1
Role Configurations and Pathways: A Latent Structure Approach to Studying Learning in the Life Course
Flora Macleod and Paul Lambe - University of Exeter, UK
A paper presented at the Annual Conference of the British Educational Research Association (BERA), 6-8 September 2007 as part of the Symposium “Continuity and Change in Lifelong Learning: Insights from the ‘Learning Lives’ Project”
BERA SIG: Post Compulsory Education and Lifelong Learning
Symposium 8688, Session 8.07, Saturday 8 September 9.00-10.30
contact details:
Dr FJ Macleod ()
Dr PJ Lambe ()
Learning Lives Project
University of Exeter
School of Education and Lifelong Learning
Heavitree Road
Exeter, EX1 2LU
England, UK
© 2007 Work in progress; please do not quote without permission
Role Configurations and Pathways: A Latent Structure Approach to Studying Learning in the Life Course
Flora Macleod and Paul Lambe – University of Exeter, UK
Abstract
Longitudinal data from the British Household Panel Survey were used to examine the pathways taken by 345 females (born 1966-71) from early to mid adulthood. Stage one, of a two-stage latent class analysis, identified three distinct classes at age 20-25, 25-30, and 30-35 and two at age 33- 38. Expected membership of each class was related to how marriage, parenthood, and work roles configured in relation to one another and in relation to likelihood of being an adult education participant and likelihood of being geographically mobile at four time points (1991, 1996, 2001, 2004). Our second stage analysis revealed the existence of different pathways through these role configurations over time and their association with the likelihood of take-up of adult formal learning opportunities. Our results show that members of the three sub-groups identified at age 20-25 had a propensity to follow different pathways. The parent-orientators had the highest likelihood of following path I and were very unlikely to follow either paths II or III. The work-orientators, on the other hand, were most likely to follow path III and were least likely to follow path I. The smallest sub-group, the multi-taskers were most likely to follow path II but were highly unlikely to follow path I. These findings, together with other our emerging findings from the Learning Lives project, including the combining of our quantitative and qualitative (case study) datasets, are being used to explore the place of learning in the lives of adults.
Introduction
With traditional markers of adulthood such as transitions into work, home, marriage and parenthood being deferred in western countries (Skrede, 1999) and a greater emphasis on individual agency in selecting diverse pathways into adulthood (Larsen, Wilson & Mortimer, 2002), the lifecourse is said to have become a much less collectivized experience (Beck, 1992; Giddens, 1991). On the one hand, this individualization and diversity in the paths to adulthood, over time and across countries, is being increasingly reflected in the transition to adulthood literature (e.g. Shanahan 2000). On the other hand, the literature has pointed to the persistence of age norms and age related behaviour (Schoon & Bynner, 2003) . It is therefore of interest to ask how far life chances, in terms of participation in formal learning, can be predicted from early circumstances. It is also legitimate to ask what is the place of formal learning in adult’s lives and how does it interrelate with other competing demands?
The purpose of this paper is model the life course using role state changes as turning points and interpret the consequences of differing role configurations at age 20-25 for the likelihood of participation in formal learning opportunities later on. Specifically the paper asks: Do different individuals take different pathways from early to mid adulthood in terms of their involvement in the roles that have traditionally marked out adulthood? And, if so, with what consequences for their likelihood of participating in formal adult learning opportunities? What is the relationship between early and later patterns of participation?
Theoretical and conceptual framework
We draw on Elder’s (1985) conceptualization of the life course as a series of interlocking trajectories of social roles over time. Movement through social institutions normally involves taking on an institutionally defined role such as being a student, an employee, a husband or wife, a mother or father. Each of these role configurations gives an indication of the extent to which a particular individual is embedded in a given social institution over time. Trajectories are seen as longitudinal involvement in, or connection to, social institutions such as work, marriage and parenthood. Entry into (or exit from) these, often time ordered and age structured institutions, is normally characterized by an event such as a marriage ceremony, a divorce, or a birth of a first child. The specific event that moves an individual into (or out of) a life course institutional context is referred to in this paper (and by Elder, op cit) as a transition. Transition thus indicate when a particular trajectory of a given social role begins, ends, and how long it lasts.
