1

Death clustering and the intergenerational transmission of death risks in nineteenth century Belgium and Sweden

Mattijs Vandezande Centre of Sociological Research, Catholic University of Leuven, Belgium

Sören Edvinsson Centre of Population Studies, Umeå University, Sweden

Paper prepared for the Social Science History Conference Boston, 17-21 November 2011

Provisional paper, please do not quote without the author’s permission

Death clustering, parental competence and the intergenerational transmission of death

risks

In 1990 Monica Das Gupta published her now often quoted paper on child mortality in rural

Punjab, India. What made the paper so noticed, is her observation on the spread of child

deaths across families. Deaths have a strong tendency to occur in a limited number of

families. Later, these families have become known as “high-risk families” (Edvinsson et al.

2005). At the time of her paper, Das Gupta already cited examples of several other developing

regions were the existence of clustering was suggested. Her paper was followed by many

others, who found evidence of death clustering in several regions of the developing world,

such as – among others – Mali (Hill & Aguirre 1990), Guatemala (Guo 1993), Brazil (Curtis

et al. 1993), Bangladesh (Zenger 1993; Saha & van Soest 2011), and Kenya (Zaba & David

1996; Omariba et al. 2008), as well as in populations of the past, such as Sweden (Lynch &

Greenhouse 1994, Edvinsson et al. 2005) or the Netherlands (Janssens 2010). Although all of

these studies found a similar disproportionate spread of child or infant deaths among families,

and several of them have put forward possible causes, no full explanation of the phenomenon

is agreed upon. However, there seems to be some kind of general awareness that clustering is

very important for the study of child mortality, since it might alter our understandings of the

factors determining child deaths.

2

Much of the effort so far has gone towards explaining what makes high-risk families

so different than others. As Das Gupta already suggested in 1990, and confirmed in her later

study of 1997, the diverse degrees of individual parental (in)competence probably lies at the

basis of clustering. While education and income might improve this parental personal

abilities, they do not take away the variation among individuals, especially not in the lower

social strata. A hypothesis like this is difficult to prove, especially in historical populations.

In order to shed light on this issue, we argue that an intergenerational view might be

very useful, especially for populations were little background information is known. The risk

of starting a high-risk family might be inherited, if attitudes and practices of childcare are

learned from one generation to the other. If husband and wife are parenting a high-risk family

because of (the lack of) their own childcare competence, their children who survive childhood

might – at least in some cases – socialise that ‘faulty’ behaviour, and they themselves might

begin a high-risk family in the next generation.

Our previous findings clearly show that there is an intergenerational transmission of

death risks, at least for infants. Infants were found to have increased mortality risks, if there

was as least one family member in the previous generation that did not survive infancy. This

observation was still valid after keeping different kinds of socioeconomic factors constant,

and for both the maternal (Vandezande 2010) and paternal bloodline (Vandezande 2011).

Moreover, age-specific analyses pointed towards postneonatal deaths, as caused by, for

example, lack of breastfeeding. This made a biological hypothesis, the intergenerational

transmission of faulty genes, less likely. If parental competence really lies at the basis of

clustering, it might be revealed through its intergenerational transmission.

The question remains how the intergenerational transmission of infant death risks is

explanatory for the clustering phenomenon. This is addressed in the underlying paper. To

further strengthen our findings, and to rule out a local element or peculiarity, several regions

are studied simultaneously: the industrializing city of Antwerp, Belgium (object of study in

our previous research), and several rural and more urban regions in northern Sweden,

populated by both Swedish settlers and the Sámi, a group of indigenous semi-nomads.

Preliminary analyses already suggested that, for several Swedish regions, death risk

was transmitted through generations along the maternal line as well. This paper will therefore

be devoted to the direct comparison of clustering in its first part, and the impact of

intergenerational death risk transmission on it in its second part. The following will only

discuss infants, so the first year after birth. All mentions of ‘children’ should therefore be

interpreted as individuals in their first twelve months of life.

