Navigation – Plan du site
12ème Colloque Européen de Géographie Théorique et Quantitative, St-Valéry-en-Caux, France, 7-11 septembre 2001

Deriving supply-side variables to extend geodemographic classification

De nouveaux indicateurs d'offre pour étendre une typologie géodémographique
James Debenham, Graham Clarke et John Stillwell


Cet article souligne les avantages considérables d'une intégration d'indicateurs d'offre complémentaires dans les systèmes géodémographiques. Le terme "d'offre" dans ce contexte pourrait suggérer le nombre de consommateurs déjà présents dans une zone. D'autres variables comme l'offre d'emploi et de logements sont tout aussi importantes pour comprendre la demande. Nous suggérons que la description d'une zone par ses caractéristiques d'emploi donne une meilleure information sur la chaîne de revenu, tandis que l'offre de logements pourrait être un facteur crucial dans la formation des ménages qui à son tour retentira sur la structure démographique. Avec l'exemple de la région du Yorkshire et de Humberside dans le nord de l'Angleterre, nous indiquons comment une série de variables d'offre liée à la main-d'œuvre peut être réunie et utilisée à côté d'un ensemble d'indicateurs pour engendrer une nouvelle classification. Les modèles d'interaction spatiale sont ajustés pour sélectionner des variables qui tiennent compte de la taille des zones autonomes et de leurs bassins d'emploi.

Haut de page

Texte intégral


1Geodemographics involves the "classification of small areas according to their inhabitants" (Rothman, 1989, p1) and geodemographic systems are built upon the principle that two people who live in the same neighbourhood are more likely to have similar characteristics (and consumption behaviour) than two people chosen at random. The widespread commercial use of geodemographics as a mechanism for analysing consumption patterns may be attributed to the fact that they produce what Beaumont and Inglis (1989) refer to as "actionable information – information viewed strategically as an important resource or asset" (p. 587). One of the defining features of geodemographic systems hitherto has been the demand-side nature of the variables included in the systems, i.e., demographic and social characteristics of the population that result in different propensities to consume different products or services. As the most reliable and comprehensive source of socio-demographic information on small areas in Britain, the Census of Population has been very important. Whilst all the major proprietary systems now make use of non-census variables (such as county court judgements, credit applications or the electoral registers), geodemographic typologies are still almost entirely based upon the demand in a given area that is determined by the characteristics of the resident population on census night.

2Our contention is that the traditional systems pay no attention to the supply-side characteristics of the market that also vary spatially, and therefore we argue that existing systems might not be totally fulfilling the criteria of business need. We suggest that no indication is given of the economic, social or environmental conditions that might influence the consumption of retail goods and services in an area, let alone how these conditions might change. Yet it is clear that areas with good employment opportunities, housing provision and environmental conditions are likely to be areas where demand for goods and services is buoyant. In contrast, areas lacking in jobs, with low levels of housing development and poor environmental quality are much less likely to be areas identified as having the potential for business exploitation, unless they are likely targets for gentrification or policy-focused regeneration. Thus, it seems appropriate to extend the traditional framework of geodemographic systems to include other variables that indicate the potential of an area. This means drawing upon information about the level of employment, the provision of housing and the condition of the environment in an area, as well as details of the existing infrastructure of retail and service facilities. Furthermore, there is a strong argument that the suite of variables used to classify areas should include the dimension of change over time. The process of population decentralisation, for example, could render a once attractive looking investment totally unworkable in 15 to 20 years. Alternatively, spending power might be drastically reduced if an area suffers the closure or downsizing of a major employer or industrial establishment. These examples indicate that, in addition to static or cross-sectional measures pertaining to demand and supply, it may be appropriate to take into account temporal dynamics in the spatial system through the inclusion of variables such as population and employment change.

3In this paper, we focus on the labour market and utilise a suite of variables that represent characteristics of the local economy in the areas used for classification. Only static variables are defined in this instance. It is important to acknowledge that conventional classifications in business geodemographics, such as MOSAIC or Superprofiles, include census variables indicating the employment characteristics of the residential population in a small area. The variables are usually residence-based since the Census is a survey of households. However, the Annual Employment Survey (formerly the Annual Census of Employment) provides information about the jobs provided in a zone and therefore it becomes possible to define the characteristics of the zone according to the people who work there rather than those who live there. Workplace-based statistics indicate the nature and pattern of the jobs and the establishments that are available across the zonal system, regardless of where the employers or employees are living. Moreover, indices of specialization can be computed that provide summary measures of the employment structure of zones and give some indication of the extent to which a zone is dominated by one industrial sector.

