Service sector employment in Great Britain 1841–2011

ABSTRACT The most detailed published census data on occupations and industries from nine British censuses between 1841 and 2011 have been computerised, and original methods developed both to re-district them from constantly changing historical districts to the 380 Local Authority Districts reported on by the 2011 census and to re-classify them to the 20 ‘Sections’ of the 2007 Standard Industrial Classification. This large dataset is then used to reveal new insights into the long-run evolution of geographical unevenness in the British economy. Initial analysis reported here maps the total percentage in the service sector for 1841, 1931, 1991 and 2011, with original data for key sectors presented for 1841, 1931 and 1991. While counts for individual districts and ‘Sections’ must be used carefully, overall trends appear robust: the localities of Britain have not just been moving steadily towards economies dominated by the service sector, they have steadily been becoming more similar.


Introduction
How has Britain's economy changed over the last two centuries, and how has this affected the relative prosperity of different areas? A conventional narrative would emphasise the initial rise but then the long decline of mining and manufacturing, associated with the north and 'outer Britain' (Champernowne, 1938), and the rise of service sector employment associated with London and the south east. A nearconsensus among politicians of the need for 're-balancing' and 'levelling-up' to benefit those 'left-behind' industrial areas suggests the problem is a divergence in industrial structure.
It should be easy to test this, as Britain has been gathering data on how people and localities earn their livings since the first census household enumeration in 1841, but revisions to both reporting geographies and classifications greatly complicate studying long-run change; and besides, the evolving British space-economy is fundamentally complex. This paper is the first presentation of truly long-run change in Britain's local economies at sub-county scale, and our map necessarily focuses on just one dimension, the almoststeady growth of the service sector, from that first enumeration in 1841 through 1931 and 1991 to 2011. Justifying this focus, today three-quarters of British workers are in services, and services have employed more workers than manufacturing since 1931.
The underlying research is more comprehensive. Its starting point was table WP605 in the 2011 census outputs. This lists, for 380 local authority districts (LADs), the number of workers in each of twenty industrial 'Sections' as defined in the 2007 version of the Standard Industrial Classification (SIC2007; Office for National Statistics, no date), plus a more detailed seven-way breakdown of Section M, 'Manufacturing'. We then replicated both the geography and classification used in WP605 for earlier censuses to create a time series with data at, generally, twenty year intervals.
The census has long been an essential source for the economic geographer. However, this generally means presenting data from just the most recent census, or discussing change by comparing two adjacent censuses. For example, Coates and Rawstron (1971) compare just 1951and 1961while Manners et al. (1972 include some longer runs, but only for total population or total employment. Buxton and MacKay (1977) include standardised counts of employees by industry from 1931 to 1971, but only national totals (pp. 28-9). Similarly, Mitchell and Deane (1971) standardised national data from all censuses 1841 to 1921, to 1911 census categories.
The only study presenting geographically detailed and standardised industry data over a long period is Lee (1979), who worked with all censuses 1841 to 1971 but was limited in several ways. Firstly and most importantly, the most detailed geographical areas covered were counties, which was probably unavoidable without using the GIS-based methods discussed below but means the performance of town and country cannot be distinguished, while most of the major conurbations each forms part of more than one county. Secondly, Lee was still forced to separate results for Registration Counties, up to 1911, and Administrative Counties thereafter. The differences were substantial, so Lee was unable to create continuous time series. Lastly, Lee inevitably re-classified to the 1968 version of the Standard Industrial Classification, and could not include the censuses from 1981 to 2011.
Some of these limitations were overcome by Martin et al. (2014) working with Cambridge Econometrics. They constructed consistent time series for city regions, defined as groupings of local authorities, derived mainly from the annual Business Register and Employment Survey and its predecessors; but only for 1981 to 2011, and the data are not generally available.
This article mainly concerns sources and methods. We can address a more detailed geography than Lee and a much longer period than Martin et al because we previously reconstructed the changing administrative boundaries of Britain, including the c. 16,000 parishes. This is outlined in the next section, and the following section then describes our general methodology for redistricting from one geography to another.
The fourth section describes the specific methods used to re-district and re-classify the data for the three 'historical' data points in our map: 1841, 1931 and 1991. Our time series spans nine censuses, but the selected dates are sufficient to demonstrate the range of challenges and solutions.
The final section discusses the trends in service sector employment displayed on the Main Map.

