Associations of parental age with offspring all-cause and cause-specific adult mortality

People are having children later in life. The consequences for offspring adult survival have been little studied due to the need for long follow-up linked to parental data and most research has considered offspring survival only in early life. We used Swedish registry data to examine all-cause and cause-specific adult mortality (293,470 deaths among 5,204,433 people, followed up to a maximum of 80 years old) in relation to parental age. For most common causes of death adult survival was improved in the offspring of older parents (HR for all-cause survival was 0.96 (95% CI: 0.96, 0.97) and 0.98 (0.97, 0.98) per five years of maternal and paternal age, respectively). The childhood environment provided by older parents may more than compensate for any physiological disadvantages. Within-family analyses suggested stronger benefits of advanced parental age. This emphasises the importance of secular trends; a parent’s later children were born into a wealthier, healthier world. Sibling-comparison analyses can best assess individual family planning choices, but our results suggested a vulnerability to selection bias when there is extensive censoring. We consider the numerous causal and non-causal mechanisms which can link parental age and offspring survival, and the difficulty of separating them with currently available data.

. Conversion of ICD codes to descriptive categories. Table S2. Primary analyses of offspring mortality and maternal age with alternative adjustment sets. Table S3. Primary analyses of offspring mortality and paternal age with alternative adjustment sets. Table S4. Analyses decomposing the difference between the primary and sibling-comparison analyses of offspring mortality and maternal age. Table S5. Analyses decomposing the difference between the primary and sibling-comparison analyses of offspring mortality and paternal age. Table S6. Analyses of offspring mortality and maternal age, conducted separately for offspring who were the oldest child in their family and for others. Two-variable analyses of later children were also conducted with mutual adjustment for mother's age at the birth of (i) the child in question, and (ii) her first child. Table S7. Analyses of offspring mortality and paternal age, conducted separately for offspring who were the oldest child in their family and for others. Two-variable analyses of later children were also conducted with mutual adjustment for father's age at the birth of (i) the child in question, and (ii) his first child. Table S8. Primary analyses of offspring mortality and maternal age, conducted separately for offspring born up to 1969 (when parental age was decreasing) and for offspring born from 1970 (when parental age was increasing). Table S9. Primary analyses of offspring mortality and paternal age, conducted separately for offspring born up to 1969 (when parental age was decreasing) and for offspring born from 1970 (when parental age was increasing). Table S10. Sibling comparison analyses of offspring mortality and maternal age, conducted separately for offspring born up to 1969 (when parental age was decreasing) and for offspring born from 1970 (when parental age was increasing). Table S11. Sibling comparison analyses of offspring mortality and paternal age, conducted separately for offspring born up to 1969 (when parental age was decreasing) and for offspring born from 1970 (when parental age was increasing). Table S12. Primary analyses of son's outcomes and maternal age, with and without additional adjustment for maternal survival to the offspring's 35 th birthday. Table S13. Primary analyses of son's outcomes and paternal age, with and without additional adjustment for paternal survival to the offspring's 35 th birthday. Table S14. Primary analyses of offspring mortality and maternal age, with and without additional adjustment for offspring birth weight and birth length. Table S15. Primary analyses of offspring mortality and paternal age, with and without additional adjustment for offspring birth weight and birth length. Table S16. Associations between offspring mortality and maternal age from primary and sibling comparison analyses, including the potentially biased subset of offspring who died between 1961 and 30 th June 1991 (N=56,901). Table S17. Associations between offspring mortality and paternal age from primary and sibling comparison analyses, including the potentially biased subset of offspring who died between 1961 and 30 th June 1991 (N=56,901). Table S18. Primary analyses of offspring mortality and maternal age, with follow-up split by categories of offspring age. Table S19. Primary analyses of offspring mortality and paternal age, with follow-up split by categories of offspring age. Table S20. Primary and sibling-comparison analyses of