4The addition of workplace characteristics therefore adds a further dimension to the classification of small geographical areas based on residential characteristics. However, whilst these ‘stock’ variables provide evidence of how the levels of population and employment variables vary spatially, they do not give any indication of the relationship between residential areas and workplaces. Some zones may attract large numbers of workers from across a wide area ; other zones may provide jobs in their workplaces for those living within the locality ; other zones may have no dwellings or no workplaces at all. Consequently, we suggest that a further set of variables should be derived that measure the interaction that each zone has with other zones in the same system. These variables reflect concepts such as zonal self-containment and catchment size, both of which require careful definition and modelling based on Special Workplace Statistics (SWS) from the Census of Population.

5Thus we contend that employment stock variables, and variables that indicate the degree of commuting interaction that zones experience with those that surround them, may prove to be measures that would enhance the segmentation that geodemographics systems deliver. The aim of this paper is to build upon these ideas by proposing a series of variables (Figure 1) that might be included along with those representing population characteristics when constructing a new zonal classification that is based on a postal geography, the spatial units preferred by businesses.

6Following a discussion in Section 2 of the residence-based variables and the spatial units adopted, Section 3 of the paper focuses on employment, showing the types of data that are available for direct incorporation as well as a synthetic index that represents the extent of specialisation in the industrial structure of each zone. The use of postcode sectors as the spatial units presents some problems when the interaction variables are computed since the Census journey-to-work data sets are only available for flows between and within Census wards. In Section 4 of the paper, we therefore explain why spatial interaction models are used and how they are calibrated. Destination-specific distance decay parameters are mapped to show their spatial variation and used to estimate flows between postcode sectors that are subsequently incorporated in the calculation of zone self-containment and catchment size.

7In Section 5, we present a classification system that includes both demand variables and those variables that represent the labour market and the interactions between residence and workplace zones. The K-means method of clustering is explained and the clusters that are identified are mapped and interpreted. Finally, some conclusions are contained in Section 6.

The spatial units and the demand variables

8Businesses usually require geodemographic systems that relate to postal geographies and the variables proposed here are held in the database at the postal sector level. Postal geography is different to census geography. It is based upon the administrative structure of the Royal Mail postal service, as opposed to the arbitrary statistical boundaries used by the Office for National Statistics (ONS) for the dissemination of small area census data and by Central Government for the demarcation of local government boundaries. Postal geography is now more widely used as a basis for study in the UK, especially in the private sector (Martin, 1995, Raper et al., 1992), although any analysis using a long time-series can be hampered by regular alterations of boundaries on the ground as development occurs. Postal sectors are areal units created by the aggregation of point postcodes, for which no geographical boundaries exist. The first five characters of a UK postcode (e.g., LS2 9JT) reveal the postal sector (LS2 9). Postal sectors aggregate into postal districts (LS2), and then to postal areas (LS). We use a selection of 784 postcode sectors whose centroids fall within the boundary of Yorkshire and the Humber (Figure 2).

Figure 2 : Postal sector and local authority boundaries in Yorkshire and the Humber

9A set of 51 demand variables were used (Table 1), many of which (variables 9 to 50) have been extracted from the 1991 SAS available from MIMAS. One key problem here is that census data is not held at the up-to-date postal sector level, but only for the postal sectors as they were on census night in April 1991 (MIMAS, 1999). The conversion of census data to modern postal geographies has long been recognised as a problem (Raper et al., 1992) but the situation has been facilitated by the development of a series of look-up tables from the All Fields Postcode Directory (AFPD) for moving between 25 different administrative and statistical geographies (Simpson and Yu, 2001). The SAS data required to compute these variables were obtained at the enumeration district level and converted using this method. All the variables were computed as proportions of the total population/households unless otherwise stated.

Table 1 : The suite of residence-based variables

10The only other non-census variable is the unemployment rate (variable 51). This is taken from the Computerised Claimant Count for July 1999, made available through NOMIS. Unemployment data is available from the census tables but this is now 10 years old so the Claimant Count data provides a more up to date picture. The variable here is measured as a percentage of the Experían mid-year estimates of the working age population.

11Whilst the Census and Experían data sets provide useful information about the population and household characteristics of postcode areas, data on residential property transactions are also available from the Experían database that indicate the numbers of houses sold per quarter and the average price at the postal sector level as documented by HM Land Registry. Data are available from the beginning of 1995 to Quarter 2 of 2000. These data on housing turnover and value (Table 2) may be considered as representing the housing market since they reflect both demand and supply. House prices are a clear indication of the buoyancy of an area yet they can also provide a reasonably good measure of affluence, hence their inclusion in the set of residence-based variables

12It surprising that this data is not used more in mainstream GDIS, although spatial variability in the data may preclude this. The data are only the average values of the houses sold in an area, not an accurate survey of the full housing stock. Some areas will see many transactions while others may have very few and this will affect the averages. The data is disaggregated by four housing types, detached, semi-detached, flats or maisonettes and terraced houses so variables 52 to 56 not only detail the total number of sales in the 12 months running from mid 1999 to mid 2000 but also the breakdown by housing type. Variables 57 to 61 do the same for the average value of those transactions.

