1. Epidemiology and Global Health
Download icon

The mother’s risk of premature death after child loss across two centuries

  1. Unnur A Valdimarsdóttir  Is a corresponding author
  2. Donghao Lu
  3. Sigrún H Lund
  4. Katja Fall
  5. Fang Fang
  6. Þórður Kristjánsson
  7. Daníel Guðbjartsson
  8. Agnar Helgason
  9. Kári Stefánsson  Is a corresponding author
  1. University of Iceland, Iceland
  2. Harvard TH Chan School of Public Health, United States
  3. Karolinska Institutet, Sweden
  4. Brigham and Women's Hospital, Harvard Medical School, United States
  5. deCODE Genetics, Iceland
  6. Örebro University, Sweden
Research Article
  • Cited 0
  • Views 520
  • Annotations
Cite this article as: eLife 2019;8:e43476 doi: 10.7554/eLife.43476

Abstract

While the rare occurrence of child loss is accompanied by reduced life expectancy of parents in contemporary affluent populations, its impact in developing societies with high child mortality rates is unclear. We identified all parents in Iceland born 1800–1996 and compared the mortality rates of 47,711 parents who lost a child to those of their siblings (N = 126,342) who did not. The proportion of parents who experienced child loss decreased from 61.1% of those born 1800–1880 to 5.2% of those born after 1930. Child loss was consistently associated with increased rate of maternal, but not paternal, death before the age of 50 across all parent birth cohorts; the relative increase in maternal mortality rate ranged from 35% among mothers born 1800–1930 to 64% among mothers born after 1930. The loss of a child poses a threat to the survival of young mothers, even during periods of high infant mortality rates.

https://doi.org/10.7554/eLife.43476.001

Introduction

The death of a child is arguably one of the most stressful events that a parent can experience. Grief is the natural emotional response to such loss and has been described across human populations (Lindemann, 1944; Bowlby, 1980; Parkes, 2014; Boyden et al., 2014; Einarsdóttir, 2004). Intense or prolonged grief can induce behaviours and physiological states that are detrimental to health and survival. Thus, child loss has been associated with elevated risks of psychiatric symptoms (Kreicbergs et al., 2004), psychiatric hospitalizations (Li et al., 2005) and sickness absence (Wilcox et al., 2015), cardiovascular disease (Li et al., 2002a; Li et al., 2003a), and some cancers (Fang et al., 2011; Huang et al., 2013; Li et al., 2002b). Studies also suggest that parents, particularly mothers, who lose a child face increased risk of death (Li et al., 2002a; Li et al., 2003a; Levav et al., 1988; Li et al., 2003b; Qin and Mortensen, 2003; Agerbo, 2005; Rostila et al., 2012; Rostila et al., 2015; Espinosa and Evans, 2013; Harper et al., 2011; Chen et al., 2012; Schorr et al., 2016), both from suicide and natural causes. The aforementioned studies were all conducted in affluent developed populations, where infant mortality rates are low and life expectancy is high. In contemporary developing nations, infant mortality rates (number of deaths during first year per 1000 live births) >50 are common, whereas in developed nations they are usually lower than eight (United Nations, 2015). Parent-child attachment in pre-industrial and developing populations has been a matter of debate, with some historians arguing for parental reluctance to invest emotions in children where mortality rates are high, while others argue that parents' attachment and grief after child loss is similar across human populations, past and present (Magnússon, 2010; Woods, 2003). One way to shed light on this issue is to study the impact of child loss on parent mortality in the same population before and after a demographic transition that dramatically altered infant mortality rates.

Here, we examine the association between child loss and parent mortality in Iceland from 1817 until 2015, a period that spans a drastic transition from a poor agricultural society to an affluent industrial society. One marker of this drastic change is that infant mortality rates in Iceland (per 1000 births) dropped from one of the highest known values in history, 238 in 1881–1885 (Jónsson and Magnússon, 1997), to one of the lowest, 2.07 in 2005–2010 (Statistics Iceland, 2019). We used a comprehensive genealogical database to assess whether the relative rate of parent mortality after child loss changed during this transition, and to determine whether this rate was further affected by factors such as parent’s sex and age at loss, the number of children as well as the age of the lost child. In line with some previous studies (Li et al., 2003b), we used parent mortality as a measure of the cumulative health response to the emotional and physiological strain of child loss. While mortality obviously does not capture the entire range of such reactions, it has the advantage of being reliably recorded at different times in a genealogical database.

Results

We followed all Icelandic parents, born 1800 or later, from the birth of their first child (earliest from 1817) until death or to the end of 2015 (see further in Materials and methods). From 1817 to 2015, 64 044 parents lost at least one child (Supplementary file 1 - Table 1). The proportion of parents who experienced child loss decreased from 61.1% among parents born 1800–1880 to 5.2% among parents born 1931–1996. To adjust for potential confounding by socioeconomic status or other familial factors that might affect the risk of mortality, our primary analyses were focused on the 47,711 parents who lost a child and had at least one sibling who did not lose a child (N = 126,342). Compared to their siblings who did not lose a child, parents who lost a child were more likely to be women, have more children, be younger at first childbirth and older at cohort entry, and have a longer follow-up time (p<0.05; Table 1). Similar patterns were observed for mothers and fathers who lost a child, respectively (Supplementary file 1 - Table 2).

Table 1
Descriptive characteristics of parents born from 1800 to 1996 who lost a child by death during their life course, and their siblings who did not lose a child, N (%).
https://doi.org/10.7554/eLife.43476.002
Parents with lossParents without lossP for difference
Total number47 711126 342-
Sex
Female25 125 (53)63 328 (50)3.65e-21
Male22 586 (47)63 014 (50)
Total number of children276 819531 736-
Number of children, mean (SD)5.80 (3.19)4.21 (2.71)0
Number of children
11775 (4)16 012 (13)0
2–417 766 (37)64 908 (51)
5–921 947 (46)39 019 (31)
10+6223 (13)6403 (5)
Age at first child birth, mean (SD)25.55 (5.09)27.16 (6.11)0
Age at matching*, mean (SD)38.76 (15.39)37.35 (14.60)8.84e-67
Age at matching*
13–3017 963 (38)52 434 (42)1.84e-70
31–5019 980 (42)51 348 (41)
51–758123 (17)19 480 (15)
76+1645 (3)3080 (2)
Age of deceased child, mean (SD)9.44 (14.95)--
Length of follow-up, mean (SD)31.79 (18.36)25.45 (18.03)0
Age at death70.81 (16.66)71.80 (16.20)1.40e-21
  1. * Age at loss of first child for parents who lost a child; the same age for their siblings who did not lose a child or the age when the siblings became parents, whichever came later.

Table 2
Hazard ratios (HRs) and 95% confidence intervals (CIs) of premature mortality after loss of a child among young parents followed up to age 50, by time since loss and demographic characteristics, compared to their siblings.

