1Disease mapping has been a crucial tool in visualizing the geographical distribution of diseases, with John Snow’s iconic ‘1854 cholera outbreak ’ map in London serving as the landmark example. In 1875, geographer Alfred Haviland paved the way for disease mapping beyond infectious diseases by shedding light on spatial patterns of cancer mortality in Victorian England and Wales [Haviland 1875]. Internationally, the practice of mapping cancer grew considerably during the twentieth century in line with increased public and medical awareness [Walter 2001; Koch 2022]. Current research shows that cancer is not only determined by internal risk factors such as sex, age, or faulty cell division, but also by external risk factors often related to space. For instance, exposure to environmental pollution, area deprivation, and access to medical care are contextual risk factors that influence cancer rates. Additionally, individual lifestyle choices and cultural preferences, such as smoking and dietary habits are also known to have a strong impact on cancer mortality [WHO 2024]. At present, cancer kills more than 27.000 people in Belgium every year [Statbel 2024]. To acquire a better understanding of current mortality disparities, several researchers have proposed to examine the historical roots of diseases [Link & Phelan 1995; Mackenbach 2020].
2However, for Belgium, comprehensive spatial cancer studies are scarce, particularly before the 1960s. Most of the Belgian research is limited in scope as can be inferred from an overview of frequently referenced cancer mapping studies in Table 1. Each of these studies is confined to research periods shorter than twenty years or uses varying geographic units, complicating comparisons over time [Delanghe 1972; Ramioul &Tuys 1977; Lemonnier 1981; Ryckeboer et al. 1983; Arbyn et al. 2002; Flemish Cancer Registry 2006; Grimmeau et al. 2007; Hagedoorn et al. 2016; Hagedoorn 2018; Renard 2019; Digital Atlas of Belgium 2023]. Since none of these studies cover the twentieth century entirely, our understanding of which specific areas were most affected in the past remains incomplete. Moreover, without information on the early and mid-twentieth century, we cannot detect rising (nor falling) cancer patterns over the long run. Such trends have been established in, amongst others, Sweden, England, Wales and Spain [Mackenbach 2020]. In this article, we therefore do not only examine the various reasons why this particular omission exists, but also address it by mapping cancer mortality throughout the twentieth century while simultaneously offering some alternative approaches for future research.
Table 1 : Overview of studies on cancer mapping in Belgium after 1960.
References: see bibliography.
- 1 Choropleth maps use varying shades or colors within geographic areas to represent data values, such (...)
3The publication of ‘Cancer mapping’ by the International for Research on Cancer (IARC) in 1989 was the first attempt of a critical evaluation of cancer atlases worldwide [Boyle et al. 1989]. More recently, international studies have recognized similar issues in cancer mapping as those identified decades ago. Concerns have been raised regarding data collection – such as cause-of-death certifications and changes in the nomenclature of the International Classification of Diseases which led to misclassifications – as well as the selection of appropriate geographical units [D’Onofrio et al. 2016]. Additionally, the problems of data quality, availability, and reliability have been repeated [Rican & Salem 2009; Sahar et al. 2019]. The value of choropleth maps has also come under scrutiny, with alternative methods like tile or geo-faceted maps suggested for a more accurate representation of the cancer burden [Kobakian et al. 2020].1 While not all these hurdles apply to the Belgian context, this article identifies five major challenges relevant to Belgium and which we have developed further: i) changing administrative boundaries, ii) small numbers, iii) evolving registration and classification practices, iv) data availability and compatibility, and v) visual presentation over time. Each section of this article addresses one of these specific challenges. By alternately using different boundaries, mortality indicators, and visualization approaches, we offer a multi-faceted spatial research approach towards cancer mortality in twentieth-century Belgium. Lastly, we discuss how these interventions might affect the visualization of cancer trends.
- 2 Mergers in the twentieth century: 2629 municipalities in 1910 to 2638 (1920), to 2671 (1930), to 26 (...)
4Flanders and Wallonia, the two principal regions of Belgium, were formally established as distinct geographic entities in 1980. Prior to this, Belgium functioned politically as a unitary state. Flanders, located in the northern part of Belgium, encompasses the provinces of West Flanders, East Flanders, Antwerp, and Limburg, along with the northern two districts of Brabant. The southern or Walloon region includes the provinces of Hainaut, Namur, Liège, and Luxembourg, as well as southern Brabant, specifically the district of Nivelles. The language border was established in 1962: Dutch in the northern provinces, French in the southern part of Belgium. After World War I, Belgium annexed a very small part of Germany, commonly referred to as the east cantons. For most of the twentieth century, Belgium was divided into nine provinces and forty-one districts or ‘arrondissements’ (see Figure 1). These administrative districts were subdivided into municipalities, the smallest administrative subdivisions of Belgium [Vrielinck 2000]. In 1910, there were 2629 distinctive municipalities. Due to various mergers only 589 municipalities remained in 1983 [Algemene Volks- en Woningtelling 1991].2
Figure 1 : Administrative map of Belgium, 1961 [Ronsijn 2014].
![Figure 1 : Administrative map of Belgium, 1961 [Ronsijn 2014].](docannexe/image/14597/img-2-small480.jpg)
5Flanders and Wallonia exhibited a persistent contrast in mortality rates over the last two centuries. These patterns are undeniably linked to socioeconomic disparities between the regions. Wallonia experienced early industrial and economic development in the nineteenth century due to the specialization in the steel industry and machine-building sector. Flanders, by contrast, faced impoverishment due to the collapse of textile (linen) industries [Vandenbroeke 1981, Vanhaute 2007]. Moreover, differences in infant feeding practices (early weaning), mothers’ educational levels, and water quality infrastructure impacted the contrasting mortality rates. For years, mortality in Flanders was significantly higher than in Wallonia, especially for infants [Masuy-Stroobant 1984].