offspring mortality and maternal age, made separately for men (sons) and women (daughters). Table S21. Primary and sibling-comparison analyses of offspring mortality and paternal age, made separately for men (sons) and women (daughters). Figure S1. Recorded all-cause and cause-specific deaths by year of death. Figure S2. Hazard ratios for offspring mortality against classes of maternal or paternal age. Table S1. Conversion of ICD codes to descriptive categories. Mortality was recorded using ICD 7 (1961)(1962)(1963)(1964)(1965)(1966)(1967)(1968), ICD 8 (1969-1986), ICD 9 (1987-1996) and ICD 10 (1997-2012. ICD codes are in alphabetical (not numerical) order such that, for example, 451>45099. Alzheimer's disease includes all deaths recorded as dementia.   193-19399 191-19299 191-19299 C71-C7299 Oesophageal cancer 150-15099 150-15099 150-15099 C15-C1599 Thyroid cancer 194-19499 193-19399 193-19399 Table S2. Primary analyses of maternal age and offspring mortality with alternative adjustment sets. Adjustment set (a) was empty, (b) comprised offspring sex and date of birth, (c) additionally included educational and occupational SEP in both parents and (d) additionally included birth order. Adjustment set (e) additionally included paternal age as a linear term. Hazard ratios (95% CI) are from Cox proportional hazards regression of 2,658,132 male and 2,546,301 female offspring with age as the time axis. Maternal and paternal associations in model (e) were compared using P-values from a Wald test.
Primary analysis hazard ratio per five years of mother's age at offspring's birth with adjustment set:  Table S4. Analyses decomposing the difference between the primary and sibling-comparison analyses of offspring mortality and maternal age. Primary analyses used Cox proportional hazards regression of 2,658,132 male and 2,546,301 female offspring. Age was the time axis and robust standard errors were clustered by maternal identity. Adjustment set (e) (offspring sex and date of birth (DOB), maternal and paternal occupational and educational SEP, offspring birth order, and paternal age) was used. The secular trend per five years of offspring DOB was assessed using a similar model but without maternal age. The primary analysis was repeated with adjustment for maternal, not offspring, DOB to account for confounding, but not mediation, by secular trends. To examine whether the restricted dataset used for sibling-comparison analyses was representative of the main dataset, the primary analysis was repeated on this subset. Finally, the siblingcomparison analysis used Cox regression stratified by maternal identity and was restricted to offspring in families with discordant outcomes. All family-level confounding was intrinsically adjusted for and adjustment for offspring DOB or paternal age were not possible. Explicit adjustment was therefore limited to offspring sex and birth order.  Table S5. Analyses decomposing the difference between the primary and sibling-comparison analyses of offspring mortality and paternal age. Primary analyses used Cox proportional hazards regression of 2,658,132 male and 2,546,301 female offspring. Age was the time axis and robust standard errors were clustered by paternal identity. Adjustment set (e) (offspring sex and date of birth (DOB), maternal and paternal occupational and educational SEP, offspring birth order, and maternal age) was used. The secular trend per five years of offspring DOB was assessed using a similar model but without paternal age. The primary analysis was repeated with adjustment for paternal, not offspring, DOB to account for confounding, but not mediation, by secular trends. To examine whether the restricted dataset used for sibling-comparison analyses was representative of the main dataset, the primary analysis was repeated on this subset. Finally, the siblingcomparison analysis used Cox regression stratified by paternal identity and was restricted to offspring in families with discordant outcomes. All family-level confounding was intrinsically adjusted for and adjustment for offspring DOB or maternal age were not possible. Explicit adjustment was therefore limited to offspring sex and birth order.  Table S6. Analyses of offspring mortality and maternal age, conducted separately for offspring who were the oldest child in their family and for others. Two-variable analyses of later children were also conducted with mutual adjustment for mother's age at the birth of (i) the child in question, and (ii) her first child. Primary analyses used Cox proportional hazards regression of 2,412,961 oldest offspring and 2,791,472 later offspring. Age was the time axis and robust standard errors were clustered by maternal identity. Adjustment set (e) (offspring sex and date of birth, maternal and paternal occupational and educational SEP, offspring birth order, and paternal age) was used.
