Two Methods for Studying the Developmental Significance of Family Structure Trajectories

Abstract

Objective:

The objective of this research note is to use both sequence analysis (SA) and repeated-measures latent class analysis (LCA) to identify children’s family structure trajectories from birth through age 15 and compare how the two sets of trajectories predict alcohol use across the transition from adolescence into young adulthood.

Background:

Contemporary family scholars have studied the influence of changes in family structure, often referred to as family structure instability, on child and adolescent development. Typically, this research has focused on either the number or type of transitions children have experienced, but statistical advances are increasing the viability of more complex person-centered approaches to this issue, such as SA and LCA. The choice to use one approach or the other, however, is often discipline specific and relies on different assumptions and estimation techniques that may produce different results.

Method:

The authors used data from the National Longitudinal Study of Youth–Child and Youth Cohort (N = 11,515) to identify clusters (using SA) and classes (using repeated-measures LCA) that represented children’s family structure trajectories from birth through age 15. Using two multiple-group random slope models, the authors predicted alcohol use across adolescence and young adulthood (ages 16–24) among the clusters (Model 1) and classes (Model 2).

Results:

The SA identified five clusters, but the LCA further differentiated the sample with more detail on timing and identified eight classes. The sensitivity to timing in the LCA solution was substantively relevant to alcohol use across the transition to young adulthood.

Conclusion:

Overall, the SA is perhaps more suited to research questions requiring exclusive group membership in large, comparative analyses, and the LCA more appropriate when the research questions include timing or focus on transitioning into or out of single-parent and stepfamily homes.

The links between family structure—the living arrangements of young people, generally defined in terms of their parents’ formal and informal unions—and children’s developmental outcomes have long been a source of debate among policymakers. In recent decades, family scholars have documented these links and what they mean for young people, families, and society at large (e.g., Cherlin, 2009; McLanahan, 2004). Within this theoretically grounded and policy-relevant literature, one key innovation has been the conceptualization of family structure instability—rather than family structures themselves—as a context of the development of maladjustment and problematic functioning across the early life course (Wu & Martinson, 1993). This conceptualization necessitated the operationalization of family structure changes, with initial research counting the number of family structure transitions a child experienced over time followed by attempts to capture more complex longitudinal configurations of various family structures (Crosnoe & Cavanagh, 2010). Different operationalizations, however, are based on their own assumptions and may lead to divergent results, which can challenge attempts to draw general conclusions about the nature of family instability and its developmental significance. The fact that different approaches may track disciplinary lines further complicates this challenge.

This research note, therefore, attempts to provide some assistance for making sense of different operationalizations of family instability across the life course and choosing an approach that is appropriate to particular research questions. Among the benefits of person-centered analyses is the ability to group individuals together based on multiple shared characteristics. Specifically, this study compares the results of efforts to measure family structure trajectories from birth into adolescence with two person-oriented statistical methods for capturing longitudinal configurations of categorical variables: (a) sequence analysis (SA), which is more common in the sociological tradition of family studies (Abbott, 1995; Fasang & Raab, 2014), and (b) latent class analysis (LCA), which is more common in the psychological tradition (Lanza & Rhoades, 2013). This comparison also extends to analyses of the links between such family structure trajectories and a key behavioral dimension of the transition from adolescence into young adulthood: alcohol use. This comparative framework is tested with data from the National Longitudinal Survey of Youth 1979–Children and Young Adults cohort (NLSY-CYA; https://www.nlsinfo.org/content/cohorts/nlsy79-children).

Conceptualizing and Operationalizing Family Structure Histories

Early family structure research on children focused on between-group differences (e.g., two-parent vs. one-parent families), followed by research highlighting the value of within-group approaches (e.g., diversity among single-parent families or stepfamilies; Colemen, Ganong, & Fine, 2000; Marks & McLanahan, 1993; Heatherington, 1992). More recent studies on family instability have built on both approaches, usually by examining the number of family structure transitions—regardless of family structure type—that children experience during key developmental periods. This research has shown that more family structure transitions tend to be negatively associated with young people’s academic, psychological, and behavioral adjustment (Cavanagh & Fomby, 2012; Osborne & McLanahan, 2007; Turney & McLanahan, 2015).