6After the First World War, socioeconomic trajectories started to shift as Flanders became wealthier and economic growth in Wallonia stagnated. The 1960s and early 1970s symbolized rapid industrial progress in Flanders. For instance, the development of chemical and petroleum industries increased considerably in the flourishing port areas of Antwerp and Ghent-Terneuzen. At the same time, coal mines in Wallonia suffered from exhaustion and had to close. Additionally, the steel and machine-building industry struggled to compete internationally due its historically high wages. The situation only exacerbated after the oil crisis in the 1970s and led to deindustrialization and high rates of unemployment [Buyst 2009; Mérenne-Schoumaker et al. 2010; Smeyers & Buyst 2016]. As before, changes also arose on other levels. For example, in Wallonia people consumed fattier food (e.g., butter, fries) and ate, in general, fewer vegetables, fruit and fish. Combined with increased use of alcohol and tobacco in (eastern) Wallonia, this led to higher mortality rates than in Flanders [Grimmeau et al. 2010; Eggerickx et al. 2020].
7Life expectancy in Belgium, as in other Western countries, increased significantly over the past two centuries: from approximately 40 years in 1850 to 46.7 years in 1900, and 77.7 years in 2000 [Human Mortality Database 2024, Devos 2010]. However, life expectancy varied significantly at the district level. Around 1900, life expectancy at birth was 11 to 15 years higher in the Walloon district of Philippeville than in the Flemish district of Ostend. By the late twentieth century, this trend had completely reversed. In 2010, average life expectancy in the districts of Flanders was higher than in Wallonia, ranging between 4 and 6 years [Eggerickx et al. 2020]. The rise in life expectancy was accompanied by an epidemiological transition, in which so-called degenerative diseases (cardiovascular diseases, cancers) gradually replaced infectious diseases (cholera, tuberculosis, etc.) as the dominant causes of death [Omran 1971]. Drivers in this process were better hygiene, rising living standards, and medical progress. Historical demographer Isabelle Devos estimates that overall cancer mortality in Belgium rose from 4% in 1910 to more than 27% in 1995 [Devos 2006].
8Currently, cancer is the leading cause of death in Belgium and one in four causes of death can be attributed to cancer [Statbel 2024]. In Figure 2, we compare cancer percentages for all Belgian municipalities over five distinct periods. In the first half of the twentieth century, we observe a trend of high variation between the more than 2600 municipalities; the coefficient of variation, for example, was 1.08 in the early twentieth century and only 0.14 by the end which suggests declining variability. In the first three periods, we still note a relatively high amount of municipalities with no registered cancer mortality (in the first period Q1 is equal to 0%). By the 1990s the differences between municipalities had diminished significantly. However, it is important to consider that the number of municipalities also had reduced to 589, potentially misleading interpretations of variation. Aside from variation and outliers, the overall trend was one of increasing cancer mortality for all municipalities. The median percentages rose from 2.9% at the beginning to 25.3% at the end of the twentieth century, which is in line with earlier findings on the national scale.
Figure 2 : Boxplots of cancer percentages in relation to total mortality (men and women combined) for all Belgian municipalities between 1905-1991.
On the far left : cancer percentages for 1905-09-10. On the far right : cancer percentages for 1989-90-91.
Source : Le Mouvement de la Population et de l’Etat Civil, Belgium 1905-1971 & Demobel 1989-1991 (Descriptives in Appendix).
9Data on causes of death are mainly derived from Le Mouvement de la Population et de l’Etat Civil, often shortened to Le Mouvement. This register comprises yearly data on population statistics (births, marriages, migration, deaths) between 1841 and 1976. This register also lists causes of death differentiated by age and sex at the municipal level. Although the register was conceived in 1841, cause-of-death data is only available from 1851 onwards [Bracke & Vanhaute 2005; Preneel 2010]. Since then, every municipality in Belgium has been obliged to keep a cause-of-death register, alongside the civil registration of deaths. The cause of death was preferably determined by a physician, although this was not always the case in rural areas; family members or neighbors sometimes stated the supposed cause of death to the local authorities. in large cities ‘médecins vérificateurs’ were appointed to ensure the quality of death registration [Velle 1985; Devos forthcoming]. Nevertheless, the earliest records on cancer mortality by municipality (in the 1850s) are markedly incomplete and inconsistent. In the following decades, cancer was not included in Le Mouvement as a separate category.
- 3 Cancer mortality refers to the number of deaths caused by cancer within a specific period, usually (...)
10Aggregate cancer mortality data at the municipal level for Belgium are available from 1903 onward.3 Most likely infectious diseases were the more apparent threat. Since 1903, the category ‘cancer’ has been conditional on revisions following the International Classification of Diseases (ICD), on changes in coding practices and on the quality of diagnosis (see third section for more information) [Meslé & Vallin 1996]. Data on cancer per type (lung, breast, stomach, etc.) are not available before 1954 at the national level and not before 1969 at the municipal level.
11More recent data – after 1969 – are acquired through Demobel, the demographic database of Statbel, the official Belgian statistical office. Demobel contains individual-level data such as births, deaths, migration and causes of death, supplemented by data from the population censuses up to the present day. The mortality indicators calculated for this article are acquired by combining cause-of-death registers and population censuses, containing the age-specific composition of the Belgian population. The first Belgian population census was organized in 1846 and became a decennial tradition. Mandatory population censuses were held in the twentieth century: 1900, 1910, 1920, 1930, 1947, 1961, 1970, 1981 and 1991 [Lokstat 2023].
- 4 The age-specific census data are also based on the de iure population, which means that the nominat (...)