Primary analysis hazard ratio per five years of mother's age at offspring's birth among:  Table S7. Analyses of offspring mortality and paternal age, conducted separately for offspring who were the oldest child in their family and for others. Two-variable analyses of later children were also conducted with mutual adjustment for father's age at the birth of (i) the child in question, and (ii) his first child. Primary analyses used Cox proportional hazards regression of 2,454,558 oldest offspring and 2,749,875 later offspring. Age was the time axis and robust standard errors were clustered by paternal identity. Adjustment set (e) (offspring sex and date of birth, maternal and paternal occupational and educational SEP, offspring birth order, and maternal age) was used.
Primary analysis hazard ratio per five years of father's age at offspring's birth among:          Table S16. Associations between offspring mortality and maternal age from primary and sibling comparison analyses, including the potentially biased subset of 39,558 offspring who died between 1961 and 30 th June 1991. Primary analyses used Cox proportional hazards regression of 2,682,440 male and 2,561,551 female offspring. Age was the time axis and robust standard errors were clustered by maternal identity. Adjustment set (e) (offspring sex and date of birth (DOB), maternal and paternal occupational and educational SEP, offspring birth order, and paternal age) was used. The secular trend per five years of offspring DOB was assessed using a similar model but without maternal age. The primary analysis was repeated with adjustment for maternal, not offspring, DOB to account for confounding, but not mediation, by secular trends. To examine whether the restricted dataset used for sibling-comparison analyses was representative of the main dataset, the primary analysis was repeated on this subset. Finally, the sibling-comparison analysis used Cox regression stratified by maternal identity and was restricted to offspring in families with discordant outcomes. All family-level confounding was intrinsically adjusted for and adjustment for offspring DOB or paternal age were not possible. Explicit adjustment was therefore limited to offspring sex and birth order.  Table S17. Associations between offspring mortality and paternal age from primary and sibling comparison analyses, including the potentially biased subset of 39,558 offspring who died between 1961 and 30 th June 1991. Primary analyses used Cox proportional hazards regression of 2,682,440 male and 2,561,551 female offspring. Age was the time axis and robust standard errors were clustered by paternal identity. Adjustment set (e) (offspring sex and date of birth (DOB), maternal and paternal occupational and educational SEP, offspring birth order, and maternal age) was used. The secular trend per five years of offspring DOB was assessed using a similar model but without paternal age. The primary analysis was repeated with adjustment for paternal, not offspring, DOB to account for confounding, but not mediation, by secular trends. To examine whether the restricted dataset used for sibling-comparison analyses was representative of the main dataset, the primary analysis was repeated on this subset. Finally, the sibling-comparison analysis used Cox regression stratified by maternal identity and was restricted to offspring in families with discordant outcomes. All family-level confounding was intrinsically adjusted for and adjustment for offspring DOB or maternal age were not possible. Explicit adjustment was therefore limited to offspring sex and birth order.  Table S18. Primary analyses of offspring mortality and maternal age, with follow-up split by categories of offspring age. Cox proportional hazards regression was applied to 2,658,132 male and 2,546,301 female offspring. Age was the time axis and robust standard errors were clustered by maternal identity. Adjustment set (e) (offspring sex and date of birth, maternal and paternal occupational and educational SEP, offspring birth order, and paternal age) was used. Causes of death were omitted when there were fewer than five deaths in any age category. P-values from two tests for proportional hazards are reported. The first tested the correlation between log(age) and the scaled Schoenfeld residuals for parental age in the combined primary analysis of all ages of offspring. The second was a Wald test comparing the HR within each age class.  Table S19. Primary analyses of offspring mortality and paternal age, with follow-up split by categories of offspring age. Cox proportional hazards regression was applied to 2,658,132 male and 2,546,301 female offspring. Age was the time axis and robust standard errors were clustered by paternal identity. Adjustment set (e) (offspring sex and date of birth, maternal and paternal occupational and educational SEP, offspring birth order, and maternal age) was used. Causes of death were omitted when there were fewer than five deaths in any age category. P-values from two tests for proportional hazards are reported. The first tested the correlation between log(age) and the scaled Schoenfeld residuals for parental age in the combined primary analysis of all ages of offspring. The second was a Wald test comparing the HR within each age class.  