Stan Openshaw, Marcus Blake School of Geography, Leeds University
Colin Wymer, Department of Town Planning, Newcastle University
The explosion of spatial information occasioned by the GIS revolution and the ease by which individual databases can now be aggregated to small zones (i.e. census ED's, postcode sectors, wards) emphasises the importance of being able to simplify the resulting multivariate complexity. The Small Areas Statistics (SAS) can be regarded as providing a good example of a generic problem of world wide significance. An ability to apply a multivariate classification procedure to reduce census data, for the 145,716 1991 census EDs in the UK, to a relatively small number of major types of residential area is extremely useful. For instance, most commercial geodemographic classifications segment Britain's residential areas into about 50 types based on a mix of census and non-census data for large numbers of small areas as a means of adding value to data that otherwise could not be used in an area profiling and targetting context.. The objective here is somewhat simpler in that only census data are of interest, although the challenge is harder in that the aim is to provide for research and academic purposes the best possible classification of 1991 census EDs for Britain. Previously only the 1981 Super Profiles classification (obtainable from the Essex Data Archive) was freely available for academic purposes.
In general, the quality of any area based census classification reflects three major factors. First, the classification algorithm that is used, second, the manner and extent to which knowledge about socio-spatial structure is used and is represented in the classification, and third, the sensitivity of the technology to what can be condensed to be the geographical realities of the spatial census data classification problem. In much previous classification research the quality and utility of the end product has been regarded as being critically dependent mainly upon the performance of the classification algorithm that is employed. Under this assumption, it is quite reasonable to believe that better classifications can only be produced by the application of improved classifiers; see for instance, Openshaw and Wymer (1994) and Evans and Webber (1994). There is, however, one major flaw in this approach; namely, there are limits to the degrees of improvement likely to be possible solely by developing or using improved classifiers. Maybe the possibility exists to become more intelligent in the way these classifications are developed and thus seek a quantum leap in performance rather than just a marginal gain of dubious value; Openshaw (1994a). Indeed, this goal of seeking to inject more intelligence into the total classification process rather than to continue down a narrowly focused purely classifier algorithm route is very important. It may also help explain the anomaly whereby the differences in end-user performance between a hastily cobbled together classification and one rigorously produced after a massive expenditure of research effort are not particularly noticeable (see Charlton et al, 1985). In short, it has long been apparent, but perhaps not sufficiently clearly recognised, that the really critical limitation on the performance of small area census classification is not the classification algorithm per se but the other steps in the process. Developing Artificial Intelligence based classifiers is only part of a wider process of being more intelligent in how we go about building better spatial data classifications.
The classification of census data usually involves the following steps:
Step 1. Decide upon the purpose of the classification
Step 2. Select variables and data to meet this purpose
Step 3. Apply a classifier to the data
Step 4. Evaluate the results and select the most appropriate number of clusters
Step 5. Label the clusters
Step 6. Embed the results in some easy to use end-user system that allows the classification to be linked to the postcode geography.
It is noted that the critical stages are highly subjective and operational decisions made there may substantially determine the utility of the results (Openshaw and Gillard, 1978). Clearly there is no easy way of turning the whole process into an optimisation problem since there is no single global function that can be optimised which would simultaneously meet all the goals. For example, there is no simple relationship between optimising a statistical measure of classification performance such as the within cluster sum of squares and the end-users perception of classification performance in a particular context. Indeed, the quality of the results is dependent in some little understood way on the performance of the classifier in step 3, on the usefulness of the results in step 6, on the extent to which the step 5 cluster labels "make sense" or correspond to what is known about socio-economic and demographic structure, and on the degree to which the variables and data selected in step 2 deliver results that are perceived to be useful at the end of the process. Note also that the perceivers of usefulness of the results may be a number of different users and not just one possessed of a homogeneous point of view.
