|Home | About | Journals | Submit | Contact Us | Français|
Using high-quality data from Norwegian population registers, we examine the relationship between family disruption and children’s educational outcomes. We distinguish between disruptions caused by parental divorce and paternal death and, using a simultaneous equation model, pay particular attention to selection bias in the effect of divorce. We also allow for the possibility that disruption may have different effects at different stages of a child’s educational career. Our results suggest that selection on time-invariant maternal characteristics is important and works to overstate the effects of divorce on a child’s chances of continuing in education. Nevertheless, the experience of marital breakdown during childhood is associated with lower levels of education, and the effect weakens with the child’s age at disruption. The effects of divorce are most pronounced for the transitions during or just beyond the high school level. In models that do not allow for selection, children who experienced a father’s death appear less disadvantaged than children whose parents divorced. After we control for selection, however, differences in the educational qualifications of children from divorced and bereaved families narrow substantially and, at mean ages of divorce, are almost non-existent.
In the latter half of the twentieth century in almost all industrialized countries, increasing proportions of children experienced family disruption, with divorce replacing parental death as the main cause. These trends have, in many countries, provoked concern because correlational evidence suggests that the changes are associated with disadvantages for the children involved. Although most research on family disruption and its consequences comes from the United States, evidence from several nations has demonstrated that, on average, children who experience a family disruption fare poorly across a wide range of adolescent and adult outcomes, including educational attainment, economic security, and physical and psychological well-being (Amato and Keith 1991a, 1991b; McLanahan and Sandefur 1994; Rogers and Pryor 1998; Sigle-Rushton and McLanahan 2004). Although many outcomes have been linked to family structure in childhood, the negative association between family disruption and educational attainment may be especially important (Hobcraft 2000). Poor educational outcomes may initiate processes that lead to other kinds of disadvantages and contribute to persistent differences in physical and psychological health (Dalgard et al. 2007; Krokstad, Kunst, and Westin 2002), relationship stability (Lyngstad 2004), and economic well-being (Statistics Norway 2005) later in adulthood.
In this analysis, we use high-quality register data from Norway to assess the effects of family disruption on children’s educational outcomes. We also use a method not previously employed in such studies. To control for selection into divorce, we estimate a system of simultaneous equations to investigate the effects of family disruption on educational attainment.
Norway provides an excellent case study for several reasons. Divorce rates have increased quickly and steeply, but in a context where economic inequality remains relatively low (Aaberge et al. 2000; Statistics Norway 2008b; United Nations Development Programme 2006). The Norwegian economic context is particularly salient because U.S. research suggests that economic disparities are responsible for much of the observed differences between children living with two parents and children living with only one parent (McLanahan 1999; McLanahan and Sandefur 1994; Page and Stevens 2004). If the economic vulnerability that accompanies divorce is an important explanatory factor in countries like the United States, the association between divorce and educational attainment might be smaller in the Norwegian context. Perhaps in line with this thinking, it is noteworthy that there is little political concern in Norway about the number of children who experience a parental disruption. Although only a few empirical studies examine the links between family structure and children’s outcomes in Norway, there is some evidence of more school problems and lower academic achievement among children who have experienced divorce (Lauglo 2008; Størkersen et al. 2005). Moreover, effect sizes are of the same magnitude as seen in U.S. studies (Breivik and Olweus 2006; see also Pong, Dronkers, and Hampden-Thompson 2003). Nonetheless, the Norwegian studies have employed standard multivariate statistical techniques, with controls for a few socioeconomic variables, and some have been based on samples that are of moderate size and not nationally representative.
In this section, we discuss the various reasons that children who experience a family disruption might have poorer average educational outcomes than children raised by two biological parents. We also discuss the possibility of selection bias and provide an overview of the ways in which selection has been addressed in the extant literature.
Children who experience a family disruption might have lower average educational attainment relative to children in stable, two-biological-parent families, either because of deficits resulting from the absence of a parent in the same household or because of other deleterious changes that accompany the process of family disruption (Amato and Keith 1991a; Rogers and Pryor 1998). Disruption leads to an immediate reduction of time and parenting inputs from the nonresident or deceased parent. Consequently, those who experience a parental death or divorce may experience problems in school and be less likely to take further education (Amato and Keith 1991b). Because fathers tend to maintain some contact after divorce, deficits might be greater for bereaved children. However, contact with the father after a divorce tends to diminish over time (Hanson, McLanahan, and Thomson 1996; Hetherington and Kelly 2002; King 1994), suggesting that parenting deficits increase over time. Hence, we might find stronger effects on educational attainment for those who experience a parental divorce at the youngest ages (Ermisch and Francesconi 2001; Krein and Beller 1988; and, for Norway, see Reneflot 2007).
