Spatiotemporal historical datasets at micro-level for geocoded individuals in five Swedish parishes, 1813–1914

Abstract

This paper presents datasets that enable historical longitudinal studies of micro-level geographic factors in a rural setting. These types of datasets are new, as historical demography studies have generally failed to properly include the micro-level geographic factors. Our datasets describe the geography over five Swedish rural parishes, and by linking them to a longitudinal demographic database, we obtain a geocoded population (at the property unit level) for this area for the period 1813–1914. The population is a subset of the Scanian Economic Demographic Database (SEDD). The geographic information includes the following feature types: property units, wetlands, buildings, roads and railroads. The property units and wetlands are stored in object-lifeline time representations (information about creation, changes and ends of objects are recorded in time), whereas the other feature types are stored as snapshots in time. Thus, the datasets present one of the first opportunities to study historical spatio-temporal patterns at the micro-level.

Design Type(s)	data integration objective • observation design
Measurement Type(s)	Geographic Area Level Statistic
Technology Type(s)	digital curation
Factor Type(s)	geographic feature • Period
Sample Characteristic(s)	Sweden • human dwelling • building • wetland • road • stream • lake • railway

Machine-accessible metadata file describing the reported data (ISA-Tab format)

Background & Summary

Studying how micro-level geographic factors have influenced human living conditions over long time periods provides many new and important insights. However, in the field of historical demography studies have primarily been limited to the macro level due to a lack of individual-level and longitudinal geocoded historical databases to enable studies at the micro-level. Currently, only a few historical longitudinal demographic databases have geocoded individuals to more precise locations, such as streets or property units¹. In most longitudinal historical populations, individuals are linked to a less precise position such as villages, parishes or other larger administrative units. Most historical demographic databases that are geocoded on a more detailed level (e.g., ref. 2) only cover a snapshot in time, which is incompatible with longitudinal analyses.

The datasets presented here adds geographic information to the Scanian Economic Demographic Database (SEDD). The SEDD covers an area of five rural parishes (approximately 130 km²) in southern Sweden (Fig. 1). The SEDD includes detailed demographic and economic information from 1646 to the present and is the result of project collaborations between the Centre for Economic Demography (CED) and the Regional Archives in Lund³. The SEDD has been extensively used in historical research (e.g., refs 4–6). All individuals in the SEDD are followed from birth or in-migration to death or out-migration. The database contains continuous information on the family and household structure, and on dates of births, marriages, and deaths occurring in the parishes. Characteristics such as socioeconomic status, as well as causes of death, are known for the period 1813–1914. The primary sources for the SEDD are vital registers, catechetical examination registers and annual poll-tax registers. However, the SEDD population had not been previously geocoded on a detailed level. In a recent study, we geocoded a subset of SEDD with approximately 53,000 individuals for the time period 1813–1914 (ref. 7). This geocoding enabled the first longitudinal studies of geographic factors at the micro-level⁸.

Figure 1: The five parishes (Hög, Kävlinge, Sireköpinge, Halmstad and Kågeröd) covered in the datasets.

Three types of datasets for the five rural parishes are provided in this paper: 1) property units (which the individuals in the SEDD are linked to); 2) historical wetlands; and 3) additional historical geographic data, such as infrastructure and buildings. The primary sources of the geographic data are historical maps and cadastral dossiers that encompass the five parishes for the period 1757–1914, as well as textual sources, such as annual poll-tax registers. The property units and wetlands are stored in temporal representations of longitudinal object-lifelines⁹. That is, for each property unit, we know when it was created, reformed, and ceased to exist; for the wetlands, we know when they were drained. The reason for storing wetlands in object-lifelines is to enable accurate proximity analyses for the whole study period. Proximity to wetlands and still water may indicate exposure to various diseases such as water-borne diseases and malaria (until the 20th century, malaria was a problem in Europe, and wetlands likely provided habitats for mosquitoes that transmit malaria¹⁰). Therefore, the possibility of estimating the exposure to wetlands is necessary for researchers to understand causes of mortality in historical societies. These wetlands were gradually drained during the 19th and 20th centuries; thus, their lifelines were estimated to enable estimations of dynamic geographic variables. Besides, the drainage of wetlands substantially changed the landscape and likely increased the agricultural productivity of the land. Thus, it is possible to study how these interventions affected the individual-level economic conditions of the farmers. Finally, the additional historical geographic data (roads, buildings and water) were digitized from several series of historical maps as temporal snapshots at single points in time. Consequently, we have annual information about the shape of each property unit, which enable us to accurately trace the residential histories of individuals in time and space. Combined with the historical geographic data and other modern geographic data, such as elevation and soil conditions, the datasets can be used for spatio-temporal analysis, as well as the inclusion of geographic factors in longitudinal analyses of historical demographic data. Thus, they present opportunities to address novel research questions not only in the fields of historical demography and economic and agrarian history but also in many other fields, such as epidemiology, medicine and geography.

