Marital Transitions and Children's Adjustment - August 2000

2.1 Sample

The sample for these analyses is based on the families interviewed at Cycle 1 and Cycle 2 of the National Longitudinal Survey of Children and Youth (NLSCY). Across all analyses there were only three exclusion criteria (which involved very few cases). We excluded children who were not living with at least one biological, step or adoptive parent. In addition, we excluded those children for whom the person most knowledgeable about the child and who provided the information (PMK) was neither a biological, step- nor adoptive parent. Third, we excluded those children living with two adoptive parents. The first criterion was established because of our uncertainty about the nature of the families involved; in addition, combining "looked after" children or children in alternative care arrangements would compromise comparability with other studies of family type. The second exclusion criterion was based on our concerns about the reliability of respondent reports and because virtually all previous family studies included data from a biological, adoptive or step-parent. The third criterion was established in order to assure comparability with other studies connecting family type and child well-being. The last requirement resulted in a small, but non-trivial number of families being excluded. There were 137 adoptive families with data at Cycle 1 and Cycle 2 (we inferred that children living with a biological and adoptive parent had been adopted by the stepparent).

The central analyses make use of the longitudinal data for children on whom behavioural data are available. However, there were preliminary analytic questions that required the use of larger subsets of the data. For example, before assessing the effects of parental separation and family transitions, it was first necessary to document the frequency of family change. Thus, for analyses of the rates of change in family type, we included data on all families interviewed at Cycle 1 and Cycle 2. We wanted to estimate the rate and predictors of change on the largest number of families available (N=8,139 families).

In contrast, for analyses of the effects of family type and transitions on children's behavioural development, we had to exclude children who did not have behavioural outcome data. For parent report, this meant excluding children younger than 4 years or older than 9 years at Cycle 1. Although there were self-report data available on children over 11 years of age, there are considerable differences in parent and child reports of behavioural and emotional problems, and it was decided that we could not "equate" the two in our analyses. Thus, the sample size for longitudinal analyses involving parent report of behavioural adjustment was 6,095 children. For those analyses on the effects of separation from Cycle 1 to Cycle 2 we had to exclude those families headed by one parent at Cycle 1. This resulted in a maximum sample size of 5,234 children (4,175 families).

For those analyses based on teacher data the sample size dropped to 3,027 children. Finally, for analyses based on the effects of separation, i.e., when one-parent families at Cycle 1 were eliminated, the sample size was 2,598 children (2,129 families).

The above discussion makes clear that each research question required a different sample. The drop in sample size, which was in many cases substantial, was a result of issues related to the design of the study rather than to missing data as such. Chief among the design considerations was the decision not to include parent report data on children over 11 years of age. This resulted in a loss of important information. The absence of teacher report data for children not yet in school was, of course, unavoidable. In fact, with the exception of the teacher report data, it should be noted that the rate of "truly" missing data was minimal, typically less than 5% for most variables (importantly, the rate of missing data for parent reports of child outcomes used in this report was less than 5%). The only exception to this was teacher data, for which the rate of missing data was substantial at Cycle 1 (a rationale for this, and the greater rate of missing teacher data at Cycle 1 compared with Cycle 2, is given in the published information about the NLSCY design).

However, despite the different sample sizes, the rate of family type (see definitions below) was essentially invariant in each of the central subsamples (i.e., the largest sample of families available at both cycles; the subsample on which parent reported outcomes were available at both cycles; the subsample on which teacher reported outcomes were available at both cycles). That is, when single-parent families were included, the rates of family type were: biological families, 76%; simple stepfamilies, 8%; complex/stepmother stepfamilies, 2%; single-parent families, 14% (the rationale for this categorization is given below). In addition, across the parallel subsamples noted above, the overall rate of parental separation among 2-parent families at Cycle 1 was comparable, slightly less than 5%. Furthermore, the average number of children per family, approximately 1.2, was constant across family type and subsamples. These findings indicate that the subsamples of families were very similar in key respects. This is to be expected because, with the exception of the teacher datafiles, exclusion criteria were based simply on a child's age (i.e., both the oldest and youngest were excluded).

