The Instability of Divorce Risk Factors in the UK

The Instability of Divorce Risk Factors in the UK∗ Tak Wing Chan Department of Sociology University of Oxford

Brendan Halpin Department of Sociology University of Limerick

August 8, 2005 Abstract We examine the stability of divorce determinants in the UK between 1960 and 1989. Using retrospective marriage history data, we show that the effects on divorce rate of educational attainment, premarital cohabitation, and spouse’s previous marital status have all undergone significant changes. This is in sharp contrast to results recently reported for the US (Teachman, 2002). We also confirm an unexpected finding of B¨ oheim and Ermisch (2001) and Chan and Halpin (2002): that in the UK children are now associated with higher divorce risks. The destabilising effect of children, we further show, is partly related to premarital birth.

1

Introduction

It is well known that the divorce rate in the UK has increased very sharply since the 1960s. Figure 1 shows the trend for England and Wales, where the annual average between 1950 and 1964 was 2.4 divorces per 1,000 marriages. This increased sixfold to over fourteen divorces per thousand in 1993 (Office of Population Censuses and Surveys, 1985, 1990, 1993).1 The divorce rate in We thank Yvonne ˚ Aberg, Josef Br¨ uderl, Lynn Prince Cooke, Jaap Dronkers, John Ermisch, John Goldthorpe, Juho H¨ark¨onen, Peter Hedstr¨om, Stephen Jenkins, Matthijs Kalmijn, Rob Mare, Mike Murphy, Wendy Sigle-Rushton, Jay Teachman, Michael Wagner, and participants of the third conference of the European Research Network on Divorce in K¨oln, the GHS User Group, and seminar audiences in Colchester, London, Oxford, Paris and Urbana-Champaign for helpful comments on early drafts. 1 The divorce rate of recent years can be found on the web site of the UK Office for National Statistics: www.statistics.gov.uk. ∗

1

12 10 8 6 4 2

Divorce decrees per 1,000 married population

14

England and Wales has since declined slightly to about thirteen per thousand in the late 1990s, but it is still among the highest in the European Union (Office for National Statistics, 2004, Table 2.13).

1950

1960

1970

1980

1990

2000

year

Source: Office of Population Censuses and Surveys (1985, 1990, 1993)

Figure 1: Crude divorce rate in England and Wales. What is also well established are some of the risk factors associated with divorce. For example, Murphy (1985a,b) analyses data from the 1976 Family Formation Survey and the 1980 General Household Survey and reports that young brides, those who were pregnant at the time of the wedding, those who got married in civil ceremonies, childless couples as well as couples with four or more children face significantly higher divorce risks. He also reports an association between divorce and social class (measured in terms of the UK Registrar General classification), though its magnitude is rather modest.2 Murphy’s findings are largely consistent with other studies done in the UK (e.g. Berrington and Diamond, 1999) and elsewhere (e.g. White, 1990; Bracher et al., 1993; Jalovaara, 2001). What is not clear, however, is whether the effects of various divorce risk factors have changed over time. Recently, some UK researchers (B¨oheim and Ermisch, 2001; Chan and Halpin, 2002) have reported a change in the direction of the association between children 2

‘The maximum gross difference between the relative risks in Social Classes II to IV is similar to that of a change of one year in age at marriage’ (Murphy, 1985a, p.459).

2

and marital stability. Using data from the British Household Panel Survey (BHPS), they suggest that in the 1990s British couples with children face significantly higher divorce risks than similar but childless couples. This is an intriguing and potentially important result as it contradicts established empirical findings and widely accepted theories on marital stability (e.g. Becker et al., 1977). In this paper we have two objectives. First, we wish to put the unexpected finding of B¨oheim and Ermisch (2001) and Chan and Halpin (2002) to further empirical tests using a large and nationally representative data set. And if their finding could indeed be confirmed, we shall take some initial steps to explore why such a change has taken place. Our second goal is to ascertain whether the effect of other divorce risk factors have changed over time.

2

Stability and change in family formation and dissolution

It is unusual in the social sciences to see large shifts in the relationship between basic variables, let alone a reversal in the direction of their association. But marriage and the family is one domain in contemporary societies that is truly in a state of flux. We have seen, for example, large increase in the rates of premarital cohabitation and out-of-wedlock birth, the closing of the gender gap in education, and a rise in the employment of married women with young children. Many of these factors are well known determinants of divorce. But why might their effect be changing? A possible explanation is that the direct and indirect costs of divorce have changed. When divorce was relatively expensive, only the rich and resourceful could afford to divorce. But as divorce becomes cheaper, more people are able to do so. Such a change would imply a flattening of the divorce gradient of those variables which index resources, such as income or educational attainment (Goode, 1993, p.vii). But apart from the changing costs (and benefits) of divorce, we believe, broadly speaking, two types of processes might be at work. For convenience, we shall refer to them as composition effects and contextual effects.

2.1

Composition effects

An example of composition effects concerns premarital cohabitation. It is a well established finding that marriages which began as cohabiting unions are less stable than those which did not (DeMaris and Rao, 1992; Axinn and Thornton, 1992; Lillard et al., 1995). As many researchers have observed, 3

this finding could plausibly be interpreted as arising from a process of selfselection. That is, individuals with non-traditional attitudes about marriage and the family are more likely to cohabit before marriage; but they are also more likely to divorce. Thus, the association between premarital cohabitation and marital instability is, under this view, spurious. To the extent that this is true, because the rate of premarital cohabitation in the UK has risen from 3 per cent in the 1960s to over 70 per cent in the 1990s (Berrington and Diamond, 2000),3 the cohabitants should have become a much less selfselected group, and thus the association between premarital cohabitation and marital instability should have attenuated over time. However, as the level of premarital cohabitation gets very high, an opposite self-selection process might begin to operate: only those who are very conservative about marriage and the family would marry directly. If this group is especially unlikely to divorce, we would expect the association between cohabitation and divorce to become stronger again. A similar argument could be made in relation to marrying a divorcee.4 When this was very rare, the individuals involved are likely to be a selected unconventional group. However, as divorce and remarriage become more common, this group should become less self-selective. Accordingly, the association between remarriage and divorce would attenuate over time.