One solution is to create 50 or 100 classifications based on different numbers of clusters (and variables), and then select whatever works best in a particular application. Various cross classification validation methods can be used to automate this decision, see Openshaw (1994b). However this `intelligent geodemographic' approach is probably more relevant to applied commercial uses of census classifications than in a research context, although the principles are transferable. It is also clearly an extreme response and that before commending its universal adoption it is necessary to investigate whether or not the needs of many research users would be better satisfied by developing a single good classification. This again re-emphasises the design of the classification algorithm (Step 3).
The conventional best practice is to employ a K-means nearest neighbour classifier to spatial census data that has been orthonormalised. Table 1 outlines some of the problems associated with this approach. The next question now is whether or not it is possible to greatly improve this technology? There are three ways of becoming smarter in developing a census classification.
(1). Improve the classification procedure by switching to superior algorithms, for example, simulated annealing approaches and neuroclassification methods;
(2). develop a means of incorporating knowledge into the classification process; and
(3). discover how to make the cluster assignment or allocation stage more sensitive to the nature of the task.
Until fairly recently, insufficient computing power was available to allow much or any progress to be made on any of these themes; for instance, a key requirement in designing new classifiers is computational tractability when faced with 145,716 or so zones to process. There has been, therefore, a tendency to simply continue using old methods dating from the early 1960's on larger and larger data sets. As a result this has lead to a failure to evolve new approaches that are really needed to ensure that the census classification challenge is being properly addressed. This has also been a failure to properly appreciate the complexity of the problem. The relative ease by which virtually any multivariate classifier when fed with virtually any set of census variables produces plausible results has tended to disguise the inherent difficulty of the task.
Some of these problems might be avoidable by using a more spatially sensitive census data classifier that can also provide a good representation of the natural levels of fuzzyness that seem to characterise spatial data in general and census data in particular. One of the key characteristics of census data is that EDs vary in size and thus the level of the precision and resolution of the data varies geographically. This aspect is often ignored. The problem is that this variation is not random but is spatially structured as it tends to reflect systematic urban rural differences and population density. This is partly due to a small number of problems which result in the most extreme results being found in the smallest areas which will tend to be more homogeneous and rural rather than urban. On the other hand, the largest EDs tend to be urban and often have very mixed characteristics, but their size produces data values which are much more accurate than is the case for small areas. The conventional classifier gives equal weight to each ED and thus will tend to focus on the more extreme results that will tend to represent best those areas for which the census data are least reliable. This is the opposite of what the geographer might wish to happen. As a result many of the larger urban EDs with mixed characteristics will tend to be "poorly" classified and there may well be a number of different possible allocations. Of course this may turn out to be a geographical fact of life with census data; however, it is worth investigating whether or not it might be reduced by using either much larger numbers of clusters in a conventional classifier or by switching to a more data sensitive classification process.
Another feature of UK census data concerns the mix of 100 and 10 percent data coding and the data blurring employed by the census agency to ensure confidentiality. These effects are partly handled by taking into account ED size variation but they also operate on an individual variable level. A small area may have highly accurate data for some variables and highly uncertain data for others. It would seem important that these data uncertainties are taken into account by the classifier rather than simply ignored.
It is with these factors in mind that Openshaw (1994) argues that the use of an unsupervised neural net based on Kohonen's self organising map (1984) provides the basis for a much more sophisticated approach to spatial classification that reduces the number of assumptions that have to be made and neatly incorporates many of the sources of data uncertainty; see Openshaw (1994a). A basic algorithm is as follows:
Step 1. Define the geometry of the self-organising map to be used and its dimensions. Here a grid with 8 rows and 8 columns is used.
Step 2. Initialise a vector of M weights (one for each variable) for each of the 8 by 8 neuronal processing units.
Step 3. Define the parameters that control the training process: block neighbourhood size, training rate, and number of training iterations.
Step 4. Select a census ED at random but with a probability proportional to its population size.
Step 5. Randomise the vector of M variable values to incorporate data uncertainty computed for each variable separately (optional).
Step 6. Identify the neuron which is "closest" to the input data.
Step 7. Update the winning neuron weights and those of all other neurons in its block neighbourhood or vicinity.
Step 8. Reduce slightly the training parameter and the block neighbourhood size.