In addition to deficits associated with the direct loss of a coresident parent, there may be other disadvantages, such as economic vulnerability (McLanahan 1999; McLanahan and Sandefur 1994; Page and Stevens 2004; and, for Norway, see Breivik and Olewus 2006), inadequate parenting (Amato 2000; Hetherington 1999; but see also Thomson, Hanson, and McLanahan 1994), and increased levels of acute stress and conflict (Kelly and Emery 2003). Even if experiences of disadvantage are short-lived, when parental divorce takes place close to key turning points in the educational career—such as the point where mandatory schooling ends—the disruptive effect on educational progression could be substantial (Jonsson and Gähler 1997; see also Frost and Pakiz 1990). Consequently, the experience of disruption during adolescence might have a stronger effect than the same experience at earlier ages.
To the extent that bereaved families receive better economic and social support than divorced families, we might expect children who experience a parental divorce to fare worse. The gap between the two may be less severe in Norway, where the welfare state works to mitigate the economic consequences of family disruption (Bratberg and Tjøtta forthcoming) without differentiating, to a great extent, between causes of disruption, as happens in other countries like the United States (Biblarz and Gottainer 2000; Skevik 2003). There may be less stress and more support offered after bereavement than after divorce, where the parents may begin or continue to engage in conflict and where ties to wider kin networks may be strained (Hetherington 1999; Johnston and Campbell 1988; Kelly 2002). Hence, the effects of death or divorce may be qualitatively different, with children from bereaved families faring somewhat better.
The effects of family disruption may be more pronounced for some educational transitions than others, but whether divorce is likely to have the greatest effect at earlier or later transitions is ambiguous. Early educational transitions take place closer to the disruption, which means that the stress level is relatively high, but this may be offset by the aforementioned decline in father involvement over time. Moreover, to the extent that disruption leads to a reduction of economic resources, one would expect effects to be largest at the later transitions, for which, even in a context like Norway where educational costs are highly subsidized, economic and opportunity costs of tertiary education are higher (OECD 2007: Table B3.2b, Chart B5.2).
Because our sample is large, we are able to estimate precisely the effects of a parental divorce or death for different academic transitions. We are also able to explore whether the parameter estimates differ by age. Although our data contain a good deal of high-quality longitudinal information for a large, representative sample of individuals, the data come from administrative sources and do not contain rich measures of the characteristics of the mothers and children in our sample. We do not have access to the kinds of measures that would allow us to test directly different explanations for the association between disruption and educational outcomes briefly outlined above. Instead, we aim to identify whether the effects of disruption are large and significant net of selection on mother-specific unobserved variables. In other words, we hope to identify whether, net of selection, there is a significant association that needs to be explained.
An important (but difficult to refute) explanation for the association between parental divorce and child outcomes is selection (Sigle-Rushton and McLanahan 2004). Proponents of this view argue that the relative disadvantage observed among children whose parents divorce is spurious and due to differences between the kinds of people who dissolve their relationships and those who remain together. In Norway, as in other countries, divorce and separation tend to be more common among the lower educated, and children from these families have, on average, poorer educational outcomes (Lyngstad 2004). Other, less readily observed personality characteristics might be important as well (Amato 2000). Although there has been little work on the association of family disruption and child well-being in Norway, more research has been carried out using Swedish data. Results from recent studies that estimate sibling-difference models using samples of full siblings (Björklund and Sundström 2006) or both full and half siblings (Björklund, Ginther, and Sundström 2007) suggest that when fixed effects are included, selection explains almost all of the differences in education and earnings between children who experience a parental divorce and those who do not.
In the absence of experimental study designs, it is impossible to ascertain how children would have fared if their parents had not divorced. However, second-best strategies have been developed to try to deal with selection. One strategy has been to exploit longitudinal study designs and control as much as possible for pre-divorce circumstances using standard regression models (see, e.g., Sigle-Rushton, Hobcraft, and Kiernan 2005; Sun and Li 2002) or, in more recent applications, propensity score techniques (see, e.g., Amato 2003; Frisco, Muller, and Frank 2007). Given our data source, it is unlikely that the set of variables we have available would control sufficiently for the selective factors.
Another approach compares children who share parents (or a parent) but who vary in their exposure to different family structures. Usually this is operationalized as time spent in different family types (Björklund et al. 2007; Björklund and Sundström 2006; Ermisch and Francesconi 2001; Ginther and Pollak 2004; Hao and Xie 2001). For example, if parents separate when one child is 8 years old and the other child is 6, the younger sibling can be said to have experienced two more years of father absence before the age of 16. This approach allows researchers to include family- or parent-specific fixed effects but relies on the assumption that the effect of divorce differs by the length of exposure. Moreover, sibling models assume that parents treat their children exactly the same and that children respond similarly to family-wide risk influences, both of which are highly unlikely (Carbonneau et al. 2002; Jenkins, Rasbash, and O’Connor 2003). Finally, when the outcome is discrete (as in our application), fixed-effects models are identified only from the subsample of families in which there is within-family variation in the outcome.1 Consequently, the fixed-effects approach excludes families with only one child and those in which all children have the same outcome. Taking into account the limitations of both these methods in the context of our application, our approach to handling selection, described in the Methods section below, is to model children’s educational attainment jointly with mothers’ risk of divorce or separation.