2.2

Contextual effects

It is also possible that as remarriage becomes more common, the stigma associated with them declines, and couples in higher order marriages receive greater acceptance and support from their friends and relatives, leading to a weakening of the association between remarriage and divorce. But note that this argument refers not to a composition effect, but rather to the effect of a changing context, viz. there are more remarriages around. Along a similar line, Wolfinger (1999) argues that as divorce becomes more common, children with divorced parents suffer less stigma and trauma than did similar children in the past. Through various psychological and social pathways, this might then lead to a reduction of the excess divorce risks faced by people with divorced parents. Empirically, he shows that in the US intergenerational transmission rate of divorce risks has declined by almost 50 per cent between 1973 and 1996. 3

See also Figure 3 below. For trends of cohabitation in the US, see Smock (2000) and Bumpass and Lu (2000). 4 Although the focus of this paper is women in their first marriage, their spouses could be divorcees.

4

A further example of contextual effects concerns women’s employment (Becker et al., 1977; Tzeng and Mare, 1995). As South (2001) points out, the marital strain experienced by a working wife might be higher when few married women work, because under such a condition she is more likely seen as violating traditional gender role expectations. Furthermore, institutions and policies that support women’s employment, such as flexible working schedules or childcare facilities, are likely to be less developed. These and other arguments imply a steeper divorce gradient for women’s employment when the overall level of female employment is low. However, South also points out that the contextual effects might work in the opposite direction. For example, better childcare facilities and flexible schedules might help divorced women combine paid work and family obligations. The lack of these facilities in an early period might have made even working wives reluctant to divorce. As these facilities develop with rising level of female employment and, further, if working wives are in a better position to take advantage of such facilities should they choose to divorce, the association between women’s employment and divorce might be stronger when the overall level of female employment is high. Thus, there is no clear theoretical prediction. Empirically, it is the second contextual effect that is supported by his data. We have noted two mechanisms through which the effect of divorce risk factors might change. But in evaluating them against empirical results, three points should be kept in mind. First, given the nature of the data that is available to us, our ability to distinguish between composition and contextual effects is rather limited. However, there is one circumstance under which the two mechanisms give quite different predictions: as a behaviour or phenomenon changes from very rare to very popular, composition effects would imply a non-monotonic trend in the relevant parameter (see the discussion on premarital cohabitation above). This is not true for contextual effects. Secondly, it is likely that multiple mechanisms are at work. Thus, if we do find evidence for composition effects, it does not follow that contextual effects are absent. Thirdly, both composition and contextual effects are driven by changes in the relative size of different subgroups (e.g. cohabitants and non-cohabitants) in the population. As we shall see below, there are other changes in some divorce predictors which would require more involved explanations. We shall discuss these in Section 5.5 5

A third class of processes, namely diffusion effects, can also explain shifts in the effect of divorce determinants. For examples of diffusion of demographic behaviour through social networks, see Lesthaeghe and Surkyn (1988), Montgomery and Casterline (1996), Axinn and Yabiku (2001), Kohler et al. (2001), Behrman et al. (2002), and Rindfuss et al. (2004). However, because of the nature of the data that is available to us, we are not able to examine this class of processes in this paper.

5

2.3

Empirical research

While there are good reasons to believe that the effect of many divorce risk factors might be changing, the empirical evidence is far from clear. Teachman (2002) analyses data taken from five rounds of the US National Survey of Family Growth. After examining data on first marriages formed between 1950 and 1984 with proportional hazard models, he concludes that, with the exception of race, the effects of major sociodemographic predictors have not changed by historical period.6 However, as noted above, recent research in the UK suggests that the association between children and marital stability has changed. B¨oheim and Ermisch (2001) draw their sample from the first eight waves of the BHPS (1991–98). They consider married and cohabiting couples with at least one dependent child in the household. Using discrete time event history models, they show that ‘the risk of partnership dissolution increases with the number of children’ (B¨oheim and Ermisch, 2001, p.205).7 Chan and Halpin (2002) use the same BHPS data, but they consider married couples only (with or without children). Controlling for a similar set of covariates, they also report an association between children and higher divorce risks. They then check this finding with data collected in the Family and Working Lives Survey (FWLS), a retrospective life-history survey conducted in 1994–95. Using the Cox proportional hazard model, but controlling for fewer covariates, they report comparable results.

3

Data and methods

Our data come from the General Household Survey (GHS), which is a repeating general purpose survey conducted by the UK Office for National Statistics. The GHS has been running on an annual basis, almost without interruption, since 1971. Each year, about 9,000 households are sampled and face-to-face interviews were conducted with all individuals in these households aged 16 or over. The GHS is a cross-sectional survey, but since 1979 respondents were asked quite detailed retrospective questions on marriage and the family. This allows us to reconstruct their complete marriage and 6

Teachman (2002) notes that overall blacks have higher divorce rates than whites, but this gap has been narrowing over time. 7 B¨oheim and Ermisch (2001) control for age at marriage, marriage duration, age of the youngest child and various characteristics of the respondents and their partners, such as age, religion, ethnicity, educational attainment, employment status, income and financial surprise.

6

fertility histories. In this paper, we pool together relevant GHS data from 1989 to 2000, and focus on women’s first marriage.8 Because of the retrospective nature of the GHS data, almost all of our covariates, e.g. year of marriage, age at marriage, premarital cohabitation and educational attainment, are time-constant in nature. There is just one set of time-varying covariates, which measures parity.9 We are not able to examine the dynamic effect of those factors which might themselves be changing over the course of a marriage, such as women’s employment or family income. Also, we have only one piece of information about the husbands: whether they were previously married.10 Because the GHS is a household survey, there is additional information about the husband of those women who are currently married, such as their age or educational attainment. But such information is not available for women whose first marriage had already dissolved. This precludes us from including these covariates in the analysis. These limitations notwithstanding, the GHS is still an invaluable data source. It gives us information on the first marriage of 31,381 women who got married between 1960 and 1989. The size of this data set and its temporal coverage allow us to examine the stability and change of the effect of divorce risk factors over a period in which family behaviour has been changing most rapidly. Table 1 reports the basic descriptive statistics of the covariates. Although the overall means and proportions reported in this table are fairly self-explanatory, we hasten to add that they mask a great deal of change over time. Thus, for example, while the mean age at marriage of all brides in our sample is 22.5, there is a well known trend towards late marriage, one which has accelerated since the mid-1970s (see Figure 2). Furthermore, the variation in the age at marriage seems to have increased over time. A similar point can be made about premarital cohabitation. Overall, 22 per cent of all brides in our sample lived with their future husband as a couple before marriage.11 But the level of premarital cohabitation has varied 8