Step 9. Repeat steps 4 to 8 a very large number of times.
If Step 4 is replaced by a sequential selection process and Step 5 is ignored then the algorithm is essentially the same as a K means classifier; with a few differences due to the neighbouring training which might well be regarded as a form of simulated annealing and it maywell provide better results and avoid some local optima. However, from a geographical perspective Step 4 is extremely important because it provides a means of explicitly incorporating spatial data uncertainty into the classification process. The method also provides a very natural means of handling cluster fuzzyness without having to impose an arbitrary metric; since the distance between the best and the next best neurons can be readily measured.
The simplicity of the self-organising map approach readily lends itself to ad hoc modification designed to improve the quality of the geographic representation offered by the classification. There are various ways of meeting this objective. the simplest is to select an ED, as in the standard algorithm described previously, but then to use a distance weighted average value for the k nearest ED's. This neighbourhood in geographic ED space is gradually reduced as the block neighbourhood in the self-organising map's topological space is also reduced, slowly over many millions of iterations. The logic is to incorporate some notion of local geographical neighbourhood structure into the classification. Here the geographic neighbourhood is limited to the 10th nearest neighbour of each ED.
Another way of attempting the same objective is to change the updating mechanism (see Steps 6 and 7 in the basic algorithm) to update neurons assigned to the k th nearest geographical neighbours of the ED being used for training at any particular instance, irrespective of whether these neurons are within the block neighbourhood of the winning neuron. Experimentation suggests that the `OR' rule is slightly better than the `AND' rule. Equally, restricting the neuron updating to only the geographical neighbour related neurons also yielded slightly poorer results. However, the resulting classifications seemed to offer levels of descriptive resolution equivalent to conventional cluster systems with many more cluster in them.
The principle disadvantage of neuroclassification concerns the computationally intensive nature of the method. If the technique is to properly handle and represent the 150,000 cases then large numbers of training iterations (Step 3) are required. In a census application an ability to represent the data is much more important than any generalisation to unseen data; since there is none. This requires many millions of training iterations; indeed runs of up to one billion iterations have been investigated. In practice this means that parallel implementations are required and a parallel supercomputing version is under development. However, it is worth noting that a conventional classification of 150,000 ED's may well require 200 passes through the data. This does not seem much but nevertheless it would be equivalent to 30 million training iterations and this conventional classifier is much harder to parallelise or vectorise in any worthwhile manner
Finally, Tables 2 and 3 briefly summarise some of the strengths and weaknesses of a spatial neuoclassification approach.
Following the census classification process as described in section 2, a set of 85 broadly representative 1991 census variables were derived; see Blake and Openshaw (1994) for a full description. These variables are listed in Appendix 1. A conventional iterative relocation procedure is then used to create flock of cluster systems with between 2 and 2,000 cluster in them. The CCP (Census Classification Program) software is described in Openshaw (1983) and is still available at MCC for research uses. Figure 1 shows a plot of the average percentage within cluster sum of squares for these classifications. The resulting is very smooth and would apparently confirm the general view that somewhere between 40 and 70 clusters is needed to provide a useful classification of Britain's residential areas; indeed most 1991 census geodemographic systems offer less than 60 cluster solutions. However this disguises the fact that some variables are much better represented than others; for example, variables such as older couples (35-54) without children and couples aged 55-74+ (denoted as A and B in Figure 1) are not well represented. It is this application or data specificity that Openshaw (1994a)'s Intelligent Geodemographic Targetting System (IGT/1) attempts to exploit. In the present context it merely means that a general purpose census classification with a fixed number of clusters will not satisfy all purposes equally well, but maybe in a research context with the cluster codes being used as a simple index summarising multivariate complexity, it is of a little consequence.
For current purposes the neuroclassifier is run with an 8 by 8 matrix of neurons for the 85 variables listed in Appendix 1. A total of 200 million training iterations were used. Step 5 was omitted to reduce the run-time on a workstation to 5 days. The labels that were derived for the resulting clusters are listed in Appendix 2. Comparisons with conventional classifications suggest that the differences appear to be slight in a qualitative sense. Quantitative comparisons are more difficult because it is not clear as to what the performance measure should be.