Workplace-based Indicators

The supply side of jobs

13One of the key determinants of the potential of an area for investment is the structure of its labour market. Traditional geodemographic systems might include the number of employees in certain industries that live in a particular place but they do not take into account the actual provision of jobs in that area. It can be argued that employment is one of the key factors underpinning retail consumption levels because it provides the income that creates the opportunities to spend. The conventional view of the labour market is based upon the notion that the workforce sells their supply of labour to employers who demand this factor of production to make goods and services to sell for a profit. However, it may be possible to view this situation another way. Since income from employment provides the means to consume retail goods, then it follows that the population will demand the jobs that the workplaces supply.

14A large supply of jobs in an area may suggest a buoyant local economy while the specific industry in which these jobs are provided might indicate likely income levels. However, the dependence of an area on one industry might indicate a vulnerability to decline if national or even international economic conditions promote a collapse of such industries. The Office for National Statistics recently published figures showing that UK manufacturing output in the second quarter of 2001 fell by 2% from the previous quarter (ONS 2001b), prompting fears that the industry was heading into a recession. If such a situation were to arise and jobs were threatened in the same way as they were in the early 1990s then it would be reasonable to expect the investment potential of areas that are heavily dependant upon manufacturing to be depleted. The workplace-based variables proposed here are designed to test for such dependencies, while also building up a picture of the employment characteristics of an area.

15Basic labour market data is provided using the Annual Employment Survey (AES) that has been downloaded from the National Online Manpower Information System (NOMIS). Since 1995, the AES has been the principal source of data on employee jobs and is the result of an annual survey of a maximum of 125,000 business enterprises. This sample data is used to estimate employment in non-responding or non-sampled businesses (NomisWeb Reference Centre, 2001). Before 1995, the AES was known as the Census of Employment and was undertaken biannually. AES data for 1998 was obtained at the ‘Section’ level of the UK Standard Industrial Classification (SIC92). There are 17 SIC Sections in total but only 15 were included as two had no employees in the region (Figure 3).

16Two sets of 15 variables are derived from this data set : the proportion of jobs by SIC section in each postal sector (variables 62 to 76) and the proportion of employment units by SIC and postal sector (variables 77 to 91). These variables are designed to characterise the provision of labour in each postal sector. The number of employment units is used in conjunction with the number of jobs to try to highlight areas dominated by small or large employers. Areas with a large number of employees yet only a small number of employment units in a given SIC section might indicate a dependency on a large employer. However, it should be noted that this comparison is not perfect because the term ‘employment units’ does not refer to an individual business per se but the individual site from where a business operates. Some confusion is therefore possible in areas of high ‘job density’ such as industrial estates or business parks.

17In addition to these two basic indicators, the share of the regional workforce in each section is also calculated for each zone along with the total proportion of employment. This may help identify particularly large employment zones.

Share of employment in large companies

18In addition to the employee analysis, the AES also provides a workplace analysis that gives the number of data units and employees in each postal sector broken down by the size of the unit and by industry (SIC). Therefore a variable was included that measured the proportion of units employing over 300 employees, as were 13 further variables that quantify proportions of jobs in large employment units (over 300 employees) in each of the 15 SIC sections - with the exception of Agriculture and Fishing (sections A and B) which have no units employing over 300 employees in the region. These variables are therefore used to inform of the presence of particularly large employment units in an area, providing a more definitive indication than the relationship between the number of jobs and the number of units described above as it is possible that these might get separated in the segmentation process.

19Essentially this indicator will show where areas have a high dependency upon large employers in particular industries and will serve as a warning should that industry or company be in decline. Past experiences in the region have shown the devastating effects of the loss of a major employer to the local economy such as the decline of the steel industries in and around Sheffield, the closure of the coal mines in the Yorkshire coalfields and the loss of textile jobs in Leeds and Bradford. The downsizing of Vickers in Leeds is one example of employment change exerting a major influence on the local community. Such dependency upon a certain industry can be quantified using a pair of indicators that have also been built into the classification system.

Index of specialisation

20Stillwell and Palmer (1986) describe the index of specialisation as a variant of the index of dissimilarity, which is often used in population geography to measure such things as the residential segregation of ethnic populations within cities (Rees and Birkin, 1983). The index of specialisation for zone j can be defined as one of two indices, SPj(A) and SPj(B) that represent the characteristics of employment in the workplace. The first indicates the extent to which the structure of employment by sector k in zone j, Ejk, differs from the employment structure of the entire system (excluding zone j) and is defined as :

21The second index compares each zone with all the other zones (excluding zone j) and is derived from the first index as :

22These two variables are designed to pick out any zones that are heavily dependent upon one industry. To a certain extent, these two variables are just a compression of the basic labour market variables. However, it is possible that, with all the other variables, the clustering algorithm might be pulled away from the structure of employment provision in each zone. It is therefore hoped that these variables might draw the segmentation process back to areas with a high dependency on particular industries. This may serve to give warning of a dependence upon a particular industry and thus warn of the likely impacts of closures or downsizing. However, the impact of closure depends to a large extent on where the workforce lives.