We stratified by sibling groups and additionally adjusted for birth year and sex. IR, incidence rate, per 1000 person-years.

https://doi.org/10.7554/eLife.43476.003
OverallWomenMen
N (Crude IR)HR (95% CI)N (Crude IR)HR (95% CI)N (Crude IR)HR (95% CI)
Time since loss
0–4 years1711 (10.2)1.19 (1.11–1.27)907 (10.3)1.52 (1.35–1.71)804 (10.0)0.96 (0.85–1.07)
0–4 years*1595 (9.5)1.10 (1.03–1.18)832 (9.4)1.40 (1.24–1.58)763 (9.5)0.89 (0.80–1.01)
5–9 years1145 (7.9)1.13 (1.05–1.23)508 (6.5)1.22 (1.06–1.41)637 (9.5)1.02 (0.90–1.16)
10–19 years1654 (8.2)1.15 (1.07–1.23)764 (6.7)1.33 (1.18–1.50)890 (10.0)1.09 (0.97–1.22)
20–39 years388 (6.2)1.12 (0.98–1.29)214 (5.3)1.28 (1.03–1.59)174 (7.9)1.16 (0.91–1.48)
Child’s age at loss
03357 (8.9)1.14 (1.08–1.20)1672 (8.0)1.43 (1.31–1.57)1685 (9.9)0.94 (0.86–1.03)
1–52016 (9.4)1.16 (1.09–1.24)964 (8.1)1.26 (1.13–1.41)1052 (11.0)1.12 (1.00–1.25)
6–17269 (6.7)1.17 (0.99–1.38)130 (5.6)1.31 (0.99–1.74)139 (8.3)1.32 (1.00–1.73)
18+66 (5.4)1.44 (1.05–1.97)22 (2.8)0.96 (0.54–1.71)44 (10.3)1.99 (1.24–3.20)
Number of alive children at loss
02390 (9.7)1.20 (1.13–1.28)1135 (8.3)1.39 (1.24–1.56)1255 (11.4)1.15 (1.04–1.28)
1–32793 (8.4)1.12 (1.06–1.18)1379 (7.5)1.37 (1.25–1.51)1414 (9.5)0.98 (0.89–1.07)
4+525 (8.1)1.12 (0.99–1.26)274 (7.3)1.18 (0.97–1.44)251 (9.1)0.94 (0.77–1.14)
Sex of the lost child
Female3075 (8.6)1.14 (1.08–1.20)1483 (7.5)1.29 (1.17–1.41)1592 (10.0)1.06 (0.97–1.16)
Male2633 (9.2)1.17 (1.10–1.24)1305 (8.2)1.44 (1.30–1.59)1328 (10.4)1.01 (0.92–1.11)
Age at loss
13–303459 (8.5)1.13 (1.08–1.19)1815 (7.5)1.33 (1.22–1.45)1644 (10.1)1.02 (0.93–1.12)
31–401945 (9.5)1.19 (1.12–1.27)873 (8.7)1.43 (1.27–1.61)1072 (10.2)1.04 (0.93–1.15)
41–50304 (8.7)1.14 (0.98–1.32)100 (6.1)1.11 (0.81–1.52)204 (11.0)1.12 (0.90–1.40)
Age at first childbirth
13–21988 (6.0)1.20 (1.09–1.32)653 (5.4)1.29 (1.13–1.47)335 (7.7)1.15 (0.95–1.40)
22–241507 (8.5)1.09 (1.01–1.17)829 (8.0)1.27 (1.12–1.44)678 (9.2)0.97 (0.84–1.11)
25–271520 (10.3)1.20 (1.11–1.30)665 (9.5)1.35 (1.17–1.56)855 (11.1)1.08 (0.96–1.23)
28+1693 (10.8)1.16 (1.08–1.24)641 (9.9)1.55 (1.34–1.80)1052 (11.3)1.01 (0.91–1.13)
  1. * A sensitivity analysis excluding parents dying on the same day as their child.

Parent mortality rate after child loss across two centuries

We used a conditional Cox proportional hazards model to contrast the mortality rates of parents who lost at least one child to their respective siblings who did not (Figure 1). Throughout the entire study period, we observed an elevated hazard ratio (HR) of overall mortality rate among parents who lost a child (1.04; 95% confidence interval [CI] 1.02–1.06), which was mostly driven by maternal (HR 1.07; 95% CI 1.04–1.10) rather than paternal mortality rates (HR 1.02; 95% CI 0.99–1.05; p-for-difference=0.0096). Figure 1 shows that there are no obvious temporal trends in the hazard ratios, except for an increase in mortality rate after child loss for parents born after 1930 (p-for-difference = 0.002 when compared to parents born 1800–1930). Similar results were obtained from the population-based matched cohort and a subset of this cohort with identifiable siblings, separately, (Figure 1—figure supplement 1), albeit with greater HRs revealing the importance of the sibling-controlled design to adjust for unmeasured familial confounding factors. The HRs changed minimally after exclusion of parent-child pairs with the same date of death (Figure 1—figure supplement 2).

Figure 1 with 2 supplements see all
Hazard ratios (HRs) and 95% confidence intervals (CIs) of parental mortality after loss of a child by birth cohorts (every 20 years until 1900, 1901–1930, and 1931–1996), in the sibling cohort.

We estimated HRs from stratified Cox proportional hazards model using age as underlying time scale. We stratified by sibling groups and additionally adjusted for birth year and sex. IR, incidence rate, per 1000 person-years.

https://doi.org/10.7554/eLife.43476.004

Child loss and premature maternal mortality rate

As the average life expectancy of Icelanders increased considerably from 1817 to the present, we performed separate analyses of parent mortality before and after the age of 50 years (Figure 2). Before the age of 50, mothers who lost a child experienced increased mortality (HR 1.36; 95% CI 1.27–1.45) during the whole study period with the HRs consistently elevated across all seven birth cohorts (Figure 2a). For mothers, the relative mortality rate increase ranged from 35% (born 1800–1930) to 64% (born after 1930), yet was not statistically significantly different between these two cohorts (p=0.0641). In contrast, no such excess mortality was observed for fathers who lost a child in any of the birth cohorts (HR for entire observation period 1.04; 95% CI 0.97–1.11). Interestingly, the excess maternal mortality rates before the age of 50 after child loss seem limited to loss of a child younger than 18 years, while the elevation in paternal mortality rates are more notable after loss of an adult child (Table 2). The difference between mothers and fathers in excess mortality rate before the age of 50 is particularly marked after loss of infants, representing 59% of child losses during the study period. These results were robust to the exclusion of mothers who died within a week after giving birth (Figure 2—figure supplement 1), making it unlikely that the higher mortality rate in mothers is explained by delivery or postpartum complications that resulted in the deaths of both newborn and mother (Nour, 2008).

Figure 2 with 6 supplements see all
Hazard ratios (HRs) and 95% confidence intervals (CIs) of parental mortality after loss of a child, by age bands at follow-up (age 14–50 and 51+) and birth cohorts (every 20 years until 1900, 1901–1930, and 1931–1996), in the sibling cohort.

We stratified by sibling groups and additionally adjusted for birth year and sex in Cox proportional hazards model. IR, incidence rate, per 1000 person-years.

https://doi.org/10.7554/eLife.43476.007

Later-life mortality rate after child loss

After the age of 50 and across the entire study period, neither mothers (HR 1.02; 95% CI 0.99–1.05) nor fathers (HR 1.01; 95% CI 0.98–1.05) suffered increased mortality rate after child loss. Separately analyzed by time since loss and demographic characteristics, we did not observe significantly increased mortality in response to child loss for parents after 50 years of age (Supplementary file 1 - Table 3). However, when broken down into the seven smaller birth-cohort periods, we observed significantly increased mortality in parents older than 50 who lost a child and were born after 1930 (HR 1.11; 95% CI 1.03–1.20; p=0.0063), particularly among fathers (HR 1.21; 95% CI 1.08–1.37; p=0.0009; Figure 2b). Curiously, this recent increase in mortality after the age of 50 years is associated with child loss that mostly occurred before the age of 50 (Figure 2—figure supplement 2). As a result, we postulate that the recent increase in parent mortality rate after the age of 50, particularly in fathers, may be due to an increased risk of long-term social isolation after child loss. In support of this notion, we indeed found very high mortality rates for parents born after 1930 who lost a child and had no living children at the age of 50, regardless of the sex of deceased child (Supplementary file 1 - Table 4).

In a sensitivity analysis, restricting to parent-sibling pairs born within five years from the parents who lost a child, comparing same-sex siblings, or adjusting for the number of children alive at time of matching, we observed similar estimates of parental mortality rates after child loss (Figure 2—figure supplements 35). The only notable difference was that, in the birth cohort 1931–1996, the impact of child loss on maternal mortality after 50 years of age became statistically significant. We further restricted unexposed siblings to those who were already parents at the age when the index sibling lost a child, which also yielded similar results (Figure 2—figure supplement 6).