Our analysis is based on data up to 2003 from the Norwegian Population Register, the Population Censuses, and Statistics Norway’s Educational Registration System. These data include everyone living in Norway after 1960. For each person, there is information about marital status on January 1 each year beginning in 1974 and the spouse’s identification code (allowing information on spouses to be merged); the highest educational level achieved as of October 1, 1970, and annually from 1980; dates of any in- and out-migration; and dates of death. Parents are identified for almost everyone born in Norway after 1953.
The analysis file includes ever-married women (those at risk of divorce) (1) who married for the first time after 1973 and thus have complete marital history records, (2) whose oldest child was born after 1974 and had turned 16 by 2003 (and thus was old enough to have made the first educational transition), and (3) to ensure that we have complete educational records, who did not live outside Norway when any of their children were younger than 16 years. The mothers of these children were born between about 1930 and 1970, but we nevertheless restrict the analysis to mothers born in 1935–1970 (leaving out 0.02% of the mothers of children born in 1974–1987) because, due to a limitation in the registration system with respect to the parent-child linkage, we have complete birth histories only for those born after 1935. For each of these women, we extracted information on her education and year of birth/death/emigration; the dates of birth, education in 2003, and father’s education for the first five children born between 1974 and 1987; and the start and duration of up to five marriages and the education of each spouse. The restriction to five children is purely a matter of convenience, but it leads to the omission of only 0.7% of births between 1974 and 1987. Fathers were identified for 97.7% of the children. The analysis is based on a 50% random sample of the linked files. This sample contains 197,638 children from 113,980 mothers who had 129,189 marriages.
The outcome variables for the analysis are defined as follows.
Marriage durations. The marital history data contain the years in which marriages start and end, but not the exact dates. If the marital status is never-married on January 1 of year t and married on January 1 of year t + 1, we can infer only that the couple married at some point during year t. Similarly, if a status separated, divorced, or widowed is seen for the first time on January 1 of year t*, we know that a disruption took place some time during year t* – 1. In that case, the duration is set to t* – t – 1. If no such disruption is indicated, the duration is, of course, 2003 – t. The duration of marriages ending in death are censored at the time of death.
Some events are ignored because we know only the marital status at the beginning of each year. Most importantly, we treat a couple who separates and reunites within one year as continuously married. There is no information about how common this behavior is in Norway. In principle, it is also possible that a person who is married at the beginning of a year divorces and remarries within the same year. Another possibility is that a person marries and separates or divorces within the same year. These are ignored in our analysis but are extremely rare because a separation period of one year is usually required before a divorce can be obtained (Statistics Norway 2008a).
A more pernicious problem is that we are unable to measure cohabiting unions that are not eventually formalized as marriages, and for those that are formalized, we identify the partnership only when they marry. For children whose biological parents never married, we do not observe whether they never lived with both biological parents or were born into a cohabiting union that subsequently dissolved. However, very few children born between 1974 and 1987 fall into this category. Only 16% of the births were out-of-wedlock in the cohorts we analyze (as compared with 55% currently; Statistics Norway 2008c), and many of the unwed mothers subsequently married the biological father. In total, only 8% of the children in our sample had a mother who never married before 2003, and thus they were dropped from our sample. Many of those children probably had cohabiting parents and experienced their parents’ disruption. A further 4% had a mother who married someone other than the father of the child; it is likely that most of these women were cohabiting at birth and subsequently experienced a disruption. Another problem is that we do not account for those cohabiting unions that form after a first disruption and dissolve before turning into marriages. For this reason, we probably underestimate children’s experience of family transitions.
Children’s education. Educational level is coded into five levels, according to the 2000 standard (Statistics Norway 2001), as follows: Level 1 represents compulsory education only (10 years of schooling); Level 2, lower secondary education (11–12 years); Level 3, higher secondary education (13 years); Level 4, some college or university education, up to and including a Bachelor’s degree (14–17 years); and Level 5, all college education taking five or more years, such as a Master’s degree (18 or more years). The small proportion of children who had unknown levels of education or who had not yet completed compulsory school were omitted from the analysis.
Not all children in the cohorts we consider had completed high school by the end of 2003. For example, 4% of those born in 1974 had only compulsory education by age 29, while 18% had 1–2 years of high school education (calculated from the data used in the analysis). Among those who had graduated from high school, a large proportion did not take further education. Among the 1974 cohort, 35% had no more than a high school education by age 29, 35% had completed some college or university education, and 8% had earned a Master’s degree or its equivalent.
In this article, we model children’s educational attainment jointly with mothers’ risk of divorce or separation using a multilevel (random-effects) simultaneous equation or multiprocess model. This model allows us to identify the association of divorce and educational attainment net of time-invariant unobserved differences between families, allowing us to compare the effects of different kinds of family disruption with greater confidence.
