Kinship networks & health: theory & modeling advances

Ashton M. Verdery

Duke Network Analysis Center

May, 2025

Outline

• Social networks & health
• What network?
• Importance of family & kinship networks
• Survey data, strengths & limits
• Key gaps in current knowledge
• Less used tools for kin networks
• Microsimulation
• Analytic approaches
• Resources for kin network analysis
• Microsimulation in R
• Analytic methods in R
• Conclusions

Social networks & health – what network?

Networks literature heavily emphasizes friendship ties

• E.g., Web of Science
• ~75k articles w/ social networks in title, abstract, keywords
• ~6k (8%) also have family/kinship
• E.g., Social Networks journal
• 1,371 articles ever published
• 122 (9%) also have family/kinship
• Growth since 1990s, now steady

Searches conducted 5/4/2022: WoS total: https://www.webofscience.com/wos/woscc/summary/114acb43-cfa0-40f1-bd04-fc9f9e39867e-35891e86/relevance/1; WoS refined: https://www.webofscience.com/wos/woscc/summary/74944c8d-fa61-47c9-a861-ad867aa1bcde-35892aff/relevance/1; WoS counts for SN journal from https://www.webofscience.com/wos/woscc/summary/62268c7f-a6a7-48fa-8fbb-5545ad1490fc-3589cb01/relevance/1 and https://www.webofscience.com/wos/woscc/summary/2e215662-52f1-4e3a-ad16-12ac19c495ea-35a781c3/relevance/1

Driving with a blindfold

Family ties are central to of many well-known network findings, especially those relevant to health, but we rarely focus on them

Health

Search for Abortionist

• 50% of searchers asked family members for help

Lee 1969

“Social Isolation in America”

• 69% of US nominates more kin than non-kin

McPherson et al. 2006

Framingham Heart Study

• 55% of ties in this network are family ties

Christakis & Fowler 2007

Not Health

“Small World Networks”

• 14% of letters were sent to family members

Travers & Milgram 1969

The Strength of Weak Ties / Finding a Job

• 31% of jobs found through family

Granovetter 1974

Christakis and Fowler 2007, NEJM, “Spread of obesity”

55% of ties in this network are family/kin

Social networks & the life course

General Social Survey; Verdery calculations

Who do we talk to?

“who are the people with whom you discussed matters important to you?”

Verdery graph; data from American Time Use Survey analysis by Our World in Data: https://ourworldindata.org/grapher/time-spent-with-relationships-by-age-us?country=~USA

Who do we spend time with?

Results from time use diaries

What do we know about kin networks?�Kinship as culture

What else?

• Not a lot, other possible insights are obscured
• Reasons:
• Theoretical focus on social networks/“peer effects”
• Split between “family” & “social networks” literatures
• Household-based nature of data collection
• An assumption that (non-household) family ties are less relevant in advanced industrial economies
• What we look at is constrained by data available, which constrains what we look at, which constrains…

​

Ashton M. Verdery

7

Some questions we can answer with survey data / interviews, etc.

Example: Is COVID Bereavement Worse?

Leveraging a large survey in Europe that COVID interrupted, we apply a difference-in-difference analysis comparing recently widowed people who lost spouses before the COVID outbreak to those who lose spouses to COVID in order to test whether COVID bereavement is worse for mental health.

Wang, Verdery, Smith-Greenaway, Margolis, Bauldry. 2022. JGSS.

Example of extended kin importance

Ashton M. Verdery

9

Emerging understandings of kin networks influence on social life, example:

Ashton M. Verdery

10

Ashton M. Verdery

11

Survey data are inadequate

Ashton M. Verdery

12

Kinship Demography

Entering a golden age

• Public interest surging
• Rapid population aging
• Growth reversals/decline
• Disinvestment / underinvestment in public goods (e.g., pensions)
• New techniques available
• More computation for microsimulation
• New formal techniques
• Accessible versions of both
• More empirical estimates

Articles per year with “kinship” and “demography”

in the title, abstract, or keywords

in the Web of Science™ database

3 broad approach for kinship networks

• Estimate with empirical data (already covered)
• Surveys
• Administrative data
• Simulate the network
• Have agents do things (be born, marry, have children, divorce, die, etc.)
• Track kinship networks as they evolve
• Analytical tools
• A series of analytic methods for estimating expected numbers of different kinds of relations (examples)

