Identifying Matrilineal Kin Networks in the United States 1900–1940

Abstract

National-level analysis of historical women’s life courses and kinship networks is understudied due to data limitations. Previous analysis of patrilineal kin propinquity in the United States relied on surnames to identify the probability of non-random isonymy (same surname) as a proxy for nearby kin. One limitation of this approach is losing matrilineal kinship networks due to women’s changing surnames upon marriage. Using linked census data from IPUMS MLP, LIFE-M, and Census Tree for 1900–1940, this project aims to identify birth surnames for married women to better measure matrilineal kinship networks in the United States. Results using linked census suggest non-coresidential matrilineal kinship networks underwent a similar decline as proximate patrilineal kinship networks in this time period, but nuances suggest that we should not treat these declines similarly. Matrilineal and patrilineal kin propinquity differed in the life course patterns for currently married women, and accounting for both matrilineal and patrilineal kin propinquity are important to assess the full kinship network. The rates differ in some cases between the three linked datasets, suggesting that underlying biases in the data explain some of the differences. However, the trends are similar in all geographies and methods suggesting that the identified trends are not due to random chance. While this analysis does not directly identify kin, the data produced here allows other research to identify relations from other censuses with data quality flags indicating how relationships are determined. This marks a revolutionary improvement in national-level analyses for women’s life course research in the early twentieth century.

Introduction

A lack of data has complicated studying women’s full life courses using historical demographic data in the United States due to women’s changing surnames preventing linking between pre- and post-marital states. Relatedly, we are unable to study the national prevalence of matrilineal kinship networks due to this data limitation. This is important for studies on fertility, mortality, migration and family structure as kinship networks influenced and were influenced by all of these demographic outcomes in many different spatial and temporal contexts (Dillon et al. 2024; Engelhardt et al. 2019; Hacker and Roberts 2017, 2019; Hacker et al. 2021; Harton et al. 2023). While recent research leveraged the IPUMS United States full count census data to estimate patrilineal kin propinquity in the United States between 1790–1940, it did not identify matrilineal kin proximity (Nelson 2020). This paper uses three different linked datasets to identify matrilineal kin, focusing on currently married women with surnames changed upon marriage.

The importance of this work for studying the life course of women cannot be understated. One of the largest data limitations in historical demography is the loss of information on women after marriage and the inability to study early life effects on later life outcomes at a national scale for these women. With linked census data, researchers can explore these kinship links to study intergenerational mobility, demographic outcomes, and socioeconomic inequalities for ever-married women. This paper discusses the creation of kinship links using linked census data, analysis of patrilineal and matrilineal kinship networks for married women, and some limitations and biases in the data. While the analysis for this paper focuses on the early twentieth-century United States, the method also works for linked nineteenth-century United States census data. These new kinship data marks a revolutionary improvement in national-level analyses for women’s life course research in the United States in the early twentieth century.

Background

Demographic processes and kinship networks shape each other intergenerationally, with fertility, mortality, migration, and marriage patterns determining the prevalence and extent of kinship networks, while at the same time kin provided child care, affected migration decisions, and exchanged economic resources which shaped demographic processes in turn (Daw, Verdery, Margolis 2016; Hacker et al 2021; Harton et al 2023; Nelson 2019; Newson et al. 2005, 2007; Newson & Richerson 2009). Due to data limitations, our understanding of these intricate relationships between demographic processes and non-coresidential kinship networks are not well understood for the nineteenth and early twentieth century United States. While local studies often contain an amazing amount of detail on kinship networks, these studies tend to be highly selected populations and difficult to generalize broadly (Billingsley 2004; Gjerde 1985; Nelson 2019; Ostergren 1988). This is particularly true of non-whites, migrants (harder to track over time) and women (changed their surname upon marriage and also difficult to track over time).