Mapping Britain's historical boundaries
From 1971 onwards, the census provides detailed Small Area Statistics but this study is based mainly on the published census reports from 1971 and earlier. These reported statistics almost entirely for areas with a broader administrative or electoral role: for c. 630 Registration Districts between 1851 and 1911; then for c. 1500 Local Government Districts (LGDs) from 1911 to 1971; plus basic head counts at parish level.
Both district geographies were defined mainly as groupings of parishes, so historic boundary construction began by computerising all the parish boundaries on the second edition of the Ordnance Survey County Administrative Diagrams, published in the late 1900s at a scale of two miles to one inch; except that additional accuracy was achieved for London by also using 1:10,650 topographic maps. This initial mapping was then checked to ensure that the set of parishes exactly matched the parishes listed in table 5 in the Registration Areas reports of the 1911 Census of Population. This pivotal census reported on both Registration Districts and Local Government Districts, so linkage to the census listing enabled parishes to be assembled into both types of district for 1911.
Construction of digital boundaries for dates on either side of 1911 was by modifying the 1911 parish boundaries to reflect the inter-censal boundary change listings in census reports. Creating digital boundaries for dates post-1911 was straightforward, as the change listings were comprehensive and the Diagrams were revised at approximately ten-year intervals. Working backwards from 1911 was more difficult, but the first edition of the Diagrams appeared around 1900. The earliest printed set of boundary maps, titled Sanitary Districts showing Civil Parishes, appeared in 1888, but only at four miles to one inch scale, as exemplified in Figure 1.
The 1881 and 1891 Census reports provide extensive lists of the changes resulting from the Divided Parishes Acts of 1876 and 1882, which essentially created the Civil Parish system, but as these changes were mostly the elimination of 'detached portions', where part of parish A was entirely surrounded by parish B, knowing that such a portion existed is of little assistance in placing it on a map. Unpublished maps in the National Archives, in classes RG18 and IR105 were used to add as many of these areas as possible.
The above covers England and Wales. Scottish reporting geographies were different, and often limited to counties, four cities and selected other towns. Further details of boundary reconstruction are given in Gregory et al. (2002), while Southall (2012) describes the integrated system supporting the present research.

Redistricting a historical geography to a modern one
Reporting of the 2001 and 2011 censuses used a specially-defined geography of Output Areas which is substantially more detailed than any current administrative geography, which the census office aims to hold constant. This is important as one consequence of current procedures for statistical disclosure control is that the 2011 outputs include no data at all from earlier censuses reported for 2011 geography, as comparison with earlier reporting would sometimes enable characteristics of particular individuals and households to be inferred.
Most earlier census reports do contain such listings, and in particular tabulate populations at parish-or ward-level from the previous censuses. Their listings of inter-censal boundary changes generally enable calculation of reasonably good estimates of other characteristics from the earlier census for the later geography based only on the statistics in the reports. However, this approach cannot be applied to long-run analysis.
Our consistent method requires Geographical Information Systems (GIS) technology and digital representations both of the boundaries of modern LADs and of the various historical reporting geographies, as described above. Figure 2 helps explain the basic method: the thicker blue boundary line represents a single LAD, while the thinner black lines show boundaries at some past census. Clearly, the whole of historical Bigtown County Borough identified as 'A' on the diagram falls within the modern Unitary Authority of Bigtown ('F' on the diagram), so all of its 100,000 population gets assigned to the UA. In addition, 45 per cent of the area of Arcadia Rural District ('B') is now within the Unitary Authority, so assuming we know nothing about the more detailed distribution of population we assign the same percentage of Arcadia's population. This continues for the other three Rural Districts (Camelot 'C', Thule 'D' and Westernesse 'E') which partly lie within the area of Modern Bigtown. Our final estimate of Bigtown Unitary Authority's population is made up as shown in Table 1.
Obviously, assuming that the population of the historical districts was evenly distributed over their areas is problematic, but there are two saving graces. Firstly, we have gone to considerable effort to ensure that the historical reporting geographies are, wherever possible, considerably more detailed than the modern LAD geography. Where the 'historical' units are wards from between 1981 and 2001, the method mostly means simply assigning all data for a ward to the local authority which now contains it. Between 1911 and 1971, we generally have to work with whole local government districts, but for example the 1800 local government districts in England and Wales in 1931 are a substantially more detailed geography than the current 348.
Secondly, while most variables to be redistricted are available only at district level, we do have available for every census some limited data at Civil Parish-level, and for every census between 1881 and 1971, except 1901, we have parish-level digital boundaries. Our full methodology, therefore, involves first re-allocating the district-level data to parishes pro rata to the most relevant characteristic we know for the parishes: the total population, the male or female population, or the number of households; and only then do we reallocate from parishes to LADs pro rata to area. Pre-1881 our results are inevitably more approximate, because of a lack of parish boundary maps.
Although the above account breaks the redistricting process up into a series of steps, in practice they have been used to build a library of Geography Conversion Tables (Simpson, 2002), each of which converts directly from a particular historical geography to the modern LAD geography, using a choice of population or household weightings. Building these tables is computationally demanding, but the actual conversions are quick.
One final detail is that small inaccuracies in the historical boundary mapping, and different definitions of the coastline, might lead to significant populations in coastal areas being redistricted into the sea, and lost from the national population as a whole. Corrections have been built into the conversion table construction process to ensure that all historical data are assigned to a modern local authority.
The overall approach to both re-districting and reclassification taken here is to work with the most detailed possible data but then to rely on reallocation pro rata, usually to either area or population. We are exploring the application of the Bayesian areal interpolation method (Mugglin et al., 2000) to model within-area distributions, but pro rata methods will always be easier to communicate.