Given the relatively low rate of missing data (with the exception of the teacher data), we used a mean substitution method of replacing missing values for explanatory and outcome variables. For those explanatory variables with greater than approximately 10% missing data (e.g., marital satisfaction) we also defined a dummy variable scored '1' for missing (i.e., those cases for whom a series mean was substituted were scored '1') and '0' for not missing. This dummy variable, when entered into a regression analysis alongside the new explanatory variable (i.e., the one with missing values assigned to the series mean), provides information on whether or not the cases assigned missing values differ from those without missing values on the outcome variable. This procedure also adjusts the estimate of the explanatory variable so that it is not biased by the missing values assigned. This is a standard method of dealing with missing data in developmental research.

2.2 Measures

2.2.1 Definition of family type

We made the following definitions of family type based on prior research and empirical considerations:

a) Biological families² are those in which all children are biologically related to both parents;

b) Stepfather families are those in which at least one child is biologically unrelated to the father, but all children are related to the mother;

c) Stepmother/Complex stepfamilies are those in which at least one child is biologically unrelated to the mother, but all children are related to the father (stepmother); or those families in which at least one child is biologically unrelated to the father and at least one child is biologically unrelated to the mother (Complex stepfamily);

d) Single-parent families are families headed by a non-married, non-cohabiting adult.

It is important to note that two forms of stepfamilies defined above may also include children who are biologically related to both parents (i.e., a child of the new union). The above definitions also do not consider whether or not the partners are married or cohabiting, or have been previously married. These factors are considered separately from family type in our analyses. We initially distinguished between stepmother stepfamilies and complex stepfamilies, and among all stepfamilies according to whether or not there was also a child of the current union living in the home. However, with this more specific categorization there were too few cases to provide reasonable estimates for all analyses. The relative rarity of "atypical" forms of stepfamilies, i.e., stepfamilies other than stepfather stepfamilies, has been noted in investigations of U.S. and U.K. community samples (Haskey, 1996; O'Connor et al., 1999a; Reiss et al., 1994).

For analyses of the Cycle 1 and Cycle 2 outcomes and the change between Cycles, we analyze reports of child adjustment separately for parent and teacher ratings. The models predicting parent reports (cross-sectional and longitudinal) suffer from methodological problems arising from rater bias. That is, information on the predictor and outcome variables were provided by the same respondent. However, this is not the case for predicting teacher reports of behavioural/emotional problems. Using a cross-informant design (i.e., predicting teacher-rated outcome from parent-reported risk factors) is critical if we are to be certain that shared method variance is not inflating the connection between risk and adjustment. The problems on relying on a single reporter for all information are serious and well-known. Therefore, we were especially interested in the degree to which the findings for teacher-reported outcome replicate findings for the parent-reported outcome.

2.2.2 Central outcome variables

Aggression and emotional problems reported by PMK and teacher. For the PMK assessment of aggression at Cycle 1 and Cycle 2, we used the conduct disorder and physical aggression subscale based on factor analyses carried out by Statistics Canada (ABECS09, BBECS09). For the PMK assessment of emotional problems at Cycle 1 and Cycle 2 we used the emotional disorder-anxiety (ABECS08, BBECS08). For the teacher assessment of aggression at Cycle 1 and Cycle 2 we used the conduct disorder and physical aggression subscale (AETCS28A, BETCS28A). For the teacher assessment of emotional problems at Cycle 1 and Cycle 2 we used the emotional disorder-anxiety (AETCS28E, BETCS28E).

Children's aggressiveness and emotional problems were the central outcome variables used in analyses. They are the behaviour scales most often included in research on children's adjustment to family transitions. The results for the remaining two behaviour scales, hyperactivity and indirect aggression, did not offer new insights into the risk and protective factors for children's adjustment. They were therefore dropped from our central analyses.

Both the aggression and emotional symptom scales were skewed, with relatively few cases at the high extreme. After considering several alternatives to transforming the raw data, we collapsed the top 5% of scores. This is analogous to defining a "threshold" point for severe disturbance, which in the current case was defined by the 95th percentile. A consequence of this procedure is that we are not accounting for individual differences in the extreme high end of the distribution. This procedure was performed for both parent and teacher reports. Alternative approaches were considered (log or the square root of the raw score) but were no more effective in producing a normally distributed variable.

