Page 24 - IJPS-4-1

Transition to first marriage in China

background factors such as hukou status, father’s education, and employment conditions, as well as regional dummies.
Model 3 introduces an interaction term between hukou and gender.
We can glean the following points from this analysis. First, compared with the youngest cohort (born 1976–1985),
older cohorts had a lower risk (meaning starting later) of entering marriage (e.g., by 17% for the Cultural Revolution
cohort), if other things being equal; males had a lower risk of getting married than females by over 30%; higher educational
attainment was correlated with lower hazard ratios. Second, urban hukou respondents had a lower risk of entering marriage
by 20% or more; fathers’ higher education had similar effects with respondents’ own educational attainment on marriage
timing.In terms of father’s employment status when respondents were at age 14, compared with those whose fathers
were farmers (reference group), those whose fathers were full-time employed had a lower hazard of entering marriage
by 6% and those whose fathers’ employment status categorized as “others” had a lower hazard by 11%. In combination,
we observed that urban hukou, father’s relatively higher education, and stable employment were correlated with youths’
lower risk of marriage suggesting some level of modernization of urban middle-class families in their marriage formation
patterns. Warranting future inquiry, the puzzle of the significant lower marriage risk of those with fathers categorized as
“others” (mostly the marginalized workforce in urban labor market) in employment status, relative to farmers’ children,
may be related to growing economic pressure for the urban poor to establish a household in cities (e.g., Choi and Peng’s
[2016] analysis of the undesirable plight of second-generation rural migrant men in urban marriage market).

5.3. Trends of educational homogamy

Table 4 describes the educational homogamy pattern in general and by marriage cohorts. The row variable represents
husband’s education, and the column variable represents wife’s education. The cells on the main diagonal display
homogamy statistics. For the whole analytic sample, if summing all the percentages on the main diagonal, we see 56.6%
of marriage unions with identical educational levels. Statistics in lower rows of Table 4 and Figure 2 describe educational
homogamy patterns across marriage cohorts. As seen, about 60% of marriages contracted during Cultural Revolution era
constitute homogamy; this number dropped to 52% in early reform cohort and kept increasing thereafter. The share of
educational homogamy among the late reform cohort (married during 2001–2015) reaches 63%. The general scenario
seems to suggest that the economic reform in its early era has a liberalization effect on respondents’ selection of spouse,
as reflected in the steep decrease of homogamy proportion in this marriage cohort.
However, as Xu et al. (2014) demonstrated, such reliance on raw statistics from the contingency table (with 3*3*4 cells)
could yield misleading results, because China’s educational structure has undergone considerable transformations during
the past four decades, which may affect the opportunity structures for homogamy. For example, the very high homogamy
statistics for the Cultural Revolution cohort might be a result of the marginal distribution of wives’ and husbands’ education
(46% of marriages were between low-educated men and women).
To get more accurate estimation, I fit a series of log-linear models: The conditional independence model (Model 1)
as the baseline model and five log-multiplicative layer effect models (Models 2-6) with different design matrices for the
Row-Column association. Table 5 presents the goodness-of-fit statistics for each model, including the degree of freedom,
the log-likelihood ratio Chi-square statistic (L ), the Bayesian Information Criterion, and the index of dissimilarity (D).
2
According to multiple criteria, the log-multiplicative layer effect full interaction model (model 6) fits the data best (L
2
is 81.3; DF = 23; BIC is −146.9; Δ = 1.8, indicating that <2% of observations need to be reclassified to make all observed
cell counts exactly equal to the fitted cell counts). Models 1 and 2 are ill-fitted to test the two hypotheses: (1) There is no
association between husband’s education and wife’s education in each cohort and (2) there is no relative association of
educational attainments among heterogamous pairings. Likewise, Models 3 and 4, despite relative improvement of model fit
by L and Δ scores, are still less satisfactory to support their hypotheses. The quasi-symmetry model (Model 5) significantly
2
improves by all four goodness-of-fit measures. For example, the dissimilarity index shows that only 1.9% of all cases were
"Wrongly" classified under this model. However, Model 6 is preferable over Model 5 according to L and Δ scores. For
2
an additional three degrees of freedom, the L reduced by 14.7 (P < 0.01). In other words, not only is there a significant
2
association between husband’s and Wife’s educational levels but also their homogamy patterns vary significantly by cohort.
Figure 3 shows the change of the Ф parameters (denoting the cohort-specific strength of association between husband’s
and wife’s education, with ±1 indicating perfect agreement or disagreement and 0 indicating no relationship) based on
estimates from Model 6 in Table 5, which serve as the index of the strength of educational homogamy. Very different
from Figure 2, we see steady growths of the strength of homogamy from the Cultural Revolution cohort on, from 0.42 to
0.56 for the most recent cohort. This finding is consistent with some previous research which documented higher levels
of homogamy as China’s economic development accelerates after reform (Han 2010; Xu, Ji and Tung, 2010; Xu, Li and
Yu, 2014).

18 International Journal of Population Studies | 2018, Volume 4, Issue 1