Historical U.S. demographic research that does focus on the role of kin on demographic processes has focused on coresidential kin (Esteve & Reher 2021, 2024; Gratton and Gutmann 2010; Ruggles 1994a, 1994b, 2007, 2011). While studying coresidential kin is an important aspect of demographic outcomes, it excludes the locality of residence for married couples when they leave their households of origin. An example could be elderly adults who may not live with children but live near children. These types of residency could be important for health or socioeconomic outcomes. While United States residence patterns were typically defined by neolocal residency, individuals often lived near both patrilineal and matrilineal kin (Fawver 2006; Nelson 2020; Ruggles 2015). U.S. kinship generally (but not universally) emphasized multilineal kinship networks rather than patrilineal or matrilineal which is lost when we focus only on coresidential kin (Parsons 1943).

Nelson (2020) and Smith (1989) established that U.S. Families experienced a long-run decline in patrilineal kin propinquity in the United States since 1790, but even in 1900 nearly 30% of individuals lived near another family with the same surname non-randomly. This matters because geographic proximity of kin played important roles in reproductive outcomes, kin assistance, and migration decisions (Engelhardt et al. 2019; Hacker et al. 2021; Koylu et al. 2021). Unfortunately, research on non-coresidential kin proximity experienced data limitations by focusing on the patrilineal line, ignoring the common (but not universal) bilateral descent kin structure in the United States. Matrilineal kin networks at a national scale have been understudied due to these data limitations. With new linked census datasets, we can now start overcoming some of these limitations from a data perspective.

DATA

To analyze matrilineal kinship networks, modern researchers require linked census data to identify the kinship networks for married women. For this comparison of linked census data, I use three datasets; the IPUMS Multigenerational Longitudinal Panel (IPUMS MLP), the Longitudinal Individual Family Electronic Micro-database (LIFE-M), and Census Tree. All of these datasets are available for download via their public websites.¹ IPUMS MLP is a supervised learning method that probabilistically links census records. It links cases in two steps; first by linking individuals using a large number of features for the highest accuracy, and then in the second step linking remaining household members that are present in both censuses. The IPUMS MLP data used in this paper also links cases from 1900–1940 to the Social Security Numident, capturing a subset of potential matrilineal links. The Social Security Numident records contain parental names on Social Security application records, allowing for the identification of birth surnames for women (Helgertz et al. 2021, Ruggles et al. 2021).²

Data from the LIFE-M project (Bailey et al. 2022) and Census Tree (Price et al. 2021; Buckles et al. 2023) are also analyzed to validate the results. LIFE-M is limited to persons from administrative records in Ohio and North Carolina, but using administrative birth, marriage, and death records to link persons provides more potential links for matrilineal kin.³ By using these administrative records, LIFE-M directly observes name changes for women which makes unique from the other two linked datasets. The individuals and families are linked over multiple steps to a variety of data sources using a supervised machine learning algorithm. The combination of a variety of sources allows for high accuracy in linking but also contains a wealth of information on various socioeconomic and demographic features of interest. Because LIFE-M focused on linking between administrative and vital records, sometimes the LIFE-M data only links to one census in the IPUMS full count data. To maintain similar data structures as IPUMS MLP and the Census Tree data, only LIFE-M data with links to two or more censuses is used here.

The Census Tree data were developed by combining multiple datasets to increase the number of links while adjudicating links between methods. The Census Tree data benefits from combining links from their own linking method using optimized distributed gradient boosting (XGBoost) with IPUMS MLP, the Census Linking Project (CLP), crowdsourced genealogical data from FamilySearch, and FamilySearch hints generated by a proprietary machine-learning algorithm. Because of the inclusion of genealogical data, Census Tree data allows for identifying matrilineal links beyond what IPUMS MLP currently identifies. The current limitation in Census Tree is selecting only non-conflicting links from their crosswalks⁴ and that the provenances of some of the proprietary FamilySearch hints data is unknown. Links from the Census Linking Project (CLP) were not used because the Census Linking Project only links men (Abramitzky et al. 2020). Census Tree links were selected by dropping CLP only links, and the remaining links were combined with all conflicting links removed.