Trajectories and transitions, according to Elder (op cit) are useful conceptual tools in helping understand and describe the life course. They can be used not just to indicate stability or change over time, but also to map out broader life course dynamics by showing the ways in which trajectories interconnect and unfold in unison over time. The shape of one trajectory influences, and is influenced by, the shape of another as occurrences in one trajectory influence occurrences in another. Traditionally, for example, marriage has normally preceded parenthood and both have tended to follow school completion and job stability. But the timing of transitions may vary, influenced by events in other life course trajectories. For example, if an individual drops out of school early then that individual may get married and/or have children earlier than anticipated. Thus, according to Elder, it is necessary to give attention to the timing of occurrences in different but interlocking or conjoining trajectories if we are to understand and describe the life course. We take this theoretical position as our starting point in this paper.
Modeling the life course
We follow Macmillan and Eliason (2003) by adopting a two-stage latent class model approach to identify configurations of social roles over time and life paths that link these role configurations over the life course. Their methodological approach views individuals as being probabilitistically distributed across various role configurations and life paths.
Macmillan and Eliason’s research approach assumes underlying categories. That is, it seeks out sub-groups or collectives within populations. Groups include those whose identification has been subjectively (internally) ascribed and those who have been ascribed by others (externally) e.g. using social categorizations such as social class (Jenkins, 2000). Latent class analysis, like all classification systems, involves a process of dividing a large heterogeneous group into smaller homogeneous groups where members are similar to each other while different from individuals in other groups (Gordon, 1983). These classifications can be a useful way of tracking individual life course pathways (e.g. Richters, 1997).
Latent class analysis identifies cases using: (i) a model-based method which means that a statistical model is assumed to underlie the population and this statistical model is used to identify groups of individuals with respect to a categorical latent (hidden) variable; and (ii) a probabilistic approach which means that a case is assigned to the group with the highest probability of association. The maximum likelihood method is used to estimate latent class models. It is assumed that latent class models are homogenous and individuals within a class have the same probability distribution. Further, the latent class is assumed to be the force relating the within-class observed variables. Put another way, the correlation between within-class variables is assumed to be zero. A major advantage of latent class modeling is that a mixture of categorical and continuous variables can be used without any estimation threats.
Data and measures
We use data from the British Household Panel Survey. These data were collected from a national sample of 5000 households resulting in10,000+ adults (16+) when they were first interviewed in 1991. Since then, these individuals have been surveyed annually. The longitudinal structure of the data allows us to model the life course over an extended period of time. Here we focus on whether or not a respondent participated in post initial phase formal education, whether or not they moved area, whether they were in the workforce, whether they were married, and whether they were parents at four time points (1991, 1996, 2001 and 2004). The three roles, work, marriage and parenthood, are seen as key markers of the transition into adulthood (Booth, Crouter & Shanahan, 1999; Shanahan, 2000).
We operationalised involvement in post initial phase formal education simply as whether respondents answered ‘yes’ or ‘no’ to the following question:
Apart from the full-time education you have already told me about) Have you taken part in any other training schemes or courses at all since September 1st [the previous year] or completed a course of training which led to a qualification? (interviewers were instructed not to include leisure courses)
In settling for this definition we fully acknowledge its limitations in that it represents only a very particular subset of all lifelong learning. However, we judged this question represented a meaningful basis for our analysis as it tapped into participation in education potentially leading to a wide range of qualifications available to adults in the UK from the lowest to the highest.