2.2.3 Family level and child level factors

Family-level and child-level factors. For empirical and conceptual purposes the variables were categorized as either family level or child level influences. We made this distinction because of our interest in describing both between-family and within-family variation. Of course, this distinction is required in multilevel model analyses. Empirically, the distinction between the two kinds of variables is unambiguous. Those variables for which siblings necessarily receive the same score (i.e., by virtue of living in the same home they must receive the same score) are considered family-level variables. And those variables for which siblings could receive different scores are considered child-level variables. We make clear in our analyses that the "level" at which variables are measured may not equate with the kinds of effects they may have on children. For example, we examine below the question of whether variables measured at the family level have effects only at the family level.

Family-level variables. Several risk factors were measured at the family level, that is, siblings within the same family received the same score for that measure. Specific risks included socio-economic status (variables connected with socio-economic status, such as parental income and education), parental depression, urban setting, family size. These risks indexed psychological, social and economic risk conditions. Unless otherwise noted, these risks were included as continuous variables in regression, repeated measures, and multilevel modeling analyses.

An additional set of variables included in the model indexed developmental risks tied to the parent(s). In this list of factors was the number of previous relationship transitions, the couple's marriage/cohabitation status, and whether or not the parents cohabited prior to marriage. These risks are particularly interesting because, in the vast majority of cases, they precede the current family type or even the child's birth. Empirically, these risks are defined as family-level risks because children in the same family would be assigned the same score. However, they are conceptually very different from the set of family-level risks identified in the previous paragraph.

Socioeconomic status. This variable was calculated by Statistics Canada and was based on the education and occupation of the PMK and spouse (if relevant) and household income (AINHD08). Occupation was coded using the Pineo socio-economic classification.

Parental depression Cycle 1. This was measured using a modified version of the CES-D (Radloff, 1977). The PMK was asked about depressive symptoms including mood, sleeping, crying, depressive cognitions and poor appetite (ADPPS01). There are 12 items in the scale. The range is from 0-36. The internal consistency of the scale was good (Crobach's alpha=.82).

Urban setting Cycle 1. Interviewers made a coding of the size of the size of community in which the family lived (AGEHD01). Urban codes ranged from (1) which was an urban area with a population of over 500,000 people to (5) which was an urban area with a population of less than 15,000. There was also a code (6) for rural area. We created a dummy variable of urban/rural in which all urban areas were categorized together (codes 1-5) and contrasted to rural.

Previous relationship transitions. Data from the custody files was used to create a score for the number of previous marital or live-in relationships experienced by the mother and father prior to the current union.

Cohabitation status. In addition to distinguishing whether the couple heading the family was married or cohabiting, we also included information on whether the couple cohabited before marriage. The latter information was available from the custody datafiles.

Child-level variables. Several risk factors were measured at the child level, that is, siblings within the same family had (potentially) unique scores. Child-level risks were included in the models if there was evidence for its association with children's behavioural adjustment. Specific child-level risks included age, gender, parenting quality, friendship quality and violence in the home. Although there is evidence that children receive similar levels of parenting, suggesting that it might operate in a family-wide manner in some cases, there is also evidence that child-specific or differential parenting underlies within-family differences in child behavioural disturbance (Reiss et al., 1995). Accordingly, we used the child-specific measure of parent-child relationship quality.

Violence in the home. The PMK was asked whether and how often the child had witnessed violence between two adults in the home (APRCQ28). "How often does NAME see adults or teenagers in your house physically fighting, hitting or otherwise trying to hurt others." This is rated on a 4 point scale from often (1) to never (4); thus, higher scores index less violence. Importantly, this is measured separately for each child in the family.

Ineffective and positive parenting. In the NLSCY the PMK was asked to rate him/herself on a five point scale on a range of parenting variables describing affection in the parent child relationship, positive interaction, punishment and hostility. This was factor analyzed and three factors emerged: hostile/ineffective (APRCS04), consistency (APRCS05) and positive involvement (APRCS03). The hostile/ineffective (hereafter referred to as "ineffective") scale was made up of the following items: annoyance, anger, disapproval, lack of praise, difficulties managing the child, parental moodiness affecting punishment and ineffective punishment. Internal consistency of this scale was good (Cronbach's alpha=.71). The positive involvement scale was made up of: praise of the child, talk or play focusing attention on the child for 5 minutes or more, laughing with the child, doing something special together that the child enjoys, playing sports or hobbies together. Internal consistency for this scale was adequate (Cronbach's alpha=.81). The parental consistency scale was not used in the analyses presented in this report.