Discussion

The study period spans two centuries of drastic demographic transition in Iceland, from a poor developing nation with a particularly high infant mortality rate, to a modern affluent nation with one of the lowest rates of infant mortality in history. Although we observed a more modest elevation in overall parent mortality rates after child loss before 1930 compared to thereafter, the most striking finding is the consistent increase in premature maternal deaths in response to child loss throughout the 199 years of observation.

This study leverages a unique nationwide data source of genealogies with virtually complete follow-up over a 199 year period spanning an extreme change in indicators of societal development, for example infant and child mortality rate. Although based on an unusually large set of data, our study has several limitations that should be noted. First, we have no information about the causes of death for children or parents. Such data could be used to help verify the causal relationship between deaths of children and parents and to uncover patterns in the causes of death among parents (e.g. suicide or cardiovascular deaths) after child loss. Some impact of reverse causation, for example the ill-health of parents increasing the mortality risk of their children, cannot be ruled out. However, the health and welfare of children was even more dependent on their parents, particularly fathers, in pre-demographic transition (pre-welfare state times) societies. The relatively stable effect estimates of parent mortality rates after child loss across calendar periods and the consistently lower impact of child loss on paternal mortality argues against a strong influence of reverse causality. Second, we did not have direct information about the socioeconomic status of parents. However, as socioeconomic status tends to cluster in families, our sibling-controlled analysis is likely to adjust for much of the confounding due to this source. Third, it is possible that the fragmented registration of births in earlier times (particularly in the case of infant deaths) resulted in misclassification of some parents who had truly lost a child and, thus, somewhat conservative estimates of excess parent mortality rates in earlier times. Last, in the sibling cohort, parents who lost a child differed somewhat from those that did not lose a child with respect to sex, number of children, age at first birth, age at entry to the cohort, length of follow-up time and age at death. We therefore made every effort to adjust for these covariates in our analyses. We adjusted for age at follow-up and sex in all analyses and presented separate results for the different birth cohorts. We further performed subgroup analyses by age of parents at loss (or matching), parent age at first childbirth, follow-up time and number of children alive at loss, and demonstrated that the association between child loss and parent mortality is observed across the range of these covariates. Indeed, parents who lost a child may have to live longer to experience such an event, however, such survivor bias (if any) due to a selection of survivors to the exposed group would rather yield conservative estimates of parent mortality after a child loss. On the other hand, the siblings used for comparison that were not yet parents at the age as when the index parents lost a child may have a survival advantage, which could yield exaggerated estimates of parent mortality after child loss. However, we observed similar results in a sensitivity analysis that was restricted to unexposed siblings who were already parents at the same age as when the index parents lost a child.

The large increase in the rate of premature maternal mortality after child loss in 19th century Iceland suggests that mothers formed strong emotional bonds with their children – in spite of very harsh conditions and a strong expectation of child loss. This is consistent with accounts of intense grieving among mothers who lost a child in a contemporary developing society (Einarsdóttir, 2004). In contrast, the paternal mortality response to child loss is much weaker throughout the observation period and only statistically significant for older fathers (i.e. parent mortality after 50 years of age) born after 1930.

What could account for this stable sex difference in parent mortality rate after child loss? One possibility is a difference in the attachment of mothers and fathers to their children, which could result in a greater emotional response of mothers and concomitant psychological and physiological consequences. Throughout the 199-year study period, mothers may have had greater opportunity than fathers to form early emotional attachment to their infants through gestation and breastfeeding. Although historical accounts suggest that breastfeeding may have been rarer in 19th to early 20th century Iceland (Garðarsdóttir, 2005) than in the present, it does not seem to have prevented mothers from forming strong emotional bonds with their children as illustrated by the elevated risks of maternal death after child loss. It is possible that females are born with a tendency to form stronger emotional bonds with children than males. Although our results cannot disentangle the impact of nature and nurture in this context, it is interesting to note that the mortality response of fathers to child loss before the age of 50 has not significantly increased despite the greater contribution of fathers to the postnatal care of children in recent decades (Huerta et al., 2014).

Another possible explanation for our findings is a sex difference in the emotional response to traumatic events in general. Women have been reported to have a greater risk of mental disorders after exposure to emotional trauma than men (Tolin and Foa, 2006). The intensity of such responses may result in a multitude of other mental and physical ailments which in turn can affect the risk of mortality through, for example suicide (Wilcox et al., 2009; Panagioti et al., 2012; Gradus et al., 2010) and cardiovascular disease (Song et al., 2019; Edmondson et al., 2013; Gradus et al., 2015). Previous studies in contemporary developed populations have reported evidence for greater risk of psychiatric hospitalizations (Li et al., 2005) and mortality (Li et al., 2003b) after child loss in mothers than in fathers. Interestingly, an opposite sex difference is seen for mortality rate after spousal loss, where mortality rate elevation is more pronounced among widowers than widows (Moon et al., 2011; Shor et al., 2012). However, this difference may in part be due to a greater dependence of older males on social and emotional support provided by spouses. A similar mechanism is indicated in our results, namely the increased mortality of older parents after child loss in more recent times.

Our results show that even in the poor agricultural population of 19th century Iceland, with one of the highest infant mortality rates measured in humans, there was a considerable impact of child loss on premature maternal mortality rates. This contradicts claims that mothers in poor cultures adapt to high infant mortality rates with reduced emotional investment (as reviewed in Woods, 2003). Indeed, one implication of our findings is that child loss is likely to constitute a major threat to the survival of mothers in societies with high infant mortality rates. Taken together these findings highlight the importance of the mother-child bond and the extensive health threats that arise when it is broken.

Materials and methods

Data sources and study participants

Request a detailed protocol

The data source for the current study is the deCODE Genetics genealogy database, containing information on kinship, dates of birth and death of Icelandic population largely since settlement (Gudbjartsson et al., 2015). The database was constructed by compiling information from church records, various annals and censuses from 1703 to 1930 as well as the contemporary registers of the total population. The database contains an almost complete record of the ancestors of contemporary Icelanders back to around 1650 and, after 1880, is based on death records from a geographically comprehensive set of parish records. In terms of linkages between parents and offspring, the completeness of the genealogical database across the entire observation period is 99%. The database has been extensively evaluated for the accuracy of such linkage using genetic data (Sun et al., 2012; Helgason et al., 2015; Halldorsson et al., 2019).

We conducted a historical cohort study of parents who were born from January 1st 1800 through December 31st 1996 and were living in Iceland. In total 323,510 individuals were eligible for inclusion; 27,704 individuals who emigrated after 1906 were not included in the study population. Individuals entered the cohort from the birth year of their first child (earliest from 1817), and were followed until their own death or through 2015, whatever happened first. We excluded 12,233 (3.8%) parents who were lost to follow-up.

This study was reviewed by the National Bioethics Committee and Data Protection Authority of Iceland (Approval No. VSN-12–125 and VSN-16–156) that in accordance to Icelandic law (Act no. 44/2014 on scientific research within the health sector and Act no. 77/2000 on the protection of privacy in processing of personal data) permit large-scale, population-based register studies, such as this one, to be carried out without informed consent.

Ascertainment of exposure to child loss and covariates

Request a detailed protocol

All parents entering the cohort were followed for exposure to child loss, indicated by the date of a child's death being earlier or on the same day as the parent’s date of death. Information about mortality in the Icelandic genealogy database are obtained from parish records, from 1735 to 1970, and thereafter from the Population Register held by Statistics Iceland. We further obtained information about the parents’ sex and age when having their first child, the total number of children and their sex. In cases where parents experienced child loss, we obtained information about parents’ age when losing their first child and the child's age at the time of death.

Follow-up and ascertainment of parental mortality

Request a detailed protocol

We followed all parents from the year of entry to the study (when first becoming parents). The person-time of parents was defined as unexposed from entry to the cohort until the loss of their first child, death, or end of the observation period (end of 2015), whichever happened first. Person-time was defined as exposed (for parents who lost a child) from the date of their first child loss until death or end of the observation period (end of 2015).