Re-classifying historical statistics of occupation and industry
Some census questions have a very limited range of possible answers, so demographic variables have been classified very consistently over long periods.
However, the presentation of information about how people and localities make their living has been subject to almost constant change, reflecting fundamental change in how the economy operates. These challenges and our solutions are best explored for individual years. Source tables within the census outputs are listed in Table 2.

1841
This was the first census to send enumerators door-todoor, and consequently the first to ask people their 'occupation'. Census officials had not prepared for the sheer range of answers, and the results were presented as alphabetical lists of particular occupations: our transcription includes 3654 distinct terms. 3646 of these have been assigned to one of the 89 Divisions in SIC2007; bizarrely, this includes '51 Air transport', as the 1841 tables include a single 'Aeronaut', in Hackney. 2310 occupations (63%) fall into SIC2007 Sections A, B or C (Agriculture, Mining and Manufacturing). Despite these being occupation labels, assignment to industry was generally unproblematic because so many people's jobs were about directly growing, digging up or making some physical commodity. These 'makers' then often also sold the same commodity, but are classified wholly into manufacturing.
The census geography of 1841 was similarly ad hoc, reporting on counties and the larger towns, which were often not well-defined administrative units. For example, one area is labelled as 'OLD SWINFORD PARISH, INCLUDING STOURBRIDGE TOWN, BUT EXCLUDING AMBLECOAT'. At least one town is listed within every county. In order to map and redistrict these data, a GIS was constructed tailored to this specific dataset based on the parish boundaries but involving significant guesswork. Fortunately, these  towns were generally so small relative to their modern extents that their data will all be reallocated to a single modern district.

1931
By the early twentieth century, more complex organisational structures required a distinction between individual occupations and 'employer's business', leading to separate 'Industry' tables presenting the latter. Motor transport meant more people were living in a different district from where they worked; and selling commodities was increasingly separate from making them. However, the census was slow to adapt, so although the 1911 and 1921 censuses reported industry data, they are limited and problematic. The 1931 industry tables are more detailed and a good basis for re-classification, but unlike both 1921 and 1951 (and later), journey-to-work data was not gathered so the industry tables are based on place of residence. While the 1931 industrial classification is detailed, 285 of the 439 detailed 'groups' (65%) still fall into Agriculture, Mining and Manufacturing in SIC2007. Further, from 1911 onwards the census was reporting on the more detailed Local Government District geography which in 1931 comprised 1800 areas, including many small towns as separate Urban Districts. Consequently, 1931 Industry Table 2 provides a full listing only for counties and large towns, while Table 3 provides a 'condensed' listing for all other districts.
We fully computerised both tables, then for each 'group' in the full classification deducted the total for all large towns from the county totals, then allocated the remainder to the smaller districts pro rata to numbers in the 'condensed' category the group fell into. We, therefore, had at least an estimated number in all groups in all districts, which were used to compute totals for all SIC2007 Divisions for all 1931 districts, which were then reallocated to 2011 districts by the vector overlay method described above.