Ashton M. Verdery

14

Demographic microsimulation

Start with individuals of various ages & other attributes, they get married (to people in simulation or new entrants), or give birth, or die or emigrate at time-varying demographic rates

Example of microsimulation

View of traditional societies as dominated by extended kinship networks, e.g., China below

We simulate China’s demographic history from 1700-2100, decently matching historical & projected rates

Ashton M. Verdery

16

17

but the downslope will be brutal

Verdery 2019 “Modeling the future of China’s changing family structure”; Eberstadt & Verdery 2021 “China’s Shrinking Families” Foreign Affairs; Eberstadt & Verdery 2022, AEI Press

Contra assumptions, peak family in China is right about now… not in the past

Thinking through the structural & relational architecture of China’s kinship & its diffusion implications

• This presages a substantial and growing caregiving & elder support challenge
• Filial burdens in China growing for middle-aged groups, who are increasingly less likely to have siblings and other relatives with whom to share the load

Ebserstadt and Verdery 2023. monograph.

Not just China, a general feature

Demographic transitions produce connectivity bursts that differ in size, duration, & period of onset but always happens (b/c mortality decline)

Verdery, 2015. Links between demographic & kinship transitions. PDR.

Other work replicates these ideas using new mathematical demography tools

Ashton M. Verdery

20

Similar underlying finding when focused on the future, though they see less of a post-1950 bump and a peak about 20 years earlier because they assume 1950 rates are constant (cumulating into more surviving) whereas we use 1900-1950 rates, which had higher mortality pressures.

Advancing Caregiving Research

• Evidence for a U.S. “caregiver crisis” is based on a decline in the ratio of adults aged 45-64 (potential caregivers) to adults aged 80+ (those requiring care)

​

• Kinship models allow us to go beyond age group ratios and examine how family structure, health needs, & other aspects of population change will contribute to changes in caregiving availability and demands

Advancing Caregiving Research

Data: Verdery and Margolis, PNAS, 2017; Chart: Lazaro Gamio / Axios

Wu, Margolis, Patterson, Verdery.. 2024. Demography.

• Take family associations with caregaps from HRS (by race, health, age, sex, etc.)
• Take demographic projected changes along the same categories (family, race, age, etc.)
• Estimate expected changes in care gaps if associations stay constant

Early pandemic, there were broad perceptions that COVID-19 was killing only those “who were going to die anyway” and “who cares about them?”�

"It affects elderly people, elderly people with heart problems, if they have other problems, that's what it really affects, that's it.“

Donald Trump, 9/21/2020

Secondarily affected populations

& bereavement considerations

from kinship network modeling

Other data fusion approaches

About 9 people lose a close family member for each COVID-19 death

• 4.0 lose a grandparent
• 2.2 lose a parent
• 2.0 lose a sibling
• 0.5 lose a spouse
• 0.2 lose a child

People of all ages suffer losses, but of different types

Interpretation:

Each death leaves about 0.75

siblings ages 60-69 bereaved

Simulate COVID deaths on top of projected network consistent with infection fatality rates and different infection scenarios

Some impacts

Ashton M. Verdery

25

Size & scope of the U.S. population secondarily affected by COVID deaths

10+ million bereaved in U.S. as of May 2023, means

• ~4.5 million lost a grandparent
• ~2.5 million lost a parent
• ~2.2 million lost a sibling
• ~560 thousand lost a spouse
• ~224 thousand lost a child

​

Other uses: COVID deaths among parents of children 0-18 in U.S.

• First to recognize the huge number of children losing parents to this disease; changed national narrative

https://interactive.guim.co.uk/uploader/embed/2022/03/covid-family-chart/giv-825XfYE2J0DNkou/main-chart-inArticle_620.png

Benefits & Downsides of Microsimulation

• Benefits:
• Long-run perspective, including looking at future
• Macro-scale, whole populations, whole kin sets
• Readily mergeable with agent-based models, or survey estimates via data fusion approaches, etc.
• Can give us some bounds around empirical N=1
• Downsides
• Hard to implement in software (not true anymore)
• Data requirements are onerous (long run rates)
• Needs ground-truthing & assumption verification / the data are made up, which leads to uncertainty
• Not readily computable/reportable statistics

Old Models for Kinship Estimates

Keyfitz & Caswell 2005

Old method calculations:

Expected numbers of female kin alive in a stable population for 40-year-old woman using 1965 U.S. rates

Example Application�A recent debate

Application:�How common is bereavement from drug overdoses?

