 less dwellings in OS Mastermap, possibly due to underestimation of Xats in converted buildings and HMOs in the address layer as discussed previously. OS Mastermap overestimates the numbers of terraces and to a lesser extent semi-detached houses and underestimates the numbers of detached houses and Xats by around 2%. The two sample Chi-square test is signiWcant at the 0.1% level, with the diVerences in the observed and expected numbers of detached houses having the largest contribution to the statistic. At ward level, four out of the 29 wards (14%) had Chi-squared statistics that were insigniWcant at the 5% level, suggesting a good match, and these were all suburban wards with a large percentage of detached and semi-detached dwellings. In the remaining wards, one (3%) was signiWcant at the 5% level, Wve (17%) at the 1% level and the rest (66%) at the 0.1% level which were characteristic of inner-city areas with a high degree of sub-divided properties. Hence two-third of wards had highly signiWcant statistical diVerences in the numbers of diVerent dwelling types estimated by OS Mastermap and recorded in the census. In order to investigate reasons for the diVerences in estimated dwelling type in OS Mastermap and counts in the 2001 census the remainder of the analysis was conducted at the Wner level of the OA. It must be noted at this point that because a Chi-square test is undertaken for each OA and there are 991 OAs then there is a real risk of committing a Type I error. This is when a signiWcant Chi-square statistic is calculated by chance Table 4 Comparison of dwelling type estimates in OS Mastermap and the 2001 census OS MM total Detached Semi-detached Terrace Flat UnclassiWed Total residential
15,796 41,166 46,464 22,740 888 127,054
X2 D 311 (p D 0.001).
OS MM (%) 12.4 32.4 36.6 17.9 0.7
Census total 18,218 40,548 43,931 25,351 128,048
Census (%)
DiVerence total
DiVerence (%)
14.2 31.7 34.3 19.8
¡2422 618 2533 ¡2611
¡1.8 0.7 2.3 ¡1.9
¡994
S. Orford, J. RadcliVe / Comput., Environ. and Urban Systems 31 (2007) 206–227
217
indicating a poor goodness-of-Wt between the numbers of dwelling types in both datasets for an OA when in fact there is a good correspondence (i.e. incorrectly rejecting a true null hypothesis). At the 95% conWdence level, we would expect around 49 OAs to be aVected in this way meaning that in the following results, the numbers of mis-matches are likely to be overestimated. Table 5 is a summary of the signiWcance levels (p-values) of the Chi-square statistics measuring the goodness-of-Wt between the numbers of dwelling types in both data sources for the 991 OAs in CardiV. Two-thirds of OAs (652) had an insigniWcant Chi-square statistic (greater than the 5% level) showing a good degree of similarity between the two measures, although 10% also had a Chi-squared statistic with a signiWcance of 0.1%. In Fig. 2 the goodness of Wt between the two measures varies across CardiV with a high degree of mis-match in OAs on the periphery of the city as well as in the heterogeneous housing stock of the inner-city areas. The majority of OAs with a good degree of similarity tend to be located in the post-war suburbs although clusters of mis-matched OAs within these areas are evident. The OAs in Table 5 have been further divided into rows based upon the absolute diVerences in the total number of dwellings in the two data sources in each OA. As discussed previously, a perfect correspondence in the number of dwellings in the two data sources is not expected due to diVerences in the deWnition of dwelling, problems associated with the enumeration of the 2001 census and issues of disclosure at OA level and problems of the enumeration of addresses in OS Mastermap. Around two-thirds of OAs (677) have an absolute diVerence of Wve or less in the number of dwellings in each data source showing a good correspondence between enumerated properties. However, almost 20% (183) of OAs have an absolute diVerence in the number of dwellings of eleven or more, with 28 OAs having an absolute diVerence of more than forty dwellings in the two data sources. Large absolute diVerences in the total number of dwellings at OA level in the two data sources could aVect the calculation of the Chi-square statistic, particularly if the diVerences are biased towards a particular dwelling type, such as Xats. The most likely eVect is that the Chisquare statistic will show a poor goodness-of Wt in those OAs where such large absolute diVerences in the numbers of dwellings exists. Table 5 SigniWcance of Chi-square statistics between dwelling type estimates in OS Mastermap and 2001 census at OA level, by diVerence in the number of dwellings in the two data sources Absolute diVerences in the number of dwellings in the two data sets by Output Area
Number of Output Areas Total
Chi-square p-values >5%
5%
1%
0.1%
0 1–5 6–10 11–15 16–20 21–30 31–40 41+
94 583 131 67 26 45 17 28
69 426 89 32 7 18 7 4
12 68 17 18 7 6 4 2
7 57 14 11 6 6 0 6
6 32 11 6 6 15 6 16
Total
991
652 66%
134 13%
107 11%
98 10%
218
S. Orford, J. RadcliVe / Comput., Environ. and Urban Systems 31 (2007) 206–227
Fig. 2. Chi-square goodness-of Wt measures between dwelling type estimates in OS Mastermap and the 2001 census at the OA level .