Relationship with friends and siblings at Cycle 1. The PMK was asked about the quality of the child's relationship with their friends and their sibling at Cycle 1. The wording of the questions were as follows: "During the past 6 months how well has NAME gotten along with other kids, such as friends or classmates -excluding brothers or sisters?" (ARLCQ06). "During the past six months how well has he/she gotten along with his/her brother (s)/sister (s)?" (ARLCQ09). Each of these questions was rated on a five point scale from very well, no problems, to not well at all, constant problems. Higher scores thus index more problems in the relationship. These items were adapted from the Ontario Health Study.

2.3 An overview of the data analytic approach

Prior studies of the risk and protective factors for children's behavioural and emotional problems typically include only one child per family. As a result, the effects attributable to family-level factors (e.g., parental psychopathology), individual child-level factors (e.g., age, gender), and the interaction between the two are completely confounded. A novel feature of this study is the use of multilevel modeling, an analytic approach that capitalizes on the nested or hierarchical structure of family data. This approach partitions variation attributable to each "level" in the data structure. That is, we are able to distinguish between risk and protective factors that operate at the family level (which explain why families differ from one another, or between-family variation) from those that operate at the individual child level and explain why individual children differ from one another (which we term within-family variation).

Multilevel modeling (Bryk & Raudenbush, 1992; Goldstein, 1995) is designed for hierarchically organized data at a potentially infinite number of levels, such as children within classrooms within schools, or, as in the present case, children within families. Three features of the multilevel model results are highlighted. First, we present the fixed effects associated with the predictor variables. These estimates and standard errors are interpreted as in a regression model; an estimate that is approximately twice its standard error has a significant (p < .05) association with child behavioural and emotional problems.

The novel feature of multilevel modeling, the partitioning of variance into each "level" of the data, is also provided. Error variance is decomposed into family-level ("between- family") and individual child-level ("within-family") variability. These are referred to as "random effects". Estimates for the fixed and random effects are simultaneously calculated using a maximum likelihood procedure, the value of which is reported. It is important to note that the estimates included in the random effects part of the tables are not interpreted in the same way as the estimates for the fixed effects. The estimates in the random effects section are estimates of variance (with associated standard errors) rather than traditional regression coefficients.

In addition to providing potentially new insights into the risk and protective factors associated with behavioural and emotional problems following family transitions, the use of multilevel modeling handles the analytic problems arising from correlated errors when multiple children from the same family are included in analyses. Analyzing these data using conventional statistical tools and programs would result in biased standard errors and potentially misleading findings.

Throughout the results section we consider both statistical significance of the findings as well as the magnitude of the findings, or effect size. Effect size, d, was defined by Cohen (1968) as the mean difference between groups divided by the pooled standard deviation. As a general rule, effect size values of .2, .5, and .8 indicate small, medium, and large effects, respectively. Given the large sample size for most analyses (although the number of complex/stepmother families is relatively small) findings of a trivial effect could nonetheless be statistically significant.

Analyses using the multilevel method are based on weighted data using the weighting procedure in the most recent version of MLwiN (beta version 1.10.0001; Goldstein et al., 1998; Rasbash, Browne, Healy, Cameron, & Charlton, 1999). In this case, because the weights were assigned to individual children in the NLSCY, the weighting procedure in MLwiN analyzed the data with child-level weights (referred to in this case as Level 1). There was no comparable set of weights for the family-level analyses, notably analyses on the rates of family type change. Longitudinal weights are used in all analyses involving (only) the longitudinal sample. The cross-sectional (Cycle 1) weights were used in the final set of cross-sectional analyses on the Cycle 1 data.

• ²Henceforth, we use the term "biological families" simply to denote that in these two-parent families all members are biologically related to one another. We prefer this term to the somewhat more (or at least equally) awkward terms such as "intact", "nuclear", "non-divorced", and "non-stepfamily" families.