Adopted daughters and adopted daughters-in-law in Taiwan: a mortality analysis

Adoption is sometimes considered paradoxical from an evolutionary perspective because the costs spent supporting an adopted child would be better spent on rearing one's own. Kin selection theory is commonly used to solve this paradox, because the adoption of closely related kin contributes to the inclusive fitness of the adoptive parent. In this paper, we perform a novel test of kin selection theory in the context of adoption by asking whether adopted daughters-in-law, who contribute directly (i.e. genealogically) to the perpetuation of their adoptive families' lineages, experience lower mortality than daughters adopted for other purposes in historical Taiwan. We show that both classes of adopted daughter suffer lower mortality than biological daughters, but that the protective effect of adoption is stronger among daughters who were not adopted with the intention of perpetuating the family lineage. We speculate as to the possible benefits of such a pattern and emphasize the need to move beyond typological definitions of adoption to understand the specific costs and benefits involved in different forms of caring for others' children.


Introduction
Adoption-the transferral of offspring from biological to nonbiological parents [1]-arises under numerous circumstances, and differences among these circumstances are likely to affect the degree of alloparental investment in and subsequent welfare of adopted children [2]. Kin  H1: There is a difference between ADIL and AD in mortality. If this difference is due to kin selection, ADIL should have higher survivorship than AD because ADIL perpetuate the adoptive family lineage.

Material and methods
The data used in this study were digitized from household registers collected by the Japanese colonial administration between 1895 and 1945 [33]. The administration was assiduous in collecting accurate demographic data [14,[34][35][36]. A single register was kept for each household, which a senior household member was required to update at his/her local police station for all major demographic changes, including births, deaths, marriages, adoptions and migrations. When a new household was formed, all relevant information from the previous register was transcribed to a new register and cross-referenced. Households were visited periodically by police officers-at least twice a year [14]-who would check family composition against the current version of the register. Punishments for failing to report an event were potentially severe [35,37,38], and applied not only to the head of household where the omission occurred, but also to the appointed head of a community of neighbouring families. This collective responsibility ensured extremely high rates of reporting.

Sample selection
Our final sample for analyses of mortality included 31 066 women (of whom 1897 were adopted; see electronic supplementary material) after applying several exclusion criteria. All were born after 1905 and their births were recorded in the database (i.e. they did not enter the database via immigration). Additionally, the identity of the adoptive parents had to be known in order to identify the presence of adoptive siblings in the database, so girls with unknown adoptive parents were also excluded. Girls were further excluded if they had been adopted multiple times, adopted after the age of 10, or adopted into their household of birth, all of which were atypical forms of adoption. Finally, adoptions by a close relative (biological parents, aunts, uncles, cousins or grandparents) were removed from the analysis, because close-relative adoption is extremely rare in our data, comprising only 423 cases before exclusions (1.51% of 21 522 adopted and 0.30% of 140 324 total children, including boys). Electronic supplementary material, figure S1 shows frequencies of close-relative adoptions by receiving kin.

A rule for determining type of adoption
The registers do not explicitly label daughter-in-law (DIL) adoption differently from other forms of adoption. Wolf and associates [14,15] argued based on ethnographic observation that parents with a son who adopted a girl were doing so with the likely goal of providing a bride for their son. Rather than assuming this connection a priori, we developed an algorithm to identify putative DIL adoptions as follows. We ran a logistic regression on all adopted daughters with known marital outcomes. The outcome variable, whether the adopted daughter married her adoptive brother, was coded as 1 if such a marriage was observed and 0 otherwise. We then identified the set of a girl's characteristics that best predicted whether she would subsequently marry her adoptive brother using AIC comparisons. The model confirmed Arthur Wolf's ethnographic insights (table 1), so we used the following rule to identify a DIL adoption: if a girl's adoptive parents had a biological son alive at the time of adoption who was at most 15 years older than the girl, she was adopted as a DIL; other cases were coded as AD.