Before 1989 information on cohabitation before first marriage was not available for respondents who have been married twice or more. Since the multivariate analyses reported below include the covariate of premarital cohabitation, our analysis is based on a smaller data set of GHS 1989–2000. 9 In this paper, we consider children born to the respondents only. There is some information on step, adopted and fostered children in the GHS. But the relevant GHS questions refer to step, adopted and fostered children who are currently living with the respondents. Since step children from dissolved first marriages are probably not living with the respondents any more, such information is, for our purpose, incomplete, and therefore not used in this paper. 10 This piece of information is supplied by the respondents. 11 Because cohabitation with someone other than the respondent’s future husband is not

7

Table 1: Descriptive statistics of covariates. mean 22.5 1.9

Age at marriage Number of children

s.d. 4.0 1.3

min. 14.5 0.0 7.9% 13.7% 21.9% 8.2%

Education Degree A-level O-level No qualifications

18.3% 8.9% 24.6% 48.2%

25 15

20

age (years)

30

35

Premarital birth Premarital conception Premarital cohabitation Husband was a divorcee

max. 56.0 15.0

1950

1960

1970

1980

1990

2000

year of marriage

Figure 2: age at marriage of brides (mean, first and third quartiles) by year of marriage. Source: pooled GHS data 1989–2000.

8

80 60 40 20

percentage

premarital cohabitation

0

divorcee

1950

1960

1970

1980

1990

2000

year of marriage

Figure 3: Proportion of brides who cohabited with their future husband before marriage, or married a divorcee by year of marriage. Source: pooled GHS data 1989–2000. dramatically across marriage cohorts—from an average of two per cent for those who married in the 1950s and 1960s to about three quarters in the late 1990s (see Figure 3). Over the same period, the proportion of women marrying a divorcee also rose from three per cent to about 17 per cent. The decline of ‘shotgun marriage’ is evident in Figure 4, where we see an upward trend in the incidence of premarital birth and a decline, since the 1970s, of premarital conception (defined as giving birth to the first child within eight months of marriage). Specifically, about two per cent of women marrying in the 1950s had a premarital birth, but a quarter of the brides of the 1990s did. Level of premarital conception hovered between 15 and 20 per cent during the 1950s and 1960s. But it has since declined such that less than ten per cent of women marrying since the mid-1990s gave birth to their first child within eight months of marriage. Finally, Figure 5 reports the trends in educational attainment. We see an almost linear decline in the proportion of brides with no qualifications (from over 80 per cent to 20 per cent). At the same time, the shares of brides with A-levels or university degree increase by a factor of ten, reaching 30 per cent counted in the GHS, this is an underestimation of the extent of premarital cohabitation.

9

30 25 15 5

10

percentage

20

premarital conception

0

premarital birth

1950

1960

1970

1980

1990

2000

year of marriage

Figure 4: Proportion of brides with premarital birth or premarital conception by year of marriage. Source: pooled GHS data 1989–2000. and 34 per cent respectively in the late 1990s. These trends are, of course, consistent with what we know about the general expansion of education and the closing of the gender gap in educational attainment.

4

Results

In the analysis that follows, we focus on first marriages formed between 1960 and 1989. We censor all marriage spells at ten years if they were still intact by then.12 Figure 6 shows the proportion of marriages that are still intact by duration across marriage cohorts. Reflecting the accelerating divorce rate over this period, the survival curves for recent marriage cohorts are invariably steeper than those for previous cohorts. For example, just over ten per cent of those who married between 1960 and 1964 had divorced within ten years. But for the 1985–89 marriage cohort, this figure rose to 26 per cent. 12

Censoring the spells at ten years implies that there is relatively little difference between the period perspective and the cohort perspective that we adopt in this paper. We have repeated our analysis without censoring at ten years, and the results are largely the same. Details are available from the authors on request.

10

80 40 20

percentage

60

no qualifications

degree

O−levels

0

A−levels

1950

1960

1970

1980

1990

2000

year of marriage

0.9 0.8

proportion still married

1.0

Figure 5: Educational attainment of brides by year of marriage. Source: pooled GHS data 1989–2000.

0.7

1960−−64 1965−−69 1970−−74 1975−−79 1980−−84 1985−−89

0

24

48

72

96

120

months

Figure 6: Survival function by marriage cohort. 11

We then fit the Cox proportional hazard model to the data. Our baseline model contains the following covariates: a linear term and a quadratic term for age at marriage (centred at age 21), year of marriage (centred at 1960), and dummies for having one child, two children, three or more children (hereafter as the children dummies), whether the youngest child is less than four years old (hereafter as the preschooler parameter), whether the husband was a divorcee, premarital cohabitation, and educational attainment (with no qualifications serving as the reference category). This model is fitted to all first marriages in our data, and the parameter estimates are reported in the first column of Table 2. It can be seen that, considered over the entire period (i.e. 1960–89), women who marry late face lower divorce risks. For example, compared to those who married at age 21, the divorce hazard of those marrying at age 22 is, on average, 15 per cent lower (1−e−.170+.006 ). The positive coefficient of year of marriage reflects the rising divorce rate over time. Women marrying divorcees are 29 per cent (e.257 − 1) more likely to divorce, and premarital cohabitation raises marital instability by 71 per cent (e.539 − 1). The relationship between divorce risks and educational attainment is non-monotonic. Compared with the reference category of women with no qualifications, graduates have lower divorce risks, but women with A-levels have higher divorce risks (though the A-level parameter is marginally insignificant with p = .08). Having a preschooler is associated with substantially lower divorce risks. Controlling for all of the above, the dummies for having one child, two children, three or more children are all positive, though only the last of these is statistically significant at the 5% level. These estimates, apart from those concerning children, are not all that exceptional. But are they stable over marriage cohorts? To answer this question, we first report in Table 3 the fit statistics of our baseline model (model A) and those with an additional term postulating a linear interaction effect between year of marriage and one covariate (or one set of covariates, see models B to I). On the basis of the likelihood ratio test, it is clear that models B, C, D, E, H and I fit the data better than does the baseline model, while for model G the improvement in fit is just outside the conventional 5% level. When we allow all covariates to interact with year of marriage, the fit of model J is again significantly better than that of the baseline model.13 Thus, in sharp contrast to the US (Teachman, 2002), there is rather 13

We would arrive at the same conclusion if BIC is used as the criterion of model selection, except that models F and G would also be regarded as fitting the data better than model A.