Population-based matched cohort

Request a detailed protocol

Since life expectancy has changed dramatically across the study period, we performed population-based matched cohort analysis, based on 64,050 parents who lost a child during follow-up, who were matched on birth year (±1 year) and sex to parents who did not lose a child. To this end, we randomly selected three parents (born before 1920) or five parents (born 1920 or later) who had not lost a child at the time when the index parent lost a child (i.e. the reference time). Six parents who lost a child but could not be matched to any parents without loss were excluded. The resulting data set comprised 64,044 parents who lost a child and 218,824 matched parents who did not lose a child. We followed all participants from the reference time until death or December 31st, 2015, whichever came first.

Siblings-based cohort

Request a detailed protocol

To allay the concerns of potential confounders shared by family members (such as socioeconomic status, lifestyle and genetic components), we restricted our primary analysis to a sibling-based cohort. Of note, although the siblings used as reference in this analysis did not lose a child, they did lose a niece or nephew which, if truly influencing the rate of mortality, may lead to conservative estimates. Among 64,044 parents with loss, 54,532 (85.1%) had at least one sibling defined as sharing the same biological mother. Each parent with loss was paired with three available siblings on average (range 2–16) who had not lost a child at the same age when the index parent lost a child (i.e. the reference age), whereas 6821 parents were excluded because they only had siblings who had experienced child loss. We followed all participants from the reference age, or the age when the siblings became parents, whichever came later, until the age at loss of a child, death, or December 31st, 2015, whichever occurred first. Specifically, siblings who did not lose a child started to contribute to the follow-up at the same age as the parent who lost a child or at a later age when they became parents. The sibling-based cohort comprised 47,711 (74.5%) parents who lost a child and 126,342 matched siblings without loss.

Statistical analysis

Request a detailed protocol

We first compared the characteristics of parents who lost a child (exposed) with those who did not (unexposed) across birth cohorts for the following covariates: sex, mean age at loss, mean child's age at the time of death and total number of children. In this analysis, we used a chi-square test for the categorical and a t-test for the continuous covariates. We then calculated the crude mortality rates (deaths per 1000 person-years) in both exposed and unexposed cohorts, overall and stratified by parental sex, birth cohorts, and/or age at follow-up.

Although we performed parallel analyses on the population-based matched cohort, the primary analysis was focused on the siblings-based cohort which provides a better control of unmeasured confounders shared by siblings. We used conditional Cox proportional hazard regression models to estimate hazard ratios (HRs) and 95% confidence intervals of mortality for exposed parents relative to unexposed parents, while stratifying on the matched pairs in analyses of population-based matched cohort or sibling groups in analyses of siblings-based cohort. We used time since the reference time and age at follow-up as the underlying timescale for population-based matched cohort and sibling-based cohort, respectively. All models were additionally adjusted for birth year for the population-based matched analysis (because ±1 birth year was allowed in matched pairs), and adjusted for birth year and parental sex in the sibling-based analysis. Of note, we did not match on parental sex in the sibling-based cohort but adjusted for it in all models and thereby accounting for the effect of sex on mortality rate across the sibling groups. Matched pairs or sibling groups who were invariant with regard to the outcome were included, although they did not contribute to the estimates. Guided by the timing of the dramatic societal change and to ensure sufficient statistical power for each period, the same analyses were performed for seven birth cohorts separately (every 20 years from 1800 to 1900, 1901–1930, and 1931–1996). Because of the better control of potential confounders, we only performed the subsequent analyses in the sibling-based cohort.

To investigate potential bias due to common external causes of death (e.g. pandemics or accidents taking the lives of both parents and children) we performed a sensitivity analysis excluding all parent-child pairs with identical date of death. Due to differential impact across birth cohorts, the subsequent analyses were always separated by birth cohorts. All analyses were performed on men and women separately.

As the average life expectancy of Icelanders increased considerably from 1817 to the present, we performed separate analyses of parent mortality before and after the age of 50 years, that is 13-50 and ≥51 years. To reduce the potential differences between siblings we, in sensitivity analyses, restricted to parent-sibling pairs born within 5 years from the parents who lost a child (N = 96,041, 55.2%) as well as to unexposed siblings who were already parents at the same age as when the index parents lost a child (N = 135,719, 78.0%). To further address the potential influences of parent sex and number of children, in another sensitivity analysis, we only compared same-sex siblings and additionally adjusted for number of alive children at the time of loss/matching, respectively.

Parent mortality before and after age of 50 years was further analyzed in subgroups of age of the deceased child (0, 1–5, 6–17, or ≥18 years old), number of children alive at loss (0, 1–3, or ≥4), sex of the deceased child, parental age at loss (13–30, 31–40, or 41–50 years old), and parental age at first childbirth (13–21, 22–24, 25–27, or ≥28 years).

The excess risk of mortality beyond age 50 among parents born 1931–1996 was further explored with respect to the age of parents when the loss occurred (13–30, 31–50 or ≥51 years); as well as sex of deceased child and number of children by age 50.

All statistical analyses were performed using R, version 3.3.1. Survival analysis was performed with the packages ‘survival’ and graphs were plotted with the package ‘ggplot2’. All R-codes used in the primary analyses are available (Source Code File 1).

References

  1. 1
  2. 2
    Attachment and Loss
    1. J Bowlby
    (1980)
    Loss, Sadness and Depression, Attachment and Loss, 3, New York, Basic Books.
  3. 3
  4. 4
  5. 5
  6. 6
    Mother Love, Child Death and Poverty in Guinea-Bissau
    1. J Einarsdóttir
    (2004)
    Tired of Weeping, Mother Love, Child Death and Poverty in Guinea-Bissau, 2nd edition, Madison, Wisconsin, University of Wisconsin Press.
  7. 7
  8. 8
  9. 9
    Brjóstagjöf og áhrif hennar á lífslíkur ungbarna á Íslandi 1910-1925
    1. Ó Garðarsdóttir
    (2005)
    Ljósmæðrablaðið 83:19–27.
  10. 10
  11. 11
  12. 12
  13. 13
  14. 14
  15. 15
  16. 16
  17. 17
  18. 18
    Hagskinna: Icelandic Historical Statistics
    1. G Jónsson
    2. MS Magnússon
    (1997)
    Reykjavik, Iceland: Hagstofa Íslands.
  19. 19
  20. 20
  21. 21
  22. 22
  23. 23
  24. 24
  25. 25
  26. 26
  27. 27
    A Social History of Iceland
    1. SG Magnússon
    (2010)
    99–111,  Wasteland with Words, A Social History of Iceland, London, Reaktion Books.
  28. 28
  29. 29
    An introduction to maternal mortality
    1. NM Nour
    (2008)
    Reviews in Obstetrics & Gynecology 1:77–81.
  30. 30
  31. 31
    The Price of Love: The Selected Works of Colin Murray Parkes
    1. C Parkes
    (2014)
    London: Routledge.
  32. 32
  33. 33
  34. 34
  35. 35
  36. 36
  37. 37
  38. 38
  39. 39
  40. 40
  41. 41
    Statistical Yearbook 2012, Fifty-Seventh Issue
    1. United Nations
    (2015)
    New York, United States: United Nations.
  42. 42
  43. 43
  44. 44

Decision letter

  1. M Dawn Teare
    Reviewing Editor; Newcastle University, United Kingdom
  2. Eduardo Franco
    Senior Editor; McGill University, Canada
  3. Susan C Alberts
    Reviewer; Duke University, United States

In the interests of transparency, eLife includes the editorial decision letter and accompanying author responses. A lightly edited version of the letter sent to the authors after peer review is shown, indicating the most substantive concerns; minor comments are not usually included.