We employ a multilevel model because a woman may have more than one marriage or more than one child over the course of the observation period. The simultaneous equation model consists of two components: a hazards model for marital dissolution and a sequential probit model for educational progression. Each component includes woman-specific random effects or residuals that allow for the influence of unmeasured time-invariant characteristics of women on each outcome (commonly known as unobserved heterogeneity). These residuals may be correlated across processes, thereby allowing for the possibility that the risk of divorce and educational outcomes may be influenced by a common or correlated set of unobserved characteristics. The direction and magnitude of this residual correlation provides information on the nature and extent of selection on time-invariant maternal characteristics. For example, a negative correlation would suggest that women with an above-average risk of divorce tend also to have children with poorer educational outcomes.
As in sibling fixed-effects models, we control for time-invariant unobserved factors at the family level (in our case, the mother level) rather than the individual child level. Our models control for stable characteristics of the mother that are associated with divorce or educational success but not time-varying environmental conditions that may also have a strong influence, particularly at key transition points in the life course. In addition, our models assume that the mother-specific effect is the same for all of her children. To the extent that children have different temperaments and/or parents treat children differently, failure to allow for unmeasured child-specific variables means that some bias will likely remain.
As in previous research, we assume that parental death is exogenous. A parental death censors the duration to marital dissolution and is introduced as an explanatory variable in the education process.
Denote by hij(t) the hazard of marital dissolution in year t of marriage i of woman j. A multilevel continuous-time event-history model allowing for unobserved heterogeneity between women may be written
The log-hazard of dissolution is assumed to depend on the marriage duration at year t through a function f(t), the baseline log-hazard rate. Here, we assume that f(t) is a piecewise-linear spline with nodes spaced at biannual intervals up to 10 years. Covariates wij(t), with coefficients α, may be time-varying or characteristics of a particular marriage (e.g., the education or age of spouse i).
Among women with the same observed characteristics (values of wij(t)), there will be some at high risk and others at low risk of divorce. The woman-specific random effect vj accounts for unobserved heterogeneity that is attributable to time-invariant unmeasured characteristics. For each woman, vj represents the value of a collection of unobserved attributes drawn at random (from a normal distribution with variance ) at an early age, staying with her throughout her adult life, and affecting her risk of separation in any marriage she forms.
Children’s educational qualifications are modeled using a sequential probit model (see, e.g., Brien and Lillard 1994; Upchurch, Lillard, and Panis 2002). Denote by yij the highest qualification achieved by the end of the observation period by child i of woman j, which is measured by an ordered categorical variable with five levels (as described above). We view the observed yij as the result of a sequence of binary transitions made from one education level to the next. For five levels, a child can make up to four sequential transitions, where the transition from level r (r = 1,…,4) is possible only for those for whom we observe yij ≥ r. The transition from level r is indicated by a binary variable , coded 1 if the child attains level r + 1 and coded 0 if he or she stops at r.
One advantage of breaking down yij into a set of binary transitions is that sample children (aged 16–19 in 2003) who have not yet completed their education by the end of observation period can be retained in the analysis and will contribute up to the level they are observed. The sample used to estimate the probability of progressing beyond a given level is restricted to children who were old enough to have made that transition. Specifically, the probability of progressing from compulsory to lower secondary school is estimated for children aged 16 years or older; lower to higher secondary is estimated for those aged 18 years or older; higher secondary to Bachelor’s degree, for those aged 19 years or older; and Bachelor’s to Master’s or a higher degree, for those aged 23 years or older. A rather young lower age restriction is used for the estimation of the equation for the third transition because Level 4 captures a wide range of university courses, from a short course to a full Bachelor’s degree.
The sequential probit model is defined in terms of a set of continuous latent variables or propensities underlying the observed binary responses , where if and otherwise. A multilevel model that allows for unobserved heterogeneity between mothers can be written:
where is a vector of potentially endogenous indicators of family disruption with coefficients γ(r), and is a vector of background characteristics of the child and the mother with coefficients β(r). Rather than focusing on one specific educational transition—such as graduating from high school or graduating from college—this model specification allows us to examine the educational career as a whole and to identify the stages of the educational system at which the family disruption variables have stronger and weaker associations.
As in the model for marital dissolution, Eq. (1), we include a woman-specific random effect uj, which here represents unobserved time-invariant characteristics of the mother that affect the probability of progressing beyond level r (r = 1,…,4) for each of her children. These random effects have transition-specific coefficients or “loadings” λ(r).2 Thus, although the same unmeasured mother characteristics are assumed to influence progression at all levels of education, their effects may differ across the four educational transitions. We assume that the uj follow a normal distribution with mean zero and variance . The model also includes residuals that are specific to a particular child and transition and are assumed to follow independent standard normal distributions.