12

Table 2: Cox proportional hazard model fitted to episode data of first marriage (baseline model). 1960–89

1960–64

1965–69

1970–74

1975–79

1980–84

1985–89

age at marriage

-.170** (.006)

-.198** (.018)

-.216** (.015)

-.162** (.013)

-.166** (.013)

-.155** (.013)

-.166** (.014)

age at marr. sq.

.006** (.000)

.022** (.004)

.012** (.002)

.007** (.001)

.006** (.001)

.005** (.001)

.005** (.001)

year of marriage

.043** (.002)

.075* (.032)

.034 (.024)

.041 (.022)

-.003 (.021)

.030 (.021)

.038 (.026)

husband a divorcee

.257** (.048)

.593** (.197)

.307 (.171)

.095 (.126)

.335** (.102)

.189* (.094)

.178 (.107)

pre. cohabitation

.539** (.035)

1.014** (.205)

.786** (.138)

.660** (.088)

.423** (.074)

.467** (.064)

.524** (.074)

O-levels

-.043 (.033)

-.177 (.140)

.013 (.090)

.013 (.078)

-.088 (.071)

-.121 (.070)

.023 (.083)

A-levels

.091 (.051)

.724** (.194)

.070 (.169)

.019 (.125)

.095 (.109)

.039 (.104)

.037 (.117)

university

-.097* (.044)

.011 (.151)

.281** (.104)

.101 (.091)

-.228* (.098)

-.336** (.103)

-.272* (.118)

one child

.095 (.052)

-.173 (.172)

-.395** (.130)

.100 (.113)

.065 (.116)

.174 (.116)

.499** (.140)

two children

.008 (.058)

-.496** (.189)

-.435** (.140)

-.250 (.130)

-.136 (.133)

.246 (.126)

.744** (.149)

three + children

.231** (.073)

-.586* (.228)

-.289 (.176)

-.057 (.172)

.264 (.167)

.509** (.157)

1.018** (.189)

youngest child < 4 -.369** (.045)

-.016 (.145)

-.200 (.109)

-.287** (.099)

-.462** (.104)

-.414** (.099)

-.585** (.124)

# respondents # divorce Log-likelihood

4669 478 -3929.15

5796 845 -7128.66

5903 1030 -8718.99

5182 1098 -9115.61

4888 1121 -9160.67

4943 816 -6424.15

31381 5388 -53915.56

Notes: * p < .05, ** p < .01, standard errors in parentheses.

13

Table 3: Model fit for the baseline additive model and models involving an interaction between each predictor variable and year married. Model A B C D E F G H I J

term interacting with year of marriage baseline age at marriage age at marriage squared age at marriage & age at marriage squared children dummies youngest child < 4 husband was a divorcee premarital cohabitation education all

log-likelihood -53915.559 -53913.339 -53908.176 -53908.166

p — .035 .000 .000

χ2 1917.79 1922.23 1932.56 1932.58

df 12 13 13 14

BIC -2020.893 -2033.925 -2044.255 -2052.867

-53906.709 -53914.429 -53913.908 -53911.676 -53897.214 -53878.035

.000 .133 .069 .005 .000 .000

1935.49 1920.05 1921.09 1925.56 1954.48 1992.84

15 13 13 13 15 23

-2064.369 -2031.745 -2032.785 -2037.255 -2083.359 -2190.454

clear evidence of significant shifts in the determinants of divorce in the UK. To represent these changes, and allowing for non-linearities, we divide our sample into five-year marriage cohorts and fit the Cox model to each cohort separately. The results are reported in columns 2 to 7 of Table 2. It can be seen that the magnitude of the two age-at-marriage terms generally become smaller across cohorts. But the linear term remains negative and the quadratic term positive throughout, and they have retained their statistical significance. As expected, year of marriage becomes insignificant within the five-year bands, except for 1960–64. But we see much greater changes in other parameters. For example, the effect of marrying a divorcee generally weakens over the marriage cohorts,14 such that by the 1985–89 cohort, the last cohort for which we have a full ten-year observation window, its effect was insignificant. This shift could be explained in terms of both composition and contextual effects. Likewise, the parameter for premarital cohabitation also weakens across cohorts. For the 1960–64 cohort premarital cohabitation was associated with a 176 per cent (e1.014 − 1) increase in divorce risks, but by 1985–89, when over half of the couples cohabited before marriage, this dropped to a still large but much reduced 69 per cent (e.524 − 1). Again, this decline could be explained in terms of both composition and contextual effects. But note that the cohabitation parameter weakens over the 1960s and 1970s, then it seems 14

In relation to this and some other parameters, the 1970–74 cohort is an outlier. We believe this might be due to the fact that this is the cohort most immediately affected by the 1969 Divorce Reform Act, which came into effect in 1971 (see the spike in Figure 1).