We defined ‘move’ as a geographical move within the UK, that is, a move out of the area in which they had been living. We defined work as either being in the workforce irrespective of whether they were self-employed or employed on a part or full time basis. This meant we tapped into those who were out of the workforce not just for reasons of unemployment, but also because of illness, disability, caring for dependents, or other such reason. We defined marriage by its legal status meaning that the single, divorced and those who were in a state of co-habitation were classified as ‘not married’. Finally, we defined parenthood in terms of whether each respondent had dependent children living with them, irrespective of whether they were the birth parent or not. Our data did not allow us to differentiate.
While more complex operationalizations of these five social roles are possible using our dataset, these were well-suited to our interest in modeling interlocking pathways of roles over time. For all of these states, we examined their joint occurrence at four time points, 1991, 1996, 2001, and 2004. These made it possible for us to assess the degree to which a given role was achieved by a specific age. We were also able to check whether roles appeared or disappeared within each timeframe as a check against our delineation of interconnected pathways that make up early adulthood. Our choice of time points was made on theoretical and empirical grounds. Since our interest was to broadly map the life course over an extended period of time, we took the first and last points of data collection available to us (1991 and 2004 covering a 14-year period) and two points in between, 5 years from the first point (1996) and 5 years from that point (2001).
Our analytic sample in this particular paper consists of female original sample members only who were aged 20-25 in 1991 (n=345), with birth dates ranging from 1966 to 1971[1]. Whilst we accept that there is no precise way of determining which years constitute ‘early and mid adulthood’, we believe that with this group we are able to examine movement through this period. We see this time of life as a period of restructuring of relationships and commitments in relation to family, work and parenthood. We increased our sample size and statistical power by focusing on six adjacent birth cohorts (1966-1971). This increased our ability to consider greater homogeneity in the structure of the life course. At the same time, these cohorts are closely related in historical time and are thus likely to have experienced unique cohort or period effects that could also have impacted upon the structure of the life course (Elder, 1998).
Analysis and results
Stage 1: latent role configurations
We began our analysis by examining the goodness of fit of statistical models for the first stage analysis of latent role configurations. Table 1 shows seven columns that includes the model chi-square statistic (X2) and, in parenthesis, its significance emphasizing local independence, the likelihood chi-square statistic (L2), degrees of freedom (df), the index of dissimilarity (D), Raftery’s (1995) Bayesian Information Criterion (BIC), and Akaike’s (1974) Information Criterion (AIC). The row panels correspond to a specific age or stage, beginning with ages 20-25 (the 1991 or BHPS wave 1 data collection point) and concluding with ages 33-38 (the 2004 or wave 14 data collection point). In each age panel, four models were compared. These were a null (one-class) model, a two-class model, a three-class-model and a four-class model. The models indicated the number of classes of latent role configurations that effectively characterized the sample at each age. As the models were not hierarchical, model selection was based on overall goodness of fit.