Eqs. (1) and (2) together define a multilevel multiprocess model. The equations are linked in two ways. First, the family disruption indicators in (2) are prior outcomes of the marital dissolution process in (1). Second, we allow for the possibility of a nonzero correlation between the unmeasured woman-specific components uj and vj. Specifically, uj and vj are assumed to follow a bivariate normal distribution with correlation ρuv. A value of ρuv that is significantly different from zero would suggest that at least one element of is endogenous with respect to educational transitions.
The presence of uj in all four educational transition equations in (2), and the correlation between uj and vj, mean that Eqs. (1) and (2) must be estimated simultaneously. We estimate the model via maximum likelihood using the software package aML (Lillard and Panis 2003).
In order to estimate a simultaneous equation model, it is usually necessary to impose some identification conditions on the exogenous covariates in the model, represented by and wij(t) in (1) and (2) above. In the present case, however, exclusion restrictions are not required for model identification because some women have both multiple children and multiple marriages. After the mother-level random effects and their correlation are accounted for, the remaining variation in between siblings represents the effect of experiencing disruption on the probability of progressing to the next level of education, adjusting for selection on mother-specific unobserved variables. Nevertheless, the dissolution equation does contain several variables that are not included in the educational transition equations.
Explanatory variables in the model for marital dissolution. The model for divorce or separation includes a range of explanatory variables specific to marriage i: an indicator of previous marriage, the presence of children from a previous relationship (for both the woman and spouse i), an indicator of whether the woman had a premarital birth with spouse i, a time-varying count of the number of children fathered by spouse i, the woman’s age at marriage and the age difference between the woman and her spouse, and the education level of each partner. Table 1 shows descriptive statistics for variables included in the final dissolution equation.
Indicators of family disruption. Marital dissolution is represented by a dummy variable indicating whether the biological parents separated and the child’s age at the time of the separation. Children whose parents did not marry during this period (but whose mother married someone other than the biological father between 1974 and 2003)—representing about 4% of our sample—are indicated by a dummy variable and coded 0 on the disruption indicator. A corresponding set of variables are created to indicate disruption following the father’s death.
Marital reunions are not uncommon in Norway. For 10% of our sample children, the mother is coded as formally separated from and later married again to the same person. These annulled separations are treated as disruptions in the analysis, but we indicate them by a dummy variable to explore whether a reunion ameliorates the effect of separation. Finally two variables are defined to indicate whether a child had a stepfather and, if so, whether that marriage was dissolved.
Other explanatory variables in the model for educational transitions. The covariates of prime interest in the analysis of children’s educational transitions are the indicators of family disruption. We also control for the child’s sex, the child’s age in 2003, the number of older and younger siblings, and the highest level of education achieved by each of the parents (see Table 2 for descriptive statistics). Unfortunately, other potentially important determinants of education are not available in these register data.
Although each of the four sequential probit equations is estimated for the subsample of children who are old enough to make a given transition and have achieved the previous level in the sequence, an additional control for age is included to allow for between-child variation in the timing of transitions and the fact that the probability of making a transition increases with age. In each probit equation, age effects are modeled as a step function in which the age categories used are permitted to differ across transitions.
Although a study by Black, Devereux, and Salvanes (2005) showed that family size had little impact on educational outcomes in Norway, we include the number of siblings present at the time a child is completing compulsory education as an indicator of competition for household resources. The number of younger siblings is defined at age 17 rather than 16 (the age when compulsory schooling ends) to allow for the fact that the mother may be pregnant at the time of transition.
Based on evidence from previous studies (d’Addio 2006), we expect parental education to be a strong predictor of children’s education. We represent parents’ education by a composite categorical variable.
Unfortunately, we had access only to a measure of gross labor market earnings. For a variety of reasons, we were concerned about including this measure in our models. In contrast to countries like the United States, guaranteed child maintenance and greater levels of support to mother-only families (Kjeldstad and Rønsen 2002; Skevik 2003) mean that the earnings of single mothers in Norway may not always be the main part of their income package. Hence, we were concerned about the validity of an earnings measure, and we decided not to include it in our main model specification. We did, however, examine whether excluding this variable affected our results substantially (see the Conclusion for further discussion).
The final model specifications were reached after preliminary analysis in which models were fitted separately for marital dissolution and children’s education—that is, the single-process models represented by Eqs. (1) and (2) with zero correlation assumed between the mother-level random effects uj and vj.
Following preliminary analysis using more flexible specifications that showed fairly monotonic effects, we decided to treat age at divorce or separation as a continuous variable with linear effects. The child’s age at the time of death was not found to be significant for any transition or any of the different age specifications and is therefore not included in the models presented below.
Because previous research suggests that the educational attainment of girls and boys may be differentially associated with family disruption (Amato and Kieth 2001b), we considered interaction effects. Specifically, we tested whether the experience of a parental separation or death affects boys and girls differently. We also considered interactions between gender and the child’s age at separation, but none were significant.