Acceptance summary:

This is an interesting and valuable analysis that uses the historic registry data in Iceland to ask whether losing a child increases the risk of death for the parents. As the authors state, several papers have shown loss of a child is associated – in observational studies – with a higher risk of parental death, particularly for mothers. However, the novelty in this study is using a sibling comparison design over a 200-year time frame with a large sample size. Hence they are able to answer the question "is the phenomenon of interest – increased risk of parental death following the death of a child – particular to industrialized human populations in which child death is quite rare." The answer from this paper is clearly 'no' – even in a context in which child death affects more than half of parents, the authors detect a raised risk of death among mothers and (in a more specific context) fathers.

Decision letter after peer review:

Thank you for submitting your article "The mother's risk of premature death after child loss across two centuries" for consideration by eLife. Your article has been reviewed by two peer reviewers, and the evaluation has been overseen by M. Dawn Teare as the Reviewing Editor and Eduardo Franco as the Senior Editor. The following individuals involved in review of your submission have agreed to reveal their identity: Susan C Alberts (Reviewer #1).

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

Essential revisions:

1) There are some puzzling features of this study that it would be helpful to understand better. Parents who lost a child and had more children, were younger at first birth, were more likely to be women and had longer follow-up that those who did not lose a child (Table 1). Was the difference in% women for parents with and without loss significant? If so, please add to the text. More generally, please explain these differences between the parents with and without loss, and consider whether the differences are an artefact of the design/implementation that might bias the estimates.

2) The difference in age between parents with loss and their siblings could be because follow-up time started when the child died for the parents with loss, and started for the sibling when the sibling became a parent. As such, for the sibling the time between their nephew/niece dying and the birth of their first child might be "immortal" time (1) that gives the sibling a survival advantage. For example, if the parent and their childless sibling both died a week after the child died, then the sibling would not be eligible for the study, making the surviving eligible siblings more strongly selected survivors than the parents with loss. Please consider whether such an issue could affect the estimates and amend the method or add a limitation explaining this point.

3) Please explain the adjustment for sex in the sibling design in more detail. Was the sibling design adjusted for sex composition of the sibling pairs, i.e., both female, both male, etc., because men and women have children at different ages, giving a different exposure time till age 50 years?

4) Given the above concerns about the handling of exposure time, would it be helpful to add a twin comparison possibly with exposure time handled in a different way?

5) Number of children could be a confounder (common cause of exposure and outcome) or a mediator, because a higher number of children affects parental survival. Obviously, the more children in a family the greater risk of a child dying. Alternatively, parents might have a 'replacement' child after a death. Would it be possible to distinguish between these possibilities and conduct the analyses accordingly? Would a marginal structural model help? Alternatively, it would be helpful to see estimates adjusted for number of children in the family.

6) The authors allude to an important point in the Discussion when they say "…we have no information about the causes of death for children or parents. Such data could be used to help verify the causal relationship between deaths of children and parents and to uncover patterns in the causes of death among parents (e.g. suicide or cardiovascular deaths) after child loss." This seems a highly salient point: Indeed, a causal relationship in the other direction is quite possible, and mutual causal arrows are possible. That is, child death may be greatly increased in cases where parents are unwell or in poor condition – implying a causal arrow from parent health to child death. Conversely, the risk of parental death after child death may be particularly high for parents in poor health. The authors need to lay out these alternative causal pathways more explicitly. Is there any data that could provide insight into these questions?

7) It would be good to know a bit more about the completeness (or lack thereof) of the dataset and whether the authors used any specific techniques to deal with this incompleteness. For instance, The authors make the statement "it is possible that the fragmented registration of births in earlier times (particularly in the case of infant deaths) resulted in misclassification of some parents who had truly lost a child and, thus, somewhat conservative estimates of excess parent mortality rates in earlier times." And "The database contains an almost complete record of the ancestors of contemporary Icelanders back to around 1650 and, after 1880, is based on death records from a geographically comprehensive set of parish records." It is difficult to get a full picture of dataset completeness from this relatively scanty information. What proportion of records are incomplete? A table with information about various types of missing data would be advisable.

8) It would be helpful if the discussion focused more on the strengths and weaknesses of the study.

[Editors' note: further revisions were requested prior to acceptance, as described below.]

Thank you for resubmitting your work entitled "The mother's risk of premature death after child loss across two centuries" for further consideration at eLife. Your revised article has been favorably evaluated by a Senior Editor and a Reviewing Editor, and two reviewers.

The manuscript has been improved but there is one remaining issue that needs to be addressed before acceptance. Reviewer #2 still has concerns about survivor bias. Comparing people who survived at least one child with those who did not may result in selection bias. Adjusting for potential confounding variables will not protect against this. Essentially the people who have child loss may have to live longer, which may obscure the effect of losing a child. The authors need to elaborate on this issue. This could be done by explaining why this is not a problem, performing an additional comparison or highlighting the issues under limitations. We transcribe below the reviewers' comments.

Reviewer #1:

The authors have thoroughly and clearly addressed the concerns and completed the essential revisions. I believe the paper is ready for publication and will make a strong contribution to the literature.

Reviewer #2:

This is an interesting study looking at the relation of offspring death with survival over 200 years in a unique data set.

My previous concern with this study was that the people who had lost a child are different from the comparison group who did not lose a child, which might cause bias, due to some form of survivor bias, i.e., selection bias. Specifically, the parents with loss are more likely to be women, younger at first birth, to have more children and to have longer follow-up. The authors have responded by saying that they addressed these issues by adjustment or stratification, which are means of dealing with confounding not selection bias.

Essentially, the study appears to be comparing people who survived at least one of their children with people who did not survive any of their children. It is easier to survive at least one of your children if you are a woman (because women live longer than men), are younger when the children are born, if you have more children and are followed up for longer which explains why the people with child loss are more likely to be women, younger at first birth, with more children and to have longer follow-up.

I remain concerned as to whether comparing people who survived at least one child with those who did not gives the effect of child loss on survival. Essentially the people who have child loss may have to live longer, which may obscure the effect of losing a child. Of course, it is possible that these issues do not matter, but it would be helpful if the authors could explain why not, or amend accordingly.

https://doi.org/10.7554/eLife.43476.018

Author response

Essential revisions:

1) There are some puzzling features of this study that it would be helpful to understand better. Parents who lost a child had more children, were younger at first birth, were more likely to be women and had longer follow-up that those who did not lose a child (Table 1). Was the difference in% women for parents with and without loss significant? If so, please add to the text. More generally, please explain these differences between the parents with and without loss, and consider whether the differences are an artefact of the design/implementation that might bias the estimates.

Given the large sample size, any differences between parents who did vs. did not lose a child is likely to be statistically significant even when the absolute difference is small. We have now added P for difference for all comparisons in Table 1 and revised the text accordingly. The P for difference across parental groups was <0.05 for all covariates, including: parents’ sex, number of children, age at first birth, age at entry to the cohort, length of follow-up time and age at death.

Although we note a statistically significant difference between sibling groups on all covariates we have directly addressed each of these factors in our analyses and, therefore, are confident that they do not account for our results. We adjusted for age at follow-up (as the underlying timescale) and sex in all analyses and present results for each of the studied birth cohorts throughout. In Table 2, we further present the association between child loss and premature parent mortality rate stratified by parental age at loss/matching and follow-up time, and sub-grouped by the number of children alive at loss. These results are not consistent with the notion that our findings are driven by differences between the groups on the abovementioned factors. Although varying somewhat, the association between loss and survival is evident across the range of these covariates. The considerable differences in average follow-up time between parents who did vs. did not lose a child (31.79 vs. 25.45 years, respectively) is likely explained by the parents who did not lose a child at the time of matching but later encountered a child loss and were then censored as end of follow-up in that matching set. This is relatively common in the earlier birth cohorts due to the higher mortality rates for children in that period. This concern is partly alleviated by Figure 2, where the follow-up of parents is restricted to 50 years across all birth cohorts. Moreover, the results in Table 2 do not suggest a considerable variation in parent mortality across extended periods of follow-up time from child loss/matching. This further argues against the possibility that the difference of average follow-up length between two groups may account for our findings.