3

Introduction of the data

The first dataset is retrieved from the Belgian “Antwerp COR*-database” (Matthijs &

Moreels 2010), a database built and maintained by the research group Family and Population

from the Centre of Population Studies, Catholic University of Leuven. This database is based

on information of population registers (“Bevolkingsregisters”) and vital registration records

(“akten der Burgerlijke Stand”), including certificates of birth, marriage and death. The

sources accessed for data construction span nearly seven decades (1846-1920). The database

comprises the Antwerp district: city of Antwerp, an industrializing and modernising world

port, its suburbs, and the surrounding countryside. The whole area is not bigger than 1000

square kilometres (600 square miles), but with its 325,000 inhabitants in 1900 it is a very

densely populated one. This huge population is mainly due to massive immigration, following

the port developments through the whole of the nineteenth century. From this area a 0.4 per

cent name sample was taken of all people living within the area and the research period,

having a surname beginning with C-O-R, or being directly kin-related to a person with such a

name. This yielded about 30,000 individuals, whose life course was followed for as long as

they stayed within the regional and temporal study frame.

The dataset which was retrieved for this research limits itself to infants, born in the

1846-1905 period within the study period, of whom we have the full reproductive life course

of the mother. Additionally, in order to be able to conduct intergenerational analyses, we did

not allow children of whom the mother had no siblings. An important bias in the retrieved

data therefore is that only those families were selected of whom two successive generations

lived in or around Antwerp. First-generation immigrants are thus always excluded. Lastly, to

ease the interpretation of intergenerational figures, we also excluded illegitimate children and

twins births. The final dataset consists of 1,718 children, born in 332 families.

This data is compared to digitized data from the Demografiska Databasen at Umeå

University (DDB), http://www.ddb.umu.se/), a huge historical dataset of Swedish life courses,

which rightfully claims to be the largest in Europe, and one of the oldest in the world.

Moreover, the quality of their source material if similar to the 19th-century Belgian sources:

both are very detailed, and very rich. The database is based on church records, mainly

catechetical registers on a household level (“husförhörlängder”), enriched with registers of

baptisms, marriages and burials. Whereas in Belgium the sources were maintained by state

officials, from 1846 onwards, the Swedish sources were maintained by parish ministers, and

often date much further back in time. No sample was taken, but for several parishes all people

4

were collected. The regions of the dataset that are fit for this analyses are the Sundsvall region

(SUL), the Skellefteå region (SKE) and the Northern Inland parishes (NIL); maps are found in

appendix. Sundsvall is a region named after a small town, mainly known for its sawmill

economy, located in the central east of Sweden. From the middle of the nineteenth century

onwards, when the first sawmills start to arrive, the region transforms from a couple of small

fisher villages, into a industrializing migrant attraction pole, increasing its population tenfold.

The region of Skellefteå is located in the north, along the coast. It is a stable rural region of

freehold farmers, with limited industry along the coastline. Both regions are much larger than

Antwerp (which has the size of Byske, the smallest of Skellefteå’s parishes), but they are

nothing compared to the vast Northern Inlands. These parishes cover the whole western part

of Sweden’s northern half. Population is extremely scarce, and is located in a small number of

villages in the east of each parish. Also registered at these parishes are the indigenous Sámi

nomads, reindeer hunters who follow their herds in a seasonal cycle from the mountains in the

west in summer, to the eastern coast in winter. Some of them settled down, mostly during

wintertime, while others turned into fishermen for major parts of the year. Although they had

distinct cultural practices, a different language and religion than the Swedes, ‘colonisation’

already took place for more than two centuries, and many Sámi settled down and intermarried

with the Swedish colonizers. At the end of the nineteenth century, the discovery of iron ore

and the subsequent opening of large iron mines – first in Gällivare, later in the more northern

Kiruna – started a constant flow of immigrants. A rapidly constructed railroad to the east

coast, the Malmbanen, helped to maintain this population flow.