It seems then that any preference for a neuroclassifier requires both a significant amount of faith and a judgement about the relative merits along the lines of Table 1 to 3. This can be back-ed up by an assessment of whether the results are plausible. Figure 2 shows the distribution of the principle residential areas types in Sheffield. This stands up well to both local knowledge and previous research (Haining, Wise and Blake, 1992) on area types in Sheffield. For example, on a broad scale, the classic east-west division found in many industrial cities can be seen, with the affluent west and south-west of the city being dominated by the affluent and climbing categories while the city centre, and east of the city has more struggling and aspiring areas, see Figure 2.
An illustration of the apparent complexity of the census data classification process is provided by allowing fuzziness to occur in the cluster assignment stage. Openshaw (1989a & b) suggests that there is a particularly easy way of incorporating spatial data uncertainty into the spatial classification process. This illustrates the two principal sources of uncertainty; fuzziness in the geography space and fuzziness in the classification space. Traditionally, neighbourhood effects in geography are regarded as a spatial phenomenon in that people who live "near" to each other tend to share some behavioural characteristics despite other differences. In geodemographic classifications these effects have been implicitly exploited at the enumeration district scale; hence why these classifications are sometimes referred to being neighbourhood classifications. However this is an extremely crude representation of a highly complex and high variable spatial phenomenon. Geographers in the GIS era should really be able to do better than this and regard spatial neighbourhood effects in an elastic fashion rather than at a discrete ED geography space. Similarly fuzziness in the classification space should be exploited rather than ignored. Areas may differ by only very small amounts in the classification but be assigned to very different clusters. This is particularly important with census data because of lack of social homogeneity of the census ED and the tendency of the classification process to focus on highly distinctive minority characteristics of areas due to small number effects. As a result, it is likely that in many classifications the distinguishing cluster descriptions are minority features that are either created by aggregation effects at the ED scale or represent a profile based on the mixture of different individual household types. It is with great regret that in the UK there is currently no data available which can be used to measure these effects. The ecological fallacy problem needs to be handled rather than ignored in census classifications. Openshaw (1994a) provides a specification of a fuzzy geodemographics system to try and handle these problems. This can be demonstrated by using the results of the neuroclassification procedure.
The first aspect to consider is the structure of the K th nearest neighbour distances in the classification. Figure 3 shows the histogram of the number of different clusters "near" to each ED. It suggests that perhaps a surprising number of EDs are "near" to more than one cluster and could in fact be assigned with only a relatively small degree of error to a different cluster all together. Figure 4 shows a map that identifies the location of these "uncertain" EDs in Sheffield. Relatively few areas seem to be without some classification uncertainty. This measure of fuzziness is however only partial in that geographic neighbourhood or distance effects are excluded. It perhaps matters less if an ED can belong to two or more different clusters if these clusters are located nearby than if they are a long way off. The converse may also be important; that is neighbouring EDs should perhaps tend to belong to the same or similar cluster types.
To illustrate the further effects of fuzziness, Table 4 provides a cross tabulation of the census EDs in Britain by different levels of uncertainty in both the geography space and the classification space for a few illustrative cluster types. It is immediately apparent that a small amount of fuzziness soon introduces a number of other EDs that could be considered as belonging to each of the clusters. In fact it seems that the all or nothing nature of the conventional census classification is hiding considerable degrees of uncertainty. A surprisingly large numbers of EDs can in fact be assigned to different clusters. This may well reflect the heterogeneity of the census ED as a geographical entity, Openshaw (1984). However, not all this fuzziness is harmful to the classification as it can be used to improve the local fit of a classification by using geographic neighbourliness as a kind of smoothing operator. In fact, the first column in Table 4 shows the distribution of nearest neighbour geographic distances for ED's in the selected cluster. The distribution varies according to the nature of the cluster. Some are very closely related; for example council multi-storey housing and others much less so; for example poor semi-detached.