To demonstrate this, three-quarters of OAs with an insigniWcant Chi-square statistic (greater than the 5% level) have an absolute diVerence of Wve or less in the number of dwellings in each data source, compared to just over half for OAs with a signiWcant Chisquare statistic. Indeed there is a clear pattern of increasing absolute diVerences in the total number of dwellings in each data source and the increasing signiWcance of the Chi-square statistic. Hence a notable reason for mis-matches in the estimated dwellings types at OA level as measured by the Chi-square statistic is diVerences in the two data sources per se, rather than errors introduced by the modelling technique. These diVerences are a result of the factors discussed previously and are diYcult to untangle at this scale of analysis. These issues are discussed further in Section 9. 7. Evaluating classiWcation errors resulting from the modelling process This section explores the mis-matches in dwelling type caused by the modelling process rather than those caused by diVerences in the data sets per se. In order to do this the 182 OAs in Table 5 that have a signiWcant Chi-squared statistic (at the 5% level or less) and a close correspondence in the total number of dwellings in each data set (no more than an absolute diVerence of Wve – i.e. the Wrst two rows in the table) were selected for a more detailed examination. Twenty-eight of these OAs contained dwellings categorized as ‘unclassiWed’ in OS Mastermap and so these OAs were omitted leaving 154 OAs in the subsequent analysis. These OAs are shown in Fig. 3 in which just over 50% of OAs cluster into four distinct groups as shown in Fig. 4 (Riverside, Splott, Ely and Pentwyn). Three of these
S. Orford, J. RadcliVe / Comput., Environ. and Urban Systems 31 (2007) 206–227
219
Fig. 3. OAs with signiWcant diVerences in dwelling type estimates in OS Mastermap and the 2001 census where diVerences in the number of dwellings within each OA is Wve or less.
areas are characterised by local authority built housing stock and the fourth (Riverside) an inner-city area with a high degree of Xats in converted buildings, HMOs and recent inWll. The remaining OAs tend to be distributed around the edge of the city. This spatial grouping of OAs into distinct areas suggests that the mis-match as a result of the modelling process may be caused by particular building styles of houses especially those originally built by the local authority in peripheral estates. The 154 OAs were classiWed based upon the principal diVerences in the number of dwelling types in each data source. For instance, if the largest absolute diVerences occur in the numbers of semi-detached and terraced houses, the OA will be given the classiWcation of ‘ST’. An example of this is in an Output Area in which OS Mastermap has 38 Detached, 26 Semi-detached, 67 Terraced and 14 Flats and the 2001 census has 36 Detached, 48 Semidetached, 48 Terraced and 14 Flats. In this case the total number of dwellings are almost the same (145 and 146 respectively) but the principal diVerences are in the numbers of semi-detached and terraced houses and hence the OA will be classiWed as ‘ST’. If there are large diVerences between three dwelling types, such as detached, semi-detached and terrace, then the OA will be classiWed as ‘DST’. OAs with large diVerences in all four dwelling types are classiWed as ‘DSTF’. These classiWcations are summarised in Table 6, which shows the number and percentage of OAs for each classiWcation, and are mapped in Fig. 4. A mismatch occurred in a number of the 154 OAs (40%) because semi-detached were classiWed as terraces in OS Mastermap (ST). In 12% of OAs detached houses were classiWed as terraces (DT) and in 10% they were classiWed as semi-detached (DS). In 28 (18%) OAs, a
220
S. Orford, J. RadcliVe / Comput., Environ. and Urban Systems 31 (2007) 206–227
Fig. 4. A classiWcation of OAs based upon major diVerences in dwelling types estimated in OS Mastermap and recorded in the 2001 census, highlighting four spatial groupings.