To this data, similar restrictions as in Antwerp were applied in order to retrieve a

dataset usable for analysis (full reproductive history of two successive generations, mother

should have at least one sibling, no illegitimates, no twins). Overall, data was obtained for 19

parishes belonging to one of the three regions (see table 1). The infants were born roughly in

the period 1790-1894, although the years differ from one region to the other (e.g. data from

Sundsvall and Skellefteå started more or less one generation later, mainly due to data quality

restrictions). The operation yielded a total of 25,559 infants, born in 5,404 families. Half of

them were born in the Skellefteå region (11,105), a third of them in the Northern Inland

region (7;994), and the rest was born in the Sundsvall region (5,387).

Table 1. Parishes in the DDB retrieval Region/Parish Period Individuals Skellefteå (SKE)

5

Skellefteå 1815 – 1894 12209 Sundsvall region (SUL) Alnö 1835 – 1893 482 Attmar 1848 – 1894 1113 Hässjö 1845 – 1895 558 Indal 1849 – 1894 901 Ljustorp 1834 – 1894 708 Lögdö 1857 – 1891 17 Njurunda 1845 – 1891 1070 Selånger 1850 – 1894 286 Skön/Skönsmo 1839 – 1892 255 Sättna 1838 – 1894 385 Tuna 1840 – 1894 622 Norrlands Inlandregion (NIL) Föllinge 1811 – 1892 1215 Gällivare 1790 – 1903 1116 Jokkmokk 1804 – 1895 1316 Jukkasjärvi 1817 – 1900 1991 Karesuando 1816 – 1894 305 Undersåker 1823 – 1893 357 Vilhelmina 1810 – 1895 2633

Three out of four regions, two of the Swedish ones and the Belgian one, are named after their

largest and most influential parish or community. Therefore the following text will use the

name of that parish or community to describe the whole of the region.

The spread of infant mortality in Belgium and Sweden

An overview of the infant mortality in all regions is given in table 2. Infant mortality is the

highest in the early industrializing regions, Sundsvall and Antwerp. Both have an infant

mortality rate of about 160 deaths per 1000 live births. In the northern areas, the mortality rate

is lower. An explanation that has been put forward in the Swedish literature is its geographical

location and its remoteness: infectious diseases, such as measles, smallpox and cholera, had

less chance to spread, let alone reach the distant and cold north. This is especially true for

Skellefteå, which was an isolated and stable region with very little migratory movements,

hence its remarkably low levels of infant mortality.

In order to get some idea of the spread of these infant deaths over families, we

followed a common strategy. We assesses the proportion of families that did not experience a

single infant death, given a certain level of mortality. These numbers of given in the second

part of table 2. A closer, comparative look reveals some interesting patterns. Although the

mortality levels in the regions around Antwerp and around Sundsvall are similar, in the

6

Sundsvall region more families are mortality-free. This means that the deaths in Sundsvall

occur in a smaller proportion of the families, suggesting deaths are more clustered in

Sundsvall. Also, the proportion of mortality-free families is similar in both Sundsvall and

Skellefteå. However, the levels of infant mortality in these regions are very dissimilar, the

highest and the lowest of all regions, respectively. One would expect that if the mortality rate

increases, the number of families without infant deaths decrease. Therefore, again, this

suggests that deaths are clustered the most in Sundsvall.

