Extended abstract prepared for the Population Association of America Annual Meeting
Note to co-authors before submission. Values marked [verify] were read off preliminary figures and must be replaced with exact output from
tabstat/tablebefore the abstract is circulated. Means for the dating outcome are exact (computed from the frequency tables). The survey’s official English name and wave/fielding dates also need to be filled in.
Research on the social diffusion of family behavior has established that childbearing spreads among siblings, friends, and coworkers, but the evidence comes almost entirely from Western moderate-fertility settings, examines one relationship domain at a time in separate datasets, and stops at childbearing—leaving the earlier and, in East Asia, more binding transition into a romantic relationship unexamined. Using a national survey of 4,000 Korean adults that asks each respondent, in parallel, about the marital, parental, and dating status of alters in three relationship domains—kin, close friends, and organizational acquaintances—we compare exposure across strong and weak ties within the same person, allowing respondent fixed effects to absorb age, gender, and reporting-style confounding. Preliminary results show a domain reversal: exposure to marriage and childbearing is highest among strong ties, while exposure to dating is highest in the weak-tie organizational domain. We discuss implications for the network mechanisms behind Korea’s fertility trough and partial rebound.
South Korea offers the clearest contemporary case of fertility collapse and, more recently, of a puzzling partial reversal. The total fertility rate fell to 0.72 in 2023—the lowest ever recorded for any country—then rose to 0.75 in 2024 and 0.80 in 2025, with the rebound widely attributed to a 14.9% surge in marriages in 2024, the largest since records began in 1970. Because childbearing outside marriage remains rare in Korea, marriage is the proximate gate to fertility, and dating is the gate to marriage. Yet the dominant policy framework, and most Korean scholarship, treats these transitions as products of individual constraints—housing costs, employment precarity, gender-role attitudes, the cost of private education—with the social network in which decisions are made left almost entirely unmeasured.
This omission matters for two reasons. First, if family formation diffuses through interpersonal ties, then aggregate fertility can move faster in either direction than individual-level incentives alone predict, and cash-transfer policies aimed at atomized individuals will be mis-calibrated. Second, diffusion is not intrinsically pro-natalist: the same network channels that once spread marriage and childbearing can spread deferral, voluntary childlessness, and the bihon (non-marriage) identity now visible in Korean public discourse. Whether Korea’s networks are currently transmitting family formation or its refusal is an empirical question that existing Korean data cannot answer.
Diffusion as a demographic framework. Bongaarts and Watkins (1996) and Montgomery and Casterline (1996) established social learning and social influence as mechanisms in fertility transitions, and Kohler, Billari, and Ortega (2002) invoked social interaction to explain the emergence of lowest-low fertility in Europe. Balbo and Barban (2020) synthesize the micro-level literature and its identification problems.
Domain-by-domain evidence, using event-history designs. The dominant empirical template is a discrete-time hazard model of the transition to parenthood with an indicator for a recent birth in the reference network:
| Domain | Representative study | Data and design | Core finding |
|---|---|---|---|
| Siblings | Kuziemko (2006); Lyngstad & Prskawetz (2010) | PSID/NLSY and Norwegian registers; hazard models | A sibling’s birth raises the rate of entry into motherhood over the following ~3 years; effects are stronger when the sibling is older, and absent for second births |
| Friends | Balbo & Barban (2014, ASR) | Add Health friendship nominations; discrete-time event history with dyad-level random effects | A friend’s childbearing raises an individual’s risk of becoming a parent; the effect is curvilinear, peaking about two years later and then decaying |
| Coworkers | Pink, Leopold & Engelhardt (2014) | German linked employer–employee data, 33,119 women in 6,579 firms | Transition rates to first pregnancy roughly double in the year after a colleague gives birth, decaying to nil after two years; social learning is the implicated mechanism |
| Multiple domains | Buyukkececi et al. (2020, Demography) | Dutch register data, IV strategy | Both colleagues and siblings matter; effects spill over across network boundaries; colleague effects concentrate in female–female pairs |
| Mechanisms | Lois & Becker (2014) | German panel | Observational learning, perceived pressure, and anticipated loss of ties all mediate the network–fertility association, with effects declining by age |
The “knowing someone” design. Closest to our data, Rindfuss, Choe, Bumpass, and Tsuya (2004, ASR) asked Japanese respondents whether they knew anyone who had cohabited, borne a child outside marriage, used childcare, or intended never to marry, separately by relationship domain (sibling, other relative, friend, coworker). They found such knowledge to be widespread and patterned by domain and by social structural position. Piotrowski et al. (2022) extended this with two waves and respondent fixed effects, linking “knowing” to attitudes and emphasizing gender differences.