number of Xats in the census were classiWed as either detached, semi-detached or terrace in OS Mastermap (DF, SF, TF), reXecting Xats in converted buildings. Only 14% of OAs had mis-matches between three dwelling types in the two data sources and 5% of OAs had mismatches between all four dwelling types (DSTF). Fig. 4 shows that three of the four groups of OAs tend to be classiWed into two or three types of dwelling mis-matches, typically ST, DT and DS in Pentwyn, ST, SF and TF in Splott and ST, DT and TF in Riverside. In Ely, however, there is a wide variety of mis-matches including several mis-matches between three and four dwelling types. The
S. Orford, J. RadcliVe / Comput., Environ. and Urban Systems 31 (2007) 206–227
221
Table 6 ClassiWcation of OAs by dwelling type diVerences Main dwelling type diVerences
No. of OAs
OA (%)
ST DT DS TF SF DST DSTF DSF DTF STF DF
62 18 16 16 9 7 8 5 5 5 3
40 12 10 10 6 5 5 3 3 3 2
Total
154
100
remaining OAs on the periphery tended to be characterised by ST and DT mis-matches. Detailed visual examination of aerial photographs from Google Earth of OAs in Splott, Pentwyn and Riverside revealed that particular building styles of detached and semidetached houses made them more likely to be classiWed as terrace in OS Mastermap. For instance, some detached houses were joined by neighbouring garages resulting in them being classiWed as semi-detached or terrace using OS Mastermap. Some semi-detached houses were joined by entrance porches resulting in them being classiWed as terraced in OS Mastermap – this particularly being the case of local authority built stock which was characteristic of a more communal style of building such as shared entrances. Indicative examples of both of these instances are illustrated in Fig. 5. In all these cases the houses were only joined at the ground Xoor level and did not share a continuous roof line. Although these joins should have been classiWed as ‘trivial’ building divisions in stage two of the classiWcation process, post-modelling analysis revealed that this did not occur due to the length of the building division exceeding the parameters used in the modelling criteria (i.e. building division is either less than 4 m in length or less than 15% of the entire length of the building outline). This resulted in the mis-classiWcation of these dwelling types. A possible solution would be a re-calibration of the parameters in stage two (to 5 m in length, for instance), but this would then increase the chances of mis-classifying genuine non-trivial building divisions which marginally exceed the existing criteria. 8. Mis-matched OAs: highlighting possible errors in the 2001 census The large and varied number of mis-matches of the 24 contiguous OAs in the Ely cluster warranted further investigation. Table 7 summaries the diVerences in dwelling type classiWcation in OS Mastermap and the 2001 census for the 24 OAs. In all but Wve OAs there are sizeable diVerences in the number of Xats with more Xats generally recorded in the census. This can be explained, however, by the method used to estimate Xats using OS Mastermap discussed previously. More interesting are the discrepancies in the number of other dwelling types, and particularly the numbers of detached dwellings recorded in the census compared to OS Mastermap. The census records substantially more detached houses in the
222
S. Orford, J. RadcliVe / Comput., Environ. and Urban Systems 31 (2007) 206–227
Fig. 5. Dwelling type mis-matches in the OS Mastermap modelling process.
OAs than is estimated using OS Mastermap in which detached dwellings are very few. There are also major discrepancies between the numbers of semi-detached and terraced houses in over half of the OAs. In order to determine whether these discrepancies were also caused by errors in the modelling process, a detailed and labour intensive comparison of the OS Mastermap classiWcations for individual properties into detached, semi-detached and terrace was made against high resolution aerial photographs from Google Earth so as to ground truth the classiWcation results. This indicated that the OS Mastermap classiWcations were over 98% correct (and any mistakes were rectiWed in Table 7) in these 24 OAs and that the modelling process had not mis-classiWed dwelling types due to problems with identifying trivial boundaries as had happened in Splott and Pentwyn. An example of a comparison is presented in Fig. 6, which refers to the OA in bold in Table 7. Zoomed in, this image clearly identiWes the 130 terraces and 10 semi-detached houses classiWed using OS Mastermap but not the 18 detached houses, 34 semi-detached houses and 76 terraces recorded in the census for this OA. Indeed, the detailed examination of the aerial photographs for all the 24 OAs revealed very few detached houses (and detached houses are arguably the easiest dwelling type to detect from aerial photography) and certainly not in the numbers recorded in the census. Therefore, checking against Google Earth and OS Mastermap, it would appear that the dwelling type information in the census for these 24 OAs is substantially incorrect, even when taking into account the problems of identifying Xats using OS Mastermap.