14

to be strengthening again. This non-monotonic pattern is consistent with the view that an opposite selection process begins to operate when cohabitation becomes very common. This lends support to the view that composition– selection is operating here. But since multiple mechanisms could be at work, one should not rule out the relevance of contextual effects. Furthermore, we note that the cohabitation parameter is never insignificant. The fact that cohabitation always has an effect in our model suggests that composition– selection is not the whole story behind the association between premarital cohabitation and marital instability. It would appear that the cohabitation experience does have some causal influence on the cohabitants which somehow makes them more prone to divorce. The changing association with educational attainment is intriguing. Recall that taken over the entire period of 1960–89, university graduates are significantly less likely to divorce than women with no qualifications (see first column of Table 2). When split into marriage cohorts, the following pattern emerges: compared to women with no qualifications, better educated women used to have higher divorce risks, but the education gradient has now reversed. For example, in the 1960–64 cohort, women with A-levels were twice (e.724 ) as likely to divorce as women with no qualifications. But for subsequent cohorts, the A-levels parameter became statistically insignificant. As for university graduates, those who got married in 1965–69 were 32 per cent (e.281 − 1) more likely to divorce than women with no qualifications. This parameter became insignificant for the 1970–74 cohort, and then turned negative and significant for subsequent cohorts, such that graduates who married between 1985 and 1989 were 24 per cent (1−e−.272 ) less likely to divorce than women with no qualifications.15

4.1

Children and marital stability

Coming to the effect of children, our result is consistent with the finding of B¨oheim and Ermisch (2001) and Chan and Halpin (2002). Table 2 shows that for marriages formed in the 1960s, the children dummies were negative in sign, and four of the six parameters were statistically significant at the 5% level. At the same time, the preschooler parameter was not significant. Thus, it would seem that for these couples, children generally reduced the divorce hazard, and the marriage-stabilising effect of children applied regardless of the children’s age. For marriages formed in the 1970s, the children dummies were insignifi15

Hoem (1997) reports a similar reversal of the educational gradient in divorce risks for Sweden.

15

cant, while the preschooler parameter turned significant. Thus, children were still associated with marital stability, but only when they were under the age of four. When we get to the 1985–89 cohort, the children dummies were positive and significant, and their effects were large. Although the preschooler parameter was also significant and its magnitude had in fact increased, the net association for having two children, or three or more children, even when they were under four, was marriage-destabilising. For example, the divorce rate of those couples with two children (at least one child being a preschooler) was 17 per cent (e.744−.585 −1) higher than otherwise similar but childless couples. B¨oheim and Ermisch (2001) and Chan and Halpin (2002) use panel data from the 1990s to show that children are associated with greater marital instability. Using retrospective GHS data, we now see that this shift was well underway in the 1980s. What might account for such a change? It is beyond the scope of this paper to address this question in detail. But we shall consider a plausible explanation and suggest several reasons as to why it is probably not an important contributing factor. The plausible explanation goes as follows. Children are usually a factor protective of marriage because they are a marriage-specific investment (Becker et al., 1977). But it has been quite convincingly demonstrated that step children receive less parental investment than biological children (see e.g. Biblarz and Raftery, 1999; Case et al., 2000). So could the change in the children parameters be due to the growing prevalence of step families? We think not, firstly because in this paper we consider children born to the respondent only (see note 9). However, we recognise that this does not rule out step relationship in the family, as children born to the respondent could be step children from the husband’s point of view. We do not have information on the children’s paternity. But step relationship might be thought more likely if the husband was previously married.16 Thus, in analyses not reported here, we have tested for interaction effects between the children parameters and husband’s previous marital status. It turns out that only one of the 24 interaction terms (4 × 6 cohorts) are significant at the 5% level. Furthermore, this significant interaction term pertains to the 1965–69 cohort, and so cannot explain the change in the 1980s.17 At the end of this section, we shall provide a third argument against the ‘step relationship’ or ‘paternity’ interpretation that is based on the timing of cohabiting spell and premarital birth. But at this point, we note that 16

For example, the husband could bring his own children from a previous marriage into the marriage in question, though this is unlikely as custody of children usually goes to the mother in the event of divorce. 17 Details are available from the authors on request.

16

two children

1.0

1.0

One child

−0.5

0.0

0.5

model 0 model 1 model 2 model 3 model 4

−1.0

−1.0

−0.5

0.0

0.5

model 0 model 1 model 2 model 3 model 4

1965

1970

1975

1980

1985

1965

1970

1975

marriage cohort

marriage cohort

Three or more children

Youngest child < 4

1985

1980

1985

0.0 −0.5 −1.0

−1.0

−0.5

0.0

1980

model 0 model 1 model 2 model 3 model 4

0.5

model 0 model 1 model 2 model 3 model 4

0.5

1960

1.0

1.0

1960

1960

1965

1970

1975

1980

1985

marriage cohort

1960

1965

1970

1975

marriage cohort

Figure 7: Estimated effects of the children dummies and the preschooler parameter under different models. Note: Model 0 contains the three children dummies and the preschooler parameter only. Model 1 is the baseline model reported in columns 2 to 7 of Table 2. Model 2 is the baseline model plus a dummy for premarital conception. Model 3 is the baseline model plus dummies for premarital conception and premarital birth (see Table 4). Model 4 substitutes the premarital birth parameter of model 3 with three separate dummies indicating whether premarital birth took place without cohabitation, before cohabitation or during a cohabiting spell (see Table 5). “•” denotes that the parameter estimate is significantly different from zero at the 5% level, while “+” denotes that it is not.

17

previous research suggests that the association between children and marital stability is multidimensional. Not only are the number of children and their age important, the timing of childbirth also matters (e.g. Murphy, 1985a; Waite and Lillard, 1991). To disentangle the effect of timing from those of number and age, we control for premarital conception and premarital birth in the model. Figure 7 shows how the inclusion of these terms affects the children dummies and the preschooler parameter. Model 0 is the null model in the sense that it contains the children parameters only. Thus, the parameter estimates of the null model show the gross association between children and marital stability. Model 1 is the baseline model reported in columns 2 to 7 of Table 2. In both models 0 and 1 the children dummies generally trend upward, while the preschooler parameter trends downward. As we saw in Table 2, for those who got married in the late 1980s, having children is associated with significantly higher divorce risks. When we control for premarital conception in model 2, the children dummies and the preschooler parameter change little, as can be seen from the close proximity of the lines for models 1 and 2 in Figure 7. However, when we further control for premarital birth in model 3, the estimates for the children dummies remained negative throughout, although they became insignificant for the last two cohorts. The preschooler parameter also becomes insignificant once premarital birth is taken into account. Thus, it would seem that the destabilising effect of children is related to the growth of premarital birth in recent years. But it should be noted that even under model 3 the general trend is for the children dummies to become less negative (i.e. less stabilising). We report the parameter estimates of model 3 in Table 4. (Figure 7 also contains a Model 4 which we shall discuss below.) Given the apparent importance of premarital birth, two questions come naturally to mind. First, who are more likely to have premarital birth? Secondly, under what circumstances is premarital birth especially likely to accentuate the association between children and marital instability? To answer the first of these questions, the lefthand panel of Figure 8 shows the association between educational attainment and premarital birth. It is striking how that association has strengthened over time. For example, in the 1960s, the odds of a woman with no qualifications having a premarital birth was about 2 to 3 times that of a female graduate. By the late 1980s, the odds ratio has risen to almost 6, which gets even higher in the 1990s.18 As regards the second question, we note that childbirth timing is related to other social changes discussed above, especially premarital cohabitation. 18

We have carried out some data smoothing for Figure 8 by averaging the contingency table of any year with those of the immediately preceding and subsequent years.