Theoretical expectations about tie strength. Granovetter’s (1973) argument that weak ties bridge otherwise disconnected clusters implies that novel information—including information about unfamiliar family arrangements, and access to unmet potential partners—arrives disproportionately through weak ties. Centola and Macy (2007) qualify this: behaviors requiring costly commitment are complex contagions needing reinforcement from multiple sources, which strong, clustered ties supply and long bridges do not. Applied to family formation, the two arguments make opposite predictions for different transitions, and no study has tested them side by side.
Three gaps.
We ask three questions.
Contributions. (a) The first systematic test of family-formation diffusion in the world’s lowest-fertility society, at the dating margin as well as the marriage and fertility margins. (b) A within-person design: because the same battery is administered for three domains, respondent fixed effects absorb age, gender, region, own family status, and individual reporting style—the confounds that dominate cross-sectional network studies—so the strong/weak comparison is identified off within-person differences across domains. (c) A composition measure (opposite-sex share by domain) that lets us distinguish an opportunity-structure account of dating from a normative-influence account. (d) An explicit treatment of negative diffusion: in a setting where fertility fell 90% in two generations, the transmitted content may be deferral rather than adoption.
Data. The third wave of the national survey of Korean attitudes toward dating, marriage, and childbearing [fill in official name, fielding dates, sampling design, weights]. N = 4,000 adults.
Network structure (ego level). - Strong-tie size: number of friends or acquaintances with whom the respondent routinely shares news and feelings (open-ended count, 0–99). Mean 3.54 (SD 3.37), median 3; 7.7% report zero; heaping at 5 and 10; top-coded at 20 for analysis (0.5% of cases). - Weak-tie breadth: extent of acquaintance within (i) workplace/school and (ii) hobby, alumni, or religious organizations, on a four-point scale from “no affiliation or no participation” to “acquainted with most members.” 16.3% report no workplace/school affiliation or activity; 37.5% report none for voluntary organizations; 11.7% report neither (these respondents are skipped out of the organizational-domain items). - Kin size: number of siblings.
Exposure (domain level, repeated three times per respondent). For each of three domains—(1) siblings and close cousins [strong, ascribed], (2) coworkers, classmates, and organization members [weak, organizational], (3) close friends and acquaintances [strong, chosen]—respondents reported the share of alters who are married, the share who have children, the share of unmarried alters currently dating, and the sex composition. We reverse-code each to a 0–2 exposure score (higher = more of the behavior present in that domain) and construct a binary “most/many” indicator. Sex composition is recoded relative to the respondent’s own sex into an opposite-sex-share score (0–4).
Structural missingness. Domain items were asked conditionally: the friend domain only of respondents with at least one confidant, the organizational domain only of those with some organizational activity, the kin domain only of those with kin. Missingness is therefore informative—it identifies the isolated—and we analyze it as an outcome in its own right as well as conditioning on tie presence.
Stage 1 — Within-person domain comparison. We reshape to person–domain format (N ≈ 12,000 domain observations from 4,000 respondents) and estimate
\[Y_{id} = \alpha_i + \beta_2 \text{Weak}_{d} + \beta_3 \text{Friend}_{d} + \mathbf{X}_i'\boldsymbol{\gamma}_d + \varepsilon_{id}\]
with respondent fixed effects \(\alpha_i\) and standard errors clustered on the respondent; \(\beta\) identifies the strong/weak contrast net of all time-invariant ego characteristics. Ordinal analogues use fixed-effects ordered logit (Baetschmann et al.); binary analogues use conditional logit. Because ego covariates are absorbed, heterogeneity enters as domain × covariate interactions (\(\boldsymbol{\gamma}_d\)), which is where the substantively interesting variation lies: whether the strong/weak gap in dating exposure is larger for younger, unmarried, or male respondents.