S. Orford, J. RadcliVe / Comput., Environ. and Urban Systems 31 (2007) 206–227
223
Table 7 Summary of dwelling type classiWcations for 24 OAs in Ely OS Mastermap
2001 Census
Total dwell Detached Semi-detached Terrace Flat Total dwell Detached Semi-detached Terrace Flat 139 135 127 142 133 128 127 115 139 134 127 133 135 133 140 124 126 142 122 122 134 135 142 123
1 0 0 0 0 0 0 0 0 0 1 2 0 0 0 0 0 0 1 0 0 0 0 0
94 93 83 84 19 70 62 100 26 116 24 91 38 104 10 34 79 78 60 26 76 75 50 51
44 18 44 45 84 58 54 4 60 18 96 40 61 28 130 49 47 64 61 28 48 60 52 28
0 24 0 13 30 0 11 11 53 0 6 0 36 1 0 41 0 0 0 68 10 0 40 44
140 137 128 140 136 125 127 115 137 137 127 131 130 133 139 128 127 139 123 120 136 134 141 125
6 11 9 17 6 9 9 12 14 10 12 10 8 8 18 10 6 9 11 6 3 5 11 10
92 85 71 75 25 72 65 82 32 104 37 83 46 93 34 41 74 74 60 34 78 81 60 47
33 15 38 30 69 36 40 7 56 17 71 31 53 26 76 30 39 47 47 24 35 34 34 23
9 26 10 18 36 8 13 14 35 6 7 7 23 6 11 47 8 9 5 56 20 14 36 45
There are several explanations as to why dwelling type information would be incorrectly recorded in the census for these 24 contiguous OAs, mostly relating to mistakes in the enumeration, collection (e.g. householder understanding of dwelling types) and entry of the data and also in the creation and the labelling of the OAs. Although unlikely to be a major issue, in some cases the householder may claim to live in a more prestigious form of housing. However, given that the total number of dwellings in the OAs recorded in the census is very similar to those recorded in OS Mastermap (and observed in Google Earth), data handling errors seem unlikely. A more interesting explanation concerns issues of disclosure in the census and data manipulation prior to dissemination as well as inaccuracies introduced by the imputation process associated with the ONC. According to Martin (2006, p. 8) small area data for the 2001 census “have been subject to unprecedented levels of disclosure control, involving pre-tabulation record swapping, random adjustments and rounding of small counts to 0 or 3”. Clearly by deWnition, it is impossible to establish fully whether the incorrect information in these OAs is a direct result of disclosure control since disclosure methods are kept conWdential. Indeed it is hard to see how rounding of small numbers could account for the diVerences observed here, although record swapping may be an explanation since the total numbers of persons and households (and by implication the total number of dwellings) within an area aVected by record swapping is not aVected. Hence further research in this area could provide some interesting insights into issues of census disclosure. It also suggests that OS Mastermap could be a better estimate of dwelling type in these OAs than would be concluded from an examination of the census alone.
224
S. Orford, J. RadcliVe / Comput., Environ. and Urban Systems 31 (2007) 206–227
Fig. 6. Comparison of an aerial photograph and OS Mastermap with dwellings classiWed into semi-detached and terrace for an OA in Ely.