Model-based Interaction Variables

23The employment variables may tell us a lot about the extent of the provision of employment in a given postal sector but there is very little indication of the level of interaction between the jobs and the population who might demand them. The impact of changes in the labour market on the surrounding areas can only be ‘guestimated’ using such indicators of employment provision because they are calculated with little reference to other zones and only use data on the geographical variation in the level of supply (Clarke and Wilson, 1994a ; 1994b). Bertuglia and Rabino (1994) maintain that a model-based approach can be employed to reduce this reliability on ‘one-dimensional’ indicators. By focusing on the performance of zones in a system rather than individually, the level of interaction and interdependence can be measured. This can be assessed by operationalising a number of ‘performance indicators’ developed by Clarke and Wilson (1994) and Birkin et al. (1994) that are specifically designed to analyse the journey to work flows in an interaction matrix.

Modelling the journey to work

24In order to compute these indicators, we need to know the volume of movement between zones in the system. No journey to work data for postal sectors exist so workflows in the region have to be simulated using a spatial interaction model. This model is calibrated using ward to ward flows obtained from the 1991 Census Special Workplace Statistics and the estimated beta values derived from this calibration used to generate postal sector to postal sector flows.

25Since we know the number of jobs provided in each zone, a simple residential location model can be used to allocate individuals who work in workplace zone j to residences in residence zone i. This residential location model takes the form of the attraction constrained spatial interaction model originally proposed by Wilson (1971). An attraction (destination) constrained model is used because the number of jobs in each zone serves as observed data on the total in-flow into each workplace. We want every job to be accounted for and this is ensured using the balancing factor, Bjk. The model takes the form :

where :
Tijk = flow from residence i to jobs in workplace zone j in industry k,
Wi = origin attractiveness factor (population at working age),
Djk = number of jobs in workplace zone j in industry k,
Bjk = destination balancing factor,
dij = distance from zone i to zone j, and
β jk = destination-specific distance decay value for industry k.

26The power distance function on the far right hand side of the equation measures the distance-decay effect of the attraction between two zones. In this case the physical distance between two zones was measured by calculating the Euclidean distance between the centroids of the two zones.

27Calibration was achieved using routines adapted from the Inter-area Migration analysis and Projection package (IMP) developed by Stillwell (1984) and explained in detail in Stillwell (1991). An iterative Newton Raphson automatic search routine is used to find a best-fit beta value on the basis of the difference between the average trip distance in an observed matrix and that predicted by an attraction constrained spatial interaction model. In this case the observed matrix was the ward-level journey to work flows.

28As we have shown, postal and census geographies are not coincident. However, postal sectors and wards are roughly similar in size, both demographically and physically. The main difference comes in the urban areas where postal sectors are much smaller than wards due to the density of postcodes. As these postcodes are mostly business addresses, these postal sectors have very small populations. Nevertheless, we made the assumption that the beta values calculated for a system of wards will be appropriate to use for predicting flows for a system of postal sectors.

29The commuting data in the 1991 SWS are dissaggregated into 11 industrial groups that are not exactly equivalent to the 15 SIC sections used so far in the creation of the basic labour market indicators. Therefore the SWS data and the AES data had to be amalgamated to create nine ‘interaction groups’ that adequately represented the characteristics of both datasets without combining dissimilar industrial groups. Figure 3 outlines the two disaggregated datasets and shows the newly created interaction groups.

Figure 3 : New interaction groups created from amalgamations of AES and SWS disaggregations

30The procedures derived from IMP allow the creation of destination-specific beta values for each of the 626 wards in Yorkshire and the Humber. This was done for each of the nine interaction groups. The spatial distribution of the decay parameters for interaction group iii (Manufacturing) is illustrated in Figure 4 and indicates how the frictional effect of distance on travel to work for behaviour is more significant in rural areas for manufacturing jobs. Lower distance decay parameters are calibrated for flows into workplaces in the larger cities.

31The ward-level beta values were assigned to the postal sectors using a point-in-polygon search in a GIS to determine which postal sector centroids fall within each ward. With these beta values assigned, the residential location model could be run to create the large array Tijk. The origin zone attractiveness factor (Wi) was created using the working ages of the 1999 mid year population estimates from the Experían postal sector data set. The number of jobs in each interaction group k in each workplace zone j, Djk, was obtained by aggregating the AES as detailed in Figure 3. The predicted flows between postal sectors can be used to compute indicators of self-containment and catchment size.

Figure 4 : Ward-based destination-specific decay parameters for Manufacturing, 1991

Self-containment and catchment size variables

32The degree of self-containment measures the extent to which jobs in zone j in industrial sector k are taken up by residents from within that zone. It is formally described as :

where Ejk is the number of jobs in industry k in zone j (equivalent to Djk in equation 3).