Methods

The creation of the datasets consisted of three process steps: (1) creating geographic snapshot data by georeferencing and digitizing historical maps; (2) transforming the snapshot property unit and wetland data into an object-lifeline representation using Supplementary data; and (3) linking individuals in the SEDD to the property units in which they lived. The digitized buildings, communications and streams remained in their snapshot representations.

Data sources

The maps were obtained in digital format from the Lantmäteriet (the Swedish mapping, cadastral, and land registration authority). The historical maps originate from four map series: land survey maps (LSMs), military topographical survey maps (MTSMs), topographic maps (TMs), and economic maps (EMs). The 150 digitized cadastral dossiers (CDs) were primarily used to document changes in the property units and to create an object-lifeline representation. As shown in Table 1 and Fig. 2, the maps were created at different scales and for different purposes. The LSM and CD maps primarily focused on documenting juridical borders, such as property units, but also documented buildings and land use, communications and the natural landscape. The MTSM and TM were constructed for military purposes and thus, focused on the natural landscape, land use, topography, communications and physical objects, such as buildings. The EMs documented property unit borders and the natural landscape, land use, communications and buildings.

Table 1 Number of map documents digitized from historical maps.

Figure 2: Four maps that show Videröra mansion at the border of Halmstad parish.

(a) LSM from 1831. (b) MTSM from 1812–1820. (c) TM from 1860. (d) EM from 1910–1915. (Source: ref. 17).

For supplemental textual data, we used the Swedish poll-tax registers that cover the study period, which contain the addresses, taxation value (of the property unit), and household and owner information of each person who had to pay taxes. These individual taxation values can be used to detect possible geometric changes in the property units and estimate the lifelines of the property units.

Georeferencing and digitizing historical maps

We georeferenced and digitized the historical maps and cadastral dossiers in Table 1. In the georeferencing process, we used orthophotos from 1940 and modern topographic web maps to identify common points in the historical maps, such as churches, terrain features, crossroads, and river junctions. The historical maps were transformed using a piecewise interpolation method (spline). Geographic objects, such as property units, buildings, roads, railroads and streams, were digitized as temporal snapshots from the maps. The georeferencing and digitalization steps were performed using ArcGIS Desktop 10.3 (http://www.esri.com/software/arcgis/arcgis-for-desktop). Historical and digitized wetlands from the MTSM series were received from the Swedish County Administrative Board of Skane.

Creating object-lifeline representation of the property units and geocoding individuals

In this step, we transformed the snapshot data into an object-lifeline model and linked individuals to the property units in which they lived. This method is described in ref. 7.

To create an object life-line representation of the property units we combined the digitized property units from the LSMs, CDs, and EMs with information from the poll-tax registers and cadastral dossiers. With this information, we learned when the property units were created, changed, and ceased to exist. The main aim of the geocoding was to establish links between the records in the poll-tax registers and the digitized property units, which was simultaneously performed with the estimation of the property units’ lifelines.