9. Conclusion This paper has examined the potential use of OS Mastermap in providing a source of information on dwelling type for individual residential addresses. The classiWcation of dwellings in OS Mastermap into four basic dwelling types showed a close correspondence with similar dwelling type information recorded in the 2001 census for two-thirds of OAs, as measured by Chi-square statistics. Where there were signiWcant diVerences at OA level
S. Orford, J. RadcliVe / Comput., Environ. and Urban Systems 31 (2007) 206–227
225
between the dwelling type classiWcations, this was shown to be caused principally by diVerences in the total number of dwellings in each data source. The reasons for such diVerences are documented in detail elsewhere (e.g. Barr, 2005; Martin, 2006); for instance, OS Mastermap has been shown to substantially underestimate Xats in certain residential environments, especially HMOs and Xats in converted buildings. OS Mastermap continues to have issues with the positional accuracy of some of its features (e.g. address points) and another shortcoming is that it does not contain attribute information referring to the date that a feature was Wrst added, resulting in a reliance on the AFPD to establish when an address was created. Flats were also problematic in the 2001 census, and the census suVered from under-enumeration in three wards in CardiV. The ONC imputation process may also have introduced inaccuracies in the numbers of dwelling types in some small areas. Because of this, however, it is likely that the modelling process is in fact more successful at the individual property level than is suggested by the OA level Chi-square analysis; particularly for those OAs with large discrepancies in the number of dwellings. The main issue with the modelling process concerned the diVerentiation between trivial and non-trivial building divisions. In a number OAs, mis-matches were due to a mis-classiWcation of semi-detached, detached and terrace houses in OS Mastermap as a result of errors in identifying trivial property boundaries. In these cases, modelling breaks in the roof line would improve the classiWcation and also remove some of the current arbitrariness in the criteria used in the modelling process. There are two potential methods that could be used to model roof lines: analysis of aerial imagery and analysis of building height information. With respect to the former, Elaksher, Bethel, and Mikhail (2003) discuss a semi-automated technique that uses aerial images to extract 3D wire-frames of buildings. The frame can be classiWed into roof and non-roof regions allowing the perimeter of the roof to be identiWed and represented as linear features. An example of the latter is discussed by Alexander, Smith, Javis, Tate, and Tansey (2006) and uses height information extracted from LiDAR data that has been incorporated into OS Mastermap. Ostensibly to construct 3D representations of roofs for visualization purposes, the research also produces roof lines that take into account breaks in height and thus could be used to identify trivial and non-trivial building divisions. In addition, LiDAR data could be used to identify bungalows – a dwelling type that does not have a separate category in the 2001 census or one that can be modelled using OS Mastermap. The research has also uncovered some discrepancies in the dwelling type information in the 2001 census for a cluster of contiguous OAs. Although these could be a result of data management errors in the dissemination of the census data, they could also be indicative of wider disclosure control, although more substantive research is needed in this area. Using identiWable features in the built environment (such as diVerent dwelling types) as a way of checking the quality of the census could be a novel and less problematic method than comparing population counts collected in diVerent registers by diVerent agencies. Understanding the incidence of census errors and disclosure control is important as they both have widespread implications for the analysis of census data. To conclude, the paper has investigated a method of assigning a dwelling type classiWcation to individual residential addresses. In terms of applications, having a dwelling type classiWcation could add value to existing address lists and address based data sources and can have important uses in academic and policy research. It can be used to create more informed sampling frames in social science surveys, particularly those in which dwelling type is an important variable of reference. Dwelling type is used frequently in Government
226
S. Orford, J. RadcliVe / Comput., Environ. and Urban Systems 31 (2007) 206–227