Modified manuscript text:

Results section: “Compared to their siblings who did not lose a child, parents who lost a child were more likely to be women, have more children, be younger at first birth and older at cohort entry, and have a longer follow-up time (P<0.05; Table 1).”

Discussion section: “Last, in the sibling cohort, parents who lost a child differed somewhat from those that did not lose a child with respect to sex, number of children, age at first birth, age at entry to the cohort, length of follow-up time and age at death. We therefore made every effort to adjust for these covariates in our analyses. We adjusted for age at follow-up and sex in all analyses and presented separate results for the different birth cohorts. We further stratified analyses by age of parents at loss (or matching), follow-up time and number of children alive at loss, and demonstrated that the association between child loss and parent mortality is observed across the range of these covariates.”

Materials and methods: “In this analysis, we used a chi-square test for the categorical and a t-test for the continuous covariates.”

See also Table 1.

2) The difference in age between parents with loss and their siblings could be because follow-up time started when the child died for the parents with loss, and started for the sibling when the sibling became a parent. As such, for the sibling the time between their nephew/niece dying and the birth of their first child might be "immortal" time (1) that gives the sibling a survival advantage. For example, if the parent and their childless sibling both died a week after the child died, then the sibling would not be eligible for the study, making the surviving eligible siblings more strongly selected survivors than the parents with loss. Please consider whether such an issue could affect the estimates and amend the method or add a limitation explaining this point.

We agree with the reviewer that the sibling design needs further clarification in the manuscript. First, in order to be at risk of losing a child, one needs to have one. Thus, our analysis of index persons who lost a child and their siblings only considers follow-up time where both of these groups were parents. We started to follow the siblings who already had a child from the same age as their index-sibling when he/she lost a child. If the siblings did not have any children at the age when the index-persons lost a child, these were not included in the analysis until they subsequently had one, and then only from the age when they became parents. Indeed, we agree with the reviewer that the time between the sibling’s nephew/niece dying and the birth of their first child is “immortal” and was therefore not included in our analysis and should not bias our estimates. Indeed, both groups (individuals with or without child loss) need to survive long enough to become parents to be considered for analysis. The reviewer notes that the parents who lost a child are a bit further into his/her life history than their siblings who are still childless at the time of the loss. However, the population cohort – where even higher hazard ratios are observed – does not have that problem. Please also refer to the results of our sensitivity analysis below (Point #4), restricting the sibling cohort to unexposed siblings born within 5 years of the parents who lost a child.

We have now clarified this important issue in the Materials and methods section.

Modified manuscript text:

“Specifically, siblings who did not lose a child started to contribute to the follow-up at the same age as the parent who lost a child or at a later age when they became parents.”

3) Please explain the adjustment for sex in the sibling design in more detail. Was the sibling design adjusted for sex composition of the sibling pairs, i.e., both female, both male, etc., because men and women have children at different ages, giving a different exposure time till age 50 years?

We did not restrict the sibling sets to all females or all males since that would inevitably have further raised the bar for selection of siblings. Instead, we matched a mother who lost a child to all her eligible siblings with children, including both her sisters and brothers, who had not lost a child at the same age. We adjusted for sex in the sibling-based analysis which accounts for the effect of sex on mortality rate across sibling sets. Moreover, we used attained age as the underlying time scale and therefore the age difference was also controlled for. We have now further clarified the sibling-based analysis in the Materials and methods.

Modified manuscript text:

“Of note, we did not match on parental sex but adjusted for it in all models and thereby accounting for the effect of sex on mortality rate across the sibling groups.”

4) Given the above concerns about the handling of exposure time, would it be helpful to add a twin comparison possibly with exposure time handled in a different way?

Unfortunately, a twin-based comparison would be underpowered due to the low number of pairs. However, we have now conducted a sensitivity analysis restricting to 96,041 (55.2%) siblings born five years before or after the index parents. The results remain largely unchanged, as presented in Figure 2—figure supplement 5.

Modified manuscript text:

Results section: “In a sensitivity analysis, restricting parent-sibling pairs born within five years from the parents who lost a child or adjusting for the number of children alive at time of matching, we observed similar estimates of parental mortality rates after child loss (Figure 2—figure supplements 5-6). The only notable difference was that, in the birth cohort 1931-1996, the impact of child loss on maternal mortality after 50 years of age became statistically significant.”

Materials and methods: “In a sensitivity analyses, we restricted to parent-sibling pairs born within five years from the parents who lost a child (N=96,041, 55.2%) to reduce the potential differences between siblings.”

5) Number of children could be a confounder (common cause of exposure and outcome) or a mediator, because a higher number of children affects parental survival. Obviously, the more children in a family the greater risk of a child dying. Alternatively, parents might have a 'replacement' child after a death. Would it be possible to distinguish between these possibilities and conduct the analyses accordingly? Would a marginal structural model help? Alternatively, it would be helpful to see estimates adjusted for number of children in the family.

We agree with the reviewer that the number of children could be a confounder or a mediator in the association between child loss and parent mortality. This is why we did not adjust for it in our primary analysis. Instead, we performed a subgroup analysis by the number of children alive at the time of loss among parents who lost a child (Table 2), where we did not observe varying effect sizes across subgroups. In accordance with the reviewer’s note, we now present results from a sensitivity analysis, adjusting for the number of children alive at loss/matching. The results are consistent with those of the main analyses, except that maternal mortality after 50 years of age, in the birth cohort 1931-1996, becomes statistically significant (Figure 2—figure supplement 6). Adjusting for or matching on children born after loss (“replacement children”) is challenging, since the parents need to survive long enough to have them. This would introduce a survival bias. We have therefore opted not to adjust for the number of any post-loss/matching children in our analysis.

Modified manuscript text:

Results section:

In a sensitivity analysis, restricting to parent-sibling pairs born within five years from the parents who lost a child or adjusting for the number of children alive at time of matching, we observed similar estimates of parental mortality rates after child loss (Figure 2—figure supplements 5-6). The only notable difference was that, in the birth cohort 1931-1996, the impact of child loss on maternal mortality after 50 years of age became statistically significant.

Materials and methods: To further address the potential effect of number of children, in another sensitivity analysis, we additionally adjusted for number of alive children at the time of loss/matching.

6) The authors allude to an important point in the Discussion when they say "…we have no information about the causes of death for children or parents. Such data could be used to help verify the causal relationship between deaths of children and parents and to uncover patterns in the causes of death among parents (e.g. suicide or cardiovascular deaths) after child loss." This seems a highly salient point: Indeed, a causal relationship in the other direction is quite possible, and mutual causal arrows are possible. That is, child death may be greatly increased in cases where parents are unwell or in poor condition – implying a causal arrow from parent health to child death. Conversely, the risk of parental death after child death may be particularly high for parents in poor health. The authors need to lay out these alternative causal pathways more explicitly. Is there any data that could provide insight into these questions?

This is a valid point that requires further clarification. We have no data on causes of death while previous studies (e.g. Li et al., 2003) suggest that child loss in modern times (after 1980) increases the risk of maternal death, particularly of unnatural causes but also natural causes. We agree with the reviewer that some risk of reverse causation (parent’s ill health increasing risk of child’s death and subsequent death of parent) cannot be ruled out. However, if deteriorating parental health was strongly associated with the risk of child death, we might expect the association between child loss and parent mortality to be particularly evident during the first years of follow-up. However, the relatively stable association between child loss and parent mortality across follow-up times (as shown in Table 2) does not indicate a strong influence of reverse causality in these data. Indeed, families (mothers and children) were extremely financially dependent on fathers, such that if fathers died or were unable to work, then families were often separated and children fostered – which usually meant very hard work and diminished living conditions for these children. Thus, if parents' ill-health was a driving factor, we would expect a stronger association between child loss and father's mortality. However, it is indeed maternal mortality rate, and not paternal mortality rate after child loss, which is consistently elevated across all follow-up periods.