METHODOLOGY

Identifying parents, siblings, and children

Analyzing linked census records enables capturing a more representative sample of the population with matrilineal kin networks compared to the cross-sectional data. Imagine a household with a married couple in 1900 that is linked to 1910 in Figure 1. In 1910, this married couple is now living with the wife’s parents with the surname of Faber. Since the wife’s record is linked to 1900, we can pull the Faber surname to the 1900 records, which allows us to identify potential matrilineal kin for these years in the absence of a matrilineal surname from 1900. We can also identify known kin with this approach. If the wife’s parents lived in their own household in 1900 nearby the married couple and are linked to the 1910 census, we can identify these people as kin. Identifying known kin is currently limited to parents, siblings, and children, although one could identify other kin such as grandparents, cousins, aunts and uncles, nieces, and nephews. The primary aim for this paper is to identify potential surnames (described more below under the kin propinquity method), and extended kin are unlikely to contribute towards that goal without increasing computation time. Future analysis that relies on direct kinship identifiers should expand to include these additional kin. Currently this analysis does not consider stepparents, and cases with conflicting parental identification are dropped from the analysis.

Figure 1:

Example of Linked Census Records and Identifying Matrilineal Kin

Table 1 provides the percentage count for each year of identified mothers in IPUMS MLP, Census Tree, and LIFE-M. As an example for MLP, in 1900 no mother is identified in any census for 42.5% of all linked census records. 47.2% of linked 1900 records are living with their mother in 1900, 9.6% are living with their mother in another census year, 0.7% of linked records had their mother identified via another relation (e.g., ego lived with a sibling in 1900, and their sibling lived with their mother in 1910). 0.6% of cases the mother was not identified because of conflicting links.⁵ This could indicate stepparents and bad links. The numbers were similar for fathers.

Table 1:

Data Quality Flag for Identified Mothers by Linking Method, 1900–1940

MLP
Year	Not identfied	In Census Year	In different census year	With other relation	Conflicting links	N
1900	41.6%	47.7%	9.5%	1.0%	0.3%	39,489,497
1910	38.3%	46.4%	14.4%	0.8%	0.1%	56,020,355
1920	37.6%	49.0%	12.6%	0.7%	0.1%	67,337,295
1930	36.6%	47.3%	15.5%	0.7%	0.0%	81,081,849
1940	37.8%	44.0%	17.7%	0.6%	0.0%	84,715,943
Census Tree
Year	Not identfied	In Census Year	In different census year	With other relation	Conflicting links	N
1900	31.7%	48.8%	16.0%	2.5%	1.0%	49,731,024
1910	28.2%	45.8%	23.7%	1.7%	0.6%	63,638,624
1920	25.3%	47.3%	25.9%	1.3%	0.3%	74,644,330
1930	24.8%	45.2%	29.0%	0.9%	0.1%	88,120,030
1940	24.1%	34.3%	40.8%	0.7%	0.1%	71,354,256
LIFE-M
Year	Not identfied	In Census Year	In different census year	With other relation	Conflicting links	N
1900	44.8%	46.9%	8.3%	0.1%	0.0%	488,802
1910	58.1%	24.6%	17.3%	0.1%	0.0%	626,022
1920	65.6%	6.5%	27.9%	0.0%	0.0%	758,910
1930
1940	55.2%	2.4%	42.4%	0.0%	0.0%	688,854

Because of the genealogical data, Census Tree captures more mothers beyond the household than MLP, but more conflicting mother links were identified as well. This could be due to several reasons. First, Census Tree has more links than MLP and on average links an individual more often between censuses, which raises the likelihood of a bad match being made. Second, since Census Tree combines multiple methods together, it’s possible that because of the intersecting nature of links that if a bad match is made in one method, it’s cascading into other methods causing further bad matches. By the 20^th century, Census Tree links improve with lower rates of conflicting mother identifications and more mothers identified in different census years. While the sample size of the LIFE-M data is smaller and less representative than MLP and Census Tree, the use of administrative records improves the number of mothers identified from a different census year and has the unique feature of having no conflicting mother links in the data. Table 1 shows 44% of linked individuals in 1940 LIFE-M data with their parents identified in the data from a different census year.