Yet, family structure trajectories can be conceptualized more holistically in terms of the number of family structure changes among different types of family structures across the full arc of childhood and adolescence. For example, experiencing a parental divorce while young, living in a single-parent family, and then having a stepparent enter the home and being born to a single parent before experiencing the entrance and exit of a stepparent across adolescence are two histories with the same number of family structure transitions that could have different developmental implications (Cavanagh & Fomby, 2019). That conceptualization is much more methodologically demanding. SA and LCA, which classify a sample by longitudinal configurations of some binary or categorical status, are two methods that can meet those demands (Barban & Billari, 2012). Comparing the two can help researchers adjudicate between them for different kinds of questions.

First, SA is an analytical method that has been used extensively by life course researchers, most often from sociology (Abbott, 1995; Aisenbrey & Fasang, 2010). Individual elements—categorical variables—describe a person’s “state” during a particular time, such as married, single, or cohabiting, to create life course sequences. For example, a person may be married for several years, followed by a long period of singlehood and then a new cohabitation, with a divorce signified by the transition between the married and single states. SA’s main purpose is to determine how dissimilar one’s life course is from another. There are several methods to conduct SA; the most common when transitions within trajectories are of interest is optimal matching (OM). OM essentially asks the following: How many insertions, deletions (indels), or replacements of states would it take to match two sequences to one another (Abbott & Tsay, 2000)? OM is useful for family scientists in part because of its flexibility. Scientists can choose to prioritize indels to focus on duration in a state or substitution to focus on timing and order. Subsequence-based metrics, as opposed to OM’s edit-based metrics, can also be used in sequence analyses, but OM has routinely outperformed when the life course sequences are related to family formation (Aisenbrey & Fasang, 2010; Han, Liefbroer, & Elzinga, 2017).

Second, LCA is a person-centered analysis that has become a popular method for developmentally oriented family researchers, particularly from psychology (Han et al., 2017). The logic for repeated-measures LCA is the same as SA—examination of a categorical variable observed over time. In contrast to SA’s dissimilarity measures, LCA is a mixture model that uses probability distributions to identify latent classes. Unlike SA, which exclusively “assigns” each person to one group, LCA proportionally assigns people based on their membership and item response probabilities (Lanza & Cooper, 2016). For example, one person could be 82% likely to be in Class 1, 18% likely to in Class 2%, and 0% likely to be in other classes (Masyn, 2013). SA, however, would exclusively assign that person to Class 1. This study focused on the differences in estimation techniques between SA and LCA, although we extracted and hard classified the LCA classes based on their most likely modal class for additional multivariate analyses.

The years between early childhood and adolescence offer a potentially valuable window for comparing these approaches as they capture a period a maximum “vulnerability” to experiencing family structure changes and allow consideration of the developmental timing of such changes (Elder, 1998). The two approaches may reveal different dimensions of family structure trajectories across this period that then differentially predict the transition into adulthood. Alcohol use is an ideal developmental domain for that comparison. Drinking is a highly normative behavior during this transition, but this elevated level subsumes substantial variation in initiation and escalation with strong implications for whether alcohol use will eventually plateau and drop as adulthood progresses or instead persist into alcohol abuse. It also has complex associations with family structure, suggesting that different family structure trajectories growing up can lead to differences in young adult alcohol use with meaningful implications for the later life course (Cavanagh, 2008; Schulenberg & Maggs, 2002).

Goals of Research Note

In the general spirit of a research note (i.e., exploring new empirical findings to guide future research in an important area), our first goal was to use both SA and repeated-measures LCA to identify children’s family structure trajectories from birth through adolescence with the same data. Do the OM assumptions of the SA and the proportional based assumptions of the LCA lead to different insights about the most common types of family structure trajectories? The second goal was to use these two sets of family structure trajectories to predict patterns of alcohol use across the transition from adolescence into young adulthood. If the two methods reveal different kinds of family structure trajectories, how do those differences inform our understanding of the transition into adulthood?