33Nine variables were created when the degree of self-containment was calculated for each of the interaction groups in each workplace zone. This indicator is specifically designed to look for postal sectors that are heavily dependent upon the supply of jobs in one industry. A high degree of self-containment will suggest a local workforce that will be very vulnerable to changes in that industry and may serve as a warning that such an area might not be so stable for investment under certain economic conditions.

34Very few zones will be totally self-contained. Most destination zones will see workers arrive from a number of different locations. Therefore, in order to predict the impact of changes in the labour market upon the surrounding area, it is important to know how far such effects will be felt. Estimating the average distance travelled by persons working in the zone will allow us to do this and give an idea of catchment size. Nine more variables were therefore calculated using the equation :

Preliminary Results : A New Geodemographic Classification

35The last two sections have proposed a number of variables to be classified using geodemographic clustering techniques in an attempt to build a more informative picture of small area populations, their labour markets and their linkages with other areas. A number of key small area data sources have been utilised, from the older, more established sources such as the SWS and the Census of Employment/AES, to the newly released Experían postal sector data. This section shows the preliminary results of a clustering exercise that makes use of the variables proposed in the previous section. A brief synopsis of the basic principles and methodology behind clustering is presented and some of the issues surrounding the choice of the optimum number of clusters are discussed. The clustering results in a new supply and demand-based area taxonomy, which is mapped and interpreted.

Basic principles of geodemographic clustering

36The overriding principle of geodemographic clustering is that systems should be mutually exclusive and collectively exhaustive ; a zone can only belong to one cluster grouping and every zone must be accounted for. The clustering process was undertaken using the procedures available in SPSS for Windows to classify an m × n data matrix (where m is the number of variables and n is the number of cases). An iterative relocation algorithm, otherwise known as ‘K-means’, was used as it is regarded as being less computationally intensive than the stepwise (hierarchical) approach also available.

37The number of clusters (K) is specified before the clustering process starts. The data is randomly split into K clusters and the distances between the cluster centres and the observation values in m-dimensional space are measured using the Pythagorean equation for Euclidean distances. Each postal sector is then allocated to the cluster it is nearest to. A new cluster centre is calculated by averaging the observations that fall in each cluster and the process is repeated until there are no more changes in the location of the clusters or some minimum average distance criteria are satisfied.

38The algorithm has two principal aims : to minimise the distance from the cluster centre for all observations belonging to a cluster and to maximise the distance between clusters. However, the clustering process depends upon a pre-emptive decision about the value of K before the classification is undertaken. There is no optimum value for K ; it depends entirely on the data being classified and also the user’s personal impression of how many typologies the segmentation process should create. It is therefore instructive to repeat the process with different values for K to find the best results. Therefore the process was run for values of K = 2 to K = 100. However, it can be argued that this only serves to add further subjectivity to the process because once all these cluster solutions have been created, the best one must be chosen.

Selecting an appropriate number of clusters

39Some of the uncertainty in selecting a value for K can be removed by monitoring the distance of the cases from the cluster centre. For each clustering solution, the average distance of each case from its cluster centre was computed and the results graphed. Figure 5 shows that the progression of change across the different schedules rises reasonably steadily until the around the 85th procedure (15 clusters) when it starts to increment a little more sharply. However, an analysis of the graph would suggest that the 8-cluster solution (circled in red in Figure 5) might be more appropriate as it is the final solution before a very rapid rise towards higher average distance values and, furthermore, represents a fall in value from the previous (9-cluster) solution.

Figure 5 : Monitoring the average distance from cluster centre with different values of k

40Another way in which the optimum number of clusters can be ascertained is by monitoring the cluster membership. While we would neither expect nor want an equal number of cases in each cluster, the better segmentations are ones that avoid having the majority of cases in one or two clusters and then a number of sparsely populated groups.

41Cluster memberships were checked for between 10 and 6 cluster solutions. The results of the 10 and 8 cluster solutions are presented in Table 3. We can see that both the solutions look reasonable, as there is a fairly good spread of membership values. However, both solutions contain clusters that have very small membership values. It happens twice in the 10-cluster solution and once in the 8-cluster solution. The problem is not avoided until the 5-cluster run, but by this time any valuable segmentation of the data set has been lost.

Table 3 : Cluster membership returns from the 10 and 8 cluster solutions

42This problem may be attributed to one of the failings of the K-means algorithm ; data outliers can seriously affect the results by drawing the cluster centres away from their most favourable locations. Clearly, in the 10 and 8 cluster solutions, there are cases that are so far away from the others in multivariate taxonomic space that the algorithm is forced to place them into their own cluster. Furthermore, because the schedule has defined a limit to K, the remaining cases must be attributed to a cluster that may be some distance away. This is why the average distance component in Figure 5 rises so sharply.