Mortality analysis-bivariate models of association
We began by calculating two bivariate measures of mortality. Age-specific mortality rates (ASMRs) were calculated for age groups until the age of 40, after which the number of recorded deaths per category was too low to make any meaningful inferences. ASMRs were calculated as the number of death events recorded for all individuals between age x and x + 1, divided by the number of person-years of exposure to risk for that interval [39]. ASMRs were calculated separately for biological daughters, ADIL and AD; however, because the time lived by adopted daughters prior to adoption was, by definition, not a period of risk for adopted daughters, any person-years lived by a girl before she was adopted were allocated to the person-years for biological daughters.        to select covariates and interaction terms included in the model [40]. As, by definition, all girls who experienced adoption had to survive until they were adopted, adoption was treated as the start of the trial period, with age at adoption included as a control (   In order to further confirm the causal effect of adoption on mortality, we ran additional Cox models that include adoption and cancellation (N = 192; see §2.3 Variables, below) as time-dependent covariates [41]. Whether or not a person was adopted or had their adoption cancelled was coded 0 or 1 for each 1-year time period in their life, allowing us to model the effect of becoming adopted (or having that adoption cancelled) as opposed to simply comparing individuals who were adopted at some point during their lives to those who were not. This circumvents the selection issues that would otherwise exist when comparing adopted with non-adopted individuals (i.e. that arise because death would often precede adoption and be allocated mistakenly to biological daughters). 1 We ran four models: the first     [20] result showing that adopted girls, taken together, were less or equally as likely to die as their biological counterparts, with an additional control for a woman's parity (see below); the second two focus on adopted girls, comparing AD (Model 2,  , table S3 for results with parity included), as the majority of mortality arose prior to childbearing ages. Finally, we include four logistic regression models in the electronic supplementary material (tables S5-S8) that explore the likelihood of ADIL versus AD surviving to age 5 and 10, respectively; these are run on the entire sample of adopted daughters or with girls dying before six months of age excluded as   indicated by the table captions. These analyses serve to verify the main results of the CPH models during the period of highest mortality and before childbearing would have begun.

Variables
We include several covariates in our inferential models that are known or hypothesized to be associated with either adoption status or the risk of mortality. The Japanese collected information on the occupation of the head of household, which we use as an indicator of socioeconomic status, lacking any other direct indicators such as household income, land ownership or taxes paid. Following [20], we collapsed over 200 original socioeconomic categories into five: agriculture, craftsman, labourer, landlord and merchant. Labourer and landlord are the most reliable of these categories for indicating wealth, for household heads classified by colonial officers as labourers had no land and no skills and those classified as landlords rented out their land. We also include whether a girl had bound feet, which causes limited mobility, and has been widely (if perhaps mistakenly) believed to be linked to marital aspirations (e.g. [42,43])for evidence that foot-binding is not linked to marriage, see [44][45][46] [45][46][47][48][49]). The 10-year birth cohort in which a girl was born is included in our models, as both the prevalence of adoption and the risk of mortality declined through time [14,15,37,50]. Because ADIL prevalence varied regionally in Taiwan (e.g. [37,50]), we include an indicator of the region in which the girl resided. We include parity to account for a possible link between fertility and mortality (including maternal mortality) [51], and to allow for different fertility trajectories for ADIL versus AD [15,52,53]. Finally, we include whether an adoption was cancelled in order to evaluate the effects of cancellation on mortality. Ethnographic insights regarding motivations for cancellation are scant but suggest that cancellations may have occurred when girls ran away or were not delivered to adoptive parents as promised, if a boy (either a biological or adoptive brother) became ill or died (thereby affecting the girl's sibling-based status within the household), or if the girl was sickly. Thus, there are a priori reasons to anticipate a positive relationship between cancellations and mortality; hence our inclusion of cancellation as a covariate.