Finally, one of the objectives of the present research is to constructed a geodemographic profiling system that researchers can easily use. Using Microsoft Visual Basic an easy-to-used windows based system called GB Profiles `91 has been developed. This allows the classification of the underlying ED of every unit postcode in Great Britain to be accessed. Its primary use is to allow the academic community easy access to the results of the neuroclassification research.
It has two modes of use, an interactive single postcode search(or Single Search Mode - SSM) which instantly provides the cluster information on the screen and a multiple postcode search (Multiple Search Mode - MSM) which allows the user to batch process postcodes stored in a file. Some of the Windows associated with the MS mode are shown in Figure 5. The Search Setup Window allows the user to select a particular classification and determine which mode of operation to use, SS or MS mode. If MS mode is selected and a file loaded then this is stored in the list box of the Search Window where postcodes can either be added or removed. When these are processed a record is kept of postcode which have failed to be found and those which are duplicates. These statistics are provided in the Search Statistics Window. The results of the search are stored in a set of arrays which can be viewed on screen or saved to a file. Further information on the frequency distribution of the clusters found and a more detailed description of the clusters is also provided.
The underlying data structure is modular and this will allow different classifications to be loaded and then selected from the interface. Modules that provide photgraphic images and summary statistics are also developed.
The paper has argued that the use of a neuroclassifier provides a much more flexible and potentially superior means of generating census classifications. However, the substantially improved results are unlikely until it is possible to improve all aspects of the classification process so that the classification better represents both the complex nature of spatial data and incorporates meta knowledge that exists about the nature of residential areas in Britain. A start has been made but the really definitive results have yet to be produced.
References
Blake, M. & Openshaw, S., 1994, `Selecting census variables for use in classification research', Working Paper, School of Geography, Leeds University.
Charlton, M., Openshaw, S., & Wymer, C., 1985, `Some new classifications of census enumeration districts in Britain: a poor man's ACORN', Journal of Economic and Social Measuremnt, 13, 69-98.
Evans , N., and Webber, R., 1994, `Advances in geodemographic classification techniques for target marketing', Journal of Targeting, Measurement and Analysis for Marketing, 2, 313-321.
Kohenon, T., 1984, Self-organization and associative memory , Springeer-Verlag, Berlin.
Openshaw, S., 1983, `Multivariate analysis of census data: the classification of areas', in D. Rhind (ED) An Census User's Handbook, Methuen, London, 243-264.
Openshaw, S., 1984, `Ecological Fallacies and the analysis of areal census data', Environment and Planning A, 16, 17 - 31.
Openshaw, S., 1989a, `Learning to live with spatial databases', in M. Goodchild & S. Gopal (ED's) The Accuracy of Spatial Databases, Talyor & Francis, London, 264 - 276.
Openshaw, S., 1989b, `Making geodemographics more sophisticated', Journal of the Market Research Society, 31, 111 - 131.
Openshaw, S., 1994a, `Developing smart and intelligent target marketing systems: part I', Journal of Targeting, Measurement and Analysis for Marketing, 2, 289-301.
Openshaw, S., 1994b, `Developing smart and intelligent target marketing system', Working Paper 94/3. School of Geography, Leeds University.
Openshaw, S., 1994c, `Neuroclassification of spatial data', in D.C. Hewitson and R.G. Craneleds, Neural Nets: Applications in Geography, Kluwer, Boston, 53-70.
Openshaw, S. and Gillard, A. A., 1978, `On the stability of a spatial classification of census enumeration district data', in P.W.S. Batey (ED) Theory and Methods in Urban and Regional Analysis, Pion, London, 101-119.
Openshaw, S., & Wymer, C., 1994, `Classification and regionalisation', in S. Openshaw (ed), Census User's Handbook, Longmans, London.
Table 1: Problems with a conventional classification procedure
1. Use of a correlation matrix which acts as a linear
filter.
2. Use of principal component scores which use Z score
transformation of the data, emphasizing non-normal
distributions and affected by spatial dependency.