Stage 2 — Ego-level models of own status. For the
respondent’s own dating status (unmarried subsample), ever-married
status, and parity, we estimate logit and ordered logit models on
strong-tie size, weak-tie breadth, domain-specific exposure, and
opposite-sex composition, controlling for age (quadratic), gender,
education, employment, income, region, and coresidence with parents.
Cross-equation coefficient tests (suest) formally assess
whether weak ties matter more for dating than for marriage and
fertility—the paper’s central claim.
Identification. We do not claim causal contagion. Manski’s (1993) reflection problem is unresolved in cross-sectional ego-reported data, and homophily and shared context are plausible alternatives for any ego-level association. Stage 1 is nonetheless informative in a way that ego-level correlations are not: the domain contrast is a within-person comparison, so any explanation must be domain-specific rather than person-specific. We frame Stage 2 associations as exposure patterns and reserve causal language for the wave-4 panel design described in Section 7.
Exposure differs sharply by domain, and the ranking reverses across outcomes (Figure 2; person-fixed-effects tests via Wilcoxon signed-rank on the 0–2 scores):
| Domain | Marriage exposure | Childbearing exposure | Dating exposure |
|---|---|---|---|
| (1) Kin — strong, ascribed | 1.21 [verify] | 1.25 [verify] | 0.955 |
| (2) Organization — weak | 1.13 [verify] | 1.20 [verify] | 1.088 |
| (3) Friends — strong, chosen | 1.09 [verify] | 1.09 [verify] | 1.005 |
Marriage and childbearing exposure are highest among kin and lowest among chosen friends. Dating exposure inverts this: it is highest in the weak-tie organizational domain and lowest among kin. Equivalently, among respondents with alters in each domain, the share reporting that “several” unmarried alters are currently dating is 27.2% for organizational ties, 25.4% for friends, and 21.8% for kin. The reversal is consistent with the complex-contagion/opportunity-structure split we hypothesized: commitment-heavy transitions cluster where ties are dense and obligations are ascribed, while the low-commitment, opportunity-dependent transition tracks the bridging domain.
An age-composition alternative must be taken seriously—siblings and cousins are on average older than organizational contacts, mechanically raising kin marriage and parenthood exposure. Two features of the design speak against it as a full account. First, it cannot explain the reversal: if kin were simply older, kin dating exposure would be lower for the same reason kin marriage exposure is higher, but then friends (age-homophilous, mostly unmarried in the younger half of the sample) should show the highest dating exposure, and they do not. Second, the wave-4 instrument (Section 7) adds alter ages, which will let us test this directly.
Isolation is common and asymmetric. 7.7% of adults report no confidants at all, and 37.5% report no voluntary-organization activity; 11.7% report neither workplace/school nor organizational engagement. Because the weak-tie domain is where dating exposure is highest, the population effectively excluded from that domain is the population most plausibly excluded from partner-search opportunity—a concentration of disadvantage invisible in aggregate marriage statistics.
Opposite-sex composition matters, but only for chosen ties (Figure 3). Among friend networks, dating exposure is higher where the network is opposite-sex-heavy (0.92 vs. 0.84); among kin the association reverses sharply (0.71 vs. 0.89), and among organizational ties it is flat to slightly negative (0.90 vs. 0.95). The kin reversal is mechanically interpretable (opposite-sex-heavy kin networks in a low-fertility society tend to be small), which is precisely why the wave-4 module must collect kin-network size and age structure. The friend result is the theoretically live one: chosen, mixed-sex networks are where partner search appears to operate.