The next step is identifying the birth surnames for currently married women. Because linking strategies do not require exact surname matches, one factor to account for is multiple surnames. For example, if you have a linked individual linked across five censuses, one linked record could in theory have five different surnames due to enumerator or data transcription error. To limit this issue, I compare each identified parental surname to each other by computing a Jaro-Winkler similarity score. A score of 1 means the strings match exactly with a lower number indicating less similarity between strings. Surnames that have a Jaro-Winkler similarity greater than 0.8 are considered to be the same surname while a score of less than 0.8 represents a different surname. Any linked records that have three or more different surnames across linked censuses are dropped from analysis. While some of these cases could represent legitimate surnames such as a mother remarrying, it could also represent an incorrectly linked record. While data from LIFE-M and Census Tree also go through a similar approach, we can directly observe name changes for many women in LIFE-M, which allows for capturing a larger part of the potential universe as seen in Table 2.

Table 2:

Sample Size Linked Currently Married Women by Dataset and Geography, 1900–1940

Sample Size Linked Ever Married Women by Dataset, 1850–1940
	United States
	Linked Currently Married Women			Linked Currently Married Women with birth surname
Year	MLP	LIFE-M	Census Tree	MLP	LIFE-M	Census Tree
1900	7,936,776		9,567,972	598,380		2,183,826
1910	11,126,596		12,686,524	1,168,393		3,110,833
1920	13,930,028		15,731,986	1,872,192		4,102,178
1930	17,467,057		19,526,075	2,816,431		4,830,295
1940	19,311,159		17,226,252	3,756,186		4,536,709
	Ohio and North Carolina
1900	707,320	97,813	836,084	51,744	18,544	216,522
1910	932,435	197,459	1,064,153	97,303	49,859	296,819
1920	1,172,800	331,047	1,301,539	158,315	110,253	374,973
1930	1,439,961		1,626,017	235,371		441,115
1940	1,564,673	278,075	1,460,896	311,641	121,161	409,634

Table 2 shows the sample size of linked ever married women for each dataset and the sample size once parental identification is completed and conflicts are removed. Census Tree identifies the largest number of birth surnames for currently married women, but in terms of the total universe, LIFE-M identifies the largest proportion of possible birth surnames. As an example, although MLP and Census Tree identify more currently-married women links in 1940 Ohio and North Carolina relative to LIFE-M because of LIFE-M sampling restrictions (1.6 million and 1.5 million respectively to 278,000), because we directly observe name changes in the LIFE-M data, we capture a larger percentage of linked married women with birth surnames (20% and 28% respectively to 44%).

Kin Propinquity 1900–1940

As a proxy for kinship, I calculate the probability of non-random isonymy (same surname) for both patrilineal and matrilineal kin as defined by Nelson (2020). This kin propinquity method relies on the sequential ordering of enumeration within the United States censuses. Households next to each other physically were enumerated in order, so distance on a census page can be used as a proxy for number of households away (Grigoryeva and Ruef 2015; Kenzer 1987; Logan et al. 2011; Owlsey 1949). The further away households are from one another, the less likely this assumption holds, however the majority of same patrilineal surname families are found within the first 7 households away in 1940. With surname distributions and household distances between same surname families, researchers can calculate the probability these same surname families non-randomly appear near each other within the census. A high probability score assumes that same surname families living near each other are likely related kin. This kin propinquity method works for data where the ordering of families or households is representative of the actual physical distance between families and households.

Equation 1 from Nelson (2020) describes the calculation of the probability of non-random isonymic (same surname) families.

$$P\left(K_{rsi}\right)=\{\begin{array}{cc}1-{\left(\frac{F_{rsi}-1}{T_{rs}-1}\right)}^{D_i},& F_{rsi}>1\\ 0,& F_{rsi}=1\end{array}$$ (1)

The probability measure uses the surname distribution (F_rsi-number of families with surname i of race r within enumeration district s divided by T_rs, the total number of families of race r within enumeration district s) and the number of households between families of the same surname (D_i) to measure the probability the same surname occurs non-randomly within the census. D_i is defined in both directions on the census page, so if the household next door had the same surname as the ego, D_i is equal to two, because one has to search above and below the ego household on the census pages. D_i is adjusted for households at the beginning and end of the enumeration district. The same equation applies for both patrilineal and matrilineal kin.