3. All or nothing nature of the classification assignment.
4. Single move heuristic which might become stuck in
sub-optimum locations.
5. Global function that is being optimised but with no
basis for knowing whether the results are better than
random.
6. Imposes arbitrary structure (viz. minimum variance) on
the data.
7. No way of handling data outliers and variations in
data precision.
8. No means of including prior knowledge into the
classification process
Table 2: Some of the benefits of a neurocomputing spatial classifier
1. Use of raw data removes the need for an
orthonormalising linear filter.
2. The self-organising nature of a Kohonen map allows
structure to emerge rather than be imposed from the
top.
3. Incorporation of data uncertainty into the
classification.
4. Simplicity and greatly reduced number of source code
lines.
5. Possible to incorporate prior knowledge into the
classification process making it more intelligent.
6. Fuzziness of the results are preserved in a
particularly easy to use form.
7. Reduction in importance of knowing precisely how many
clusters are needed.
8. Cluster interpretation is easier because the
classification takes place in the data space rather
than in some transform space.
9. Non-linear technology.
10. Less likely to be trapped in a local sub-optimum.
Table 3: Some of the problems of a neurocomputing spatial classifier
1. You need to prove that the potentially superior
technology yields improved results by comparison with
conventional benchmarks.
2. Extensive computer run times are needed requiring the
use of parallel supercomputing to adequately train on
large data sets.
3. A number of design aspects are entirely subjective, in
particular; the number of training iterations, the
architecture of the net, the updating process, and the
choice of metric for the classification.
4. The current absence of an intelligent framework for
using the results.
5. Lack of experience with the technology.
Table4: Neural net classification analysis of fuzziness
Cluster No 1: Multi-ethnic council tenants
Members = 2869
Cluster Similarity
Distances
Geog. Dist. 0.00 0.25 0.50 0.75 1.00 1.50 2.00 3.00
0.000 1 0 0 0 0 0 0 0
100.000 84 106 63 33 18 24 7 1
200.000 393 448 320 221 129 138 48 9
300.000 526 669 617 458 330 438 143 42
400.000 422 577 758 578 417 587 274 91
500.000 232 484 698 607 542 817 377 144
750.000 366 775 1461 1533 1405 2344 1211 541
1000.000 141 604 1095 1330 1403 2476 1525 633
2000.000 309 1313 2971 3908 4352 8432 5555 2601
3000.000 127 684 1841 2499 2568 4895 3486 2039
.gt.3km 269 3396 12412 12605 9490 12150 6648 4081
Cluster No 5: Poor semi detached housing
Members = 2920
Cluster Similarity
Distances
Geog. Dist. 0.00 0.25 0.50 0.75 1.00 1.50 2.00 3.00
0.000 0 1 0 2 0 0 0 0
100.000 27 45 42 29 13 15 4 1
200.000 104 187 181 127 70 78 18 3
300.000 192 228 293 267 172 181 67 9
400.000 195 285 399 339 295 361 115 18
500.000 155 228 364 374 362 463 153 52
750.000 339 516 881 1011 1003 1433 629 217
1000.000 155 365 656 936 1015 1731 927 357
2000.000 505 1107 2003 2738 3387 7123 4640 2337
3000.000 351 1011 1680 2116 2586 5793 4313 2650
.gt.3km 897 3808 7585 8437 8883 20244 16240 10546
Cluster No 6: Council multi-storey housing
Members = 3958
Cluster Similarity
Distances
Geog. Dist. 0.00 0.25 0.50 0.75 1.00 1.50 2.00 3.00
0.000 7 3 2 0 0 0 0 0
100.000 1374 1269 621 253 150 82 27 16
200.000 994 1692 1486 765 483 449 161 106