Table 2. The spread of mortality in Antwerp and three Swedish regions NIL SKE SUL DDB TOTAL ANTW births 7,994 1,1105 5,387 24,486 1,718 infant deaths 1,159 1,188 876 3,223 275 mortality rate 14.50% 10.70% 16.26% 13.16% 16.0% families 1,716 2,245 1,443 5,404 332 - with deaths 761 838 558 2,158 145 % of total 44.35% 37.34% 38.70% 39.93% 43.7% - without deaths 955 1,407 885 3,246 187 % of total 55.65% 62.66% 61.30% 60.07% 56.3%

Although this kind of suggestive figures is often given or cited in the death clustering

literature, they are essentially meaningless, at least without information on the underlying

distribution of family sizes. The more children there are in a family, the higher the risk that

there will at least be one infant death. If fertility is high in a population, and family sizes are

large, it is more difficult to find a family without deaths, than a low-fertility population, were

family sizes are small. To be able to make statements on the level of clustering in a

population, information on both the infant mortality rate and the distribution of family sizes is

necessary. This reasoning was implicitly present in Vandezande (2010), but explicitly

described in the (yet unpublished) papers by Derosas and Broström (2010), and much more in

depth by Holmberg and Broström (2010). They argue that clear statements on clustering can

only be made if the spread of infant mortality between families in a population is compared

with what is mathematically expected for that given population, if there would be no

clustering. The larger the difference between the actual infant death spread, and the expected

spread in an ideal world without clustering, the more important clustering is. Basically that

means they created an indirect measure, not putting a value on clustering itself, but on how

important clustering is.

7

Measuring clustering

According to Holmberg and Broström, we need to assume an ideal world in which there is no

clustering present, where infant deaths are distributed randomly across families. In such a

world, infant deaths in a certain family follow a binomial distribution, with parameters n

(number of trials) and p (overall chance of success). Using the binomial density function, the

chance of a given number of successes can be calculated. In order to make this clear, we can

compare this to a dice game. Imagine an experiment where one needs to throw five dices, and

try to throw as much sixes as possible. The number of trials, in stochastic theory, here stands

for the number of dices that are thrown: five. The overall chance of throwing a six is, usually,

one out of six. With this information it is possible to calculate the chance of, let’s say, no

sixes at all, which is 0.40. This means that if you would try throwing all five dices 100 times,

you would throw no sixes at all in – on average – 40 of your throws. The same applies with

infant deaths in families. The chance of having no infant deaths in a family of five children,

given an overall mortality rate in the population of one out of six (16.66%), is 0.40. This

means that if such a population would contain 100 families with five children, 40 families

would have no infant deaths. At least, this would be true under the assumption of no

clustering. The same can be calculated with families of all possible sizes, every time using the

same overall mortality rate. This would result in a expected proportion of all families which is

mortality-free, a number which varies depending on the distribution of family sizes in the

population. The difference between this expected and the real proportion of mortality-free

families is an indicator of clustering. Moreover, Holmberg and Broström state that their

indicator allows to compare clustering across populations, something which was not possible

before. The bigger the difference between observed and expected mortality-free families, the

more clustering in the population.

8

Figure 1. Distribution of families by family size, by region

0%

2%

4%

6%

8%

10%

12%

14%

16%

18%

20%

22%

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

family size

perc

enta

ge o

f tot

al fa

mili

es

NILSKESULANTW

This methodology is put into practice here with the Swedish and Belgian regions. At

first, information on the family sizes in those regions is necessary. This information is given

in figure 1. The figure shows large regional fertility differences. Skellefteå has a very high

fertility, with an almost equal proportion of families with 1, 2, 3, up to 7 children. Sundsvall

on the other hand has an extremely low fertility, with most families having only 4 or less

children. Antwerp, being on the eve of the start of the fertility transition, also has a quite low

fertility. The family sizes in the Northern Inland parishes are quite large, but not as large as in

Skellefteå. This information is used to calculate an expected proportion of families that is

without infant deaths. The results are found in table 3. Clearly Antwerp has by far the biggest

clustering, and not Sundsvall, as was suggested in table 2. The 56 per cent of all Antwerp

families that is mortality-free, is almost 10 per cent more than expected. The 61 per cent

mortality-free families in Sundsvall is only 7.62 per cent more than expected, meaning there

is less clustering in Sundsvall than in Antwerp. Our intuitive feeling of clustering in table 2

was wrong, based on the underlying fertility patterns, or more specifically, the distribution of

family sizes, in these regions. There is less clustering in the Northern Inlands, and relatively

little clustering in Skellefteå.