One important note in this research is patrilineal and matrilineal kin propinquity are defined at the married couple level, which means patrilineal identifies the husband’s potential kin while matrilineal identifies the wife’s potential kin. To assess the full kinship network, we can also analyze multilineal kin propinquity, defined as living near potential patrilineal or matrilineal kin. The analysis focuses on currently married women because LIFE-M identifies currently married women and not as many widows, divorcees, etc. Further, the analysis only looks at North Carolina and Ohio within the LIFE-M data while MLP and Census Tree data are analyzed within North Carolina and Ohio as well as at the national level.⁶ Finally, because LIFE-M does not link to the 1930 census, any results for LIFE-M and 1930 are the trend between 1920 and 1940 from Census Tree and MLP. To establish the missing data in the figures, I implemented a dotted line between 1920 and 1940 for the LIFE-M data. Table 2 describes the sample sizes of currently married women by method between 1900 and 1940 for the United States and Ohio/North Carolina. To control for representativity issues in the linked census data, I calculate inverse probability weights as described by Bailey et al. (2023) and Abramitzky et al. (2020).

In general, the three methods have different strengths for studying the underlying populations. Census Tree performs particular well in the occupational structure of husbands, with the largest different between the full underlying population related to overidentifying farmers and under identifying laborers. LIFE-M and Census Tree both do well in identifying rural farm, rural non-farm, and urban status of linked cases. While all three linking methods underidentify non-whites in the linked data, Census Tree performs slightly better than MLP, while LIFE-M only identifies 1–2% of the linked population as non-white compared to the true population rate of 10%. LIFE-M also overidentifies younger persons in the data, likely due to their reliance on birth, death, and marriage records for linking cases. Finally, while MLP still underidentifies interstate migrants and foreign-born relative to the full population, MLP performs far better than both Census Tree and LIFE-M in identifying migrants in their linking algorithm. These are all important features to consider as will be shown in the results section.

RESULTS

Aggregate kin propinquity rates

The kin propinquity rates for patrilineal kin for each linking method can be compared to the full population of currently married women in Ohio and North Carolina as a baseline to determine the general accuracy of the patrilineal kin measure. When unweighted, LIFE-M and Census Tree showed higher patrilineal kin propinquity rates than MLP. Further all three methods show significantly higher kin propinquity when unweighted, likely reflecting that those with proximate kin are more likely to be linked. When looking at the aggregated patrilineal kin propinquity rates, they are approximately 5 to 11 percentage points higher than the true population patrilineal kin propinquity rate. When comparing the weighted results of patrilineal kin propinquity for each method to the cross-sectional patrilineal kin propinquity rate for all currently married women in Ohio and North Carolina, MLP and Census Tree perform the closest to the population rate relative to LIFE-M. LIFE-M overstates patrilineal kin propinquity by seven to ten percentage points. Both MLP and Census Tree are within one percentage point of the true population for both Ohio/North Carolina and national-level analyses.

While all three methods show a general decline in matrilineal kin, their absolute rates differ from each other, even when the results are weighted (Figure 2). This suggests that the biases of the underlying linked data could be affecting the absolute rates. Census Tree data was more likely to have non-migrants in their data relative to MLP, and non-migration is associated with higher kinship network density which could explain why Census Tree has higher rates of matrilineal kin propinquity relative to MLP (Nelson 2020). Multilineal kin propinquity rates show similar trends, but with different absolute rates because of the higher patrilineal rates from LIFE-M and higher matrilineal rates from Census Tree in both Ohio/North Carolina and national-level analyses.