Identifying DIL adoptions
Using adopted women with known marital outcomes, we found that the presence of an older boy in the adoptive household was the single best predictor of later being married to a boy in the adoptive household (table 1). An adoption being cancelled was the single best negative predictor for being married to a brother in the adoptive household. Consistent with expectations, birth year was also negatively associated with the odds of being married to an adoptive brother, as were the age of adoption (a younger age of adoption was more likely to be associated with ADIL status), having bound feet, and residing in areas where the prevalence of minor marriage was known to be low (see [20], electronic supplementary material, table S1). Applying this predictive method to our sample of adopted women with known marital outcomes, we found that not having a boy present in the adoptive household at the time of adoption correctly predicted marrying outside of the adoptive family 74.2% of the time, while having at least one such adoptive brother at the time of adoption correctly predicted that an individual would marry one of her adoptive brothers 94.8% of the time. On eliminating girls whose adoption was subsequently cancelled, the predictive accuracy was 80.0% and 94.9%, respectively. These values were consistent enough to confirm Arthur Wolf's rule (see §2.2, above); thus, we identified as ADIL any adopted daughters who had an adoptive brother at the time of adoption and whose adoption was not cancelled.

Analyses of mortality
Age-specific mortality rates (figure 1; electronic supplementary material, tables S1 and S2) show that, contrary to expectations, ADIL had higher mortality than AD, although both categories of adopted daughters appeared to fare better than biological daughters. ADIL showed higher mortality particularly between the ages of 1 and 5 (figure 1), the period where most of the adoptions were occurring. Although mortality rates for AD are slightly higher than for ADIL during the first year of life, this is based on a small sample because few person-years were lived by adopted daughters in that age category. From age 5 to age 15, rates are relatively similar; mortality appears to rise again for biological daughters and ADIL relative to AD during the childbearing years; these results must be interpreted with caution given small sample sizes at later ages. We include parity as a control in our subsequent analyses except as otherwise noted to allow for the possibility that differential fertility drives differences in mortality during adulthood.
Cox proportional hazard (CPH) models using time-dependent covariates confirm that ADIL fared better than biological daughters with respect to mortality, but worse than AD. In particular, Model 1 (table 2) confirms the protective effect of adoption described by Mattison et al. [20], with adopted girls having a 29.7% lower hazard of mortality at any given time than biological daughters raised in their natal households. Models 2 and 3 (tables 3 and 4) replicate these results with AD and ADIL, respectively, and the protective effect of adoption is much stronger for AD (B = −0.51) than for ADIL (B = −0.26). Comparing AD directly with ADIL reveals a similar trend: AD have lower mortality than ADIL (table 5), though, due to small sample size, this result is attenuated with the addition of parity as a covariate (electronic supplementary material, table S3). Importantly, Model 3 (table 4) also reveals that, for ADIL, having an adoption cancelled is extremely detrimental with respect to mortality: an ADIL whose adoption was cancelled had a 1.98 higher hazard  (table 5). Highlighted regions indicate 95% confidence intervals.
of dying than a girl who never experienced adoption. This increased probability of mortality suggests strongly that adoption, per se, is protective against mortality. Sickliness of girls whose adoption was cancelled seems an unlikely explanation of their higher mortality given that only six cases were observed to die soon after (within 30 days of) cancellation. All of the analyses except one point to AD having higher survivorship than ADIL. In a direct comparison that excludes parity as a covariate (Model 4, table 5), AD have approximately 32% lower hazard of mortality than ADIL, controlling for other covariates. Figure 2 displays these differences graphically in terms of survivorship. By 40 years after adoption, 83.6% of the adopted daughters were still alive, whereas only 77.9% of adopted daughters-in-law were. Adding parity as a time-dependent covariate to the comparison of AD and ADIL results in non-significance. We strongly suspect that this is an issue of sample size. Four logistic regression models comparing the probability of surviving to age 5 or age 10 (see electronic supplementary material, tables S5-S8) support lower survivorship of ADIL compared to AD. The effect is the strongest to age 5, in line with mortality differences being largest at early ages ( figure 1 and electronic supplementary material, tables S5-S8). Finally, in all models, covariates' associations with mortality are fundamentally in line with expectations. Girls born at later parities suffered higher mortality, as did daughters of labourers. The north-south gradient in mortality is also apparent in these data. Interestingly, having bound feet is also protective against mortality, though the magnitude of this effect is much smaller than the effect of adoption in all models. Finally, a woman's parity shows an inverse association with the hazard of mortality in all models; i.e. women with higher realized fertility survived longer than those with lower fertility. This effect was apparent for all classes of daughter.