9

Table 3. Observed versus expected percentage of mortality-free families, for four regions

NIL SKE SUL DDB TOTAL ANTW births 7994 11105 5387 24486 1718 infant deaths 1159 1188 876 3223 275 mortality rate 14,50% 10,70% 16,26% 13,16% 16,0% % mortality-free families (obs) 55,65% 62,66% 61,30% 60,07% 56,3% % mortality-free families (exp) 50,88% 59,25% 53,68% 55,02% 46,6% difference 4,77% 3,41% 7,62% 5,05% 9,7%

Holmberg and Broström go much further than this. Not only the proportion of

mortality-free families can be calculated, but also the proportion of families with one, two or

more infant deaths. If there are more mortality-free families than expected, there must be less

families with infant deaths than expected to make up this. In other words: in which families

are infant deaths clustered? The same calculation can be used, which was done so for all

regions in this paper. The figures for Antwerp are show in table 4. There are indeed much less

families with one or two infant deaths than expected. However, in the case of the families

with many deaths, i.e. 4 and more, they are much more frequent than expected. Only the

families with three infant deaths are about as frequent as we could expect (the difference

between observed and expected is very small). Table 4 thus shows us that infant deaths are

distributed very unequally across families. Two situations are much more common than we

expect, namely (a) families where there are no deaths, and (b) families with a lot of deaths. As

a consequence, if parents lost one of their infants, their other children are more likely to die

during infancy as well.

Table 4. Clustering of infant deaths in families in Antwerp

deaths observed expected difference0 56.33% 46.59% 9.731 24.10% 32.86% -8.762 10.24% 13.93% -3.693 4.52% 4.81% -0.29

4+ 4.82% 1.81% 3.01tot 100.00% 100.00% 25.48

The differences between observed and expected in table 4 can be plotted in a graph,

see figure 2. The interpretation stays, so a positive value means there are more such families

than expected, a negative value means these families are underrepresented. For Antwerp, this

results in a V-shaped graph, similar as the one that was produced by Holmberg and Broström.

10

A table like the one for Antwerp is made for the other regions as well, and these results are

plotted next to the graph for Antwerp.

Figure 2. The shape of clustering in four different regions

-0,1

-0,08

-0,06

-0,04

-0,02

0

0,02

0,04

0,06

0,08

0,1

0,12

0 1 2 3 4+

n° of deaths

diffe

renc

e in

obs

erve

d vs

exp

ecte

d pr

opor

tion

of fa

mili

es

ANTWSKESULNIL

Figure 2 shows that clustering has a very similar shape in all four regions. There are

always more families without infant deaths than expected, which is compensated by an

underrepresentation of the families with one and two infant deaths. Families with four or more

dead children are overrepresented. Only families with three infant deaths are about as frequent

as expected, making such families an inflection point in the clustering graph. Remind that,

due to the used methodology, these graphs take into account the different distribution of

family sizes in the four regions. It is very remarkable that the nature and shape of clustering is

almost identical in all regions, despite the huge geographical and cultural distance between

them (e.g. between modernizing Antwerp city dwellers, and a traditional semi-nomadic

population like the Sámi). This observation strengthens the presumption that clustering is a

universal phenomenon.

The extent of clustering does differ across populations. This was already discussed

with table 3. It can also be derived from table 4, and is visible in figure 1, where we can

actually see where differences in amount of clustering originate from. The sum of the absolute

values of all differences between observed and expected (since the normal sum, by definition,

is zero), is a measure of how much the overall spread of infant mortality between families

differs from an ideal world, without clustering. For Antwerp, this measure is 25.48 (table 4).