Analysis in progress: Stage 1 fixed-effects models with domain × cohort and domain × gender interactions; Stage 2 models of own dating, marriage, and parity; cross-equation tests of the weak-tie/dating hypothesis; and sensitivity analyses treating structural missingness as a substantive category.
The present measures cap what can be claimed in three ways, and each has a concrete remedy in the fourth wave, currently in design: (a) exposure is measured ordinally with no denominator, so we cannot separate “many alters, few married” from “few alters, most married”—wave 4 will collect domain-specific alter counts; (b) exposure is contemporaneous, so recent-event contagion of the kind identified in the hazard-model literature cannot be detected—wave 4 will add 12-month event recall (“how many people in this domain married / had a child in the past year”); and (c) only positive content is measured, so negative diffusion is unobservable—wave 4 adds parallel items on hearing about the difficulties of marriage and parenting. A compact ego-network module (up to five alters with age, sex, domain, contact frequency, family status, and alter–alter acquaintance) will convert the instrument from a perception battery into an ego-network dataset while adding roughly four minutes of interview time.
Balbo, N., & Barban, N. (2014). Does fertility behavior spread among friends? American Sociological Review, 79(3), 412–431.
Balbo, N., & Barban, N. (2020). Social contagion effects in fertility: Theory and analytical strategy. In Analytical Family Demography. Springer.
Bongaarts, J., & Watkins, S. C. (1996). Social interactions and contemporary fertility transitions. Population and Development Review, 22(4), 639–682.
Buyukkececi, Z., Leopold, T., van Gaalen, R., & Engelhardt, H. (2020). Family, firms, and fertility: A study of social interaction effects. Demography, 57(1), 243–266.
Centola, D., & Macy, M. (2007). Complex contagions and the weakness of long ties. American Journal of Sociology, 113(3), 702–734.
Granovetter, M. S. (1973). The strength of weak ties. American Journal of Sociology, 78(6), 1360–1380.
Kohler, H.-P., Billari, F. C., & Ortega, J. A. (2002). The emergence of lowest-low fertility in Europe during the 1990s. Population and Development Review, 28(4), 641–680.
Kuziemko, I. (2006). Is having babies contagious? Estimating fertility peer effects between siblings. Unpublished manuscript.
Lois, D., & Arránz Becker, O. (2014). Is fertility contagious? Using panel data to disentangle mechanisms of social network influences on fertility decisions. Advances in Life Course Research, 21, 123–134.
Lyngstad, T. H., & Prskawetz, A. (2010). Do siblings’ fertility decisions influence each other? Demography, 47(4), 923–934.
Manski, C. F. (1993). Identification of endogenous social effects: The reflection problem. Review of Economic Studies, 60(3), 531–542.
Montgomery, M. R., & Casterline, J. B. (1996). Social learning, social influence, and new models of fertility. Population and Development Review, 22(Suppl.), 151–175.
Pink, S., Leopold, T., & Engelhardt, H. (2014). Fertility and social interaction at the workplace: Does childbearing spread among colleagues? Advances in Life Course Research, 21, 113–122.
Piotrowski, M., Rindfuss, R. R., Tamaki, E., Choe, M. K., Tsuya, N. O., & Bumpass, L. (2022). Social networks and family change in Japan revisited. Population, 77(4).
Raymo, J. M., Park, H., Xie, Y., & Yeung, W.-J. J. (2015). Marriage and family in East Asia: Continuity and change. Annual Review of Sociology, 41, 471–492.
Rindfuss, R. R., Choe, M. K., Bumpass, L. L., & Tsuya, N. O. (2004). Social networks and family change in Japan. American Sociological Review, 69(6), 838–861.
Rosenfeld, M. J., & Thomas, R. J. (2012). Searching for a mate: The rise of the Internet as a social intermediary. American Sociological Review, 77(4), 523–547.
Figure 1. Distribution of strong-tie size (number of confidants), top-coded at 20. Figure 2. Mean exposure to marriage, childbearing, and dating by relationship domain. Figure 3. Dating exposure by opposite-sex composition, within relationship domain.