12An important difference between the historical cause-of-death registers in Le Mouvement and the more recent ones in Demobel is the fact that they do not refer to the same population. Le Mouvement registered events of the de facto population, which is the (f)actual population comprising registered and unregistered inhabitants present on the territory. In other words, it registers deaths at the place of death. In contrast, the more recent data in Demobel relate to the de iure population, the legal population consisting of registered inhabitants only. In short, it registers deaths according to place of residence.4 For the early twentieth century, this can have an impact on our results as specialized medical facilities were often centered in and around the major cities, which can artificially inflate the rates of a city. However, the impact should also not be overstated. Firstly, before the mid-twentieth century most people still died at home [Colombo & Molinari 2022]. Most likely, this was due to financial restrictions. The hospitals were also known for their lack of hygiene and subsequent high death toll. Hence, hospitals were to be avoided [Backs 2001]. The fact that a majority of people die in hospitals or care homes is also a fairly recent phenomenon. In 2001, about 60% of all cancer deaths in Flanders occurred in a hospital [Cohen et al. 2006]. And secondly, previous research indicated that mortality differences in large cities, calculated based on both the legal and the actual population, were minimal [Van Rossem et al. 2018].
13To compare patterns across time and space, we also require a historical geographic information system (GIS). For this, we rely on the previous work of the UGent Quetelet Center which has created a tailored historical GIS (HISGIS), in vector format, georeferenced to the Belgian Lambert 72 map projection. It includes more than 2600 municipalities, 41 districts and 9 provinces of Belgium. All the boundary changes in the administrative layout of Belgium between 1900 and 2000 are incorporated, and can be tailored to the chosen period. In this article, the different cancer mortality indicators are calculated and mapped using various and changing geographical units (municipalities, districts, provinces) using HISGIS.
14Navigating the intricacies of mapping cancer mortality through time and space involves grappling with the problem of altering administrative boundaries, which is illustrated by Belgium’s gradual transformation from around 2,629 in the early twentieth century to a mere 589 municipalities by the end of the century [Vrielinck, 2000]. This highly dynamic landscape prompts the question of which geographical unit should be used for comparing past and present cancer mortality: historical (i.e. contemporary) municipal boundaries, current municipal boundaries, or the boundaries of more aggregated administrative units such as districts (arrondissements) or even larger provinces. Which entity proves more useful for comparing cancer mortality over time? Past studies have not offered a clear-cut answer (see Table 1). Artificial constructions, such as smoothed maps, add another layer of complexity to this discussion as they have no recognizable boundaries. In Figure 3, we illustrate each of these options by plotting the crude cancer death rates for mid-twentieth century Belgium.
Figure 3 : Crude cancer death rates, 1947-1948-1949 (number of deaths per 10,000 people).
Top left: 2670 municipalities with historical boundaries, top right: 589 municipalities with the boundaries of 1991. Bottom left: 41 districts with contemporary boundaries, bottom right: smoothed map (spatial smoothing performed by Magrit – exponential function, span = 10 km, Beta = 2).
Source: Le Mouvement de la Population et de l’Etat Civil, Belgium 1947-1949.
15Although these four maps are based on the same data, they do not necessarily convey identical information to the reader. The map with historical boundaries (top left) confirms the previously mentioned sharp variation between municipalities. It also highlights the strongest local clusters, which is of interest to us because some types of cancer are influenced by local contextual factors such as industrial pollution and radon exposure [Maamri, 2015]. Currently, the European Environment Agency (EEA) estimates that 10% of all cancer cases in Europe are caused by exposure to pollution (EEA 2022).
16Nonetheless, it is difficult to discern straightforward patterns for the whole of Belgium using maps with only historical boundaries. Even individuals familiar with the specific intricacies of Belgium’s municipal geography may struggle to identify which municipalities exhibited the highest crude rates on the map with contemporary boundaries. In contrast, using more recent boundaries (top right) provides some clarity. By spatially aggregating all 2629 municipalities into the boundaries of 1991, a more comprehensible map encompassing 589 municipalities is created. This intervention leads, nevertheless, to a spatial aggregation of municipalities with very different rates. For example, in 1983, Antwerp, a city in the north of Belgium (area demarcated in red), merged with several of its much smaller surrounding municipalities [Algemene Volks- en Woningtelling, 1997]. As can be observed on the map with current borders in Figure 3 (top right), the high cancer rates in Antwerp also elevated the rates of its neighboring villages compared to the former map using historical boundaries.
17These nearby municipalities, however, were not homogeneous and had their distinct historical dynamics compared to the metropolis of Antwerp, such as different population structures (e.g., sex and age), lower levels of urbanization, and of medical infrastructure (e.g., hospitals and doctors per capita), all of which could affect cancer mortality rates [WHO 2017]. This issue, known as the ecological inference fallacy, falsely assumes that an underlying unit has the same attributes as the aggregated group it belongs to [Buzzelli 2020]. To illustrate this, we listed the rates of the neighboring villages in Table 2. Here, it is evident that – except for the municipality of Wilrijk – Antwerp had the highest mortality rates of all its neighboring municipalities it merged with. Although rural communities like Deurne, Lillo, and Zandvliet had very little features in common with Antwerp, they are henceforth considered as parts of Antwerp on the map with more recent boundaries, reflecting similar high rates. Therefore, using this method to visualize cancer mortality risks resulting in an anachronistic perception while simultaneously losing important information at the local level.
Table 2 : Crude cancer death rates of Antwerp and its neighboring municipalities (merger of 1983), 1947-1948-1949 (number of deaths per 10,000 people).
Source: Le Mouvement de la Population et de l’Etat Civil, Belgium 1947-1949.