Data and Methods

The National Longitudinal Study of Youth, 1979 Cohort (NLSY79) is a nationally representative sample of 12,868 youth born in the United States followed annually from 1979 to 1994 and then biennially until 2014. The NLSY-CYA includes the 11,521 biological children born to women in the NLSY79 sample. NLSY79 mothers completed assessments for each child until the children were 10 years old; at age 10, the children began to provide their own information. Beginning in 1994, children aged 15 and older completed interviews modeled after those given to their mothers in the NLSY79. We used data from each available round of the NLSY-CYA. Because the NLSY-CYA children were born in various years, we restructured the data from years to age with all NLSY-CYA participants who had available data from ages 1 to 15 for the family structure analyses and for alcohol use from ages 16 to 24. Youth without family structure data in any of the 15 years were dropped from the analysis (final N = 11,515).

Measurement

Alcohol use in young adulthood.

From ages 16 to 24, the respondents reported their drinking frequency in the last year on the following scale: 8 (daily), 7 (3 to 6 days a week), 6 (1 or 2 days a week), 5 (several times a month), 4 (1–2 times a month), 3 (every other month), 2 (3–5 days in the past 12 months), 1 (1–2 days in the past 12 months), and 0 (none).

Family structure in childhood and adolescence.

The NLSY-CYA family structure variables were coded from information about the residence of the child and their mother’s romantic partner(s). The residence variables were assessed at each year and provided information on the child’s typical residence (reported by the mother). Mothers also identified each person in her household and their relationship to her. For each child, she also reported if any romantic partner living in the household was the child’s biological or adoptive father. We created a yearly indicator of children’s living arrangements from ages 1 until 15 that included living with a single parent, living with both biological or adoptive parents, living with a biological parent and a step (married) or social (cohabiting) parent, and other living arrangements (e.g., other relatives, foster care, etc.). Children were captured as living with both biological or adoptive parents if both parents reported living with the child at the survey year. To be included as a single parent at each survey year, children lived with a biological or adoptive mother only, biological or adoptive father only (1.92%), part-time with a biological mother and part-time with a biological father (0.19%), or part-time with a biological mother and part-time with other caregiver (0.03%). Some children who had a parent who was incarcerated or in the military were measured as living with a single parent. Adoptive parents must have adopted their children within 12 months after birth to be included with biological parents in all categories.

Covariates.

Demographic information from the first wave of data collection included child gender (51% girls; 49% boys), racial/ethnic identification (19% Latinx, 28% African American, 53% White), maternal education, and maternal age at birth of first child (M = 25.2, SD = 5.9). Maternal education was measured from mother’s reports of her highest degree obtained at each survey year: 1 = “no high school diploma” (11%), 2 = “high school diploma or GED” (42%), 3 = “some college” (26%), and 4 = “college or higher” (20%).

Plan of Analyses

The SA used OM distance metrics to estimate dissimilarity between family structures from year to year among children. We chose to prioritize substitutions over indels to maintain temporal order within the sequences (Aisenbrey & Fasang, 2010). Distances were calculated using Ward’s clustering method, an iteration process that combines similar sequences in increasingly large clusters (Han et al., 2017). Missing data for the SA was accounted for using multiple imputation chained equations in Stata 15.1 (StataCorp, College Station, TX; Allison, 2000). The measurement model of the repeated-measures LCA used the same family structure variables across waves to inform a latent variable of family structure states based on an expectation maximization algorithm and maximum likelihood estimation. Multiple sets of parameter estimates were iteratively tested until a maximum number of iterations were reached or the model converged (Masyn, 2013). LCA models differ from a factor analysis in that LCA is person oriented, identifying classes by people who share similar characteristics, whereas factor analysis is variable centered, identifying factors based on similar items.

Missing data in the LCA were accounted for using full information maximum likelihood. The estimation methods for handling missing data is another difference between SA and LCA that may lead to different group identification. Both estimation methods provide unbiased parameters estimates and standard errors, but multiple imputation (MI) requires user-identified variables to inform the equation. SA models were estimated in R3.4.4 using the package TraMineR (Gabadinho, Ritschard, Muller, & Studer, 2011; R Core Team, 2013). LCA models were estimated in Mplus 8.1 (Muthen & Muthen, 1998–2017). Both the SA clusters and the LCA sequences were then exported back into Stata, generating demographics statistics for the overall sample, cluster, and class (SA in Table 1; LCA in Table 2).

Table 1.