18

Table 4: Cox proportional hazard model fitted to episode data of first marriage (extended model: baseline model plus control for premarital birth and premarital conception). 1960–89

1960–64

1965–69

1970–74

1975–79

1980–84

1985–89

age at marriage

-.167** (.006)

-.189** (.019)

-.211** (.016)

-.165** (.013)

-.161** (.013)

-.147** (.013)

-.165** (.015)

age at marr. sq.

.006** (.000)

.020** (.004)

.011** (.002)

.007** (.001)

.006** (.001)

.005** (.001)

.005** (.001)

year of marriage

.042** (.002)

.073* (.032)

.030 (.024)

.039 (.022)

-.001 (.021)

.034 (.021)

.029 (.026)

husband a divorcee

.198** (.049)

.540** (.198)

.252 (.173)

.038 (.126)

.292** (.102)

.160 (.095)

.142 (.107)

pre. cohabitation

.440** (.036)

.755** (.214)

.608** (.142)

.558** (.090)

.360** (.075)

.407** (.066)

.409** (.076)

O-levels

-.017 (.033)

-.151 (.140)

.005 (.090)

.025 (.078)

-.070 (.072)

-.091 (.070)

.061 (.083)

A-levels

.129* (.051)

.784** (.195)

.098 (.169)

.055 (.126)

.114 (.109)

.073 (.104)

.099 (.118)

university

-.047 (.044)

.040 (.151)

.322** (.104)

.140 (.092)

-.187 (.099)

-.298** (.103)

-.195 (.118)

one child

-.309** (.058)

-.439* (.181)

-.670** (.139)

-.154 (.127)

-.288* (.130)

-.198 (.129)

-.243 (.165)

two children

-.527** (.066)

-.862** (.204)

-.817** (.155)

-.584** (.149)

-.587** (.153)

-.237 (.147)

-.273 (.193)

three + children

-.507** (.085)

-1.134** (.255)

-.889** (.201)

-.547** (.199)

-.377 (.198)

-.117 (.187)

-.287 (.246)

youngest child < 4 -.114* (.047)

.137 (.148)

-.012 (.112)

-.115 (.104)

-.236* (.110

-.193 (.105)

-.132 (.133)

pre. birth

.908** (.053)

1.035** (.197)

1.060** (.145)

.764** (.124)

.750** (.131)

.604** (.117)

1.090** (.128)

pre. conception

.358** (.042)

.392** (.128)

.274** (.100)

.104 (.099)

.398** (.097)

.535** (.091)

.425** (.120)

# respondents # divorce Log-likelihood

31381 5388 -53776.82

4669 478 -3915.65

5796 845 -7105.20

5903 1030 -8701.27

5182 1098 -9097.59

4888 1122 -9138.75

4943 816 -6388.95

Notes: * p < .05, ** p < .01, standard errors in parentheses.

19

20 10

15

no qual v degree

10

odds ratio

6

5

4

no qual v A−levels

2

odds ratio

8

cohabitation

no qual v O−levels

1960

1970

1980

1990

husband a divorcee

2000

year of marriage

1960

1970

1980

1990

year of marriage

Figure 8: Trends in the association between premarital birth and educational attainment (left panel), and between premarital birth and cohabitation and husband’s previous marital status (right panel). For example, in the mid-1970s, about 5 per cent of all births were registered by the mother alone, with another 5 per cent registered by both unmarried parents. By the late 1990s, although the sole registration percentage changed little, the percentage for joint-registration was over 30 per cent, the vast majority of whom would be cohabiting couples (Office for National Statistics, 2002, Chart 2.14). However, the righthand panel of Figure 8 shows that the association between premarital birth and cohabitation has in fact declined over time. Up to the late 1960s, the relevant odds ratio was about 15. Then it dropped to a still high but much reduced level of about 5. This decline is of course due to the fact that cohabitation has become very common across the board. For our present purpose, suffice it to say that there has always been a strong association between premarital birth and premarital cohabitation (see e.g. Ermisch, 2001).19 The association between cohabitation and premarital birth is relevant as the former provides a context for the latter. If a premarital birth took place during a cohabiting spell which eventually turned into marriage, it is likely that the cohabiting partner and eventual husband of the respondent 19

Figure 8 also shows a gentler decline in the association between premarital birth and marrying a divorcee.

20

2000

is the father of the child. One might argue that premarital births of this type are comparable to marital births. But for premarital births which took place before the cohabiting spell or without cohabitation, there is a higher probability that it was someone else who fathered the child.20 In a sense, this argument goes back to the question of step-relationship and paternity. To test this idea, we distinguish three types of premarital birth: those which took place (1) without cohabitation, (2) before cohabitation, and (3) during cohabitation. We replace the premarital birth dummy of Table 4 with these three dummies and repeat the analysis. The relevant results are reported in Table 5.21 Table 5: Cox proportional hazard model fitted to episode data of first marriage (modified extended model: premarital birth dummy replaced by three dummies indicating the timing of premarital birth relative to premarital cohabitation). 1960–89

1960–64

1965–69

1970–74

1975–79

1980–84

1985–89

Pre. birth without 1.056** cohabitation (.066)

1.113** (.208)

1.123** (.154)

.879** (.141)

1.051** (.149)

.856** (.168)

1.434** (.182)

Pre. birth before cohabitation

.813** (.089)

.488 (.489)

1.162** (.367)

.497* (.245)

.421 (.258)

.583** (.171)

.959** (.170)

Pre. birth during cohabitation

.709** (.086)

1.032 (.561)

.513 (.395)

.590* (.238)

.231 (.246)

.385* (.171)

.958** (.153)

Note: * p < .05, ** p < .01, standard errors in parentheses. Other parameters of the model, such as age at marriage, education, children dummies or premarital cohabitation are not shown, but are available from the authors on request.