300.000 445 1172 1390 916 635 746 343 255
400.000 248 684 1054 858 621 794 467 334
500.000 151 511 821 685 597 794 484 433
750.000 209 786 1418 1364 1260 1825 1336 1381
1000.000 121 375 751 896 877 1516 1226 1516
2000.000 190 556 1087 1546 1717 3700 3395 5074
3000.000 83 196 467 700 845 1708 1755 3115
.gt.3km 136 525 1340 2265 2997 8032 9813 20093
Cluster No 19: Well off metro singles
Members = 3849
Cluster Similarity
Distances
Geog. Dist. 0.00 0.25 0.50 0.75 1.00 1.50 2.00 3.00
0.000 4 1 1 0 1 0 0 0
100.000 1641 883 477 177 146 85 16 3
200.000 1185 1156 1068 637 429 328 129 51
300.000 309 701 971 800 565 579 270 163
400.000 158 505 689 781 544 646 331 245
500.000 91 313 577 683 546 675 343 307
750.000 136 449 1128 1316 1151 1510 960 977
1000.000 58 295 689 884 906 1341 937 1055
2000.000 116 395 1113 1738 1918 3252 2921 3711
3000.000 47 161 393 629 812 1769 1690 2823
.gt.3km 104 416 1415 2590 3532 9139 11076 23331
Figure 1: A Plot of the average percentage within cluster sum of squares
Figure 3: The Distribution of the Number of Different Clusters "Near" to each ED
Figure 2: The Distribution of Major Classes identified using the Neuroclassification procedure
Figure 4: Distribution of "Uncertain" ED's within Sheffield
Figure 5: Layout of GB Profiles `91 in Multiple Search Mode
Appendix 1: The Variables used in both the Classification Procedures
Demographic Variables
Ref# Description 10%
1 resident persons in the 0-4 age grp.
2 resident persons in the 5-14 age grp.
3 resident persons in the 15-24 age grp.
4 resident persons in the 25-44 age grp.
5 resident persons in the 45-64 age grp.
6 resident persons in the 65-74 age grp.
7 resident persons in the 75-84 age grp.
8 resident persons in the 85+ age grp.
9 resident persons who are single
10 hhlds( with residents) with children, that
have two or more adults
11 female residents who are between 16 & 45
12 resident persons that are married
13 residents who are single parents
14 resident persons who are of pensionable age
15 persons aged 16+ who are students
Ethnic Variables
16 residents who are white 17 residents who are black 18 residents who are Indian 19 residents who are Pakistani 20 residents who are Bangladeshi 21 residents who are Chinese & others
Migration Variables
22 residents that moved last year 23 residents that are pensioner migrants
Housing Variables
24 all permanent hhlds that are owned outright
25 all permanent hhlds that are mortgaged
26 all permanent hhlds that are HA rented
27 all permanent hhlds that are LA rented
28 all permanent hhlds that are unfurnished
rented
29 all permanent hhlds that are furnished rented
30 all hhld spaces that are detached
31 all hhld spaces that are semi-detached
32 all hhld spaces that are terraced
33 all hhld spaces that are purpose built flats
34 all hhld spaces that are converted flats
35 all hhld spaces that are bedsits
36 all permanent hhlds with no central heating
37 all permanent hhlds with no/shared
bath/shower/WC
38 hhlds with residents which are overcrowded
39 hhlds with residents which are very
overcrowded
40 hhlds with residents which have more than 6
rooms
41 Number of rooms per hhld
42 Rooms per person
43 Average hhld size (rooms per hhld)
44 hhlds with residents with 2 or more cars
45 Average number of cars per hhld
Household Composition Variables
46 hhlds with residents with 2 or more e.a.
persons and no children
47 hhlds with residents with a single e.a.