For the geocoding, we primarily used information from the Swedish poll-tax registers, which contains the addresses (property units) of each person who had to pay taxes. The Swedish poll-tax registers are on a household level, which indicates that only the head of the household is noted with a full name. However, each individual in the SEDD is linked to the families and households to which they belong, and most of these families and households have a correspondence in the poll-tax registers. In this project, we used a subset of the SEDD from 1813 to 1914, which contains approximately 53,000 individuals, 10,000 households and 184,500 annual poll-tax records, in which 90.4% of the individuals experiencing an event in the demographic data are linked to the poll-tax registers (96.2% for the period 1829–1914 when continuous population registers exist for all five parishes). The subset contains all the individuals living in the five parishes for the period 1813–1914. We began the geocoding in 1,813 as information about the individuals’ migration within and between parishes from this year and onwards is available. We stopped in 1914 due to personal integrity constraints.

The geocoding and creation of object-lifelines required extensive investigation for several reasons. For instance, before the land reforms (conducted between 1757 and 1849 in the parishes), all individuals lived in small villages and cultivated nearby scattered plots. After the land reforms, the self-owned farmers received a cohesive piece of land, to which they also moved. We denote these lands as property units (Fig. 3). Although the property units were usually devoted to agriculture, some of the units also contained forest lands. Throughout the study period, several of the property units were subdivided or partitioned into smaller units (consistent with rapid population growth). However, the property units did not receive new addresses; thus, multiple property units often share addresses. We denote the set of such units as an address unit (Fig. 3). With some exceptions, they are located close and adjacent to each other. The number of property units per address unit, as well as the total number of property units, was constantly increasing throughout the study period (cf. Fig. 4).

Figure 3: Property units and address units in 1871 in Halmstad parish.

Property units that share an address (identical colour in the map) constitute one address unit. For example, the two property units with the address Hasslebacken 07 constitute one address unit. Only the address units are labelled.

Figure 4: Changes of address units and property units for the period 1813–1914.

(a–d) Percentage of address units constituted by a given number of property units for the years 1813, 1850, 1880 and 1914; the values on the bars represent number of address units with the specific amount of property units. (e) Total number of property units for the years 1813, 1850, 1880 and 1914.

The geocoding on the address unit level was more straightforward as the poll-tax register contains annual information about the address unit for the head of the family. However, we also aimed to geocode on the property unit level. To mitigate the problem of units that share addresses, we used taxation values in the poll-tax registers combined with textual sources in the maps and cadastral dossiers to separate the units that share addresses. We also traced the owner histories of the property units when property units shared both taxation values and addresses⁷. For several records, extensive manual research was required to achieve a substantial number of reliable links.

Creating object-lifeline representation of wetlands

To create an object-lifeline representation of the wetlands in the study area, we combined information about 1) digitized historical wetlands from the MTSMs from 1812–1820 (provided by the Swedish County Administrative Board of Skane); 2) modern wetland and water data; 3) soil type data; 4) the EM from 1910–1915; and 5) data regarding registered joint drainage units before 1920 within the five parishes.

For each digitized wetland from the 1812–1820 MTSMs, the end date was estimated using overlapping joint drainage units. These drainage units were joint initiatives by several property unit owners with the purpose of draining a specific area and sharing the cost. These drained areas were registered during the 19th century and have been scanned and digitized¹¹. Thus, a drainage unit that covers a wetland indicates that the wetland was drained. Observations of wetlands and water bodies in the EMs from 1910–1915 were also compared with the digitized wetlands, the modern digitized wetlands and other water bodies. A soil type map¹², in which soil types such as peat and muddy sediments indicate historical wet areas, was also used.

Because of the uncertainty in several of the start and end dates of the wetlands, we also used intervals representing the minimum and maximum start and end date. For example, if a digitized wetland from the 1812–1820 MTSM was not observed in the 1910–1915 EM nor had any overlapping joint drainage unit, its maximum end date was set to 1909. Moreover, if a wetland was observed as being more similar in shape and size to a modern wetland, the maximum start date of the digitized modern wetland was set to 1910.