policy research, particularly in the areas of housing and health. For example, it is used in the Standard Assessment Procedure for Energy Rating of Dwellings (SAP) to estimate energy eYciency in the housing stock and the incidence of fuel poverty in the UK (WAG, 2005). In housing market and house price analysis it is an important structural attribute in determining relative house prices and how these relates to the supply and demand of housing bundles. In this sense it could also be used in computer assisted mass appraisals (CAMA) such as in local property tax re-valuations such as the recent council tax re-banding in Wales and Northern Ireland. Dwelling type data has an important role in the analysis of urban morphology and urban form in terms of describing and explaining urban densities and the development of cities through time. The OS Mastermap classiWcation process and the use of high resolution imagery also represent methods of creating census like information from non-census data sources, an issue of increasing importance in the UK (Martin, 2006). Indeed the OS Mastermap COU period of a month means that dwelling and dwelling type information could be available on almost a continuous basis as oppose to every ten years with the census. Although more research is required, there is no doubt that digital framework data sources such as OS Mastermap and high resolution imagery will become increasingly important as a supplement to existing sources of socioeconomic data such as the census. Acknowledgement The authors would like to thank the Ordnance Survey for supplying OS Mastermap for CardiV. References Alexander, C., Smith, S., Javis, C., Tate, N. J., & Tansey, K. (2006). 3-D Visualization of OS Mastermap: using height data from LiDAR. In G. Priestnall & P. Aplin (Eds.), Proceedings of the GIS research UK 14th annual conference, 5–7th April 2006 (pp. 306–312). Nottingham: The University of Nottingham. Barr, R. (2005). Address referencing and the future of small area population data. Census: Present and future, University of Leicester, November 2005. Available from http://www.geog.soton.ac.uk/users/martindj/cenprog/ CPF/CPFbarr.pdf. Booth, A., & Amato, P. (1992). Divorce, residential change and stress. Journal of Divorce and Remarriage, 18, 205–213. Cho, C. J. (1997). Joint choice of tenure and dwelling type: A multinominal logit analysis for the city of Chongju. Urban Studies, 34, 1459–1473. Day, L. L. (2000). Choosing a house: The relationship between dwelling type, perception of privacy, and residential satisfaction. Journal of Planning Education and Research, 19, 265–275. Elaksher, A. F., Bethel, J. S., & Mikhail, E. M. (2003). Roof boundary extraction using multiple images. Photogrammetric Record, 18, 27–40. Gallo, J. G., Lockwood, R. C., & Rutherford, R. C. (2000). Asset allocation and the performance of real estate mutual funds. Real Estate Economics, 28, 165–184. Hamelink, F., Hoesli, M., Lizieri, C., & MacGregor, B. D. (2000). Homogeneous commercial property market groupings and portfolio construction in the United Kingdom. Environment and Planning A, 32, 323–344. Lake, I. R., Lovett, A. A., Bateman, I. J., & Langford, I. H. (1998). Modelling environmental inXuences on property prices in an urban environment. Computer, Environment and Urban Systems, 22, 121–136. Longley, P. A. (2002). Geographical Information Systems: will developments in urban remote sensing and GIS lead to ‘better’ urban geography? Progress in Human Geography, 26, 231–239. Longley, P. A. (2003). Geographical Information Systems: developments in socio-economic data infrastructures. Progress in Human Geography, 27, 114–121. Longley, P. A., & Mesev, T. V. (1999). On the measurement and generalisation of urban form. Environment and Planning A, 32, 473–488.
S. Orford, J. RadcliVe / Comput., Environ. and Urban Systems 31 (2007) 206–227
227
Martin, D. (2006). Last of the censuses? The future of small area population data. Transactions Institute of British Geographers, 31, 6–18. Martin, D., & Higgs, G. (1997). Population georeferencing in England and Wales: Basic spatial units reconsidered. Environment and Planning A, 29, 333–347. Martin, D., Nolan, A., & Tranmer, M. (2001). The application of zone design methodology to the 2001 UK Census. Environment and Planning A, 33, 1949–1962. Murray, K., & Shiell, D. (2004). A framework for geographic information in Great Britain. The Cartographic Journal, 14, 123–129. Ong, S. E., & Koh, Y. C. (2000). Time on-market and price trade-oVs in high-rise housing sub-markets. Urban Studies, 37, 2057–2071. ONS (2004) Census: Local Authority population studies progress review. Available from http://www.statistics.gov.uk/downloads/theme_population/2001CENSUSLAPOPULATIONSTUDIES.pdf. Ordnance Survey (2006). OS MasterMap part 1: user guide. Southampton: Ordnance Survey. Available from http://www.ordnancesurvey.co.uk/oswebsite/products/osmastermap/userguides/docs/userguidepart1.pdf. Orford, S. (2000). Modelling spatial structures in local housing market dynamics: a multilevel perspective. Urban Studies, 37, 1643–1671. Ostrovsky, Y. (2004). Life cycle theory and residential mobility of older Canadians. Canadian Journal on Aging, 23, 23–37. Statutory Instrument (SI) (2000). The Building Regulations 2000. No. 2531. London: HMSO. Tranmer, M., & Steel, D. G. (1998). Using census data to investigate the causes of the ecological fallacy. Environment and Planning A, 30, 817–831. WAG (2005). Fuel poverty in Wales. Housing Research Report March 2005. CardiV: Welsh Assembly Government. Wolverton, M. L., Hardin, W. G., & Cheng, P. (1999). Disaggregation of local apartment markets by unit type. Journal of Real Estate Finance and Economics, 19, 243–257.