Discussion
In this study, we present a finely resolved test of the effects of adoption on mortality, using a systematic rule to identify girls who were likely to have been adopted for the purposes of minor marriage versus for other purposes. Although inspired by ethnographic evidence, our rule was verified by demonstrated associations between a girl's natal and adoptive circumstances and her subsequent marital outcome. Thus, whereas we showed previously that adopted girls on the whole fared better than never-adopted girls in terms of survivorship, in this paper, we delineate more precisely the effects of different forms of adoption on mortality. Specifically, both ADIL and AD experienced lower mortality than biological daughters who remained in the care of their natal families, but this effect was stronger among AD than among ADIL. While ADIL did indeed experience lower mortality than biological daughters (Model 3, table 4), the association between adoption and mortality was more strongly negative for AD (Model 2, table 3). This trend is upheld in a direct comparison of AD to ADIL (Model 4, table 5), despite small sample size, though this trend loses significance with the addition of parity as a time-dependent covariate (electronic supplementary material, table S3). Taken together, these results reinforce (i) the importance of the motivations for different types of adoption in identifying its likely consequences in terms of life history and (ii) the need to look beyond narrow kin selection (i.e. the focus on the coefficient of relatedness) as the main evolutionary explanation for patterns of human adoption and its life-history consequences; other costs and benefits that might accrue to adoptive families should also be considered [5,6,13,54].
Whereas the majority of research in this area has focused on broad comparisons between biological and non-biological children (e.g. [20][21][22]), we have shown that the specific type of adoption affected the risk of mortality for Taiwanese girls during the early twentieth century. In reviewing the effects of fosterage on child welfare, Castle [55] points to the importance of adoptive parental volition in mediating foster children's nutritional outcomes-children who were demanded by their foster parents had better nutritional outcomes than children who were not. Similarly, Pillai & Sharma [26] contend that fosterage, per se, is less important than the contexts surrounding fosterage in affecting child outcomes: crisis situations lead to worse outcomes than voluntary fosterage. In the historical Taiwanese case, AD were explicitly sought by their adoptive parents to remedy perceived demographic deficiencies; thus, we might anticipate that their welfare should not have suffered as much as children being adopted under more dubious circumstances. Indeed, AD were adopted into households with an average of 0.25 adoptive siblings, compared to 3.36 for ADIL, consistent with AD serving as stand-ins when biological children were lacking.
That AD out-survive ADIL is not in line with kin selection explanations [20] that anticipate preferential treatment of the latter given their direct contributions to the adoptive parents' inclusive fitness, leaving us to speculate as to what drove the differences in mortality shown here. The prevalence of in-family adoption in this dataset is exceedingly small (electronic supplementary material, figure  S1) and such adoptions were not considered in our analyses. Thus, the protective effect of adoption is not due to increased altruism towards genetically related individuals. One possibility is that adopted daughters, regardless of category of adoption, provided economic or other benefits to their adoptive families that were not readily available via other members of the household [14,56]. Given the early position in the sibset into which these girls were adopted, adopted daughters could have assisted with activities (e.g. childcare) that increased the fertility of their adoptive parents [14,15,38,39,54,57,58]. A second possibility is that adopting a daughter at a relatively young age-particularly if daughters were breastfed, as often happened in the case of ADIL-invoked a psychological response in their adoptive parents [59] that improved the treatment of adopted daughters relative to daughters who remained in their natal homes. This hypothesis seems unlikely given that cancelled adoptions were associated with higher mortality and that Wolf [38] has provided evidence from the colonial Taiwan household registers against breastfeeding as mediating the relationship between adoption and mortality. Moreover, a presumed breastfeeding mediation of mortality would not explain why biological daughters fared worse than adopted daughters or why AD fared better than ADIL, given similar distributions in the age of adoption (electronic supplementary material, figure S2).