18Similar issues arise using maps at the district or the larger provincial levels. In this article, only the district level is plotted in Figure 3 (bottom left). Positive aspects include the ability to quickly identify and communicate regional trends to a wide, inexpert audience [Kobakian et al. 2020]. For example, cancer is not only influenced by contextual factors (supra: e.g. pollution) but also by individual-level factors, related to lifestyle and an personal choices such as smoking habits, alcohol intake, diet (meat, sugar, vegetables), physical exercise, and predisposition towards disease prevention and vaccination [Benedek & Kiple 2003; 2008]. While such sociocultural trends can be mapped at the municipal level (if the data are available), they are easier to understand and compare at a higher aggregated level, such as a district, province, or region. Nevertheless, these larger geographical units are equally fallible. For instance, they might miss the previously mentioned high mortality rates in the city of Antwerp. On the map with district boundaries, only the highest rates in the province of East Flanders and the districts of Diksmuide and Liège (see Figure 1 for reference) are observable, just as the lowest rates in the province of Limburg. All other variation is obscured.
19One step further in aggregating is the practice of smoothed maps which in essence takes a broad perspective by freeing itself from administrative constraints and cancelling the residual noise of local outliers. According to cartographers Nicolas Lambert and Christine Zanin smoothed maps reveal ‘both the local specificities and its general trends’ [Lambert & Zanin, 2016]. Compared to the former aggregated maps, this method offers even greater potential as hotspots (kernels) become immediately visible. In section 3, we will use smoothed maps to compare three causes-of-death categories: ‘cancer’, ‘old age’ and ‘ill-defined or unknown diseases’ (see Figure 6). Even so, these smoothed maps remain highly artificial and anachronistic constructions. A disadvantage can be a loss of spatial precision; and in rare situations, the kernel is located in an area where there are no actual observations [Reimold et al. 2006].
- 5 Audience familiarity here means the lay audience's ability to identify and understand certain maps.
20Failing to select an appropriate unit can lead to various spatial challenges. If the chosen geographical unit is too large, certain local peculiarities (e.g., exposure to industry) within the unit might be overlooked. In contrast, when the selected unit is too small, issues of privacy or statistical significance surface. This implies that either the identity of an individual might inadvertently be exposed or that the validity of a study could be undermined [d’Onofrio et al. 2016]. Some studies have tried to identify an empirical method to define which geographical unit is better suited to compare cancer mortality. Francis Boscoe and Linda Pickle, both experts in spatial analysis of disease mapping, proposed to rank spatial units based on a predetermined list of criteria such as homogeneity of population size, homogeneity of land area, data availability, temporal stability and audience familiarity for cancer maps.5 Although they identified a preference for (American) counties as the geographical units to achieve their objectives, they emphasized that none of the units were ideal for every purpose [Boscoe & Pickle, 2003].
21Several researchers propose using the smallest unit possible that is still meaningful [Arsenault 2013; d’Onofrio et al. 2016]. However, researchers often cannot freely choose the most optimal unit, as the administrative units found in historical sources were decided upon long ago [Muir 1989]. Determining which geographical unit or boundaries are appropriate depends on a study’s objectives and research questions. As we are interested in the evolution of spatial inequalities of cancer mortality during the twentieth century, the temporal component is of particular importance.
22By comparing all maps with crude cancer death rates for the entire twentieth century, we conclude that the maps based on contemporary borders do not offer generalizable trends for the early twentieth century. However, from the 1930s onward, an overall rising pattern of cancer mortality can be observed together with specific strong clusters in the provinces of East Flanders and Liège, the districts of Ieper and Philippeville, and around the city of Antwerp. Around the 1970s, the highest rates were registered in the southern part of Belgium, whereas in the 1990s, strong clusters are again detected on both sides of the language border, particularly in the provinces of East Flanders, Hainaut, Luxembourg and certain parts of Antwerp.
23Linked to the previous challenge is the well-known problem of small numbers. Areas with small populations accordingly must deal with small numerators and/or denominators. As a result, mortality rates often show large or erratic fluctuations and are less reliable [Sahar et al. 2019; Kobakian et al. 2020]. It is hence challenging to distinguish between areas with genuinely high mortality and those areas where high rates are the result of a small population size [Wakefield et al. 2020]. Many of the 2629 Belgian municipalities in the (early) twentieth century had populations well under 1,000 inhabitants [Vrielinck, 2000]. The number of recorded deaths per year was low in these hamlets, especially for emerging diseases like cancer.
24Surprisingly, for some municipalities, we had to contend with running low and high numbers of registered cancer deaths throughout most of our observation period, even as cancer mortality increased immensely. The municipalities of Kuttekoven and Hendrieken, both in the vicinity of the rural town Borgloon (in the province of Limburg), illustrate this peculiarity by displaying 100% cancer mortality in 1969 and 1970 respectively. At this time, the average municipal share of cancer mortality to total mortality was approximately 20%. In these small communities – both populations had less than 160 inhabitants – only a handful of deaths occurred in a given year, all of which were cancer deaths.
25To counter such singularities, we decided to use temporal aggregation which involves the pooling of observations from the same municipalities over multiple years [Tsuya 2016]. We combined data from neighboring years (if the quality of the data allowed for this) and calculated an average per period. The following years, centered around census years, were selected and combined: 1905-1909-1910, 1929-1930-1931, 1947-1948-1949, 1969-1970-1971 and 1989-1990-1991. We refer in this article to these snapshots with 20-year intervals as 1910s, 1930s, 1950s, 1970s and 1990s respectively.
Figure 4 : Percentage cancer deaths in relation to the total number of deaths: the year 1910 versus the years 1905-1909-1910.
Source Le Mouvement de la Population et de l’Etat Civil, Belgium 1905-1909-1910.
- 6 In the year 1910, 1424 municipalities did not register any cancer deaths.