As readers can see, there is partial support for the above argument. While the parameters for type 1 premarital birth are always positive and significant, those for type 2 and type 3 premarital birth are significant in four and three marriage cohorts respectively. Furthermore, where a type 3 parameter is significant, its magnitude tends to be considerably smaller than the corresponding type 1 parameter (i.e. premarital birth without cohabitation), though the pattern is less clear when we compare the magnitude of type 2 and 20

We thank Josef Br¨ uderl for this thoughtful suggestion. The model in Table 5 also contains all the other parameters considered above. But since they are substantively very similar to those of Table 4, we do not report them here. Details are available from the authors on request. 21

21

type 3 premarital births.22 Having said that, we note that all three types of premarital birth are significant in the last two marriage cohorts. More importantly, the children dummies and the preschooler parameter are unaffected when the distinction of three types of premarital birth is introduced. If we go back to Figure 7, it can be seen that the lines for model 3 (where we use one parameter for all premarital births) and model 4 (three parameters) are virtually indistinguishable from each other. This result reinforces our view that paternity or step-relationship is not a key factor which led to the observed change in the association between children and marital stability.

5

Summary and discussion

In this paper, we use retrospective marriage history data to show that in the UK the effects of many predictors of divorce have changed over time. In contrast to Teachman (2002), who finds a pattern of overall stability in the US, we find substantial, systematic and quite interpretable change in the effects of education, premarital cohabitation, marriage to a divorcee, and children. As the rate of premarital cohabitation rose from less than 5 per cent in the 1960s to about 50 per cent in the 1980s, the coefficient of our cohabitation parameter almost halves in magnitude (.409/.755, see Table 4). The estimate for marrying a divorcee also weakens over time, such that it became insignificant for the 1985–89 marriage cohort. The change of these two parameters could, at least in part, be explained in terms of composition effects, i.e. as cohabitation and remarriage become more common, they become less selective and tell us less about the individuals involved. But it is also possible that as cohabitation and remarriage become more common, they become socially more acceptable, leading to a weakening of their association with marital instability (i.e. contextual effects). We have noted one unique implication of composition effects: that is, as previously rare behaviour becomes very popular, there would be selection from the opposite direction, leading to a non-monotonic trend in the relevant parameter. This seems to be the case for premarital cohabitation, though we need data from even more recent marriage cohorts to confirm our initial 22

Compared with marriage, cohabitation is a less well-defined state (Murphy, 2000), and there would be a higher degree of measurement error in the reporting of the beginning date of cohabitation spells. Note that this is not just a matter of recall error. For many cohabitants, there is a gradual process from staying over occasionally to finally moving in. Such measurement errors mean that the distinction between premarital births which happened before and those which happened during a cohabitation spell is not very clear cut.

22

finding. Moreover, as we have noted above, the fact that composition effects are at work does not imply that other mechanisms are irrelevant. It is more than likely that multiple mechanisms are operating. Among women who got married in the 1960s, it was the better educated who faced higher divorce risks. Over time, the education gradient has not only flattened but reversed, such that for those who got married in the 1980s, university graduates were less likely to divorce than women with no qualifications. A flattening of the education gradient could be explained by the declining costs of divorce or some composition–selection process.23 But a reversal of the gradient suggests that additional factors are involved. Without further investigation, we could only offer some speculative remarks here. First, it could be the case that with greater female access to the labour market, and pressure (from the housing market and elsewhere) to have two incomes, British society has become more accepting of educated and economically successful women. Thirdly, it is also possible that the expansion of higher education and the closing of the gender gap in educational attainment have made universities and colleges more efficient institutions for sorting and matching potential spouses. Whether this happened, and if so how, are questions for future investigation. Halpin and Chan (2003) recently report that both absolute and relative rates of educational assortative mating has declined in the UK between 1973 and 1995. But we note that the measures used in that paper (e.g. percentage of cases off the main diagonal in a square table cross-classifying husband’s and wife’s education, and uniform difference parameters) are global measures. That is, they pertain to all levels of educational attainment. More work on assortative mating that is sensitive to the experience of people with specific levels of education would be very illuminating. The effect of children has been one of our main concerns. Previous research has demonstrated a reversal of the familiar protective effect of children on marriages, and we replicate that result here with a different data set. We have offered several arguments to suggest that the destabilising effect of children in recent cohorts is not due to the growing prevalence of step relationship. We then demonstrate that the change is related to the growth of premarital birth. Premarital birth, we further show, is becoming ever more strongly associated with low educational attainment. 23

When very few women were going to university, they would be a very selected unconventional group. Perhaps many of them were especially independent and focussed on their career, which might explain the higher divorce risks they faced. When more women are going to universities, such composition effects should become weaker, and the education gradient flatter.

23

The last observation is important because, as we have shown, the less qualified have recently become more likely to divorce than graduates anyway. But they are also much more likely to have premarital births (see Figure 8), which means that their children are likely to have a marriage-destabilising effect. Thus, children born to mothers with no qualifications are much more likely to go through a parental divorce and suffer its consequences. Compounding this is the link between low wage and no qualifications. Overall, the welfare implications for individuals, especially the children, is considerable. Our results clearly echo some of the concerns for the ‘diverging destinies’ of US children expressed recently by McLanahan (2004). The importance of premarital birth in the change of the children effect speaks obviously to the platitude that we need to know the process of family formation before we can understand the process of family dissolution. But premarital birth is probably not the complete explanation for the observed change. As we point out in relation to model 3 in Figure 7, even when premarital birth is taken in account, the overall trend is for the children effect to become less stabilising over time. This raises the question of whether in the UK the overall level of parental investment in children has been declining as well as polarising. We intend to pursue this important question with UK survey data on family expenditure and time use (cf. Sayer et al., 2004).