person and no children
48 hhlds with residents with a married couple
49 hhlds with residents with children
50 hhlds with residents with children and no car
51 hhlds with residents with a single pensioner
52 hhlds with residents with a single
non-pensioner
53 hhlds with residents with more than three
adults
54 residents aged 16+ in hhlds who are aged
16-24 and are without children
55 residents aged 16+ in hhlds who are aged
16-24 and have children
56 residents aged 16+ in hhlds who are aged
25-34 and are without children
57 residents aged 16+ in hhlds who are aged
25-34 and have children
58 residents aged 16+ in hhlds who are aged
35-54 and are without children
59 residents aged 16+ in hhlds who are aged
35-54 and have children
60 residents aged 16+ in hhlds who are aged
55-74 or more
Socio-economic Variables
61 residents aged 16+ and over (employed & yes
self-employed) that are in SEG 1,2,3 & 4
62 residents aged 16+ and over (employed & yes
self-employed) that are in SEG 5 & 6
63 residents aged 16+ and over (employed & yes
self-employed) that are in SEG 8, 9 & 12
64 residents aged 16+ and over (employed & yes
self-employed) that are in SEG 7 & 8
65 residents aged 16+ and over (employed & yes
self-employed) that are in SEG 11
66 residents aged 16+ and over (employed & yes
self-employed) that are in SEG 16 & 17
67 residents aged 16+ and over (employed & yes
self-employed) that are in manufacturing &
mining
68 residents aged 16+ and over (employed & yes
self-employed) that are in agriculture
69 residents aged 16 and over who are self-
employed
70 residents aged 16 and over who are unemployed
71 residents aged 16 and over who are
permanently sick
72 residents aged 16 and over who are working
(employers or employees) women
73 residents aged 16+ and over (employed &
self-employed) that are women working in
manufacturing (metal etc. not other manuf.)
74 residents aged 16+ and over (employed &
self-employed) that are women working more
than 41 hours per week
75 residents aged 16 and over who work part-time
76 male workers
77 residents aged 16+ in hhlds who are female,
married and working
78 Proportion or residents aged 18 and over yes
with a (higher) degree
79 hhlds with residents with 2 or more adults
in employment
Health Variables
80 residents (S02) with LLI
81 residents (S02) economically inactive with
LLI
Travel-to-work Variables
82 residents aged 16+ and over who work at home yes
83 residents aged 16+ and over who go to work yes
by car
84 residents aged 16+ and over who go to work yes
by train/bus
85 residents aged 16+ and over who walk to work yes
Appendix 2: Labeling System for the 64 clusters identified using the Neuroclassification Procedure
Group Sub-group Name Cluster
#
Struggling Council Tenants Multi-ethnic council 1
with multiple tenants
social problems
LA rented Semis 24
Overcrowded Council 33
Housing
Council tenants in Tower 6 & 7
Blocks
Single Parents Council 29 & 34
tenants
Single Parents in Tower 28 & 30
Blocks
Unskilled Council tenants 45
Multi-ethnic, low Bangladeshi Areas 4
income areas
Indian Areas 38
Multi-ethnic Bedsit Areas 8 & 27
& 32
Poor multi-ethnic singles 62
Less Well-off Terraces 2 & 36
Terraces
LA rented terraces 10 & 35
Fading Industrial Industrial terraces 43 & 61
Areas
Industrial Council tenants 51
Less Well-off Pensioners Council tenants 17& 25
Pensioners & 31
Pensioners in converted 18 & 26
flats
Pensioners in HA rented 57
terraces
Aspiring Young Singles in Poor young singles & 3, 55
Flats Students & 60
Singles in PBFs 53
Better-off singles 14 & 54
Better-off Council Council Semis 13
Tenants
Rural Communities Rural areas 44 & 52
Armed Services Young Armed Services 12
Families
Establishe Semi-detached Semis 56
d Suburbia
Mortgaged Semis 63
Owner occupied Semis 5
Better-off Pensioner Migrants 15,
Pensioners 16, 23
& 59
Comfortable Middle Middle Class Suburbia 37
Agers
Wholly owned Semis 21
The average 20
Climbing Metro Singles Well-off singles in 14
Bedsits
Well-off singles in PBFs 19
Well-off singles in 50
converted flats
Academic centers Students in Bedsits 41
Prospering Wealthy Achievers Middle aged Managers 46 & 58
Well-off Middle Aged 9 & 47
Managers
Self-employed Managers 48
Educated Professionals 22
Wealthy Rural Rich Agriculturalists 11,
Communities 49, 39
& 64
Unclassifi 40
ed
42