The estimates from the marital dissolution model are given in Table 3. All coefficients are in the expected directions and are broadly in line with previous research. For example, the risk of divorce is higher among couples in which the male is poorly educated and those who marry young (Clarke and Berrington 1999; Kiernan and Mueller 1999; Lyngstad 2004). Moreover, second marriages are at higher risk of dissolution, and the risk is further increased if either the woman or her husband has children from a previous relationship. The presence and number of children fathered by the current spouse, however, are associated with longer marriage durations (consistent with Lyngstad ).
Estimates of the parameters associated with the mother-specific random effects in the simultaneous equation model are given in Table 4. These results suggest that there is significant unobserved heterogeneity both between mothers in the chance that their children continue in education and between women in their risk of marital dissolution. Furthermore, there is residual correlation at the mother level between a child’s propensity to continue in education and a mother’s risk of divorce, which is reflected in the estimate of a significant and negative (likelihood ratio = 336.1, df = 1, p < .001) correlation between the random effects for the education and dissolution equations, ρuv: the children of women with an above-average risk of dissolution (vj > 0) tend to have below-average chances of continuing in education (uj < 0).
The estimates of the random-effect loadings λ(r) suggest that unmeasured mother characteristics (represented by uj in the sequential probit model) are less important for transitions to university and to postgraduate study than for transitions through high school. Factors associated with poorer educational attainment at early stages may become less salient after the child leaves the parental home—especially in a context like Norway where the costs of continuing in education are highly subsidized (OECD 2007).
Table 5 shows the estimated effects of parental disruption on educational transitions from two models. In Model 2, the correlation between uj and vj is freely estimated. In Model 1, children’s education and mother’s marital dissolution are modeled as separate processes (i.e., constraining the correlation to equal zero). Both sets of estimates are net of the effects of the covariates given in Tables 1 and and22.
From either model, we would conclude that the experience of marital breakdown during childhood is associated with lower levels of education and that the effect weakens with the child’s age at disruption. However, if the negative random-effect correlation is ignored, the effects of divorce and separation are substantially overstated. The negative effects of divorce estimated from Model 1 are partly explained by a selection into lower levels of education of children whose mothers have a high risk of divorce. Similar to findings from the United States, the effects of divorce seem strongest for transitions during or just beyond the high school level—the first two transitions in our sequential probit model (Biblarz and Gottainer 2000; Mare 1995). The effects of marital breakdown are nonsignificant for the transition from undergraduate to postgraduate education. The only significant interaction between gender and divorce is for the transition from secondary school to university; the negative effect of parental divorce is substantially stronger for girls than for boys, similar to findings reported in Reneflot (2007). That said, the main gender effect is larger; thus, although divorce reduces the probability of going to college for females more than it does for males, the predicted probability of going to college for a female is higher, regardless of her family background, than that for an otherwise similar male.
Children whose parents did not marry, many of whom were born to cohabiting parents who later dissolved their relationship, have a lower chance of progressing through school and on to college, and the parameter estimates for this group are more negative than those for children whose parents marry and divorce. Although this could suggest that the dissolution of (or indeed the failure to ever form) a cohabiting union is more detrimental to children than the experience of a marital separation, selection seems the most plausible explanation for this result. Parents who choose to have an extramarital birth and remain unmarried are likely to differ from parents who marry on unobserved factors that are also associated with their children’s educational outcomes. One way to account for this form of selection would be to extend the multiprocess model by jointly modeling a woman’s decision to marry her child(ren)’s father(s) with educational transitions.
For both types of disruption, the effects appear to be strongest at earlier stages in the educational process, but there are some noticeable differences between the two kinds of family disruption. Although parental divorce is associated with a reduction in a child’s probability of going to college, the effect of a father’s death is not significantly associated with progression beyond the higher secondary level. In addition, neither type of disruption is associated with achieving a postgraduate qualification among those who successfully complete a Bachelor’s degree. Importantly, age at divorce is significantly associated with educational attainment, but age at death is not. These differences make comparing the effects of the two types of disruption less straightforward.