11

This value corresponds to the total surface covered by the Antwerp line in figure 2. From

figure 2, it can derived that Antwerp has the most clustering, followed by Sundsvall. The

difference between the two lies with the families with two infant deaths, which are more

underrepresented in Antwerp than in Sundsvall. Again, this has nothing to do with the fertility

differences between the two regions. There is much less clustering in Skellefteå and in the

Northern Inlands. The difference between the two is visible in all situations, i.e. number of

deaths in a family, since the lines are almost perfectly proportional to each other.

Clustering and the intergenerational transmission of infant mortality risks

A major benefit of the method proposed by Holmberg and Broström is not only that it allows

to compare clustering across regions. They also created a tool to test the impact of

determinants of mortality on clustering, or – stated otherwise – a way to explain clustering. In

their paper they tested for birth order, in order to invalidate the proposition that the V-shaped

graph resulted from the strong relationship between fertility and mortality. The death of an

infant facilitates the birth of the next child, due to the hypothesis of a replacement effect, or

simply the cessation of lactation (Preston 1978). In the light of clustering, this would mean

that a very large family size cannot be reached without a certain number of infant deaths.

Moreover, some authors suggest the existence of a ‘scarring effect’. The loss of several

infants can hurt, or ‘scar’ the mother, resulting in an increased mortality risk for subsequent

children (Arulampalam & Bhalotra 2006; Saha & van Soest 2011). Recalculating the expected

proportion of families with a number of infant deaths, controlling for birth order, Holmberg &

Broström showed that this indeed explains some clustering, but as a whole it is of minor

importance.

The same method will be used to test for the impact of the intergenerational

transmission of deaths. Basically, it is the same method as in the previous section, except for

the value of the overall mortality rate. Instead a new value is calculated for every family,

based on how the overall risk of infant mortality for that particular family is altered, given its

genealogical history of infant deaths. Therefore we counted the number of infant deaths in the

previous generation (along the maternal line), and used standard (poisson) regression

techniques to recalculate the mortality risk for every family individually. This gets

computationally very intensive, especially in the case of the expected number of families with

two or more deaths. Therefore, as Holmberg and Broström suggest, simulation techniques

were used. Under a binomial distribution, with a certain family size and a family-specific

12

mortality rate, a number of deaths was simulated. This was done 10,000 times, and the

resulting averages were taken. Using such large replications yielded results very close to the

ones obtained analytically (i.e. using the formula of the binomial density function). The result

is a renewed version of table 4, with new expected values: table 5.

Table 4. Clustering of infant deaths in families in Antwerp,

after controlling for intergenerational effects

deaths observed expected difference0 56.33% 46.93% 9.401 24.10% 32.71% -8.612 10.24% 13.65% -3.413 4.52% 4.71% -0.19

4+ 4.82% 2.01% 2.81tot 100 100 24.42

Controlling for intergenerational effects only explained a little clustering: the value of 25.48

from table 4 was reduced to 24.42, or almost 96 per cent of its original size. This

interpretation of percentual reduction has similarities with that of the R² in OLS regression, or

the total variability explained. Intergenerational effects in Antwerp, although clearly present,

only explain just over 4 per cent of all variation in infant deaths between families. Again, this

information can be plotted in a graph, and compared to the graph of overall clustering in

Antwerp, see figure 3. This exercise can be repeated for the other regions as well, see figures

4 (for Sundsvall), 5 (for Skellefteå) and 6 (for the Northern Inland parishes).

From the figures 3, 4, 5 and 6, it becomes clear that the intergenerational transmission

of infant deaths does not explain a lot of clustering, if any at all. Whereas in Antwerp and in

Skellefteå clustering is a little bit reduced, in Sundsvall we see no explained variation at all.

Only in the Northern Inlands the intergenerational transmission of infant death risk seems to

matter significantly. In all regions, except Sundsvall, the intergenerational effects explain why

there were so few families with one or two infant deaths, especially in the Northern Inlands,

but it is only a tiny bit of the answer.