43There are two solutions to this problem. The first is to leave the cluster solution as it is and accept two crucial inadequacies ; firstly, that some cluster groups will be nearly redundant because they either contain single cases or too few observations to draw any relevant descriptions, and secondly, that the remaining cases have not been optimally clustered. The second is to remove the offending cases and reclassify the remaining data.

Table 4 : Basic demographic and employment statistics for the poorly clustered postal sectors in the 10 cluster and 8 cluster solutions

44Table 4 shows the basic demographic and employment statistics of the zones that are poorly clustered in the 8 and 10 cluster solutions. It is clear why LS1 8 presents a problem in both segmentations. With a population of just 11 it has very small demographic statistics. LS1 8 is a particularly small postal sector in the very centre of Leeds, just 1.2 hectares in area. It is a reasonably important zone for employment with over half of its 2,652 jobs being in SIC Section J (Financial Intermediation). Nevertheless, if the intention of this new clustering system is to give an impression of how the population is served by the labour market and to monitor likely population dynamics to assess the stability of investments there, it would be sensible to remove LS1 8 because there is no significant population. However, the removal of the other postal sectors is less easy to justify. They all have significant populations and significant numbers of jobs, the removal of such a zone from the information system would hinder the comprehensiveness of the final results.

45One way to decide which postal sectors to remove is through trial and error. The intention was to remove the erroneous postal sectors one at a time and to re-run the clustering process, monitoring the segmentation through the cluster membership. Table 5 shows the results of removing LS1 8 from the data set before clustering.

Table 5 : The impact upon cluster membership of removing postal sector LS1 8

46Table 5 suggests that the removal of LS1 8 has not been a total success. Both the 10-cluster and 8-cluster solutions have still failed to segment properly ; both still have clusters with very low membership. In the 10-cluster solution there are two clusters (cluster 3 and cluster 4) with 9 and 8 cases, respectively. The 8-cluster solution also has one cluster with very low membership ; group 6 has just five cases in it. Good classification systems do not have to have a high number of cases in each cluster. As we have said, it is less desirable to have homogeneity in cluster membership because it suggests that the cases have not been properly grouped together. In many ways, therefore it helps the classification if there are clusters that contain minority of extreme cases. However, in the same way, very low cluster membership is also undesirable because often there is not enough information to make accurate comments about the characteristics of that cluster. In fact, because it is likely that such low membership clusters are made up of extreme outliers in taxonomic space, it is possible that they may not have any similar characteristics at all. In such situations it may be plausible to label them as "pseudo-clusters", rather than clusters.

47It is for this reason that it is probably wise to reject the 10-cluster solution finally and accept the 8-cluster. With two potentially meaningless clusters with low membership, it can be argued that segmentation here is worse than in the 8-cluster solution, which only has one. Furthermore, if the 8-cluster solution has segregated the five "outlying" zones from the classification then this may mean that the rest of the cases may be more optimally clustered as the cluster centres are not being dragged away towards extreme cases. Therefore we can take this value of K forward and use it in our analysis of the cluster typologies.

Evaluating the clusters

48By comparing the characteristics of the clusters it is possible to determine their key features and build up a picture of the nature of the zones that fall into that category. Cluster labels and ‘pen portraits’ can then be derived. Pen portraits are small descriptive analyses of the clusters that draw upon their main identifiable characteristics. The proprietary systems use them to attach a real world context to the cluster labels and modify them to suit the particular application of geodemographics. Conventionally an ‘index table’ is produced that provides a convenient and simple means of comparing cluster diagnostics (Batey and Brown, 1995). Index tables compare the cluster averages for a given variable against the global average (standardised to 100) across all the clusters.

49However, this can be a rather time consuming task ; with 140 variables and 8 clusters there would be 1,120 means to evaluate. Furthermore, the mean value of the data can be unduly affected by ; the non-normal distribution of census data ; the size of the cluster (the number of zones it contains) and the use of percentages for most variables, which means that extreme values of 0 or 100% are unlikely to be reached. Cluster evaluation for this system was therefore performed using Z-scores derived from calculating the standard deviations that occur above and below the global mean as follows :

ZKm = (AKm – Bm) / Sm (6)

where :
AKm = cluster mean for variable m,
Bm = global mean for variable m,
Sm = standard deviation for variable m.

50Distinguishing variables will have a value that is larger than the global mean and the standard deviation categorisations offer an assessment of by how much. Typically one would examine 1, 2 and 3 standard deviations above and below the mean when evaluating the clusters. This method produces the same number of values as the index tables, yet by ranking the Z-scores for each variable in each cluster it is possible to pick out the major characteristics a little more easily.

Analysing and mapping the new geodemographic clusters

51Figure 7 shows the geographical distribution of the 8 clusters created by the K-means algorithm. An analysis of the Z-score tables and the raw data itself allows some basic ‘pen-portraits’ to be created that synthesise the sort of areas that have been grouped together in the segmentation process.