Table S20. Primary and sibling-comparison analyses of offspring mortality and maternal age, made separately for men (sons) and women (daughters). Primary analyses used Cox proportional hazards regression of 2,658,132 male and 2,546,301 female offspring. Age was the time axis and robust standard errors were clustered by maternal identity. Adjustment set (e) (offspring date of birth (DOB), maternal and paternal occupational and educational SEP, offspring birth order, and paternal age) was used. The sibling-comparison analysis used Cox regression stratified by maternal identity and was restricted to same-sex offspring in families with discordant outcomes. All family-level confounding was intrinsically adjusted for and adjustment for offspring DOB or paternal age were not possible. Explicit adjustment was therefore limited to offspring birth order.  Table S21. Primary and sibling-comparison analyses of offspring mortality and paternal age, made separately for men (sons) and women (daughters). Primary analyses used Cox proportional hazards regression of 2,658,132 male and 2,546,301 female offspring. Age was the time axis and robust standard errors were clustered by paternal identity. Adjustment set (e) (offspring date of birth (DOB), maternal and paternal occupational and educational SEP, offspring birth order, and maternal age) was used. The sibling-comparison analysis used Cox regression stratified by paternal identity and was restricted to same-sex offspring in families with discordant outcomes. All family-level confounding was intrinsically adjusted for and adjustment for offspring DOB or maternal age were not possible. Explicit adjustment was therefore limited to offspring birth order.  Figure S1a. Recorded all-cause and cause-specific deaths by year of death among the study population (born 1932-1987). Causes of death were recorded using four successive versions of the international classification of diseases (ICD) system. The solid black line indicates the numbers of deaths expected if the proportion of deaths due to each cause had remained constant over time. Number of deaths from specified cause Figure S1b. Recorded all-cause and cause-specific deaths by year of death among the study population (born 1932-1987). Causes of death were recorded using four successive versions of the international classification of diseases (ICD) system. The solid black line indicates the numbers of deaths expected if the proportion of deaths due to each cause had remained constant over time. Number of deaths from specified cause Figure S1c. Recorded all-cause and cause-specific deaths by year of death among the study population (born 1932-1987). Causes of death were recorded using four successive versions of the international classification of diseases (ICD) system. The solid black line indicates the numbers of deaths expected if the proportion of deaths due to each cause had remained constant over time. Number of deaths from specified cause Figure S2. Hazard ratios for offspring mortality against classes of maternal or paternal age. Parental age in years was put into classes of <20, 20-24 (reference), 25-29, 30-34, 35-39, 40-44 and ≥45 and each class was plotted at its median. Points, but not lines, are transposed horizontally by +/-0.5 years for clarity. The two oldest classes were combined for maternal age. Error bars are 95% confidence intervals. Primary analyses used Cox proportional hazards regression with robust standard errors clustered by parental identity. Standard adjustment comprised offspring sex, DOB and birth order, maternal and paternal occupational and educational SEP and the other parent's age at the time of the offspring's birth. Sibling-comparison analyses used Cox regression stratified by parental identity. They were conducted on the restricted dataset and adjusted for offspring sex and birth order (adjustment for parental SEP was unnecessary and adjustment for offspring DOB and the other parent's age were impossible). Figure S2 (continued). Hazard ratios for offspring mortality against classes of maternal or paternal age. Figure S2 (continued). Hazard ratios for offspring mortality against classes of maternal or paternal age. Figure S2 (continued). Hazard ratios for offspring mortality against classes of maternal or paternal age. Figure S2 (continued). Hazard ratios for offspring mortality against classes of maternal or paternal age. Figure S2 (continued). Hazard ratios for offspring mortality against classes of maternal or paternal age. Figure S2 (continued). Hazard ratios for offspring mortality against classes of maternal or paternal age. Figure S2 (continued). Hazard ratios for offspring mortality against classes of maternal or paternal age.  1 conducted similar analyses on the same starting dataset but got hazard ratios (HR) which were rather different from the most analogous models in our analysis, particularly for maternal age. Below, we consider how their analyses differ from the most closely corresponding all-cause analyses in ours and try to partially replicate their results by modifying our analyses. Their Model 1 was a standard Cox model (which we refer to as a primary analysis) and their Model 3 was a stratified Cox model (a sibling comparison analysis).