13

Figure 3. The shape of death clustering in Antwerp,

controlling for intergenerational transmission of infant deaths

-0,1

-0,08

-0,06

-0,04

-0,02

0

0,02

0,04

0,06

0,08

0,1

0 1 2 3 4+

n° of deaths

diffe

renc

e in

obs

erve

d vs

exp

ecte

d pr

opor

tion

of fa

mili

es

ANTW controlledANTW uncontrolled

Figure 4. The shape of death clustering in the Sundsvall region,

controlling for intergenerational transmission of infant deaths

-0,1

-0,08

-0,06

-0,04

-0,02

0

0,02

0,04

0,06

0,08

0 1 2 3 4+

n° of deaths

diffe

renc

e in

obs

erve

d vs

exp

ecte

d pr

opor

tion

of fa

mili

es

SUL controlledSUL uncontrolled

14

Figure 5. The shape of death clustering in the Skellefteå region,

controlling for intergenerational transmission of infant deaths

-0,04

-0,03

-0,02

-0,01

0

0,01

0,02

0,03

0,04

0 1 2 3 4+

n° of deaths

diffe

renc

e in

obs

erve

d vs

exp

ecte

d pr

opor

tion

of fa

mili

es

SKE controlledSKE uncontrolled

Figure 6. The shape of death clustering in the Northern Inlands,

controlling for intergenerational transmission of infant deaths

-0,05

-0,04

-0,03

-0,02

-0,01

0

0,01

0,02

0,03

0,04

0,05

0 1 2 3 4+

n° of deaths

diffe

renc

e in

obs

erve

d vs

exp

ecte

d pr

opor

tion

of fa

mili

es

NIL controlledNIL uncontrolled

15

Discussion

Death clustering in families is a concept that recently has received a lot of attention, but what

does it really mean? It is all about heterogeneity, about variation that is natural, and variation

that is systematic. Therefore it is a very difficult topic to grasp intuitively. Or as a participant

to a workshop on death clustering phrased it: “clustering is not about the mean mean, it is

about the variability of the variance”. It is about what makes infant deaths for one family so

different than for the other. Because this is so closely related to aspects of fertility, most

importantly family size, researching clustering requires large methodological thresholds.

In this paper, we made an attempt to explain systematic variation of infant deaths

across families by inheritance of death risks. In previous research we found that death risks

are transmitted through generations, which left us wondering whether that created differences

between families in infant mortality risks. We found that this was not a major cause of

clustering, which leaves us with questions on the origin of the phenomenon.

However, tackling this question from an international comparative perspective proved

to be fruitful. Using a prototype method, we were able to compare clustering in several

regions. Although the extent of clustering, once made measurable, differed across regions, its

nature was the same, regardless of the socioeconomic or cultural background. Crudely

spoken, a family either has no infant deaths, or a lot of infant deaths. Clustering, and how it

looks like, is seemingly universal. Also, whenever intergenerational transmission of deaths

impacts clustering, it is affecting the shape of it in a similar way. It is part of the answer why

families with only a few deaths are underrepresented everywhere.

This leads us to the true meaning of clustering. Clustering of infant deaths between

families is all about variation between families, or heterogeneity of families. In a region with

more heterogeneity, like the modernizing city of Antwerp with its generation-long tradition of

immigration from all over the world, or the booming sawmill centres around Sundsvall,

clustering is at its strongest. The intergenerational transmission of infant mortality risks only

affects this heterogeneity, if there is large variation in how this transmission works. The fact

that intergenerational transmission explained most clustering in the Northern Inlands, is

therefore maybe not so surprising. The century-long cohabitation of both Swedes and Sámi

makes the population less heterogeneous than in Antwerp, at least in terms of the spread of

infant deaths in families. Death clustering in Antwerp, namely, is stronger than in the

Northern Inlands. However, when intergenerational patterns are involved, the Northern

Inlands are very different than Antwerp. This means that, although intergenerational

transmission is present in all regions, it creates more heterogeneity in families in the Northern

16

Inlands than elsewhere, especially in families where only one or two infants die. Infant deaths

in the family’s previous generations seems to place a heavier burden on the families

inhabiting the vast, desolated lands in the North, than on families elsewhere, be it in Sweden

or as distant as Antwerp. In the Northern Inlands, the genealogical history of infant mortality

is a distinguishing characteristic. It does not seem to make a family a high-risk family, but it

creates the difference between losing one or two infants, and losing none.