Figure 5 shows the property units and wetlands in the object-lifeline representation of the parishes Sireköpinge, Halmstad and Kågeröd. In the figure, the temporal dimension is displayed on the z-axis, and the height of each object represents its existence in time. Note that several wetlands are not visible in the figure because they are covered by property units (i.e., they disappeared before 1914). Figure 5b shows an example of the property unit Bångstorp 01, which existed for the period 1801–1885. Thereafter, it was subdivided, and a new property unit (also with the address Bångstorp 01) was created in 1885. The maximum end time of the property units is 1914, whereas the maximum end time for the wetlands is 2007.

Figure 5: Property units and wetlands (blue) in object-lifeline representations.

(a) The parishes Sireköpinge, Halmstad and Kågeröd (b) Enlarged area of some property units (including Bångstorp 01).

Data Records

The data records contain eight datasets on ESRI Shapefile format (Table 2); each contains four files with the extensions.shp,.dbf,.shx, and.prj. The datasets are stored in the Harvard Dataverse (Data Citation 1). The spatial reference for all geographic data is the official Swedish geodetic reference system Sweref 99 TM (EPSG: 3,006), which is a UTM 33N projection of the Swedish realization of the European Terrestrial Reference System 1989 (ETRS89). The attributes for each of the datasets are described in Tables 3–7. Empty cell values in the datasets represent missing information. In addition to the information in Tables 3–7, the data records contain a codebook (Codebook_SEDDGeo.pdf) that lists the classifications used in the categorical attributes.

Table 2 Overview of the datasets.

Table 3 Property Units (property_units_SEDD.shp).

Table 4 Wetlands in object-lifelines (Wetlands.shp).

Table 5 Buildings (Buildings_polygons.shp, Buildings_points.shp).

Table 6 Roads and railroads (Roads.shp, Railroads.shp).

Table 7 Streams and lakes (Streams.shp, Lakes.shp).

Links to the scanian economic demographic database (SEDD)

The geocoded demographic information from the SEDD is freely obtained from http://www.ed.lu.se/databases/sedd/sedd-public-access³.

To access the dataset, the users need to send an e-mail request to the Centre for Economic Demography (for reporting data use). The SEDD dataset is longitudinal and contains individual event information on the total population within the five parishes for the period 1813–1914 (cf. ref. 13). The structure of the dataset is a table in which each row contains the identifier of one unique individual and supplied attributes of the individual. The table also contains events of specific occurrences, such as birth, marriage, outmigration and death. Each individual has one or more episodes, which is the period of time when all variables connected to the individual are constant. These episodes are defined by start- and end-dates. Each time a variable changes a value; e.g., if a person moves to another household, a new row with a new episode is included in the table. The links to the geographic data are specified in the PropID attribute, and changes each time the individual is moving to another location. In addition, the PropLevel attribute specifies whether the links are on address or property unit level. The links in the geographic data are specified by the attributes puId and auId in the Property Units dataset. By using these link attributes in combination with date we can easily link individuals to the digitized property units, as well as link geographic context variables to the individuals in the SEDD.

Technical Validation

Positional accuracy

The evaluation of the positional accuracy of the digitized property units is based on a subset (7.5%) of all property units. The selection was made based on the criteria that the property unit borders have not changed until now. This selection enables us to use modern property unit borders obtained from the Lantmäteriet (the Swedish mapping, cadastral, and land registration authority) as reference data. This subset should be a representative selection in terms of positional accuracy. In the evaluation, two measures of positional accuracy are used.

The first measure of positional accuracy is based on the buffer-overlay-statistics (BOS)¹⁴ method. The basic idea with this method is to create buffers around a line to be evaluated and another buffer around a reference line, and then base the quality evaluation on intersections between the buffers (and their set complement). In our case, we created buffers around the historical property unit boundaries (HB) as well as buffers around the modern boundaries (RB). Then the following measures were estimated:

$$\begin{array}{}\text{(1)}& Q_{\text{BOS}}=\frac{\text{Area}(HB\cap RB)}{\text{Area}(HB\cap RB)+\text{Area}\left(HBRB\right)+\text{Area}(RBHB)}\end{array}$$

$$\begin{array}{}\text{(2)}& Q_A=\frac{\text{Area}(HBRB)}{\text{Area}(HB\cap RB)+\text{Area}\left(HBRB\right)+\text{Area}(RBHB)}\end{array}$$

$$\begin{array}{}\text{(3)}& Q_B=\frac{\text{Area}(RBHB)}{\text{Area}(HB\cap RB)+\text{Area}\left(HBRB\right)+\text{Area}(RBHB)}\end{array}$$