Figure 2:

Patrilineal, Matrilineal, and Multilineal Kin Propinquity Rates by Method and Geography for Currently Married Women Age 15+, 1900–1940

These findings while different in absolute rates, all depict the same trends in both the Ohio and North Carolina and national-level analysis. Neither MLP or Census Tree directly observe married women’s name changes, but LIFE-M directly observes name changes in their source material via marriage records.⁷ The similar kin propinquity trends in Ohio and North Carolina between the three methods and at national level validates that kin propinquity declined for both patrilineal and matrilineal kin during the early twentieth century.

Five-year age group kin propinquity rates

As seen in Nelson (2020), patrilineal kin propinquity followed a family life cycle interpretation (Figure 3). Patrilineal kin propinquity declined from age 15 to approximately age 40, then increased afterwards through age 70. This reflects the earliest years of declining kin availability (mortality of older generations, migration patterns of younger generations) and later years of increasing kin availability (declining mobility of older generations, children leaving household of origin). On average, MLP performed the closest to the cross-sectional rates. Results for LIFE-M cannot be analyzed after age 50 in 1910 because of sample size limitations. These patterns are consistent in the other census years in this study.

Figure 3:

Patrilineal, Matrilineal, and Multilineal Kin Propinquity 5-Year Age Group Rates by Method and Geography for Currently Married Women Age 15+, 1910

Matrilineal kin propinquity on the other hand did not follow the same pattern as patrilineal kin propinquity. Since the woman’s children will have the patrilineal surname and not the matrilineal surname, any children who live nearby for women old than 45 will have the patrilineal surname. Related to that is kin availability; as time goes on, fewer kin with the matrilineal surname survive (except through more distant relations) while the patrilineal surname is reproduced via sons and unmarried daughters. While the trends for all three methods are consistent, Census Tree shows higher matrilineal kin propinquity rates in all age groups relative to MLP and LIFE-M. Multilineal kin propinquity rates roughly reflect the patrilineal trends. For women with any potential kin nearby, approximately 15–20% of currently married women lived near only matrilineal kin. 20–30% of currently married women lived near both patrilineal and matrilineal kin, and the remaining 60% lived near only patrilineal kin. These patterns are similar in all of the other census years.

Given the methodological limitations of applying the matrilineal kin propinquity method in the same mode as the patrilineal measure, researchers must measure both kinship networks to obtain a full understanding of the kinship network. Further comparisons of the two networks using the age profiles of potential kin help determine differences between patrilineal and matrilineal kinship patterns. Using the 1940 national rates from the Census Tree and MLP data, there are notable differences to point out. Some of these differences are due to the limitations of the applying the kin propinquity method to matrilineal networks described above, however some of the differences demonstrate nuances between patrilineal and matrilineal networks worth exploring further.

Figures 4 exhibits the age difference between the ego and the reference person of the same surname families by type of kinship network (patrilineal, matrilineal, and multilineal) for different ages of the ego. Results to the right of the black vertical box are potential kin who are older than the ego, while results to the left are potential kin younger than the ego. Most of the potential kin are 0–19 years older for ego ages 25–29, which are indicative of same generation kin (e.g., siblings) for both patrilineal and matrilineal kin. However, for patrilineal kin, while there is a smaller spike for ages 20–40 years older (indicative of first-generation kin such as parents) we do not see a similar pattern for matrilineal kin. While this could suggest a stronger patrilocal network, we need to be cautious about this interpretation because the kin propinquity method applied here only identifies the closest potential kin. This suggests that for currently married women who have potential matrilineal kin nearby, the closest potential kin were the same generation.

Figure 4:

Age Difference Between Nearby Patrilineal, Matrilineal, and Multilineal Kin and 25–29, 45–49, & 65–69 year old Currently Married Women by Method in United States, 1940

These nuances are further reinforced by for currently married women ages 45–49. Matrilineal kin tended live near same generation potential kin (though shifted to younger same generation kin over the life course) while patrilineal kin spiked for those 20 years or younger, indicating the next generation kin (e.g., children). The results for 65–69 year old women emphasizes the mechanical limitations of the matrilineal kin measure for children. Matrilineal kin exhibited a uniform distribution of potential younger kin and a spike for patrilineal kin. This reflects children taking the patrilineal surname and why researchers need to consider all potential kin rather than solely matrilineal or patrilineal kin separately. For those concerned about the mechanical limitations of the matrilineal kin propinquity measure, alternative methods include supplementing the matrilineal rates with patrilineal rates for potential children or only comparing specific age difference groups for patrilineal and matrilineal kin.