Figure 7 : The distribution of the 8 clusters of the new classification

Figure 7 : The distribution of the 8 clusters of the new classification


52The dissemination of the 2001 Census will prove an exciting time for the geodemographics industry as the old systems are brought up to date. The tradition of geodemographics remains strong and although attempts to integrate lifestyles systems and GDIS have so far proved unsuccessful (Harris, 1999 ; Birkin 1995), it is likely that they will continue to thrive in business applications.

53This paper has put forward a selection of variables that have enhanced the area taxonomies that are created using traditional demand variables. It has been argued that longer-term stability can be measured by including variables that measure the level of supply in addition to perceived demand as well as the interaction between demand and supply. Variables that relate to the provision of employment have therefore been proposed. In addition to this a further dimension has been added in the form of variables that are not solely reliant on percentage counts. Indices of specialisation have represented the structure of employment provision over and above the percentages and some model-based indicators have been used to measure the interaction between zones in the labour market.

54Figure 7 and the subsequent cluster descriptions reveal that an interesting new classification has been created. Debenham (2002) compares a classification created with the 51 residence-based variables to the Groups presented in GB MOSAIC, a proprietary geodemographic system marketed by Experían Ltd. The results suggest that the clustering methodology used here does indeed produce classifications that are similar to the existing systems on the market today. However this paper has shown how geodemographic classification can be extended to provide additional information about small areas and to answer more questions than traditional systems currently do. The small area characteristics now include measures of the level of employment provision in each small area and, importantly, the degree to which each zone is dependent upon particular industries and how they are likely to change. The new supply variables have all significantly added to the information provided by the demand data. Debenham et al (submitted) show how a similar extended classification that uses workplace-based variables compares with GB MOSAIC in its view of Selby, a town in North Yorkshire. Both classifications agree that the postal sectors there contain largely lower-middle income households. However, the new classification provides an interesting assessment of the likely stability of these income levels as the area is heavily dependent upon the mining and manufacturing industries for employment. This apparent vulnerability to change has recently been confirmed as it has been announced that the Selby mining complex is scheduled for closure by UK Coal plc. (Shutt et al 2002).

55There is a need to fully evaluate the contributions of the new variables to the cluster segmentation in addition to any assessment already provided by the Z-score analysis. It is possible to argue that some sort of factor analysis should be used on the data before it is clustered to filter out any redundant elements. The 140 variables used here have been used in a ‘melting-pot’ type experiment that has simply proposed a number of variables and clustered them. It may be possible to argue that these results need to compared with other classifications created from different combinations of these original variables Debenham (2003, forthcoming), however, takes this work further and shows that extended classifications of this nature can be built up gradually with different suites of variables added in a cumulative fashion before re-clustering. Among other methods, principal component analysis is used to filter out variables that may be redundant or are having a deleterious effect upon a classification and these variables are removed or amended before the next classification is created. The intention is to see if similar (and better) classifications can be created using a smaller number of variables. The results suggest that while the original residence-based variables may still be the most important cluster-formative variables, the new workplace-based variables have important cluster descriptive functions.

56There is scope for a considerable amount of further work here. For instance, there is a need to consider other supply-side variables that could be included to further extend geodemographics ; such as those that relate to retail services, infrastructure and proposed housing development. With this achieved, the system needs to be extended to create a national classification, in keeping with the practice of existing geodemographic classifications.

Haut de page


Batey, P. and Brown, P. (1995) From human ecology to customer targeting : the evolution of geodemographics, in Longley, P. and Clarke, G.P., (eds.) GIS for Business and Service Planning, GeoInformation, Cambridge.

Beaumont, J.R. and Inglis, K. (1989) Geodemographics in practice : Developments in Britain and Europe, Environment and Planning A, 21 : 587-604.

Bertuglia, C.S. and Rabino, G.A. (1994) Performance indicators and evaluation in contemporary urban modelling, in Bertuglia, C.S., Clarke, G.P. and Wilson, A.G. (eds.) Modelling the City : Performance, Policy and Planning, Routledge, London.

Birkin, M. (1995) Customer targeting, geodemographics and lifestyle approaches, Ch. 6 in Longley, P. and Clarke, G.P. (eds) GIS for Business and Service Planning, GeoInformation, Cambridge.

Birkin, M., Clarke, G.P., Clarke, M. and Wilson, A.G. (1994) Applications of performance indicators in urban modelling : subsystems framework, in Bertuglia, C.S., Clarke, G.P. and Wilson, A.G. (eds.) Modelling the City : Performance, Policy and Planning, Routledge, London.

Clarke, G.P. and Wilson, A.G. (1994a) Performance indicators in urban planning : the historical context, in Bertuglia, C.S., Clarke, G.P. and Wilson, A.G. (eds.) Modelling the City : Performance, Policy and Planning, Routledge, London.

Clarke, G.P. and Wilson, A.G. (1994b) A new geography of performance indicators for urban planning, in Bertuglia, C.S., Clarke, G.P. and Wilson, A.G. (eds.) Modelling the City : Performance, Policy and Planning, Routledge, London.