A third possibility worth serious consideration is that parents neglected their biological daughters. Ethnographic materials suggest that biological daughters who remained at home were viewed negatively. Seen as expensive (because they required dowries) and as a means by which other people's lineages were perpetuated, unless they began making economic contributions at an early age [45,46,48], biological daughters might have seemed like poor investments relative to giving one's own biological daughters up for adoption to another family. Although many have emphasized the stigma and neglect or mistreatment of adopted children in Taiwan during this period, treatment of biological children up until recently was also often harsh [60]. Disobedience of children was often met by beatings, because 'rods produce filial [children]' [58, p. 158]. Moreover, although adopted children were also routinely beaten by their Taiwanese parents, the number of deaths arising due to such beatings was probably very small [58]. Comparing our age-specific mortality rates to mortality rates drawn from the UN model life tables (Far Eastern female pattern; highest mortality; electronic supplementary material, figure S3) shows that AD and ADIL track the modelled mortality better than biological daughters, whose mortality appears slightly inflated, providing somewhat indirect support for the neglect of biological daughters. This result is consistent with models of conflict among kin wherein biological kin may be disfavoured if they compete for scarce resources [61]. Moreover, the voluntary nature of adoption in this context might be expected to lead to better outcomes for adopted children [62].
A final possibility arises due to potential differences in fertility between different classes of adopted daughter and biological daughters. In line with Westermarck's hypothesis that individuals co-reared together from young ages experience incest avoidance [14,15,38,[52][53][54]63], ADIL have been shown to have lower fertility than biological daughters who marry individuals from outside their household [14,15,38,[52][53][54]63,64]. If higher fertility results in higher mortality [65], including maternal mortality, then biological daughters' excess mortality might be due simply to their higher parity. Similarly, ADIL might suffer higher mortality than AD if AD had lower fertility. Our models suggest that this possibility is not a likely explanation for these data. The differences in mortality that we report exist with and without inclusion of a woman's parity as a control variable. In fact, parity shows an inverse association with mortality, suggesting that women with higher fertility also survive longer. Phenotypic correlation (i.e. women with higher energetic reserves reproduce more and have higher survivorship) has been proposed to account for positive correlations between fertility and survivorship [66,67]. A strong test of how fertility relates to mortality would have to consider not only the pace of reproduction and maternal energy budgets [66], but also potential cultural effects, such as Wolf's findings that colonial Taiwanese women's risk of divorce-with loss of old-age support implied-only declined significantly after giving birth to multiple sons [14,15]. Such a test is beyond the scope of this paper and, given that the strongest divergences in mortality arise early in life [8], is unlikely to shed additional light on why AD and ADIL experience lower mortality than biological daughters. That adopted daughters (ADIL and AD) experienced lower mortality if their feet were bound is an interesting result that we did not anticipate. Prior work on the functions of bound feet has focused on their use as a marker of status [42,43], as a means of insuring paternity (e.g. [68]), or as a way to increase the contributions of women to household economics through handicraft production [44][45][46][47][48][49]. Any of these functions could also be associated with decreased mortality if adoptive parents felt that the benefits of adoption were increased by the binding of their adopted daughters' feet. In this dataset, bound feet may also serve as an ethnic marker, because Sinicized plains Aborigines were both unlikely to have bound feet and to practise minor marriage [12]. Regardless, our result implies that functional approaches to understanding the benefits and costs of bound feet may, therefore, be more illuminating than those focused solely on its (highly visible) costs.
Our data and analyses are subject to several important limitations. First, our measures of socioeconomic status are imperfect and do not allow for fine-grained resolution of the confounding effects of socioeconomic status on adoption and mortality. Although such effects may have been relatively weak given the ubiquity of adoption across socioeconomic differentiation [14,15], we may have underestimated the importance of socioeconomic status in mediating the relationship between adoption and mortality. However, we find no evidence that different types of adoption occurred at different rates to parents of different socioeconomic status (electronic supplementary material, table S4) and higher rates of mortality among children of labourers is consistent with their disadvantaged status in Taiwan during this time period [12,14,37]. Second, we lost a significant portion of the sample by excluding individuals whose births were not registered in the dataset and by including only cases of adoption where circumstances at birth and adoption were known (as described in electronic supplementary material). These restrictions may limit generalizability of our findings if the characteristics of girls who could be followed to adoptive households differed from those who could not be followed. Still, few datasets are able to address the contributions of adoption to health outcomes in longitudinal fashion, let alone while taking characteristics of the natal household into account.