TABLE 1 Goodness of fit statistics for Latent Role Configurations
Age / Latent Classes / x2 / L2 / df / D / BICL2 / AIC
L2
Wave a
1991
Age 20 - 25
I / 135.2218 (0.0000) / 130.7653 / 26 / 0.2189 / - / -
II / 46.8094
(0.0006) / 50.8347 / 20 / 0.1284 / - / -
III / 15.3493
(0.3547) / 16.6582 / 14 / 0.0727 / -64.1514 / -11.3418
IV / 4.3681
(0.8225) / 5.8308 / 8 / 0.0263 / -40.9175 / -10.1692
Wave f
1996
Age 25-30
I / 147.9795
(0.0000) / 154.0810 / 26 / 0.2591 / - / -
II / 34.4062
(0.0235) / 35.1945 / 20 / 0.0931 / - / -
III / 14.6233
(0.5044) / 15.1919 / 14 / 0.0604 / -66.6177 / -12.8081
IV / 8.7672 / 10.5323 / 8 / 0.0396 / -36.2161 / -5.4677
Wave k
2001
Age 30-35 / I / 84.7209
(0.0000) / 87.4237 / 26 / 0.1862 / - / -
II / 28.9069
(0.0896) / 27.1909 / 20 / 0.0774 / - / -
III / 10.6987
(0.7095) / 11.5082 / 14 / 0.0383 / -70.3014 / -16.4918
IV / 6.8623
(0.5516) / 6.5216 / 8 / 0.0326 / -40.228 / -9.4784
Wave n
2004
Age 33-38
I / 78.2217
(0.00000 / 81.2777 / 26 / 0.1760 / - / -
II / 15.1550
(0.7675) / 19.5109 / 20 / 0.0727 / -97.3600 / -20.4891
III / 7.0134
(0.9342) / 10.5958 / 14 / 0.0328 / -71.2138 / -17.2138
IV / 5.8117
(0.6683) / 7.7490 / 8 / 0.0314 / -38.2510 / -8.2510
The one class null model, if it had fitted our data well, would have indicated the observed social roles were independent of one another. Deviations away from a good fit, for this null model, would, on the other hand, indicate the degree to which our selected social roles have significant associations among themselves. At our four selected age stages the null model indicated a poor fit, implying that the social roles we had selected cohered in significant ways over this period of the period of the life span studied. In table 1, the emboldened model indicates the number of classes of latent role configurations that effectively characterized our sample at each age. This shows that, whilst our female respondents were in their twenties and very early thirties, three-class models provided a good fit to the data, whereas when they reached 33-38 a two-class model provided the best fit, suggesting less variance later in adulthood in role configurations.
Whilst table 1 enables us to interpret how our selected social roles cohere in significant ways statistically at four time points, we need to study table 2 to understand the qualitative aspects of these configurations. Examination of the 20-25 year old classifications in table 2 shows that the sample can be summarized by three distinct role configurations. Class I has a high probability (0.8012) of being employed, of being a parent (0.9500), a moderately high probability of being married (0.6683), a relatively high probability of participating in formal learning opportunities, and a very low likelihood of moving (0.0898). We called this sub-group of women ‘multi-taskers’ as they had a high probability of being involved in several task (and time) demanding roles simultaneously, those of work, family, parenting and adult education. Multi-taskers were expected to make up 19% of the sample (0.1938) at this age.
Class II, on the other hand, was expected to classify 41% (0. 4088) of the sample at this age. These women had a relatively high likelihood of being parents (0.7030), although slightly lower than the multi-taskers, were slightly less likely to be married than not (0.5498 v 0.4502), and had much lower probability of being married than the multi-taskers, a slightly higher likelihood of being out of the workforce than in work (0.5439 v 04561) and were much less likely to be in work than the multi-taskers, a low likelihood of moving (0.1646) and an extremely low likelihood of taking part in formal adult education opportunities (0.0308). Membership of this group appeared to be dominated by the likelihood of the parent role occurring alongside the relative unlikelihood of the other roles occurring such as marriage and work. We called this group the parent-orientators as parenthood seemed to characterize their life at this age.