Debenham, J. (2002) Understanding geodemographic classification : creating the building blocks for an extension, Working Paper 02/01, School of Geography, University of Leeds (available on-line http ://

Debenham, J. (2003, forthcoming) The implications of small area population and household change upon household retail consumption patterns, unpublished PhD Thesis, School of Geography, University of Leeds

Debenham, J., Clarke, G. and Stillwell, J. (submitted) The estimation of self-containment and catchment size indicators for use in small area classification. Paper submitted to the International Journal of Population Geography

Experían Ltd. (2001) Experían Ltd. Postal Sector Data documentation URL : http :// (accessed 5/2/2001).

Harris, R. (1999) Geodemographics and the analysis of urban lifestyles, Unpublished Ph.D. thesis, University of Bristol.

Martin, D.J. (1995) Censuses and the modelling of the population in GIS, in Longley, P. and Clarke, G. (eds.) GIS for Business and Service Planning, GeoInformation, Cambridge.

MIMAS (1999) Census Dissemination Unit Web Page : SASPAC system files URL : http :// (accessed 8/2/2000).

NOMIS Reference Centre (2001) ONS Guide to Regional and Local Labour Market Statistics – Annual Employment Survey.

URL : http :// (accessed 20/8/2001).

ONS (2001) First Release : Index of Production, June 2001, ONS, London

Raper, J.F., Rhind, D.W. and Shepherd, J.W. (eds.) (1992) Postcodes : The New Geography, Longman, Harlow

Rees, P.H. and Birkin, M. (1983) Census-based information systems for ethnic groups : a study of Leeds and Bradford, Paper presented at the Regional Science Association British Section Meeting, University of Leeds, 7-9 September.

Rothman, J. (1989) Editorial, Journal of the Market Research Society, 31,(1) : 1-7

Shutt J, Henderson R, Kumi-Ampofo F, Thursfield D. 2002. The Selby Coalfield impact study, European Regional Business and Economic Development Unit, Leeds Metropolitan University, Leeds.

Simpson, S. and Yu, A. (2001) Updated UK Area Masterfiles : Full report of research activities and results, University of Manchester, Manchester. (available on-line URL : http ://

Stillwell, J.C.H (1984) IMP : a program for Inter-area Migration analysis and Projection : user’s manual (revised), Computer Manual 12, School of Geography, University of Leeds, Leeds.

Stillwell, J.C.H (1991) Spatial interaction models and the propensity to migrate over distance, in Stillwell, J.C.H and Congdon, P. (eds) Migration Models : Macro and Micro Approaches, Belhaven Press, London.

Stillwell, J.C.H. and Palmer, J. (1986) User’s guide to PACE : A Program for Analysis of the Concentration of Employment, Computer Manual 25, School of Geography, University of Leeds, Leeds

Wilson, A.G. (1971) A family of spatial interaction models and associated developments, Environment and Planning A, 3 : 1-32.

Haut de page

Table des illustrations

Fichier image/jpeg, 108k
Fichier image/jpeg, 84k
Fichier image/jpeg, 164k
Fichier image/jpeg, 64k
Titre (1)
Fichier image/png, 2,5k
Titre (2)
Fichier image/png, 2,5k
Titre (3)
Fichier image/png, 1,7k
Fichier image/jpeg, 128k
Fichier image/jpeg, 236k
Titre (4)
Fichier image/png, 1,6k
Titre (5)
Fichier image/png, 1,9k
Fichier image/jpeg, 48k
Fichier image/jpeg, 40k
Fichier image/jpeg, 36k
Fichier image/jpeg, 176k
Titre Figure 7 : The distribution of the 8 clusters of the new classification
Fichier image/jpeg, 168k
Fichier image/jpeg, 136k
Fichier image/jpeg, 156k
Fichier image/jpeg, 168k
Fichier image/jpeg, 300k
Fichier image/jpeg, 281k
Haut de page

Pour citer cet article

Référence électronique

James Debenham, Graham Clarke et John Stillwell, « Deriving supply-side variables to extend geodemographic classification », Cybergeo : European Journal of Geography [En ligne], Dossiers, document 223, mis en ligne le 18 septembre 2002, consulté le 20 novembre 2019. URL : ; DOI : 10.4000/cybergeo.1671

Haut de page


James Debenham
School of Geography, University of Leeds Leeds LS2 9JT, United Kingdom

Graham Clarke

School of Geography, University of Leeds Leeds LS2 9JT, United Kingdom

John Stillwell

School of Geography, University of Leeds Leeds LS2 9JT, United Kingdom

Haut de page

Droits d’auteur

Licence Creative Commons
La revue Cybergeo est mise à disposition selon les termes de la Licence Creative Commons Attribution - Pas d'Utilisation Commerciale - Pas de Modification 3.0 non transposé.

Haut de page