Explaining the Effect of Parent-Child Coresidence on Marriage Formation: The Case of Japan

One exception is Tsuya and colleagues’ (2004) analysis of the effect of coresidence on marriage desires among single men and women in Japan. Their use of cross-sectional data, however, makes it impossible to rule out the possibility that selection into residential independence explains the observed between-group differences.

When respondents were not interviewed at a wave, we have no way of telling their marital status unless they reentered the survey at a later wave, which allows us to retrospectively fill in their marital histories. If respondents became married in the same year they missed the survey, then the lack of information for that wave should hardly affect our analysis given that the person-year observations would have been excluded from the fixed-effects models anyway. The 15 % reported here is based on the assumption that respondents remained never-married during all the missing years, if they were never married at the last complete interview. Because attrition as a result of entering marriage is likely (as indicated by an analysis of respondents who missed a wave and then reentered the survey), this assumption is likely to lead to an overestimation of the amount of eligible person-year observations that are missing.

The models also include respondents’ father’s education and their own schooling, city size, employment status, work hours, and commute time.

We also conducted an additional analysis comparing those reentering the survey after missing a wave with those with all five waves completed. We found that the association between coresiding with parents and marriage chances was in the same direction and was stronger for those reentering the survey. If those missing some interviews and not reentering by Wave 5 were similar to those who did reenter, then excluding all person-years with no interviews should only weaken the association between coresidence and marrying found in our models. In this sense, the survey attrition leads to more conservative estimates, rather than distorting the results.

In addition to age, the logistic regressions also included father’s education; respondent’s educational attainment; current school enrollment; sibship size; whether the respondent’s first job was a full-time, regular one; size of the city in which the respondent lived; and whether the respondent had ever cohabited. Other than being older, living in smaller cities (or rural areas) and having cohabited before were the two variables that increased both men’s and women’s likelihood to have entered first marriage before the first wave. Women who were more educated and still in school were also more likely to be never-married, whereas men from smaller families and whose first job was a temporary or part-time one were more likely to remain never-married. In a further analysis, we tested whether the associations between coresidence and various outcomes differed by sibship size, first job’s employment status, and city size for men and schooling and city size for women; we found no significant differences, suggesting that our underselection of men and women of certain demographic characteristics might not alter the results in a meaningful way. Moreover, although we have no information about whether those who had been married lived with their parents before marriage, the finding that this group was more likely to have had cohabitation experience suggests that the rate of independent living was higher among this group. If we consider those ever-married at Wave 1 as having a faster pace of entering a romantic relationship and first marriage, then the associations between premarital living arrangements and entry to relationship and marriage for this group would be similar to those reported in the Results section, based on what we can infer about their living arrangements. This similarity should further alleviate concern about the data limitation.

We show a full list of the reasons provided in the selected waves of JLPS in the Results section, when we present the fixed-effects models predicting individuals’ probabilities of selecting each of the reasons.

Although linear probability models can be more intuitively interpreted, they may encounter the problem of predicted probabilities falling out of the 0–1 range. In our additional checks, however, we found no such problem for the fixed-effects linear probably models presented in this article.

We used two user-written programs for Stata, bucologit (version 11.2) and feologic_buc (version 10), to implement fixed-effects ordered logit models with the “blow-up and cut” estimator. Both programs yielded results similar to ours from fixed-effects linear regression models.

We also treat 27 person-year observations for those who lived with a grandparent, but no parent, as living in the parental home, with the assumption that the grandparents serve functions similar to those of parents. The results were virtually the same when we coded these observations as not living with parents.

An additional analysis estimating variance inflation factors indicated that including both age and school enrollment does not cause excessive multicollinearity.

The reason for coding the job characteristics of the jobless as 0 can be illustrated with a simplified equation: M = b ₀ + b _1i X _i + [b ₂ + b _3j (Y _j – \( \overline{Y} \) _j )]E, where M represents marriage intentions; X _i represents i predictors relevant to both employed and nonemployed people; (Y _j – \( \overline{Y} \) _j ) indicates j job characteristics centered on the group means; and E indicates whether observations are employed (coded as 1) or not (coded as 0). When an observation is not employed, M = b ₀ + b ₁ X _i , whereas M = b ₀ + b ₁ X _i + [b ₂ + b _3j (Y _j – \( \overline{Y} \) _j )] when an observation has a job. When an observation has average job characteristics (Y _j = \( \overline{Y} \) _j ), the difference between this observation and one without a job is b _2. Compared with average workers, a one-unit increase in work hours or commuting time further contributes to b _3j amount of change in marriage intention. The original equation can also be written as M = b ₀ + b _1i X _i + b ₂ E + b _3j (Y _j – \( \overline{Y} \) _j )E, where b _3j (Y – \( \overline{Y} \) _j )E = 0 when an observation is jobless. By coding mean-centered job characteristics (i.e., [Y _j – \( \overline{Y} \) _j ]) as 0 for the jobless, we can further modify the equation to be M = b ₀ + b _1i X _i + b ₂ E + b _3j (Y _j – \( \overline{Y} \) _j ), as b _3j (Y _j – \( \overline{Y} \) _j ) would be 0 when one is jobless. In this study, we use a set of dummy variables for employment status, rather than just having a job or not, but the employment status dummy variables serve the same function as E in the illustrated equation.

Specifically, we used father’s education, a combination of father’s and mother’s education, or father’s occupation when respondents were 15 years old to approximate the economic advantages or disadvantages of respondents’ family of origin. For sibship characteristics, we tried sibship size, birth order rank, and whether respondents were the first or only son or daughter in the family.

We also considered the possibility that engaging in formal and informal partner-seeking activities has different implications for marriage transitions for those with and without romantic partners. An additional analysis including interaction terms between involvement in partner-seeking activities and relationship status, however, showed no significant interaction effects.

We contend that the comfort of the parental home, if it decreases one’s incentive to marry, as expected by the theory, should make young adults express a lower level of eagerness to marry. One may argue, however, that singles living with parents simply want to postpone marriage but still aspire to marry eventually. Because the question from Wave 1 of the JLPS asked respondents the age by which they expected to marry, we were able to conduct an ancillary analysis examining the association between coresidence and anticipated timing of entering first marriage with data from only that wave. We found no significant associations for either gender group. Thus, our results do not support the comfort-of-home hypothesis, regardless of how the hypothesis is conceptualized.

The results remained statistically nonsignificant when we compared the rates of home-leaving in the year of becoming engaged with those in any other year.

Because a fixed-effects modeling approach compares time-varying individual characteristics to the individual mean, rather than using characteristics of an earlier point to predict the current outcome, using fixed effects to control for unobserved heterogeneity is not possible for this part of the analysis.

Japanese men are marginally more likely to think they lack funds for marriage when they live with parents, but this association suggests more about men’s decreased concern about the funds needed to marry when starting to live independently than their worries about lower living standards after marriage.