1991
It might be expected that 1991 was less problematic than earlier years, but significant statistical archaeology was still required. Firstly, it emerged that although 2001 and 2011 census outputs had been placed under    1971 Table 3, 'Industry and status by area of workplace and sex' for 'County, county boroughs, urban areas with populations of 50,000 or more, conurbation centres (10% Sample), in the Economic Activity County Leaflets series. Secondly, although the 1991 Small Area Statistics provide very detailed geography, for 10,766 wards, they cover only sixty different 'Classes', insufficient detail for a re-classification. Here the solution was to also draw on data from the 1991 Census of Production (Office for National Statistics, 2016), enabling us to combine the ward-level data on 'Classes' with district-level data on 335 'Activities', a more detailed level within SIC1980, to estimate numbers in each Activity in each Ward. The Census of Production data are confidential, requiring us to obtain a special license (Notice NTC/BRES14-P0829), but ONS agreed that the results of combining them with the Small Area Statistics then aggregating them to 2011 districts were non-disclosive and could be published. Because the classification dates from 1980, it still places a strong emphasis on traditional industries: 219 of the 335 'Activities' (65%) fall into Agriculture, Mining and Manufacturing, so no change from 1841 or 1931! Re-districting is still needed because there were significant changes to district boundaries between 1991 and 2011; for example, one of the authors' birthplace was moved from Malvern Hills district to Herefordshire. A final 1991 challenge is shown by Figure 3: 1991 ward boundaries were available from the UK Data Service, but the original download for Wales was missing many polygons. The Service was fortunately able to locate a complete version, but this shows how parts even of digital datasets can go missing when they are no longer regularly used.

Simplifying the classification
Initial re-classification from all historical censuses was always to the 89 'Divisions' of SIC2007, and also retained separate counts for males and females. However, our source for 2011, table WP605, is mainly limited to 20 'Sections'. Even these are too complex to include in a single graph, so Table 3 lists how these have been simplified to seven 'Sectors' of the authors' devising. That the modern classification differs drastically from almost all historical classifications is shown by Agriculture, Mining and Manufacturing now forming only 15 per cent of all Sections.
Data for all individual LADs can be accessed on Vision of Britain (www.visionofbritain.org.uk), for   Figure 4 graphs these seven sectors over all nine censuses for Great Britain. Over a century of slowly decreasing employment in agriculture and to a lesser extent mining, contrasts with the more concentrated decline in manufacturing after 1971, and the matching expansion of services, particularly business and public ones. The regional increase in business services in London and the South East identified by Johnston and Huggins (2018) between 1971 and 2005 exemplifies this trend. Our map discussed below combines the three 'services' sections.