Descriptive statistics for the family structure history clusters

Study variables	All, N = 11,515	Consistently two-biological-parent home n = 5,427	Consistently single-biological-parent home n = 2,705	Transitioning from two-parent into single-parent home n = 1,722	Transitioning from single-parent into stepfamily home n = 1,417	Transitioning from parental care into nonparental care n = 244
Child gender (female), %	51	48	48	49	52	47
Race/ethnicity, %
Latinx	19	19	15	25	21	23
African American	28	14	56	25	28	28
White	53	67	29	50	51	48
Maternal education, %
Less than high school	11	8	15	11	14	26
High school	42	36	49	43	48	40
Some college	26	27	25	27	26	29
College degree or more	20	29	11	19	13	5
Maternal age at 1st childbirth, M (SD)	25.2 (5.9)	26.8 (5.6)	23 (5.8)	25 (6.2)	21.9 (4.8)	22 (5.4)

Table 2.

Descriptive statistics for the family structure history classes

Study variables	All, N = 11,515	Consistently two-parent biological n = 4,305	Consistently single-parent biological n = 1,696	Early transitioning into single-parent biological n = 1,205	Late transitioning into single biological n = 1,104	Early transitioning into stepfamily n = 899	Late transitioning into stepfamily n = 748	Returning to two biological parent n = 1,166	Transitioning from parental into nonparental care n = 392
Child gender (female), %	51	48	48	46	49	53	53	51	48
Race/ethnicity, %
Latinx	19	18	14	19	25	20	23	25	22
African American	28	13	59	41	19	25	34	26	40
White	53	69	27	41	56	56	43	50	38
Maternal education, %
Less than high school	11	8	14	12	7	14	12	17	27
High school	42	37	50	47	35	48	46	40	41
Some college	26	27	25	26	27	25	27	26	27
College +	20	29	8	14	30	14	14	18	4
Maternal age at 1st childbirth, M (SD)	25.2 (5.9)	26.6 (4.9)	24.6 (5.7)	23.9 (6.4)	28.5 (5.5)	22.2 (4.6)	22.1 (5.0)	24.8 (7.8)	21.8 (5.6)

Next, multiple-group random slope regression models in Mplus 8.1 examined family structure group differences in alcohol use intercepts and slopes from ages 16 to 24, controlling for child gender and mother age, education, and race/ethnicity. Covariates were added to the regression models after classes (LCA) and clusters (SA) were identified. We estimated two separate models (one by clusters and one by classes). We chose to use a multiple-group approach so that we could assess the alcohol consumption trajectories of all classes and clusters. The consequence of using a multiple-group approach is that the classes and clusters were not compared to one another but to zero. Consequently, we also conducted Wald tests to determine whether intercepts and slopes significantly differed from each other.

These methods estimate either clustered (i.e., SA) or latent (i.e., LCA) groups of people based on similarities in response patterns across a particular set of categorical items. In SA and repeated-measures LCA, groups are based on responses to a single categorical item that is repeated over time. Thus, SA and LCA take into account, but may or may not capture, observed family structure transitions; instead, the changes in states or probabilities of family structures across time suggest probable family structure transitions. We chose to label the sequences and latent classes by the predominant pattern and trajectory of family structure transitions across ages because we believe that the use of this language is in line with the conceptualization of family structure trajectories and because doing so is more easily interpretable.

Results

Family Structure Trajectories From SA

For the first goal, five family structure sequences emerged from the SA based on conventional standards for three cluster quality statistics (Hubert & Levin, 1976; Rousseeuw, 1987). Specifically, they indicated a moderate distinction between clusters (average silhouette width = 0.52), good ability to reproduce the original distance matrix (point bi-serial correlation = 0.77), and a good partition gap (Hubert’s C index =0.04). Figure 1 visually presents the proportion of youth in that type of family structure for that given year in each cluster based on information about the frequency and mode of sequences and mean time children spent in each state during the 15-year period. Table 1 presents more detailed demographic information for each sequence.

Figure 1.

Representations of the Family Structure History Clusters.

First, children in the consistently two-biological-parents home sequence (n = 5,427) were largely White (67%), and their mothers stayed married to their fathers. Mothers had the largest percentage of college degrees and had their first child later in life than mothers of children in other sequences. Children who followed this sequence experienced few family structure transitions.

Second, children in the consistently single-biological-parent home sequence (n = 2,705) spent most of their childhood with only their biological mothers, without experiencing many changes in their family structures. They were predominately African American (56%), and a majority of their mothers had a high school diploma, GED, or some college. Their mothers had their first child in their early to mid-20s (M = 23, SD = 5.8).