26We compared the percentages of cancer mortality in a single year, i.e. 1910, with the average percentages of the years 1905-1909-1910 (1910s) in Figure 4. The map with spatial aggregation shows a strong decrease in areas with no cancer deaths in comparison to the maps using only a single year of observations. Patterns thus become more visible. A first observation relates to the frequency of cancer registration. At the beginning of the twentieth century, cancer deaths were recorded in 72% of all Belgian municipalities (1,895 out of 2,629), not just in a select few, as might have been concluded if only one year of data were used.6 This suggests that awareness of cancer as a disease – for instance, methods for diagnosing certain types of cancer – was already present in rural communities and not confined solely to specialized medical centers clustered in and around the urban areas.
27Secondly, cancer percentages varied greatly according to the chosen period and geographical unit used. In the 1910s, we note that most of the highest percentages (varying between 9% and 33%) were located in the municipalities in mid- to southern Belgium, in areas characterized by strong industrialization and urbanization [Grauwels et al. 2011]. However, if we consider cancer percentages at district level, Philippeville – alongside the French border, barely reached the threshold of 6% cancer mortality, the highest percentage of all Belgian districts. From the 1930s onwards, cancer mortality increased strongly in the provinces of Antwerp and East Flanders. We hypothesize that this phenomenon is partly related to contemporary registration and classification practices (see next section). Nevertheless, strong clusters are still observed in the south of Belgium (districts of Philippeville, Verviers and Bastogne). After the 1950s, the highest percentages of cancer mortality shift to the north of Belgium. This trend of slightly high(er) cancer percentages in the north remains observable to this day. In 2020, the highest cancer mortality percentages are still registered in the northern region of Flanders (22.7%), although the regions of Wallonia (20.2%) and Brussels (18%) do not trail far behind [Statbel, 2023].
28In addition to the challenges related to selecting the appropriate geographical unit, ensuring the validity and comparability of historical causes-of-death registers is a significant concern. The value of these sources is contingent upon the ‘quality of diagnoses, on the system for registering and coding causes, on the training given in medical schools, and on practitioners’ habits and priorities’ [Shkolnikov, Meslé & Vallin 2012]. Historical demographer Alice Reid and her colleagues raised serious questions about the quality of past cause-of-death registers and suggested in their study on mortality in nineteenth- and twentieth-century Scotland the possibility that ‘the rise in mortality rates due to cardiovascular diseases and cancers (…) was not due to a rise in degenerative diseases, but was the result of improving diagnosis. [They] argue that the apparent rise (…) is not real, but an artefact caused by progressively fewer deaths from such diseases being attributed to ‘old age’ and ‘paralysis’’ [Reid et al. 2015]. Likewise, historian Timothy Gage claimed that certain cancers were misclassified under the category ‘unknown or ill-defined diseases’ in the cause-of-death registers of England and Wales in the nineteenth century. This resulted in a strong increase of cancer mortality parallel with improving diagnostic practices [Gage 1993]. In 1938, Willy Selleslags already hinted towards an association between these three categories in his essay on Belgian mortality. He suspected that part of the cancer deaths were hidden in the other two ‘bloated’ categories [Selleslags 1938].
Figure 5 : Evolution of the cause-of-death categories ‘cancer’, ‘old age’, and ‘ill-defined or unknown diseases’ (shortened to ill-defined).
Based on the percentages for all Belgian municipalities between 1903 and 1991.
(Note that the spikes in the category ‘ill-defined’ is are related to the start and end of Second World War).
Source : Le Mouvement de la Population et de l’Etat Civil, Belgium 1903-1971 & Demobel 1976-1991.
29Research into the relation between the categories ‘cancer’, ‘old age’ and ‘ill-defined or unknown diseases’ serves two interrelated purposes. Firstly, the two latter categories are an indication of the overall quality of the cause-of-death registration: if the percentage in the category ‘old age’ is high, the quality of registration is most likely poor or substandard. Often, small (rural) communities simply had no doctor present to record a cause of death. The Belgian government documented in 1910 that 59% of all municipalities had no doctor in attendance [Annuaire sanitaire 1912]. This implies that regularly the family or neighbors had to register the cause of death of the deceased. Quite possibly, family and friends used the generic ‘old age’, whereas professionals registered the supposed medical cause or put down ‘unknown’ [Alderson 1981; Velle 1985; Alter & Carmichael 1996]. If the latter assumption is correct, high(er) percentages in the category ‘ill-defined or unknown diseases’ then suggest improving quality for the cause-of-death registers in the early- to mid-twentieth century.
30Secondly, the rise in cancer mortality can, to a degree, be explained by improvements in diagnostic and registration practices. For example, Selleslags argued that doctors were sometimes negligent or reluctant to declare cancer as the cause of death. He suggested that certain diseases, like cancer, were subject to social prejudices or stigmas, leading to a widespread underreporting of cancer cases [Selleslags 1938]. Still, rapidly evolving new insights into cancer as a disease, along with technological innovations have necessitated governments to continually update the methods of diagnosing, recording and categorizing cancer since the mid-nineteenth century. Therefore, we must not only examine cancer mortality, but also delve into these other categories to assess whether there was a possible under-registration of cancer mortality in the early twentieth century.
31The trend in Figure 5 suggests an association between ‘cancer’ and the other two categories. As cancer deaths soared, the other categories slowly faded over time. Whereas ‘old age’ was widely used – between 5% and 10% – before and at the beginning of twentieth the century, the category started to decline greatly after 1950. In 1976, it stopped being a category altogether due to improvements in medical knowledge and better registration practices. In fact, the category became virtually obsolete as the current International Classification of Diseases (ICD) differentiates between more than 17.000 possible causes of death. ‘Old age’ is not one of them, although it is still used for indexing [WHO 2023]. The category ‘ill-defined and unknown diseases’ appears more resilient and remains a category in current Belgian and international cause-of-death statistics. In 2019, this category still recorded 2.9% of total mortality, or 3204 individual deaths [Statbel 2022].