Conclusion
We have shown that adopted daughters-both ADIL and AD-experienced lower mortality than biological daughters remaining in their natal household in early twentieth-century Taiwan, but that AD-who did not contribute to the perpetuation of the adoptive parents' lineage-out-survived ADIL. These findings suggest that the functions of adoption and other forms of allocare should be viewed in the context of forms of cooperation that extend beyond the narrowest version of the kin selection hypothesis. In revealing lower mortality for AD than for ADIL, this study contradicts the hypothesis that ADIL should be favoured because they contribute to biological perpetuation of the lineage, suggesting that other unexplored benefits, such as enhancement of parental fertility, may accrue to adoptive parents or alternatively that biological daughters may be neglected, perhaps because they compete for scarce resources, which are delivered instead to biological sons.
Our findings point to the importance of the specific contexts that give rise to adoption-especially the differences between adoption in industrialized and non-industrialized populations-if we are to make any general statements about its functional significance. Adoption practices vary significantly within and among populations, from Western-style adoption of non-related children [69], often to nulliparous parents, to adoption of closely related kin as a means of redistributing children to optimize household size and labour inputs [10,70]. Our study thus urges renewed attention in at least three domains of evolutionary research on adoption. First, our study shows that 'adoption' is not monolithic and that the specific forms it takes can differ and have different consequences in terms of individual life histories. Second, the cultural contexts under which adoption occurs will affect the attendant costs and benefits and must be considered in evolutionary arguments describing the adaptive value of adoption in any given time or place [3,7]. Third, our study offers the interesting possibility that biological daughters may be disfavoured relative to adopted children who are intentionally brought in to serve various roles in the adoptive family. Conflict among kin is widely recognized to create biases in parental investment [61]; extensions of such models to explaining the differences in care of biological and non-biological children [71,72] would facilitate a more nuanced understanding of whether and how parents ultimately stand to benefit from the investments they make in their children, related or otherwise.
Ethics. Data are obtained from the Academia Sinica and used in accordance with their policies. Data accessibility. The dataset underlying this study includes the full household registers from 13 sites distributed across Taiwan (all are listed in the manuscript tables) and includes two files per site: a 'static' file that includes information on dates of birth, dates of death, ID numbers of the mother and father, the sex, the same-sex sibling order, and whether the individual had bound feet; the second file for each site is a 'dynamic' file that includes information on events (we were concerned with adoption, cancellation of adoption and marriage) that occurred during the individual's lifetime and the dates of these events as well as relevant IDs (partner ID, adoptive parent ID). Providing a full dataset would both run the risk of identifying individuals involved in the study and would be in violation of the terms of our access to the dataset provided by Academia Sinica. Academia Sinica is an institution of higher learning in Taiwan (which is analogous to NIH+NAS); their programme for historical demography maintains the registers and provides access to researchers based on the specific needs of each researcher. Access is granted through Academia Sinica's application form, which is available here (http://www.demography.sinica.edu.tw/Downloads/ 2202/db/THHRD%20Application%20and%20Regulation%20-EN.pdf). There is no independent deposition of raw data allowed under their policy. To assist the editor and referees in assessing this work and associated data, we provided temporary access to the dataset that we worked with to run the analyses to editors and reviewers as a password-protected file that accompanied this submission. Furthermore, we provide a range of data in aggregate (e.g. means, variances and ranges) that might be of interest here: http://bit.ly/2lTXFuN. Additional data may be requested on an individual basis via the above Academia Sinica link.