To facilitate the comparison of the results for different types of disruption, and to demonstrate the way in which controlling for selection changes our results, we use the parameter estimates from Models 1 and 2 to construct the predicted probabilities of specific educational transitions, taking into account different family background experiences.3 Figure 1 presents the (unconditional) predicted probabilities for the transition from the lower secondary to upper secondary level for Models 1 and 2.4 When we assume that there is no unobserved heterogeneity bias in the estimated effect of parental divorce on education (Model 1), it appears that although any family disruption reduces the likelihood of making the transition from the lower secondary level, children who experience a parental divorce are slightly more disadvantaged than children who experience the death of their father. Although children who experience divorce at older ages are more similar to those who experience a death, we can see from Table 2 that the average age at which a child experiences any disruption is about 8, so there is a gap of about 5 percentage points in progression at the average age. However, controlling for selection on unmeasured time-invariant mother characteristics narrows attainment gaps between single-parent and two-parent families to some extent. Moreover, differences by type of disruption become almost indistinct. Children who experience a parental divorce at younger ages are slightly less likely than those who experience a death to make the transition from lower secondary school, but those who experience a parental divorce at older ages are somewhat more likely to make this transition. At the average age of disruption, the two effects are nearly equal. The results suggest that although family disruptions are associated with poorer educational outcomes, the type of disruption has little consequence. Nonetheless, it is striking that even in a country like Norway where there is far less economic inequality, there are large and significant differences in the educational attainment of children living with both biological parents and those living with only one.5
If income changes were the primary explanation for lower educational attainment among children who experienced a parental divorce, we might expect to find a compensatory effect when separated parents later reconcile, as well as negative effects of subsequent parental separations. Drawing on sociological (Cherlin 1978), psychological (Hetherington, Bridges, and Insabella 1998), and evolutionary perspectives (Biblarz and Raftery 1999; Case, Lin, and McLanahan 2001) on stepfamilies, we might also expect to find better outcomes among children whose parents reconcile than for those who live in stepfamilies. We find little evidence for any of these. There is no evidence of a compensatory effect in any of the models. In Model 1, we find that having a stepfather is weakly associated with a higher probability of continuing to the higher secondary level (consistent with findings reported by Reneflot ), but these effects are not significant at the 5% level once selection on unobserved mother characteristics is controlled (Model 2). Similarly, after we adjust for selection, the negative effects of separation from a stepfather become insignificant for all but the first transition.
The estimated effects of variables other than the disruption indicators on educational transitions are given in Appendix Table A1 (for Model 2 only). The effects of age, number of younger and older siblings, and parental education are in the expected directions and are consistent with previous literature.
In this article, we set out to examine the relationship between family disruption and children’s educational attainment. We posited that family disruptions in Norway, which has low levels of economic inequality, relatively generous support to lone parents, and highly subsidized educational system, might not be very strongly associated with negative outcomes for children. This hypothesis was strengthened by results from studies in Sweden, which suggest that selection explains almost all of the differences in educational outcomes by family type. In contrast to these expectations, however, our results suggest that although selection is important and works to overstate the effect of divorce on child outcomes, substantial differences between one- and two-parent families remain, even after time-invariant unobserved factors are controlled. Once selection on mother-level unobserved variables is included in the models, children who experience a parental disruption are still 6–13 percentage points less likely to successfully make the transition from lower secondary school and to complete higher secondary education (while the gap is twice as large without controls for selection). Moreover, disruption has the strongest effect early rather than later in the educational career. As a consequence, children who experience a family disruption could face a heightened risk of unemployment and social exclusion later in the life course. If most of the effects were contained at higher educational levels, the life course consequences might be less severe.
In models that did not control for selection, children who experienced a father’s death appeared less disadvantaged than children who experienced a parental divorce at younger ages. However, controlling for selection made differences in the educational attainment between children from divorced and bereaved families narrow substantially, and at mean ages of divorce, almost non-existent. These findings are contrary to work in the United States (work that frequently does not control for selection as we have) that suggests children in bereaved families fare better (Amato and Keith 1991a; Biblarz and Gottainer 2000). Here the large sample size is particularly beneficial; we have data for around 1,900 children who experienced the death of their father. Because few children growing up in wealthy countries experience a parental death, with smaller survey samples, we would not be able to identify differences between the two sources of disruption with much precision. In addition, we can be more confident about our finding for significant differences by age at the time of divorce but no differences by age at the time of the father’s death.
Although our data are not rich enough to distinguish between the different explanations for a significant effect of parental disruption and divorce, the findings point to areas that merit further research. For example, our finding that the reunion of separated parents confers no significant benefits suggests that it is the process of separation and the experience of family transitions (Sun and Li 2007; Wolfinger 2000; Wu 1996; but see McLanahan and Sandefur 1994), and not simply the loss of the parent from a household, that leads to lower educational success.
Our findings also highlight the importance of looking at different stages of the educational process. If we were to estimate a simple probit of university graduation (our third transition) using the whole sample, we would have found significant effects of divorce on that outcome. But our results suggest differences at that level are due, to some extent, to children falling out of the system earlier on. Taking into account the different transitions may be important for understanding the pathways that lead to poor educational outcomes. For example, our finding that divorce at younger ages has the largest effects suggests a prolonged and perhaps cumulative process of disadvantage for children who experience a parental divorce (but not for those who experience a parental death).
Despite its unique sample, our study is not without limitations. Because we can estimate effects of disruption of formal marriages only, we cannot estimate the effect of the separation of cohabiting biological parents. The strong negative effect of the dummy variable for “never married the father” suggests that this group of children did not fare well in terms of education. Additionally, we do not account for cohabiting unions that form after a first disruption and dissolve before turning into marriages. Our finding that the formation and dissolution of stepfamilies (as defined by formal marriage) was not significantly associated with educational attainment might be due to our inability to measure all changes in family structure. This issue has implications for our finding that the effects of parental divorce are strongest at younger ages. To the extent that children who experience a divorce at younger ages are at greater risk of experiencing multiple family transitions, some of which did not involve marriage, part of our age gradient may be due to unobserved risk of multiple family transitions. In addition to these measurement issues, we have not controlled for time-varying unobserved factors or child-specific heterogeneity. If these factors are correlated with parental divorce and educational attainment, some bias may remain.