Modified manuscript text:

Discussion: “Some impact of reverse causation, e.g. the ill-health of parents increasing the mortality risk of their children, cannot be ruled out. However, the health and welfare of children was even more dependent on their parents, particularly fathers, in pre-demographic transition (pre-welfare state times) societies. The relatively stable effect estimates of parent mortality rates after child loss across calendar periods and the consistently lower impact of child loss on paternal mortality argues against a strong influence of reverse causality.”

7) It would be good to know a bit more about the completeness (or lack thereof) of the dataset and whether the authors used any specific techniques to deal with this incompleteness. For instance, The authors make the statement "it is possible that the fragmented registration of births in earlier times (particularly in the case of infant deaths) resulted in misclassification of some parents who had truly lost a child and, thus, somewhat conservative estimates of excess parent mortality rates in earlier times." And "The database contains an almost complete record of the ancestors of contemporary Icelanders back to around 1650 and, after 1880, is based on death records from a geographically comprehensive set of parish records." It is difficult to get a full picture of dataset completeness from this relatively scanty information. What proportion of records are incomplete? A table with information about various types of missing data would be advisable.

In terms of linkages between parents and offspring, the completeness of the genealogical database across the entire observation period is 99%. The proportion of individuals without identifiable one or both parents in the database is: 2.4% in individuals born 1800-1880, 0.6% in individuals born 1881-1930, and 1.5% in individuals 1931-1996. Moreover, the database has been extensively evaluated for the accuracy of such linkages using genetic data – (Helgason et al., 2015, Sun et al., 2012, Halldorsson et al., 2019)

The individuals missing from the genealogical database are primarily children born prior to 1900, who died during the first few weeks of life and were not always recorded in censuses or parish records. This would lead to a misclassification of some parents, which in our study are considered as unexposed (parents that have not lost a child), but in fact should have been considered as exposed (parents who lost a child). We note that the impact of such misclassification on our results is in the form of a conservative bias. Thus, if there is a difference in survival between exposed and unexposed parents, such misclassification would only lead us to underestimate this difference and is therefore not a major concern for this study. We now provide this information in the Materials and methods section.

Modified manuscript text:

“In terms of linkages between parents and offspring, the completeness of the genealogical database across the entire observation period is 99%. The database has been extensively evaluated for the accuracy of such linkage using genetic data (40-42).”

8) It would be helpful if the discussion focused more on the strengths and weaknesses of the study.

Based on the points raised above we have now extended the discussion of limitations and moved that part to the beginning of the Discussion section.

[Editors' note: further revisions were requested prior to acceptance, as described below.]

The manuscript has been improved but there is one remaining issue that needs to be addressed before acceptance. Reviewer #2 still has concerns about survivor bias. Comparing people who survived at least one child with those who did not may result in selection bias. Adjusting for potential confounding variables will not protect against this. Essentially the people who have child loss may have to live longer, which may obscure the effect of losing a child. The authors need to elaborate on this issue. This could be done by explaining why this is not a problem, performing an additional comparison or highlighting the issues under limitations. We transcribe below the reviewers' comments.

Thank you for the opportunity for a second resubmission of our paper. Please find below our responses to the issues raised along with the corresponding modifications to the manuscript. We believe that we have now addressed reviewer #2’s main issue, i.e. survivor bias, and we hope you now find our manuscript acceptable for publication in eLife. Please do not hesitate to contact us with any remaining issues.

Reviewer #1:

The authors have thoroughly and clearly addressed the concerns and completed the essential revisions. I believe the paper is ready for publication and will make a strong contribution to the literature.

Thank you for the positive comment on our work and our paper.

Reviewer #2:

This is an interesting study looking at the relation of offspring death with survival over a 200 years in a unique data set.

My previous concern with this study was that the people who had lost a child are different from the comparison group who did not lose a child, which might cause bias, due to some form of survivor bias, i.e., selection bias. Specifically, the parents with loss are more likely to be women, younger at first birth, to have more children and to have longer follow-up. The authors have responded by saying that they addressed these issues by adjustment or stratification, which are means of dealing with confounding not selection bias.

Essentially, the study appears to be comparing people who survived at least one of their children with people who did not survive any of their children. It is easier to survive at least one of your children if you are a woman (because women live longer than men), are younger when the children are born, if you have more children and are followed up for longer which explains why the people with child loss are more likely to be women, younger at first birth, with more children and to have longer follow-up.

I remain concerned as to whether comparing people who survived at least one child with those who did not gives the effect of child loss on survival. Essentially the people who have child loss may have to live longer, which may obscure the effect of losing a child. Of course, it is possible that these issues do not matter, but it would be helpful if the authors could explain why not, or amend accordingly.

The reviewer raises several important issues.

The reviewer remains concerned for a potential bias based on that “people who have child loss may have to live longer” than the comparison cohort of parents who do not experience a child loss. However, such type of bias, if any, caused by a selection of “survivors” to the exposed population would rather yield conservative estimates of the increase in parent mortality after a child loss. Also, as explained in more details below, we carefully designed our study to compare the two groups of parents, with and without a child loss, from the same age (i.e. the age when the loss of the child in one group occurred) whenever possible. Thus, both groups were at the same age at the start of follow-up (i.e., at the time of matching).

Thus, we remain convinced that our study design entails few, if any, possible conditions giving rise to survivor bias as described in the paper cited by the reviewer in the last round of review (2). We followed parents who lose a child starts immediately after child loss and, due to the fact that we use parental age as the underlying time scale, the follow-up of the matched parents not yet exposed to child loss starts at the same age/time point as for the exposed parents yielding identical conditions for ascertaining mortalities among these two compared groups. The sibling controls are followed from the same age as the parent that lost a child, or from the age when becoming a parent if they are not yet parents when their sibling lost a child; the latter, according to the definition above, may give rise to survivor bias. For this reason, we reran our sibling-based comparison, limiting it to unexposed siblings who were already parents at the age as the index sibling who lost a child. The results remained largely similar (if anything, the point estimates indicated even stronger associations) and we have added this sensitivity analysis to the revised manuscript (Figure 2—figure supplement 8).

Thus, we believe that the reviewer’s concern of survivor bias should be eliminated through our study design only considering person time of unexposed siblings who have reached the exact age at the index parent’s loss. In other words, both exposed and unexposed parent’s need to survive through the same age in order to be compared. Therefore, the amount of “immortal time” is similar across these groups. We have now clarified this issue by adding the following statements to the manuscript:

Modified manuscript text:

“We further restricted unexposed siblings to those who were already parents at the age when the index sibling lost a child, which also yielded largely similar results (Figure 2—figure supplement 8).”

“Indeed, parents who lost a child may have to live longer to experience such an event, however, such survivor bias (if any) due to a selection of survivors to the exposed group would rather yield conservative estimates of parent mortality after a child loss. […] However, we observed similar results when we in a sensitivity analysis restricted to unexposed siblings who were already parents at the same age as when the index parents lost a child.”

“To reduce the potential differences between siblings we, in sensitivity analyses restricted to parent-sibling pairs born within five years from the parents who lost a child (N=96,041, 55.2%) as well as to unexposed siblings who were already parents at the same age as when the index parents lost a child (N=135,719, 78.0%).”

In contrast to the abovementioned concern for selection (survivor) bias induced by the study design where certain individuals (or person time) are selected to or excluded from a study (3), we believe that the remaining points raised by the reviewer represent conditions of confounding but not selection bias (4). Our study is indeed a nationwide complete follow-up of all parents during a 199-year period; no individual is excluded from the study on the basis of sex, age at first childbirth or total number of children. As the reviewer correctly points out, these factors may be associated with parental survival and some of them are unequally distributed across the compared parental groups and therefore need to be accounted for in the analysis. We have indeed made extensive efforts to account for these factors in our analysis.

Sex is accounted for in the population analysis by matching on sex, i.e. the rates of mortality are only compared between pairs of exposed and unexposed mothers and then fathers, separately. To preserve statistical power, we did not match on sex in the sibling analysis while controlling for sex in all models. We have now performed additional analysis confined to sister- and then brother-pairs (discordant on child loss) yielding similar estimates in the primary analysis albeit with lower precision. These results have been added to the manuscript (Figure 2—figure supplement 6).