32Both categories also show remarkable spatial differences, as indicated in Figure 6, where we use smoothed maps to compare the three categories for 1910s, 1930s and 1950s. The provinces of Antwerp and Brabant (in the north and middle of the country) reveal persistently low rates for the category ‘old age’, which could point to a practice in which ‘old age’ as a category was used less frequently. More importantly, in some of the areas on our maps where there was a notable decline in the rates for ‘old age’, we also observe a substantial increase in the rates for ‘cancer’ and ‘ill-defined and unknown diseases’, akin to communicating vessels. The most obvious example in Figure 6 is the province of East Flanders. If we compare the three snapshots, a rise in cancer mortality can be identified, a decline of ‘old age’ and a surge of ‘ill-defined and unknown diseases’ (see yellow squares). On the national level, no similar trends can be detected (Figure 5). The decline in the category ‘old age’ was not only very strong, but it also occurred earlier than in other provinces.
- 7 For more information on changes in confidentiality of cause-of-death certificates, see ‘Medical con (...)
33Selleslags ascribed this pattern to the early introduction of the confidential, anonymous cause-of-death certificate in the 1930s, which was introduced much earlier in the province of East Flanders than in the rest of the country [Selleslags 1938]. From that point onwards, the cause of death in most municipalities of this province had to be determined by a physician and reported anonymously to the authorities. To ensure privacy and sensitivity regarding the deceased’s medical history, the name of the deceased was no longer mentioned on the cause-of-death certificate, potentially reducing social bias. As a result, physicians and family members, who previously might have used categories such as ’old age’ to avoid shame, began registering ’cancer’ as the cause of death.7
Figure 6 : Smoothed maps of crude cancer death rates (per 10,000 inhabitants), 1905-1948.
Spatial smoothing is performed with Magrit (Exponential function, Span = 10km, Beta = 2).
Source : Le Mouvement et de l’Etat civil 1905-1948.
34In the early 1950s, the other eight Belgian provinces followed (West Flanders, Antwerp, Liège, and Luxembourg in 1952, Hainaut and Namur in 1953, and Brabant and Limburg in 1954). Only since 1954 has the registration of the cause of death been treated confidentially throughout Belgium [Elaut 1958, Devos 2024, forthcoming]. According to Selleslaghs, cancer mortality registration in East Flanders increased by 40% after the implementation of the confidential certificate, whereas in the other provinces the percentage remained limited to just 4% on average [Selleslags 1938]. Consequently, we assume that there was considerable under-registration of cancer mortality in Belgium before 1954. The extent of this under-registration is difficult to ascertain and will be part of our further in-depth analyses.
35The significant rise in cancer mortality during the twentieth century had many drivers. As previously discussed, changes in diagnostic and registration practices, lifestyle choices, and environmental factors all played a role (supra). The rise can also be understood within the framework of the epidemiological transition [Omran 1971]. An immediate consequence of the progress made against infectious diseases was a rising life expectancies. Ageing populations imply more cancer cases and cancer deaths, as cancer is primarily a disease that affects the elderly [Benedek & Kiple 2008; Roser & Ritchie 2019].
- 8 For both periods, the age structures for the biggest cities such as Antwerp, Brussels, Liège and Gh (...)
36Most Belgian cancer studies on spatial mortality differences – prior to 2006 (see Table 1) – used indirect standardization to facilitate meaningful comparisons over time by adjusting for population size and age structure. Calculating age-specific mortality rates or using direct standardization requires the number of deaths by age of the deceased and the corresponding population figures for each age group. These data are often not readily available at the municipal level for a portion of our observation period (i.e. 1930s and 1950s).8 In our previous analyses and maps, we therefore mostly relied on crude cancer death rates and cancer percentages. However, the crude rates do not consider sex and age and are subsequently difficult to use in comparisons over time and space. Moreover, cancer percentages can be influenced by specific events, such as the COVID-19 pandemic which caused in 2020 a significant drop in cancer mortality percentages in comparison to the previous year [Statbel 2024]. In absence of age-specific mortality rates and given the challenges posed by small numbers at the municipality level, we chose to use indirect standardization for the 1910s, 1970s and 1990s to enable comparisons over time and space. For the standard mortality schedule, we used the number of deaths by age as well as the age structure of the population of 1961 [Ronsijn 2014; Lokstat 2023]. The census data for 1961 are available and of very good quality.
37The issue with the age structure of the population extends beyond just data availability and also concerns compatibility. In our sources, the delineation of the age groups changed multiple times over the years. As high mortality rates shifted from infants and children to the elderly, so did the focus on different age groups. In the cause-of-death registers of the early twentieth century, infant and child mortality were emphasized, which led to an uneven consideration of the age groups most at risk. The registers differentiated between children under 1 year old, 1-2 years, 2-5 years, 5-7 years, and 7-15 years. Adults were divided into three age groups only: 15-21 years, 21-50 years, and above 50 years. From 1935 onwards, the older age groups were subdivided further, yet not using the same intervals as before (45-55 years, 55-65 years, 65-75 years, and above 75 years). For the population censuses, the age-related data vary. For 1905, the sources distinguish between minus 15, between 15 and 55 and above 55 years old. Combining all data is not straightforward. In this article, the year 1905 is our weakest link with only three age groups, which means we had to use these intervals to calculate and compare indirect standardized rates.
Figure 7 : Maps based on the percentages of cancer deaths (top left), crude cancer rates (top right), indirect age-standardized rates (bottom left) and smoothed maps (based on the indirect rates) for 1969-1970-1971.
Source: Le Mouvement et de l’Etat civil 1969-1970-1971.