Finally, our administrative data contain few measures that allow us to distinguish between the different reasons for why divorce is associated with lower educational qualifications. We have no measures on conflict or parenting quality, and the only measure of income we have available is gross labor market earnings. Hence, the parameter estimates for parental divorce reflect both the effects of dissolution through such factors and the effects through other pathways, net of time-invariant selection effects that operate at the mother level. Although we were concerned about using the gross income measure as a control, we estimated models that controlled for the equivilized labor market earnings of the mother and her spouse (if she was married) when the child turned 16. In these models (results available from the first author on request), the estimated effects of disruption on educational attainment fall, as might be expected, but our substantive results are largely unchanged.
Despite these caveats, the results from this study should be of interest to researchers and policymakers. In advanced welfare states where a good deal of resources and support are now provided publicly, it is nonetheless true that what happens in the private sphere remains crucial to the well-being of children. Learning more about what takes place in the home, including the implications of family disruptions, would obviously be important both to researchers and to those involved in the development of (family) policies that can make life better for our children. Although our findings do not settle the dispute about the “true” relationship between divorce and education and cannot shed much light on the underlying mechanisms, they do provide important theoretical and practical information about the relationship between family background and educational success. In particular, our findings suggest the need for future research on the processes that accompany family disruption, particularly in a comparative context, in order to identify ways that children experiencing family disruption can be better supported and encouraged to successfully complete their education.
|Variable||Compulsory to Lower Secondary||Lower to Higher Secondary||Higher Secondary to Bachelor’s Degree||Bachelor’s Degree to Master’s or Higher|
|Age of Child in 2003a|
|Number of Younger Siblings at Age 17|
|2 or more||−0.054**||0.015||−0.092**||0.013||−0.049**||0.016||−0.007||0.042|
|Number of Older Siblings at Age 16|
|2 or more||−0.226**||0.018||−0.298**||0.017||−0.230**||0.026||−0.328*||0.131|
|Education of Mother/Father|
This research was carried out while the authors stayed at the Centre for Advanced Study at the Norwegian Academy of Science and Letters. Sigle-Rushton’s time spent working on this article was supported by Grant No. RES-225-25-2001 from the UK Economic and Social Research Council (ESRC). The support from the Centre and the comments from Henriette Engelhardt and Ronald Rindfuss are greatly appreciated.
1.Björklund and Sundström (2006) constructed earnings-weighted educational outcomes that have more within-family variation than simple educational measures. Nonetheless, only children are still effectively excluded from the sample.
2.For identification, some constraint must be placed on λ(r) or the variance in order to fix the scale of the random effect. Common choices are to fix one of the loadings to 1 or to fix the variance to 1.
3.To calculate the predicted probabilities, we draw 100 random effect values from a normal distribution with mean zero and standard deviation σu. We then use the parameter estimates to calculate predicted probabilities for each of the 100 random-effect values and take the average. The predicted probabilities are calculated for a male aged 25 years in 2003 with one older and one younger sibling and parents who both obtained a medium level of education.
4.The overall patterns described with reference to Figure 1, for the transition for lower secondary to higher secondary school, are similar to those obtained for other transitions, although the gaps by family background are more narrow for the first transition, which almost all children make.
5.The parameters for parental divorce and disruption are similar in size if not slightly larger than those obtained from models estimated using U.S. data. For example, Biblarz and Gottainer (2000: Table 1) estimated logit models of educational attainment similar to our sequential probit model but without random effects. Their models controlled for family structure, gender, race, mother’s education, year, and child’s age. In models of high school completion and college completion, their parameter estimates for living in a single-mother family are −0.74 and −0.41, respectively. Dividing those estimates by 1.6 to approximate probit parameters, we obtain parameters of −0.46 and −0.26. Compared to the estimates for parental separation for Model 1 in the two middle columns of Table 5 (lower to higher secondary level and higher secondary level to Bachelor’s degree), they are both similar in size to those that we obtain for Norway. Comparisons with other studies (see, e.g., Lehrer 2006) and results from meta-analyses (Amato 2001; Amato and Keith 1991b) led to similar conclusions.
FIONA STEELE, Centre for Multilevel Modelling, Graduate School of Education, University of Bristol, Bristol BS8 1TX, United Kingdom; e-mail:ku.ca.lotsirb@eleetS.anoiF.
WENDY SIGLE-RUSHTON, Department of Social Policy, London School of Economics and Political Science.
ØYSTEIN KRAVDAL, Department of Economics, University of Oslo.