Differences between B&M and the current study: 1) B&M's Model 1 adjusted for the other parent's age (in categories) and offspring sex. The closest adjustment sets used with the primary analysis in the present analysis are (a) no adjustment and (e) offspring sex and date of birth, educational and occupational SEP in both parents, birth order and the other parent's age as a linear term. Their Model 3 adjusted for the other parents age (in categories), offspring sex and birth order. The closest adjustment set used with the sibling comparison analysis in the present analysis is (e) offspring sex and birth order. 2) B&M left-truncated follow-up at age 30; we did so at age 18. The aim in both studies was to investigate adult mortality, and mortality in early adulthood is very rare so this change is unlikely to have been influential. 3) B&M excluded people born before 1938; we used people born from 1932 (the earliest available). Their reasoning was that many people born in this early period would be incorrectly identified as first-born, because their older siblings were born before 1932 and thus not recorded in the data. Our reasoning was that while birth order may be an important covariate, people born in 1932-1938 provide a substantial proportion of recorded deaths and follow-up at older ages. 4) B&M used people born up to the end of 1960; we used all available data (people born up to 1987). We presume that their reasoning was that people born in 1960 were the last cohort who would already be 30 (their minimum age) by 1990 (the start of their follow-up). People born in 1960-1982 would turn 30 (and thus be followed up) sometime during their follow-up period of 1990-2012. 5) B&M followed up from January 1990; we followed up from July 1991. Prior to July 1991, there was missingness in the identification of parents which depended on the survival of the offspring (offspring who died were less likely to have identified parents). 6) B&M defined families by the identities of both parents; we used only the identity of the parent in question. 7) B&M excluded whole families which contained multiple births or where any child was born outside Sweden. We excluded only those individuals from multiple births. We did not have data on whether children were born outside Sweden (they may already have been removed) and did not replicate their removal of whole families containing multiple births. 8) B&M categorised birth order from 1st to ≥7 th ; we did so from 1 st to ≥4 th . We did not replicate this minor difference. 9) B&M categorised maternal age as 15-19, 20-24, 25-29 (ref), 30-34, 35-39, 40-44 and ≥45 years. Due to the rarity of very old mothers, we combined the oldest two categories. They categorised paternal age as 15-19, 20-24, 25-29 (ref), 30-34, 35-39, 40-44, 45-49, 50-54 and ≥55 years. Due to the rarity of very old fathers, we combined the oldest three categories. We did not replicate this difference.
The results suggest that differences between B&M's Model 1 and our primary analysis were mostly due to the adjustment variables, leading in our analysis to an increase in the additional hazard among the offspring of young parents and a reduction in the additional hazard among the offspring of older parents. The linear models in Tables S2 and S3 suggest that adjustment for SEP was the most important difference. There may also have been an effect of our inclusion of later (younger) offspring in amplifying the risk associated with young parents.
In the comparison of B&M's Model 3 with our sibling comparison analyses, the omission of offspring born after 1960 attenuated the additional hazard in the offspring of young mothers partially, but not completely to the null finding of B&M. None of the changes applied individually replicated the increased hazard in the offspring of older mothers except a slight increase when follow-up began at January 1990 instead of July 1991. Nonetheless, the application of all of changes 1-6 approximately reproduced the results of B&M. This was not the case for paternal age, where the definition of families by the identities of both parents in isolation approximately reproduced the results of B&M, but when combined with the rest of changes 1-6, gave very different results.