17

Bibliography Arulampalam, W. and S. Bhalotra. 2006. Sibling death clustering in India: state dependence

versus unobserved heterogeneity, Journal of the Royal Statistical Society Series A Curtis, S. L., I. Diamond and J. W. McDonald. 1993. Birth interval and family effects on

postneonatal mortality in Brazil, Demography 30(1): 33–43. Das Gupta, M. 1990. Death clustering, mother’s education and the determinants of child

mortality in rural Punjab, India, Population Studies 44(3): 489–505. Das Gupta, M. 1997. Socio-economic status and clustering of child deaths in rural Punjab,

Population Studies 51(2): 191–202. Edvinsson, S., A. Brändström, J. Rogers and G. Broström. 2005. High-risk families: The

unequal distribution of infant mortality in nineteenth-century Sweden, Population Studies 59(3): 321–337.

Derosas, R. and G. Broström. 2010. Infant death clustering in nineteenth-century Venice,

Paper presented at the international workshop for death clustering, Umeå University, Sweden, 22-23 October 2010.

Guo, G. 1993. Use of sibling data to estimate family mortality effects in Guatemala,

Demography 30(1): 15–31. Hill, A. G. and A. Aguirre. 1990. Childhood mortality estimates using the preceding birth

technique: some applications and extensions, Population Studies 44(2): 317–340. Holmberg, H. and G. Broström. 2010. On statistical methods for clustering; a case study on

infant mortality, northern Sweden 1831-1890, Paper presented at the international workshop for death clustering, Umeå University, Sweden, 22-23 October 2010.

Janssens, A., M. Messelink and A. Need (2010). Faulty genes or faulty parents? Gender,

family, and survival in early and late childhood in the Netherlands, 1860-1900, The History of the Family 15(1): 91-108.

Lynch, K. A. and J. B. Greenhouse (1994). Risk factors for infant mortality in nineteenth-

century Sweden, Population Studies 48(1): 117–133. Matthijs, K. and S. Moreels. 2010. The Antwerp COR*-database: a unique Flemish source for

historical-demographic research, History of the Family 15(1): 109–115. Omariba, D. W. R., F. Rajulton and R. Beaujot. 2008. Correlated mortality of siblings in

Kenya: the role of state dependence, Demographic Research 18: 311–336. Preston, S. 1978. The effects of infant and child mortality on fertility. New York: Academic

Press.

18

Saha, U.R. and A. van Soest. 2011. Infant death clustering in families: magnitude, causes, and the influence of better health services, Bangladesh 1982-2005, Population Studies Forthcoming (available online 14 September 2011).

Vandezande, M. 2010. Born to die. Intergenerational patterns of infant mortality, the Antwerp

district, 1846-1905. Paper presented at the 35th Annual Meeting of the Social Science History Association, Chicago, IL, USA, 20 November 2010.

Vandezande, M. 2011, submitted. Using death clustering in an attempt to reassess the impact

of men on infant mortality. An intergenerational approach, special issue in Biodemography and Social Biology.

Zaba, B. and P.H. David. 1996. Fertility and the distribution of child mortality risk among

women: an illustrative analysis, Population Studies 50(2): 263–278. Zenger, E. 1993. Siblings’ neonatal mortality risks and birth spacing in Bangladesh,

Demography 30(3): 477–488.

19

Appendices Appendix 1. Map of the Antwerp district within Belgium

20

Appendix 2. Map of the Swedish regions of Sundsvall and Skellefteå

21

Appendix 3. Map of the Northern Inland parishes, in the northern half of Sweden