Class III, constituting 40% (0.3974) of the sample. Members of this group had a very low probability of parenthood (0.0767), a low probability of marriage (0.2540), and a very low probability of being out of the workforce (0.0243). They, in common with all others in the age group, had a low probability of moving to another region of the UK (0.2464) and had a moderate probability
1
TABLE 2 Conditional probabilities for Latent Role Configurations
AGE 20-25 / AGE 25-30 / AGE 30-35 / AGE 33-38Latent Class / I / II / III / I / II / III / I / II / III / I / II
Roles
MARRIED Yes
No / 0.6683 / 0.4502 / 0.2540 / 0.9803 / 0.4662 / 0.4035 / 0.7991 / 0.4909 / 0.3181 / 0.6795 / 0.2674
0.3317 / 0.5498 / 0.7460 / 0.0197 / 0.5338 / 0.5965 / 0.2009 / 0.5091 / 0.6819 / 0.3205 / 0.7326
PARENT Yes
No / 0 . 9500 / 0.7030 / 0.0767 / 0 . 9532 / 0.8664 / 0.0408 / 0 . 9002 / 0.8020 / 0.0867 / 0.9704 / 0.0001
0. 0500 / 0.2970 / 0.9233 / 0. 0468 / 0.1336 / 0.9592 / 0. 0998 / 0.1980 / 0.9133 / 0.0296 / 0.9999
InFormED Yes
No / 0.7030 / 0.0308 / 0.5033 / 0.4411 / 0.1080 / 0.4658 / 0.4149 / 0.0895 / 0.4859 / 0.2819 / 0.3023
0.2970 / 0.9692 / 0.4967 / 0.5589 / 0.8920 / 0.5342 / 0.5851 / 0.9105 / 0.5141 / 0.7181 / 0.6977
NOTinWORK Yes
No / 0.1988 / 0.5439 / 0.0243 / 0.2320 / 0.6497 / 0.0143 / 0.1909 / 0.3896 / 0.0800 / 0.2934 / 0.0930
0.8012 / 0.4561 / 0.9757 / 0.7680 / 0.3503 / 0.9857 / 0.8091 / 0.6104 / 0.9200 / 0.7066 / 0.9070
MOVED Yes
No / 0.0898 / 0.1646 / 0.2464 / 0.3269 / 0.1719 / 0.1405 / 0.1416 / 0.1322 / 0.1095 / 0.0849 / 0.0930
0.9102 / 0.8354 / 0.7536 / 0.6731 / 0.8281 / 0.8595 / 0.8584 / 0.8678 / 0.8905 / 0.9151 / 0.9070
Latent Class
Probability / 0.1938 / 0.4088 / 0.3974 / 0.1654 / 0.3588 / 0.4759 / 0.4192 / 0.3303 / 0.2505 / 0.7507 / 0.2493
1
of participation in formal learning (0.5034). As the likelihood of the work role occurring was extremely high(0.9757) and the likelihood of occurrence of other social roles was low or extremely low at this age, we called them the work-orientators.
I sum, table 2 shows that, in our role configuration analysis, amongst women in our sample aged 20-25, three contrasting qualitative pictures were unfolding, the multi-taskers, the parent orientators and the work orientators. Beyond this starting point we were interested to investigated the respective trajectories of these sub-groups. For this it was necessary to move to latent pathways and the derivation of latent paths over time. The above identification of latent role configurations at successive stages provided us with the backdrop for our stage two latent class analysis that links latent role configurations over time. It is to this analysis that we now turn.
Stage 2: latent life paths
(The first paragraph deals with the technicalities of how our latent pathways were achieved given the sparseness of data. Some readers may prefer to move straight to the second paragraph. )
The criteria we used for the best fitting life path model were the log-likelihood based BIC and AIC statistics, with lower values indicating better fitting models. The modelling begins with a simple one latent life path model with one latent class of role configurations at each age-stage. This is equivalent to the log-linear model for complete independence across all observed variables. A series of models was then estimated adding one class at a time to the number of latent life paths and one class at a time to the number of latent classes of role configurations ( e.g. if y = latent life paths, x1= latent classes at age-stage 20-25, x2= latent classes at age-stage 25-30, x3= latent classes at age-stage 30-35, and x4= latent classes at age-stage 33-38, then one –class life path model y=1, x1=1, x2=1, x3=1 and x4=1, two-class latent life path model y=2, x1=2, x2=2, x3=3 and x4=4 etc). This process was continued until the log-likelihood based BIC statistic reached a clear minimum. Using this model with the lowest BIC statistic, a further series of models with varying permutations of numbers of classes of y and classes of x1-x4 were run and the model with the minimum BIC selected. This resulted in the three latent life path models in table 3, with three classes of role configurations at age 20-25, 25-30, and 30-35 and two at age 33-38.