Third, children in the transitioning from two-biological-parent into single-parent homes sequence (n = 1,722) lived in more family structures than children in the first two sequences. They were predominately White (50%), but the largest percentage of Latinx children (25%) were in this sequence when compared with other categories (one quarter were African American). The majority of their mothers had a high school diploma (43%), some college (27%), or a college degree (19%). Their mothers first had children in their mid-20s.

Fourth, children in the transitioning into stepfamily home sequence (n = 1,417) lived in either a two-parent-biological or a single-mother family structure in early childhood and in a stepfamily in their adolescent years. Race/ethnicity mirrored the overall sample (21% Latinx, 28% African American, 51% White). A majority of mothers in this sequence had a minimum of a high school diploma (48%) and were among the youngest when they had their first child (M = 21.9, SD = 4.8).

Fifth, children in the transitioning from parental to nonparental care sequence (n = 244) lived with either both or one biological parent early in life and then later were in nonparental care. Most lived with other relatives (e.g., grandparents) or in foster care. This sequence included a large percentage of Latinx children (23%), and most mothers did not have any post–high school education (66%). Mothers had their first child in their early 20s (M = 22, SD = 5.4).

Thus, the person-oriented approach in SA identified clusters of young people who were simultaneously similar and different. There were groups who lived in two-biological-parent homes but differed in how long and if such family structures were followed by family structure transitions. The same can be said about single-parent homes. There also were groups who mixed living with parents and nonbiological parents but differed in their exact relationships.

Family Structure Trajectories From LCA

The second method identified eight family histories. The Lo–Mendell–Rubin was significant through eight classes, as was the bootstrapped likelihood ratio test (likely due to sample size). The Bayesian information criterion, Akaike information criterion, and log likelihood continued to trend downward from the two-class model. After the four-class model, the smallest class remained stable at 3%. Based on the model fit information and previous empirical research, we chose the eight-class model. Each class still had a moderate number of people (smallest class n = 391), and this model allowed us to look at the timing of transitions and maintain a good-fitting model. Figure 2 visually presents the proportion of youth in that type of family structure for that given year in each of the eight latent classes, and Table 2 presents more detailed demographic information for each one. We chose this representation of the LCA classes (vs. the individual sequences from the SA) to highlight the different ways in which each are analyzed.

Figure 2.

Representations of the Family Structure History Latent Classes.

First, the consistently two-biological-parents home class was the largest (n ~ 4,305; note that the n for each LCA class is approximate because people are proportionally placed in each one). This class consisted of children who spent the majority of their childhood with both their biological mother and father. It had a majority of White children (69%). Children’s mothers were among the highest educated (29%) and the oldest at the birth of their first child.

Second, children in the consistently single-biological-parent home class (n ~ 1,696) mostly lived with their mother, but no biological father or stepfather. A small number experienced a transition in early childhood but subsequently remained stable. Children in this class were predominately African American (50%), and a very small percentage of their mothers had a college degree (8%). Mothers had their first child in their mid-20s (M = 24.6, SD = 4.0).

Third, children in the early transitioning into a single-biological-parent home class (n ~ 1,205) were likely to live in a two-biological-parent home in early childhood and in a single-parent home in middle childhood and beyond. Children in this class were mostly African American (41%) or White (41%). Most mothers had either a high school diploma (47%) or some college experience but no college degree (26%), and they had their first child in their mid-20s (M = 23.9, SD = 6.4).

Fourth, the children in the late transitioning into a single-biological-parent home class (n ~ 1,104) were likely to live in a two-biological-parent home in early and middle childhood and in a two-single-parent home in adolescence. Maternal education was similar to mothers in the first class, with 30% having a college degree. A disproportionate percentage of Latinx children (25%) were in this class. Mother in this class were the oldest at first childbirth (M = 28.5, SD = 5.5).

Fifth, children in the early transitioning into stepfamily family home class (n ~ 899) were likely to live in either two-biological-parent or a single-biological-parent home at birth than in a family structure consisting of a biological mother and stepparent in early childhood and beyond. Children’s race/ethnicity in this class mirrored the overall sample (more likely to be White). Children’s mothers were in their early 20s at first childbirth (M = 22.2, SD = 4.6). Most had a high school education (48%), and 13% had a college degree.