Another way to visualize these data are which generations an individual link to within each age group shown in Figure 5.⁸ If we define five generations (G0: Potential kin at least fifty years older; G1: 20–49 years older; G2: Less than 20 years age difference (ego generation); G3: 20–49 years younger; G4: 50 years or more younger) the patterns of generational overlap and availability become clearer. In the youngest years, most married women lived near potential kin of a similar age or a generation older. Living near the ego generation peaks around age 40–44 while Generation 1 (parental generation) declines over the life course. Starting around age 40–44, Generation 3 (children’s generation) increases and becomes the dominant group to live near by age 60. Kin availability described here is based on mortality, fertility, and migration outcomes and processes. Patrilineal and matrilineal kin propinquity independently show similar results, but differences related to the loss of potential children for matrilineal kin propinquity do show a higher peak for G3 (79%) relative to patrilineal kin propinquity (65%). These results are consistent for the other linking datasets, years, and geographies. One important caveat with Figure 5 is this only includes links to the reference person for same surname potential kin. For example, let’s say one links to G3 (potential children). Any G4 Grandchildren living within this family are not measured in this figure as it only considers the reference persons. Future analysis that studies kinship structures with the kin propinquity method should take that into account for measuring all potential kin relations and availability.

Figure 5:

Multilineal Kin Propinquity 5-Year Age Group Rates by Generation for Currently Married Women Age 15+ in Census Tree, 1940

LIMITATIONS

Researchers studying kinship networks with this data should consider the representivity of linked census data, the indirect measure of kinship presented here, and the cross-sectional analysis in this paper. While linked census data corrects for some of the population representativity issue of the cross-sectional data, this data does not perfectly represent the true population, particularly the genealogical data (Buckles et al. 2023; Helgertz et al. 2022; Koylu et al. 2021; Price et al. 2021; Stelter & Alburez-Gutierrez 2022). This is especially true for studying racial inequalities, where linkage rates for non-whites are lower. The linked data are selective of the wider population, and researchers should consider how the linked population (and by extension their kinship networks) are representative of the population at large. In most cases, methods such as inverse propensity weighting should be applied to help address this issue when practicable (Abramitzky et al. 2020; Bailey et al. 2023).

Based on the strengths and weaknesses of each linking method, even when weighting for the underlying demographic and socioeconomic characteristics of the data, these results (particularly absolute rates) could be driven by these biases. As an example, Census Tree is more likely to identify non-migrants and farmers, both of which exhibited higher kin propinquity rates. While the general kin proximity trends between methods are the same, the higher Census Tree rates could be due to this underlying bias. Conversely, LIFE-M overidentifies younger persons and currently married persons, which may affect life course studies depending on the research question and time frame. Researchers using linked data need to consider how these data bias their results and develop techniques to account for this bias.

The kinship analysis in this paper (kin propinquity) is not a direct measure of kinship but a measure of non-random isonymy. Since researchers can capture some direct kinship links within linked census data, future research should push for more direct measures when possible. Another limitation of the kin propinquity measure is that urban enumeration districts were geographically smaller than rural enumeration districts and kin propinquity in urban areas is biased downwards (Nelson 2020). Finally, while the kinship links in this linked data would allow for some reconstruction of life courses for individuals, this research paper focused on cross-sectional analysis. Future research would benefit from analyzing this data in a longitudinal format and tracking how kinship networks changed over the life course for individuals.