Ashton M. Verdery

Which got me thinking

• 40% of U.S. Adults, but what about children?
• But this is very difficult to measure for children
• Few surveys of children
• No internet panels, for example
• Unlikely to be friends
• Age homophily, most overdoses are among people age 35+
• Having a relative die is rare
• Let alone to a specific cause
• How would you ask them?

“America’s forgotten orphans.”Mulheron, Chapman, Smith-Greenaway, Verdery, 2022.

In recent work, I’ve been trying to understand how many children are losing relatives, such as co-residential parents, with data fusion techniques between ACS and vital statistics.

Someone else had a similar idea

• Took data on co-residential children from the National Survey on Drug Use and Health
• Merged Vital Records data on age-, sex-, etc. pattern of overdose deaths
• Estimated 321,566 children lost a co-residential parent over the 2011-2021 period

34

How many children’s lives are disrupted by overdose deaths?

• We used formal demography of kinship models to estimate children losing close relatives.
• An estimated 1.2 million children under 18 in 2020 had lost a close relative to overdose (1.6%)
• Note: period prevalence
• 2.5 times more likely to lose male than female kin
• More youth age 10-17 (2.3%) than children age 0-9(0.9%)

​

Verdery, Ryan-Claytor, Smith-Greenaway, Sarkar, Livings. In review April 3, 2024.

Worsening trends

• Our estimate is a period prevalence:
• How many kids alive and under 18 in 2020 had lost a relative?
• But that elides different aging trajectories for children born in different years
• We see substantial cohort increases as they age, suggesting these numbers will grow in coming years
• Most of the growth in relative death owes to overdose death increases

Ashton M. Verdery

36

We can go further, to answer the NY Times question about “deaths of despair” & burying a child, with attention to race and ethnic disparities

Ashton M. Verdery

Probability*1000 of losing a son or daughter to suicide by age 70, by race & year

Resources

Ashton M. Verdery

38

Diego Alburez-Gutierrez, Ivan Willi, et al., Max Planck Institute for Demographic Research

https://alburez.me/DemoKin_example/

https://github.com/alburezg/rsocsim_workshop_paa#1-setup

Resources

• This is the one for microsimulations in R
• Still in devtools
• Hope to have a workbook for 2024 SN&H workshop
• Workshop listed below at April PAA 2023 is good

Ashton M. Verdery

39

Diego Alburez-Gutierrez, Ivan Willi, et al., Max Planck Institute for Demographic Research

https://alburez.me/DemoKin_example/

https://github.com/alburezg/rsocsim_workshop_paa#1-setup

Resources

• This is the one for analytic methods in R (more being added, e.g., bereavement)
• Package 1.0 available
• Github listed below has a walkthrough use case

Ashton M. Verdery

40

Diego Alburez-Gutierrez, Ivan Willi, et al., Max Planck Institute for Demographic Research

https://alburez.me/DemoKin_example/

https://github.com/alburezg/rsocsim_workshop_paa#1-setup

Remembering this

Key distinctions

• Focus on living kin or dead?
• What are you trying to estimate?
• Stable or time-varying rates?
• Do you want to show a hypothetical implication (stable) or try to produce an actually applicable estimate (varying)?
• Interested in cohorts?
• Whether to use one or two sexes of relatives?
• Can assume male+female = female*2 (or 4, etc.), but does not work well when thinking of mortality or bereavement
• Men die younger and of different causes than women
• Transitions between groups?
• Can model groups separately (e.g., race groups) but may want to let people transition (e.g., attain education, lose employment, migrate between places)

​

Example R code: China Sex Ratios

• China’s sex ratios at birth have been highly skewed in recent years
• Kinship exchange in China is thought to be highly gendered
• i.e., sons care for parents norm, etc.
• How much do skewed sex-ratios at birth affect numbers of living kin?