38To reiterate, our main question in this section is whether we can use alternative mortality indicators in the absence of adequate data for age standardization, and if so, do the results appear similar and do they convey the same information to the reader? To illustrate this, we compared in Figure 7 maps of the 1970s using percentages, crude death rates, and indirect age-standardized mortality rates for all cancer deaths, men and women combined.
39Choropleth maps work best for relative data. As observed, the maps in Figure 7 do not have a similar appearance nor do they tell the same story. The map using percentages (top left) shows a higher amount of cancer deaths in the north of Belgium whereas the map using crude cancer rates (top right) has clusters in the south of Belgium and Antwerp. Some provinces (especially Limburg) seem slightly less affected. The pattern for the map using indirect age-standardized rates (bottom left) is scattered, which is why we added a smoothed map. Higher rates are still visible in some parts of southern Belgium and to a lesser extent around the city of Antwerp. Oddly, increased rates are also found in areas with very low crude cancer rates, such as northeastern Limburg and the Belgian coastline. In contrast to other indicators, the indirect standardized rates do not necessarily overlap with large metropolitan areas.
40When it comes to determining which maps are superior and/or comparable, there is no clear-cut answer. Similarly to selecting the appropriate geographical unit (and its boundaries), the decision of which mortality indicator to use ultimately depends on the specific research questions the researcher aims to address and which data is available. Mapping cancer percentages and indirect standardized rates serves different purposes, each with distinct benefits. Cancer percentages show the proportion of cancer deaths relative to all deaths in an area, highlighting the relative burden of cancer. This method allows for straightforward regional comparisons and is simple to interpret and communicate, making it useful when detailed population data are limited. On the other hand, indirect standardized rates adjust for differences in age structures across populations, providing a more accurate picture of the cancer risk. This adjustment allows for fair comparisons across regions with varying age structures and helps identify true patterns of cancer mortality by filtering out age-related effects. Consequently, standardized rates are essential for targeted health interventions and policy-making.
41In cases where we lack data for indirect standardization on the municipal level (like the 1930s and 1950s), we recommend projecting the age distribution from the district level onto the municipalities (except for the large towns for which we do have the data). For the 1930s, the result is fairly similar to what we had already observed in the maps using crude cancer death rates, although some differences occur around the city of Antwerp, the provinces of Namur and Liège, and the districts of Ieper and Roeselare.
42Geographer Alfred Haviland pioneered the mapping of diseases beyond infectious diseases by illuminating spatial patterns of cancer mortality in Victorian England and Wales [Haviland 1875]. He used blue to indicate high mortality clusters and red for lower ones. According to modern conventions, red implies high risk, making Haviland’s maps counterintuitive. Possibly, blue reminded him of dark cyanotic blood, while red resembled healthy, oxygenated blood [Arnold-Foster 2020]. The use of colors and shading is an established method for illustrating relative differences on a map, but it is not without risk, as Haviland’s maps illustrate [d’Onofrio et al. 2016]. Colors, shading, and classification methodologies can significantly influence map interpretability. Visual patterns can be distorted through manipulation of the interval breaks of the classes. Also, the number of classes is crucial for comprehension. Too many classes cause complexity and confusion, while too few lead to oversimplification, obscuring important patterns. Generally, four to six distinct classes are most effective [Monmonier 2005].
43In this article, we use five colors, each corresponding to one of the five classes, ranging from light yellow (lowest cancer rates) to dark brown (highest cancer rates). For all preceding maps in this article (except smoothed maps), we used the quantile method, which divides cancer mortality into classes with roughly equal numbers of deaths. This method generates a readily interpretable legend suitable for a non-technical audience. However, it can result in a skewed distribution, with most data values concentrated in one or two classes, leaving the remaining classes with few or no values.
44Another challenge is comparing data over the long term. Due to the significant rise in cancer deaths over time, the (quantile) class intervals on maps representing different periods must be repeatedly adjusted. The question arises whether we can use quantile classification methods to compare over time if the class interval boundaries are not the same. For instance, in Figure 8, we compare crude cancer death rates between the 1910s (top left) and 1990s (top right). Although both maps use the quantile classification method, the boundaries differ: 0 to 72 deaths per 10,000 people in the 1910s and 0 to 47 deaths per 10,000 people in the 1990s. Should we then use other classification methods, such as natural breaks (Jenks), equal intervals, or standard deviation, to counter this problem?
Figure 8 : Crude cancer death rates, 1910s (left) versus 1990s (right).
The top maps use quantiles, the bottom maps equal intervals.
Source: Le Mouvement et de l’Etat civil 1905-1909-1910 and Demobel 1989-1990-1991.
45The two bottom maps in Figure 8 are essentially the same maps as the maps above them, as both are based on the identical data. The difference is that the bottom maps use the same pre-defined, equal class intervals across time, with a scale ranging from 0 to 50 cancer deaths per 10,000 people, divided into five classes of 10 deaths per 10,000 people each. As a result of this approach, trends become exceedingly obscured. In the map of the early twentieth century, the number of cancer-related deaths in most municipalities does not exceed the first threshold (0 to 10 deaths). Visually, most municipalities are assigned the lowest-ranking color, light yellow. Conversely, by the end of the twentieth century, due to a saturation effect, most municipalities fall within the middling class intervals. Although the highest rates are still noticeable, the map does not convey the same message as before.