DISCUSSION

This paper establishes how to use linked census data to identify familial relations beyond the household and use that information to study kinship networks. While patrilineal and matrilineal kin propinquity declined during the early twentieth century, the declines were not similar over the life course, suggesting that these kinship networks potentially operated differently in the United States. Further, the linked datasets exhibit different absolute rates of kin propinquity, suggesting that researchers need to consider the strengths and weaknesses of each linked data set when determining which linking method to use in their analyses. All of these linked datasets benefit from weighting techniques as none of the populations are representative of the true population, but the precise weighting technique depends on the population of interest and research questions. While LIFE-M is limited to Ohio and North Carolina in its current form, using additional vital records to link individuals between censuses should be considered an avenue for future data processing and record linkage.

The results suggest that similar to patrilineal kin propinquity, matrilineal kinship networks experienced a similar decline between 1900 and 1940 in the United States. Given the mechanical limitation in the matrilineal kin propinquity method for assessing children, applying a multilineal measure helps alleviate this issue. When doing that, currently married women lived near siblings and to a lesser extent parents-in-law in the earlier years of the life course, transitioning towards children by their later 40s and early 50s. All three linking datasets while having different absolute rates, nonetheless describe similar trends lending additional evidence that the observed patterns are not simply random occurrences.

An important note is that the multilineal measure as applied in this paper assumes a bilateral descent structure. Bilateral descent was common, but not universal in the early twentieth-century United States. American Indian tribes practiced different descent customs from each other and other Americans. Their family structures were further forced to fit into Eurocentric versions of the family imposed by the census which directly informs the kin propinquity method applied here (Barr 2007; Johansson & Preston 1978; Shoemaker 1992). The institution of slavery affected Black family structure post-emancipation and the variety of local descent practices of enslaved persons from Africa suggests that researchers should carefully consider whether emphasizing patrilineal, matrilineal, or mutilineal kinship networks makes the most sense depending on their population of interest (Miller 2018; Miller-Cribbs and Farber 2008; Mills, Usher, & McFadden 1999; Ruggles 1994b).

Demographic conditions of fertility, mortality, and migration shaped kinship networks in terms of availability of kin, but also by the presence of kin networks. The presence of nearby kin influenced (both positively and negatively) whether and when to have children, the number of children available for older ages were determined by how many children a woman birthed and whether said children survived to adulthood (Hacker et al. 2021; Harton et al. 2023; Newson et al. 2005, 2007; Newson & Richerson 2009). Between 1850 and 1940, small but noticeable shifts in the closest patrilineal kin occurred in the United States, where individuals tended to live near parents at a higher rate (and vice versa for siblings) in 1940 relative to 1850 (Nelson 2020).

Historical kinship studies should also consider the effects of migration. Early migrants encouraged future migration to the same region by friends and family, and some immigrant communities persisted because of strong kinship ties. Those kinship ties were formed from present socioeconomic conditions, demographic processes, and historical kinship patterns from previous geographic origins (Castillo, Dribe & Helgertz 2024; Foner & Dreby 2011; Gjerde 1985; Ostergren 1988; Reséndez 2004; Wirth 2015). Given the previous examples based on race, ethnicity, fertility, mortality, and migration, researchers using the matrilineal measure in particular need to consider the limitations of the method (e.g., children’s surnames differing from the maternal surname), potential sample limitations, and the importance of each particular kinship network to their population of interest (Leopold et al. 2025). This is particularly important because of assumptions made by the census regarding family structure which affect the measured kinship networks (Smith 1992, Ruggles & Brower 2003).

This work demonstrates how to study kinship structures using linked census data, specifically for married women whose kinship networks were often lost when they changed their birth surname to the husband’s surname. Previous research has shown effects of patrilineal and matrilineal (when available in the data) kin on fertility and mortality. The United States unfortunately does not have as rich of data for kinship networks as other nations of the early twentieth-century. Previous research in the United States primarily focused on the effects of coresidential kin because of data availability, but the kin propinquity measure (and eventually direct kin identification) help account for this omitted variable. Future research on demographic and socioeconomic outcomes should consider the effects of both coresidential kin and non-coresidential kin living nearby. This linked data has the potential to revolutionize national-level analyses of women’s life courses, and this paper is just one example of how we can do that.