Discussion
Our final results are presented in the Main Map as a sequence covering our first and last years, 1841 and 2011, plus two key intermediate years, 1931 and 1991. For each year we map our re-classified and redistricted data, showing overall trends, and for each year other than 2011 we also map a selected major historical category using historical boundaries.
We have already noted that 1841 provides an almost absurdly detailed set of statistical data but a particularly limited reporting geography, and this is shown in the map of 'Domestic Servants': because most rural counties had only county totals plus data for the county town, there are many areas where adjacent modern districts have identical rates. However, the most interesting patterns are in the manufacturing districts of northern England, and there many more towns were separately listed.
Although 'domestic servant' is a specific occupation listed by the census, nationally servants formed 57 per cent of the broader service sector and 16 per cent of all employment, and locally it could be a far higher percentage of all employment: 43 per cent in Leamington Spa, 37 per cent in Westminster, 36 per cent in Cheltenham. These towns are not easily identified on the map, but it confirms the service sector's concentration into the south east. This sector has often been neglected simply because 80 per cent were women.
The expansion of the service sector is often presented as a hollowing-out of the economy, and that a new expansion of manufacturing is needed to 'rebalance' the economy. However, the extreme focus on manufacturing we find in Britain's manufacturing districts in 1841 clearly meant a poor quality of life for the inhabitants. Low numbers in services are obvious on both 1841 maps and, for example, Blackburn in Lancashire had 80 per cent of its 17,332 workers in manufacturing, and 65 per cent specifically in textiles; conversely, there were just 68 workers in education and 28 in health care.
Moving on to 1931, we focus on financial services, defined as Banking and Bill Discounting Houses (Group 684 in the 1931 classification), Insurance (685), Money Lending and Pawnbroking (686), Building Societies (687), Stock Broking and Jobbing (689) and 'Other Finance' (698). Unlike domestic servants in 1841, the 307,000 finance workers were only 4 per cent of all workers in services, 70 per cent of whom were in consumer services, but they had a particularly strong geographical pattern, with 42 per cent of all these workers in Britain located in London and the five adjacent counties. Because the 1931 census did not record work places, the highest concentrations are in various outer London suburbs: about nine per cent of all workers in each of Southgate (Middlesex), Coulsdon and Purley (Surrey), and Wanstead (Essex). The 1931 map for all services continues to show very low percentages in the manufacturing districts, and generally high percentages in the south east. The 2011 districts with the highest 1931 percentages of workers in all services were Rushmoor (86 per cent), Surrey Heath (75) and Gosport (74), which reflects how the military dominated public sector employment. This was apparent in 1841, where Gosport and Medway (Chatham) were in the top ten districts for services, and was still a factor in 1991, where the ward-level map shows high values for Public Services around Salisbury Plain. The two wards with the very highest values are Cottesmore in Rutland, with an RAF base, and North Tidworth in Kennet, with an army barracks.
While the maps based on the original historical data inevitably each have their own keys, the sequence of maps for redistricted and reclassified data use a common set of class intervals. Evidence of forty years of regional disparity of growth in different industrial sectors (Gardiner et al., 2013) and the influence of the pre-existing regional industrial structure (Johnston & Huggins, 2018), is shown here to extend further back in time. Given the national pattern, it is inevitable that the maps show a steady increase in the percentage working in services in almost all areas. What is more surprising is that they also show a steady reduction in the variation between districts: the population standard deviation of the percentage in services is 9.83 in 1841 (the unweighted average percentage is 30.99), rising to 12.84 in 1931 (average 42.37), but then falling to 8.90 per cent in 1991 (average 60.27) and again to 6.82 in 2011 (average 73.31).
Although census taking changed very little between the first household enumeration in 1841 and the very recent move online, the large and inevitable changes in both the classifications and geographies used to report the results make long-run comparisons of these inherently problematic. So long as reports had to be printed, there was an unavoidable trade-off between providing very detailed statistics for a few areas, or providing broad-brush counts for many small areas; and today disclosure control considerations create similar trade-offs.
However, while the detailed results for particular sectors in particular districts are often questionable, our Main Map is based on a very large body of statistics, and the years included only in the long-run graphs support the same broad conclusions. Firstly, even in 1841 almost as many workers nationally were in services as in manufacturing (2 m vs. 2.2 m). Despite this, the expansion of services is very real, but that was not through services growing disproportionally in the south-east. Instead, over the last hundred years the localities of Britain have been becoming steadily more similarand more like London, which has always been dominated by services. Martin et al. (2014, p. 39) similarly find a decline in local specialisation over 1981-2011. In other words, if the old industrial regions have been left behind, it is not because they have failed to transform their economies: they have changed much more than London has. Rather, their particular role in a national economy based almost wholly and almost everywhere on services has failed to bring prosperity.

Software
The data were held in a Postgres database, and all calculations were made by SQL scripts running within the database, using the PostGIS extension for spatial operations. All scripts have been retained, creating a full audit trail. Each component map was first created as a spatial table within Postgres, combining polygons and statistics, then exported to ESRI ArcGIS for actual plotting.