Sixth, children in the late transitioning into stepfamily home class (n ~ 748) were likely to live in either a single-parent or two-biological-parent family structure in early childhood and in a stepfamily in early adolescence and beyond. Most of their mothers had a high school degree (47%) or some college exposure (26%), and they had their first child in their early 20s (M = 22.1, SD = 5.0). By age 15, adolescents in this class experienced more family structure transitions than other adolescents in other classes (M = 4.21, SD = 2.46).

Seventh, children in the returning to two-biological-parent home class (n ~ 1,166) were likely to live in single-parent and stepfamily structures in childhood and in a two-biological-parent home in adolescence. A quarter of children in this class were Latinx, and 40% of children’s mothers had a high school diploma. This class is composed of almost 10% of the sample and represents children whose parents separate and repartner with each other for a number of reasons, often referred to as churning. Churning relationships have previously represented up to 16% of the sample, but most studies have only examined parental churning during the first 5 years of the focal child’s life (Halpern-Meekin & Turney, 2016; Turney & Halpern-Meekin, 2017). Furthermore, this class includes children who experienced parental incarceration (2.4%) and likely includes other possibilities, such as parental deployment or redeployment and work separation (e.g., commercial fisherman and migrant farmers) not measured in the data.

Eighth, children in the transitioning from parental care to nonparental care class (n ~ 392) were likely to transition from parental care early in life into nonparental care, typically relative care or foster care, during childhood and beyond. Children in this class were predominately African American (40%). Very few of their mothers had a college degree (4%), and they were among the youngest at first childbirth (M = 21.8, SD = 5.6).

Similar to the SA, therefore, this person-oriented approach identified groups of young people who were similar in some ways and different in others. The primary distinctions among classes involved type of family structure, consistency of a family structure versus family structure transitions, and the timing of transitions. Two groups might be similar on one or even two of these dimensions but differ on the third (e.g., two groups lived in stepfamilies but entry at different times, multiple groups involved transitions out of two-biological-parent structures but differed in the timing and type of the subsequent family structure).

A Comparison of Methods

A clear distinction from the SA results was the number of groups identified by LCA (five vs. eight). We chose five clusters, instead of continuing toward eight clusters to match the LCA because the fit statistics demonstrated a poorer quality of fit after five clusters. We did not want to arbitrarily run the same number of clusters and classes but, rather, identify the number of clusters and classes based on fit statistics as though we did not have the other model as a comparison. Substantively, this difference emerged in that the LCA differentiated trajectories by emphasizing the timing of family structure transitions. For example, both methods identified trajectories involving transitions into single-parent and stepfamily structures, but the LCA further differentiated these histories by whether the transition—inferred from changes in family structures—occurred early or later in childhood. Notably, the LCA also identified a new family structure trajectory, one involving a return to living with two biological parents.

Calculating an adjusted Rand index score concretely assessed how much the two solutions differed. This score did not provide information about which solution was “better” but, rather, about how much overlap there was between them (Barban & Billari, 2012; Han et al., 2017). Despite the difference in the number of groups, the 0.61 Rand value suggested that the two solutions were comparable. Given the comparable cross-classification, the decision of which solution was the best option remained in the realm of construct validity and the subsequent research question. Thus, we conducted additional comparisons of the methods in Stata 15.1.

Cross-tabulations between the sequences and classes provided insight into how people from the five sequences were distributed across the eight classes. The consistently two-biological-parent home sequence dispersed approximately 1,100 people into the late transitioning into single-biological-parent home (n = 347) and returning to two-biological-parent home (n = 773) classes. Based on this dispersion, the LCA revealed slightly more movement in the modal group of young people than did the SA. The transitioning from two-parent to single-parent home sequence split among several classes, including the early transitioning into single-parent-biological home (n = 452), late transitioning into single-parent-biological home (n = 756), and returning to two-biological-parent home (n = 362), suggesting that the sequences masked movement among children’s biological parents as well as information regarding timing into new family structures. As another example, the LCA dispersed 640 children from the consistently single-parent sequence into the early transitioning into single-biological -parent home class and 216 children into the late transitioning into stepfamily home class. Children in the transitioning from single-parent into stepfamily home sequence were largely divided among the early transitioning (n = 884) and late transitioning into stepfamily (n = 394) classes, suggesting that the LCA revealed more information about the timing of movement from one family structure into another.