That is, Q_A is defined as the relative complement of RB in HB; and Q_B is defined as the relative complement of HB in RB. The quality parameters are computed for a set of buffer sizes to provide the buffer overlay statistics (Fig. 6). A high positional accuracy is indicated by high Q_BOS values and low Q_A and Q_B values. When using a buffer size of 80 meters, we include approximately 90% of the reference boundary lines.

Figure 6: Mean BOS accuracy results for the historical property units (HB) compared to the modern property units (RB).

The error bars represent the s.d.

The second measure for estimating the positional accuracy is based on the Euclidean distance between the historical property unit centroid and the centroid of the corresponding modern property unit. The non-normal error distribution of the property unit centroids may limit some of the usefulness of the statistics used in Table 8, such as s.d., mean, and the root mean square error (RMSE). The 95% RMSE was calculated after removing 5% of the outliers, which is a more representative measure for non-normal distributions¹⁵.

Table 8 Positional accuracy—property unit centroids.

Based on the result of the two quality measures, we conclude that the positional accuracy of the property unit centroids is approximately 10–15 meters, whereas the accuracy for the property unit boundaries is slightly lower.

It was more difficult to determine the positional accuracy of the wetlands. The RMSE of the georeferenced 1812–1820 MTSMs, from which most of the wetlands had been digitized, was estimated to be 30 meters. This estimation is based on a comparison of unchanged objects, such as churches, crossroads and residential buildings in the geocoded MTSM, with the reference data World Imagery web map¹⁶ (geometrical resolution between 30–60 cm). Apart from the geometric uncertainty in these well-defined unchanged objects, there are additional uncertainties in the mapping of wetlands; for example it is difficult to define borders of wetlands and also to define what a wetland is.

Match rate and geocoding level

Match rates for the geocoding on the property unit and address levels for the five parishes are presented in Table 9. The rate is measured in percent person-years for the individuals in SEDD. The geocoding period is split into two periods to show the differences in the linking levels before and after the land reforms had been conducted within all parishes. Note that the geocoding match rate has been updated and improved in this paper compared with a previous study⁷.

Table 9 Match rates for each geocoding level and parish.

Spatio-temporal topology validation (verification of overlapping polygons in space and time and verification of empty spaces)

A spatio-temporal topology validator for the property units is implemented in Python (https://www.python.org/), version 2.7, using the ArcPy package for ArcGIS Desktop 10.3 (http://desktop.arcgis.com/en/arcmap/10.3/analyze/arcpy/what-is-arcpy-.htm). One main requirement is that property unit polygons should not overlap in space and time, that is, two polygons that cover the same area should not simultaneously exist. A pseudo code for this topology check is as follows:

For polygon i1, polygon i2

Check IF (i1 overlaps i2

OR i1 contains i2

OR i1 is within i2)

AND i1.eDate >= i2.sDate

AND i1.sDate <= i2.eDate

The ‘sDate’ and ‘eDate’ represent the start and end year; thus, the topology is evaluated for each year. Polygons that overlap are corrected by either re-evaluating their object-lifelines or adjusting their borders. The former is performed when errors are observed in the estimations of the start and end dates of the property units; the latter is conducted when minor overlaps are caused by non-perfect geometries.

The study area should not contain empty spaces, which appear when one polygon ceases to exist before another overlapping polygon starts to exist. To fill these empty spaces when no digitized geometries are available for a specific time period, new polygons are created by a subtraction method⁷. A total of 107 of the 1,159 property unit polygons were created in this manner.