Modified manuscript text:

“In a sensitivity analysis, restricting to parent-sibling pairs born within five years from the parents who lost a child, comparing same-sex siblings, or adjusting for the number of children alive at time of matching, we observed similar estimates of parental mortality rates after child loss (Figure 2—figure supplements 5-7).”

“To further address the potential influences of parent sex and number of children, in another sensitivity analyses, we only compared same-sex siblings and additionally adjusted for number of alive children at the time of loss/matching, respectively.”

Other factors mentioned by the reviewer have appropriately, as potential confounders, been addressed in Table 2 in subgroup analyses. There we e.g. demonstrate that maternal mortality rates below the age of 50 are considerably elevated across parent groups with varying number of children, with the greatest excess observed after loss of an only child. In the last round of revisions, we further performed additional analyses controlling for the number of children at the time of matching revealing virtually identical, if anything stronger, results (Figure 2—figure supplement 7).

In accordance with the reviewer’s comment, we have now added a subgroup analysis of age at first child birth in Table 2, revealing that the association between child loss and maternal mortality rates below the age of 50 is present in every subgroup of parental age at first child birth.

Modified manuscript text:

“We further performed subgroup analyses by the age of parents at loss (or matching), parent age at first childbirth, follow-up time and number of children alive at loss, and demonstrated that the association between child loss and parent mortality was observed across the range of these covariates.”

“Parent mortality before and after age of 50 years was further analyzed in subgroups of age of the deceased child (0, 1-5, 6-17, or ≥18 years old), number of children alive at loss (0, 1-3, or ≥4), sex of the deceased child, parental age at loss (13-30, 31-40, or 41-50 years old), and parental age at first childbirth (13-21, 22-24, 25-27, or ≥28 years old).”

Finally, as we explained in the last round of revisions, the discrepancy between the parental groups in follow-up time is driven by the high incidence of child loss, particularly during 20th century. Our study design mimics a prospective study, where the investigators have at every time point of the observation no information on which parent will live or die, or lose a child. Therefore, parents were randomly selected as matched “control parents” while they had not lost a child and then censored from that control group at the time of child loss (if it occurred). Due to the high incidence of child loss, many parents were later censored from the control group at the time of child loss and assigned to a new set (now as index-individuals) yielding the relatively shorter follow-up time of parents not exposed to child loss relative to the exposed. Thus, the Cox model does not generate estimates of parental mortality rates based on comparison between two distinct groups of parents, exposed and unexposed to child loss, but on the basis of hazards during exposed and unexposed person time.

1) Levesque LE, Hanley JA, Kezouh A, et al. Problem of immortal time bias in cohort studies: example using statins for preventing progression of diabetes. BMJ 2010;340:b5087. doi: 10.1136/bmj.b5087 published Online First: 2010/03/17]

2) Lévesque LE et al. Problem of immortal time bias in cohort studies: example using statins for preventing progression of diabetes. BMJ, 2010;340:b5087.

3) Rothman KJ. Epidemiology: an introduction. Oxford university press; 2012: p126.

4) Newton J. Confounding is mistakenly called selection bias. BMJ, 1998. Accessed on https://www.bmj.com/rapid-response/2011/10/27/confounding-mistakenly-called-selection-bias

https://doi.org/10.7554/eLife.43476.019

Article and author information

Author details

  1. Unnur A Valdimarsdóttir

    1. Center of Public Health Sciences, Faculty of Medicine, University of Iceland, Reykjavik, Iceland
    2. Department of Epidemiology, Harvard TH Chan School of Public Health, Boston, United States
    3. Department of Medical Epidemiology and Biostatistics, Karolinska Institutet, Solna, Sweden
    Contribution
    Conceptualization, Supervision, Funding acquisition, Methodology, Writing—original draft, Writing—review and editing
    Contributed equally with
    Donghao Lu
    For correspondence
    unnurav@hi.is
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-5382-946X
  2. Donghao Lu

    1. Center of Public Health Sciences, Faculty of Medicine, University of Iceland, Reykjavik, Iceland
    2. Department of Epidemiology, Harvard TH Chan School of Public Health, Boston, United States
    3. Department of Medical Epidemiology and Biostatistics, Karolinska Institutet, Solna, Sweden
    4. Channing Division of Network Medicine, Brigham and Women's Hospital, Harvard Medical School, Boston, United States
    Contribution
    Formal analysis, Methodology, Writing—original draft, Writing—review and editing
    Contributed equally with
    Unnur A Valdimarsdóttir
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-4186-8661
  3. Sigrún H Lund

    deCODE Genetics, Reykjavik, Iceland
    Contribution
    Formal analysis, Writing—review and editing
    Competing interests
    No competing interests declared
  4. Katja Fall

    1. Department of Medical Epidemiology and Biostatistics, Karolinska Institutet, Solna, Sweden
    2. Clinical Epidemiology and Biostatistics, School of Medical Sciences, Örebro University, Örebro, Sweden
    Contribution
    Writing—review and editing
    Competing interests
    No competing interests declared
  5. Fang Fang

    Department of Medical Epidemiology and Biostatistics, Karolinska Institutet, Solna, Sweden
    Contribution
    Writing—review and editing
    Competing interests
    No competing interests declared
  6. Þórður Kristjánsson

    deCODE Genetics, Reykjavik, Iceland
    Contribution
    Project administration, Writing—review and editing
    Competing interests
    No competing interests declared
  7. Daníel Guðbjartsson

    1. deCODE Genetics, Reykjavik, Iceland
    2. School of Engineering and Natural, Sciences, University of Iceland, Reykjavik, Iceland
    Contribution
    Supervision, Methodology, Writing—review and editing
    Competing interests
    No competing interests declared
  8. Agnar Helgason

    1. deCODE Genetics, Reykjavik, Iceland
    2. Department of Anthropology, University of Iceland, Reykjavik, Iceland
    Contribution
    Conceptualization, Data curation, Supervision, Writing—original draft, Writing—review and editing
    Contributed equally with
    Kári Stefánsson
    Competing interests
    No competing interests declared
  9. Kári Stefánsson

    1. deCODE Genetics, Reykjavik, Iceland
    2. Faculty of Medicine, School of Health Sciences, University of Iceland, Reykjavik, Iceland
    Contribution
    Conceptualization, Data curation, Writing—review and editing
    Contributed equally with
    Agnar Helgason
    For correspondence
    kstefans@decode.is
    Competing interests
    No competing interests declared

Funding

RANNIS (163362-051)

  • Unnur Valdimarsdóttir

European Research Council (726413)

  • Unnur Valdimarsdóttir

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

The study is funded by the Icelandic Research Fund- RANNIS (Grant of Excellence; nr: 163362–051) and the European Research Council (StressGene; nr: 726413). We are grateful that the librarians from the Karolinska Institutet University Library provided professional help on literature search.

Ethics

Human subjects: This study was reviewed by the National Ethics Committee and Data Protection Authority of Iceland (Approval No. VSN-12–125 and VSN-16–156).

Senior Editor

  1. Eduardo Franco, McGill University, Canada

Reviewing Editor

  1. M Dawn Teare, Newcastle University, United Kingdom

Reviewer

  1. Susan C Alberts, Duke University, United States

Publication history

  1. Received: December 6, 2018
  2. Accepted: October 18, 2019
  3. Version of Record published: November 12, 2019 (version 1)

Copyright

© 2019, Valdimarsdóttir et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 520
    Page views
  • 47
    Downloads
  • 0
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Download citations (links to download the citations from this article in formats compatible with various reference manager tools)

Open citations (links to open the citations from this article in various online reference manager services)

Further reading

    1. Epidemiology and Global Health
    2. Microbiology and Infectious Disease
    Thomas Crellen et al.
    Research Article
    1. Epidemiology and Global Health
    Harriet Mpairwe et al.
    Research Article