References Axinn, W. G. and Thornton, A. (1992). The relationship between cohabitation and divorce: Selectivity or causal influence? Demography, 29(3), 357–374. Axinn, W. G. and Yabiku, S. T. (2001). Social change, the social organization of families, and fertility limitation. American Journal of Sociology, 106(5), 1219–1261. Becker, G. S., Landes, E. M., and Michael, R. T. (1977). An economic analysis of marital instability. Journal of Political Economy, 85(6), 1141– 1187. Behrman, J. R., Kohler, H.-P., and Watkins, S. C. (2002). Social networks and change in contraceptive use over time: Evidence from a longitudinal study in rural Kenya. Demography, 39(4), 713–738. Berrington, A. and Diamond, I. (1999). Marital dissolution among the 1958 British birth cohort: The role of cohabitation. Population Studies, 53(1), 19–38. 24

Berrington, A. and Diamond, I. (2000). Marriage or cohabitation: A competing risks analysis of first-partnership formation among the 1958 British birth cohort. Journal of the Royal Statistical Society, Series A, 163(2), 127–151. Biblarz, T. J. and Raftery, A. E. (1999). Family structure, educational attainment, and socioeconomic success: Rethinking the “pathology of matriarchy”. American Journal of Sociology, 105(2), 321–365. B¨oheim, R. and Ermisch, J. (2001). Partnership dissolution in the UK — the rˆole of economic circumstances. Oxford Bulletin of Economics and Statistics, 63(2), 197–208. Bracher, M., Santow, G., Morgan, S. P., and Trussell, J. (1993). Marriage dissolution in Australia. Population Studies, 47, 403–425. Bumpass, L. and Lu, H.-H. (2000). Trends in cohabitation and implications for children’s family contexts in the united states. Population Studies, 54, 29–41. Case, A., Lin, I., and McLanahan, S. (2000). How hungry is the selfish gene? The Economic Journal, 110, 781–804. Chan, T. W. and Halpin, B. (2002). Union disruption in the United Kingdom. International Journal of Sociology, 32(4), 76–93. DeMaris, A. and Rao, K. V. (1992). Premarital cohabitation and subsequent marital stability in the United States: A reassessment. Journal of Marriage and the Family, 54, 178–190. Ermisch, J. (2001). Cohabitation and childbearing outside marriage in Britain. In L. L. Wu and B. Wolfe, editors, Out of Wedlock: Causes and Consequences of Nonmarital Fertility, chapter 4, pages 109–139. Russell Sage Foundation, New York. Goode, W. J. (1993). World Changes in Divorce Patterns. Yale University Press, New Haven. Halpin, B. and Chan, T. W. (2003). Educational homogamy in Ireland and Britain: Trends and patterns. British Journal of Sociology, 54(4), 473–495. Hoem, J. M. (1997). Educational gradients in divorce risks in Sweden in recent decades. Population Studies, 51, 19–27.

25

Jalovaara, M. (2001). Socio-economic status and divorce in first marriages in Finland 1991–93. Population Studies, 55, 119–133. Kohler, H.-P., Behrman, J. R., and Watkins, S. C. (2001). The density of social networks and fertility decisions: Evidence from south Nyanza district, Kenya. Demography, 38(1), 43–58. Lesthaeghe, R. and Surkyn, J. (1988). Cultural dynamics and economic theories of fertility change. Population and Development Review, 14(1), 1–45. Lillard, L. A., Brien, M. J., and Waite, L. J. (1995). Premarital cohabitation and subsequent marital dissolution: A matter of self-selection? Demography, 32(3), 437–457. McLanahan, S. (2004). Diverging destinies: how children are faring under the second demographic transition. Demography, 41(4), 607–627. Montgomery, M. R. and Casterline, J. B. (1996). Social learning, social influence, and new models of fertility. Population and Development Review, 22(1), 151–175. Murphy, M. (2000). Editoral: Cohabitation in Britain. Journal of the Royal Statistical Society, Series A, 163(2), 123–126. Murphy, M. J. (1985a). Demographic and socio-economic influences on recent british marital breakdown patterns. Population Studies, 39, 441–460. Murphy, M. J. (1985b). Marital breakdown and socio-economic status: A reappraisal of the evidence from recent british sources. British Journal of Sociology, 36(1), 81–93. Office for National Statistics (2002). Social Trends, volume 32. HMSO, London. Office for National Statistics (2004). Social Trends, volume 34. HMSO, London. Office of Population Censuses and Surveys (1985). Marriage and Divorce Statistics: Review of the Registrar General on Marriages and Divorces in England and Wales, 1982. Number 19 in Series FM2. HMSO, London. Office of Population Censuses and Surveys (1990). Marriage and Divorce Statistics: Historical Series of Statistics on Marriages and Divorces in England and Wales, 1837–1983. Number 16 in Series FM2. HMSO, London. 26

Office of Population Censuses and Surveys (1993). Marriage and Divorce Statistics: Review of the Registrar General on Marriages and Divorces in England and Wales, 1993. Number 19 in Series FM2. HMSO, London. Rindfuss, R. R., Choe, M. K., Bumpass, L. L., and Tsuya, N. O. (2004). Social networks and family change in Japan. American Sociological Review, 69(6), 838–861. Sayer, L. C., Bianchi, S. M., and Robinson, J. P. (2004). Are parents investing less in children? trends in mothers’ and fathers’ time with children. American Journal of Sociology, 110(1), 1–43. Smock, P. J. (2000). Cohabitation in the United States: An appraisal of research themes, findings, and implications. Annual Review of Sociology, 26, 1–20. South, S. J. (2001). Time-dependent effects of wives’ employment on marital dissolution. American Sociological Review, 66(2), 226–245. Teachman, J. D. (2002). Stability across cohorts in divorce risk factors. Demography, 39(2), 331–351. Tzeng, J. M. and Mare, R. D. (1995). Labor market and socioeconomic effects on marital stability. Social Science Research, 24, 329–351. Waite, L. J. and Lillard, L. A. (1991). Children and marital disruption. American Journal of Sociology, 96(4), 930–953. White, L. K. (1990). Determinants of divorce: A review of research in the eighties. Journal of Marriage and the Family, 52(4), 904–912. Wolfinger, N. H. (1999). Trends in the intergenerational transmission of divorce. Demography, 36(3), 415–420.

27