46As a solution – and compromise, both classifications could be presented. The maps using equal intervals can be placed as smaller maps in the top right or top left corner for comparative purposes, whereas the larger maps in the center use the quantile classification method. However, in practice, the (legend of the) smaller maps in the corner would be virtually illegible, recognizing only the strongest or weakest trends. Other classification methods are also possible: natural breaks (Jenks) or standard deviation are often used. The natural breaks method, also known as the Jenks classification, employs an algorithm to categorize values into classes defined by clear breakpoints. This approach is most effective when applied to data exhibiting an uneven distribution but not skewed toward either end of the distribution. Jenks are data-specific classifications however and are not useful for comparing various maps with changing underlying information. Lastly, the standard deviation method finds the mean value of the observations and subsequently then places group breaks above and below the mean at intervals of either .25, .5, or 1 standard deviation. As well known, the standard deviation indicates the variation in the dataset, but can be strongly affected by outliers (ArcGis Pro 2023).
47After evaluating the referenced works listed in Table 1, it is clear that there is no consensus regarding the use of colors and class intervals. Yet, the time periods of these studies were relatively short, making comparisons somewhat easier. Although other classification methods are also suited for examining the spatial evolution of cancer mortality in Belgium, the quantile method appears to offer the best results if we compare the same data using each time a different classification method. Other methods lead to less convincing results: for instance, strong cancer clusters are less pronounced (even if the hot/cold spots remain fairly visible). Also, rather than highlighting light yellow or dark brown, the maps predominantly show the middle colors of orange and light brown colors, indicating little to no variation.
48Cancer mapping studies before the 1960s are scarce for Belgium. Unlike British historiography, there is a lack of extensive long-term studies for this country. In this article we addressed this gap and simultaneously highlighted five key challenges that future cancer mapping studies must tackle. We first assessed the impact of changing administrative boundaries by comparing different approaches (historical, current, etc.) using crude cancer death rates for the 1950s. Historical boundaries may be more fitting to detect local clusters and the effect of environmental factors (e.g., industrial presence, radon vicinity). On the contrary, current boundaries and smoothed maps are probably more suited to detect regional differences and communicate the effect of lifestyle factors such as smoking and dietary changes.
49We then discussed the problem of small numbers by comparing the relative share of cancer deaths to total deaths in the early and mid-twentieth century. As a solution, we proposed to use temporal aggregation, a common practice in historical demography. This rendered five snapshots of 3-year intervals spanning the twentieth century.
50In the third section, we explored concerns regarding changing diagnostic and registration practices over time by comparing smoothed maps in the 1910s, 1930s and 1950s. We identified an association between the use of categories such as ‘cancer’, ‘old age’ and ‘ill-defined or unknown diseases’, most likely reflecting the overall quality of the causes-of-death registration. Additionally, we pointed to a significant under-registration of cancer mortality before 1954 as could be inferred from the early implementation of the confidential death certificate in East Flanders. Both assumptions will be part of our future analyses.
51Fourthly, due to data constraints, we faced problems calculating standardized mortality rates for certain periods, specifically the 1930s and 1950s. We explored and compared alternative mortality indicators for the 1970s, each revealing distinct patterns. None of them offered satisfactory solutions. For the missing years, we suggest using the age structures of the districts, excluding those of the larger cities for which we do have data. The district percentages can then be assigned to the remainder of the municipalities in that district.
52Lastly, we determined that the quantile method of classification is the most effective method for comparing cancer mortality over the long term. Comparing rates from the 1910s with the 1990s, we found that equal class intervals are not interesting as they fail to reveal meaningful trends.
53In our attempt to visualize cancer mortality in Belgium, we use a range of different methods and mortality indicators, and as expected, they did not consistently yield identical results. In addition, maps based on the same data, did not automatically convey the same information to the reader. In Table 3, we offer a summary of the observed trends over all five periods.
Table 3 : Summary of the trends for the maps with contemporary boundaries.
(*) Here, we projected the age distribution of the population from the district level onto its municipalities (except for major cities).
54Over time, cancer trends have undergone significant changes. At first, crude mortality rates were marginally higher in southern Belgium. This was particularly evident during the first two periods, which also showed high variation. However, a shift towards northern Belgium became apparent over time, although southern Belgium continued to show clusters with high rates until the 1990s. The indicator using percentages, though distorted by high child mortality until the mid-twentieth century, shows a more distinct pattern: highest percentages in the south in the early twentieth century, gradually moving towards the north of Belgium by the end of the century. At present, the highest percentages remain in the north of the country [Statbel 2024]. When adjusted for age and population size, the highest rates in the first half of the twentieth century are unexpectedly found in both urban and rural areas, and on both sides of the language border. Furthermore, no clear association with industrial centers is evident. The rates for the 1970s are distributed across various districts and provinces without any discernible pattern. In the 1990s, age-standardized rates are notably higher in the province of East Flanders and the southeastern Belgian districts near the French and German borders. This may be attributed to elevated lung cancer mortality rates linked to local male smoking habits [Vanthomme et al. 2017].
55The rapid realignment of rising and falling cancer clusters approximately every 20 years underscores the importance of long-term research to study the spatial patterns of cancer mortality in relation to demographic and socioeconomic factors. As stated, cancer is determined by both contextual and individual-level factors. In this article, we focused on describing cancer mortality patterns without linking these to potential explanatory variables. Such variables might explain why some unexpected areas (rural, or with young populations) are more affected. These variables, typically sourced from industrial, agricultural and cadastral censuses, will be included in our future analyses. Examples include the presence of unhealthy businesses, employment in specific industries, population density or urbanization levels, educational levels, area deprivation, the presence of and distance to medical facilities (doctors and hospitals), age, et cetera.
56To conclude, we must also note that this article does not address the spatial differences in male and female cancers, although these differences are vast, both historically and geographically. For simplicity, we have focused on the total population and refrained from including additional maps. However, the detailed exploration of these gender-specific spatial cancer patterns will be part of our future research.
![Figure 1 : Administrative map of Belgium, 1961 [Ronsijn 2014].](docannexe/image/14597/img-2-small580.jpg)
Flux RSS