Overall, the advantage of the SA solution was the larger sizes of each group, offering more power for future analyses, and the groups’ mutual exclusivity. The advantage of the LCA solution was the emphasis on timing into, and out of, certain family structures, which injects more evidence of instability into the seemingly more stable histories identified by the SA.

Linking Family Structure Trajectories to Young Adult Outcomes

For the second goal, multiple-group random slope regression models estimated latent growth curves of alcohol use. These models calculated both an intercept (i.e., initial level at 16) and a slope (i.e., change from 16–24) of alcohol use for each cluster and then for each class.

Beginning with the intercept, the growth curves grouped by sequence clusters indicated that young people in the transitioning from parental to nonparental cluster had the highest starting point of alcohol use (B = 4.65, p < .001), followed by those in the consistently single-parent-biological home (B = 2.77, p < .001), transitioning into stepfamily (B = 2.73, p < .001), transitioning into single-parent-biological family home (B = 1.70, p < .001), and consistently two-parent-biological family home (B = 1.00, p < .001) clusters. The growth curves grouped by latent classes indicated that young people who were in the transitioning from parental to nonparental class had the highest starting point (B = 3.99, p < .001), followed by consistently single-parent-biological home (B = 2.73, p < .001), late transitioning into stepfamily (B = 2.60, p < .001), early transitioning into single-parent-biological home (B = 2.55, p < .001), returning to two-parent-biological home (B = 2.55, p < .001), early transitioning into stepfamily (B = 2.32, p < .001), consistently two-parent-biological home (B = 0.85, p < .001), and late transitioning into single-parent-biological home (B = 0.79, p < .10) classes. The biggest differences between the results for clusters and classes were among young adults who had transitioned into a single-parent home or into a stepfamily home, with analyses based on LCA revealing timing-related differences in both.

Turning to the slope, growth curves grouped by sequence clusters revealed that the steepest gains in alcohol use was among young adults who spent the majority of their time in consistently two-parent-biological homes (B = 0.38, p < .001), followed closely by those in the transitioning into single-parent-biological home (B = 0.31, p < .001), consistently single-parent-biological home (B = 0.27, p < .001), transitioning into a stepfamily home (B = 0.26, p < .001), and transitioning from parental to nonparental care (B = 0.16, p < .05) clusters. Growth curves grouped by latent classes of alcohol trajectories revealed a similar slope for those in the consistently two-parent-biological home (B = 0.40, p < .05) as did the comparative cluster. The next highest slopes were for young adults in the late transitioning into single home (B = 0.33, p < .001), followed by young adults in the consistently single-parent-biological home (B = 0.28, p < .001), early transitioning into stepfamily home (B = 0.28, p < .001), returning to two-parent-biological home (B = 0.28, p < .001), early transitioning into single-parent home (B = 0.27, p < .001), late transitioning into stepfamily home (B = 0.26, p < .001), and parental into nonparental care (B = 0.23, p < .10).

Combining the intercept and the slope, we see some consistencies and some changes in young adult alcohol use over time within clusters (Figure 3a) and classes (Figure 3b). In both, young adults who transitioned from parental care to nonparental care started drinking at higher levels than their peers and continued doing so up to age 24. The other clusters were quite similar to each other. In all clusters, drinking increased with age, and the consistently two-parent-biological home cluster caught up to the transitioning into single-parent-biological home cluster by age 24. The various classes also demonstrated increased alcohol use during the focal window of time, but some classes diverged from similar starting levels. For example, young adults in the consistently two-parent-biological home and the late transitioning into single-parent home began drinking at similarly low levels, but those in the former increased their drinking levels at faster rates than those in the latter.

Figure 3.

(a) Growth Curve Models of Alcohol Use, as Predicted by Family Structure Histories Derived From Sequence Analysis. (b) Growth Curve Models of Alcohol Use, as Predicted by Family Structure Histories Derived From Latent Class Analysis.