Ashton M. Verdery

43

China Sex Male and Female Kin�Normal and Skewed Sex Ratios

Ashton M. Verdery

44

Example: American bereavement experiences�

Ashton M. Verdery

45

Piecewise cubic Hermite interpolation�From abridged to single year fertility data

• Take the 5 year ASFRs, divide by 5
• Create a cumulative function
• Get proportion of maximum in each year
• Generate logit of proportion of maximum
• Keep the category end points
• Replace minimum and maximum ages of childbearing with –inf/inf for min/max ages (use logit=-20, logit=12)
• Interpolate with piecewise cubic Hermit spline methods
• Set infinite logits to missing
• Convert back to fertility with inverse logits
• Decumulate to get single year rates

2000 National

2020 Each state

Details buried in methodological appendices to the Human Fertility Database & related reports

https://www.demogr.mpg.de/papers/working/wp-2018-001.pdfeq4.11 in https://www.humanfertility.org/File/GetDocumentFree/Docs/methods.pdf

Interpolating death and birth rates

• CDC WONDER suppresses death and birth counts in any cell where there are:
• More than zero
• But less than 10 deaths/births
• I interpolate these suppressed counts using multiple imputation interval regression, which constrains the predicted number to be between 1 and 9
• deaths = age + age^2 + state + sex + year + ln(population in cell)
• I then rescale the predicted values to be a proportion of the sum of the predicted values, which I multiply by the total number of suppressed deaths (so the total is equal)

​

• Example
• there are 54,410,528 in the national, unsuppressed data
• there are 54,255,260 unsuppressed deaths across all states from 2000-2020 (99.7% coverage)
• Meaning 155,268 deaths are suppressed (0.3%)
• Interval imputation yields 54,425,571 deaths, so a slight overestimate (1.0003% coverage)
• Rescaling slightly down weights the imputed counts to match the totals

Run through R Code Example

Ashton M. Verdery

48

Modeling Bereavement

Ashton M. Verdery

49

Bereavement Expectancy: Years till losing father

• Substantial variation across states
• 60 years in HI
• 54 in MS & WV
• By age 25
• 7% chance
• 13% chance

​

​

Ashton M. Verdery

50

The expectancy framework helps us understand trends:�Period-based estimates of likelihood a newborn child loses a parent to major causes of death, 2000-2019

We can also compare amount of time bereaved because of specific causes

e.g., Overdose has become the biggest single cause through which people spend time bereaved of a son

Can reveal new funding priorities & areas in need of attention

• Substantial attention is paid to scientific funding priorities for different diseases
• NIH tracks funding levels for specific diseases against classic population health metrics
• Uses this to diagnose differential investment in diseases that primarily harm men vs. women

​

• Another lens: examine spending against bereavement expectancy metrics
• How much are we spending on the diseases that kill our family members?

Verdery calculations from NIH Report on spending vs. burden of disease. https://report.nih.gov/report-nih-funding-vs-global-burden-disease

U.S. Deaths vs. spending

U.S. DALYs vs. spending

Conclusions

• Social networks literature emphasizes friend, peer, & acquaintance networks
• Exciting & have the most tools available for analyzing them
• Family networks are also important, especially at some developmental stages
• Clear link to classic & contemporary theories & networks & health framework
• Many important questions can be asked at the intersection of health and family & kinship networks
• Data are there, approaches are not lacking, & new tools now out

​

The importance of family & kinship networks for population health is only going to grow

​

THANK YOU

amv5430@psu.edu

Ashton M. Verdery

55

What do we know about these family & kin ties?

Anthropological representations of kinship networks have focused on novel representations that allow different sorts of network computations

A framework for family & kin networks

There’s huge value in structural anthropology / network anthropology literature on kinship, but I hope to take a step back & tie kinship more closely to key social networks & health considerations

Ashton M. Verdery

58

A framework for family & kin networks�Some Topics/Questions & Methods

Structural anthropology work

How do kin sets look?

Ethnography, genealogy, surveys

Demographic microsimulation

Societal caregiving demands

Secondarily affected population size

Survey analysis, qualitative methods

Agent Based Models

Person-level kin availability

Access to kin resources (e.g., wealth)

Surveys, genealogy, data fusion

Formal demography of kinship

Social support, caregiving

Bereavement

Survey analysis, demographic simulation

Formal demography of kinship