To directly compare the models using the variables from each of the two methods, we calculated Wald tests of significance. Recall that some young people in the consistently two-parent-biological cluster were distributed, in part, to the returning to two-parent-biological home class. Those in the returning to two-parent-biological home class reported a statistically higher average use of alcohol at age 16 (B = 2.55) than the consistently two-parent-biological cluster (B = 1.00) or class (B = 0.85). As previously described, the transitioning into a single-parent family cluster was broken down into two groups, by timing, in the LCA: early transitioning into single-parent family and late transitioning into a single-parent family. Young people in the earlier of these two classes reported statistically higher levels of alcohol use at age 16 than those in the late of these two classes (B = 2.55; B = 0.79), but increased their levels of drinking at statistically faster rates (B = 0.33; B = 0.27).

Overall, the sensitivity to timing among family structure trajectories in the LCA appeared to be substantively important in relation to levels of alcohol use as adolescents aged into young adulthood. Transitioning late into a stepfamily home or early into a single-parent home were associated with higher levels of alcohol use than transitioning early into a stepfamily home or late into a single-parent home. Furthermore, young adults in the returning to two-parent-biological home class behaved more similarly to their peers in either the late transitioning into stepfamily home or the early transitioning to single-parent classes than to those in the consistently two-parent-biological home. Young adults in the late transitioning into single-parent class behaved similarly to their peers in the consistently two-parent-biological class. Generally, the comparison of the two sets of analyses revealed SA as beneficial when the research question lends itself to large group differences, and the nuances of an LCA were particularly informative when studying the timing of transitions into single or stepfamily structures at different ages.

Discussion

In this research note, we compared two methods for measuring family structure trajectories over long developmental windows in more complex ways than counting family structure transitions. The main objective was to help family scholars decide which method might best work for their own research given the continued interest of the journal’s readership in family structure–related differences in human development across the life course. These analyses are not ends in themselves, therefore, but instead were intended to spur future research.

The main differences between the SA and the repeated-measures LCA occurred in the early transitioning to single-parent home, late transitioning to single-parent home, early transitioning to stepfamily home, late transitioning to stepfamily home, and returning to two-biological-parent home groups. The SA resulted in one group that transitioned into a single-parent home during the 15-year period, whereas the repeated-measures LCA resulted in two groups differentiated by the age at which they transitioned into a single-parent family. The same pattern was repeated for those who transitioned into a stepfamily family. The OM method of the SA and the EM algorithm for maximum likelihood estimation that characterizes the LCA likely resulted in these differences. Recall that OM looks for similarities in the elements of a sequence (family structure at each year) to minimize the distance required to transform them into identical sequences, placing more emphasis on order instead of timing so as to not warp time, whereas the LCA estimates the probability a person has of belonging to a latent class. It is also important to remember that the number of clusters (SA) and the number of classes (LCA) are dependent on both statistical and substantive considerations. Statistical measures are critical in understanding how people respond similarly to a set of variables. However, previous research and pertinent theories and previous research should also be taken into account. Together, fit statistics and substantive research help researchers make informed decisions regarding how many clusters or classes are appropriate.

Each method of analysis provided something different, and the decision to use SA or repeated-measures LCA should be carefully considered in the context of the research question and the outcome. A SA (or multidimensional SA, which allows researchers to introduce other pieces of information in addition to family structure) is more agile and lends itself more easily to diverse analyses for distal outcomes. For example, sequence analyses are easily exported from the R package into Stata (or another analytic package) for further analysis. The advantage to the proportional nature of the LCA includes the ability to more precisely identify underlying latent groups in the sample population. There are also analyses available, such as the auxiliary three-step method, to analyze distal outcomes with repeated-measures LCA classes as predictors.

Our results suggest that SA may be more appropriate when (a) further analyses are warranted that require group analysis with hard-classified groups (e.g., event history analyses) and (b) with large comparative analyses, particularly if comparing to a two-parent-biological family structure. Repeated-measures LCA are perhaps more appropriate when research questions are centered on more nuanced patterns such as timing, particularly when there are multiple changes between the measures (years) as with classes where children were transitioning into a single-parent home or into a stepparent home.

In the past, holistically capturing the complexity of family structure transitions was challenging. Both SA and repeated-measures LCA represent unique ways of measuring family structure over time. By taking 15 years of family structure into account in a person-centered analysis, these methods are better able to give a sense of the full history of family structure as a context of development, including the number of transitions, the timing of when transitions occurred, and the type of transition that ensued.