Content uploaded by Danilo Bolano
Author content
All content in this area was uploaded by Danilo Bolano on Jan 08, 2020
Content may be subject to copyright.
LIVES WORKING PAPER 2020 / 82
The National Centres of Competence
in Research (NCCR) are a research instrument
of the Swiss National Science Foundation
THE LINK BETWEEN PREVIOUS LIFE TRAJECTORIES
AND A LATER LIFE OUTCOME:
DANILO BOLANO
MATTHIAS STUDER
RESEARCH PAPER
http://dx.doi.org/10.12682/lives.2296-1658.2020.82
ISSN 2296-1658
A FEATURE SELECTION APPROACH
LIVES Working Papers - Bolano
A u h t o r s
Bolano, D. (1)
Studer, M. (2)
A b s t r a c t
Several studies have investigated the link between a previous trajectory and a given
later-life outcome. Trajectories are complex objects. Identifying which aspects of the
trajectories are relevant is of primary interest in terms both of prediction and testing
specific theories. In this work, we propose an innovative approach based on data
mining feature selection algorithms. The approach is in two steps. We start by
automatically extracting several properties of the sequences. Using a life course
approach, we focus here on features related to three key aspects of the life course:
sequencing, timing and duration of life events. Then, in a second step, we use feature
selection algorithms to identify the most relevant properties associated with the
outcome. We discuss the use of two features selection approaches a random forest
approach (Boruta) and a LASSO method (Stability Selection). We also discuss the
inclusion of control variable such as socio-demographic characteristics of the
respondent in this selection process. The proposed approach is illustrated through a
study of the effects of family and work trajectories between age 20 and 40 on health
and income conditions in midlife.
K e y w o r d s
sequence analysis | variable selection | life course methodology | machine
learning
A u t h o r ’ s a f f i l i a t i o n
(1) NCCR LIVES, University of Lausanne
(2) LIVES Center and Institute of Demographics and Socioeconomics, Faculty of Social
Sciences, University of Geneva
C o r r e s p o n d e n c e t o
danilo.bolano@unil.ch
* LIVES Working Papers is a work-in-progress online series. Each paper receives only limited review. Authors are
responsible for the presentation of facts and for the opinions expressed therein, which do not necessarily reflect
those of the Swiss National Competence Center in Research LIVES.
The Link Between Previous Life Trajectories and
a Later Life Outcome: a Feature Selection
Approach
Danilo Bolano1and Matthias Studer2
1NCCR LIVES, University of Lausanne
2LIVES Center and Institute of Demographics and Socioeconomics,
Faculty of Social Sciences, University of Geneva
Corresponding author: danilo.bolano@unil.ch (Danilo Bolano)
Abstract
Several studies have investigated the link between a previous trajectory and a given
later-life outcome. Trajectories are complex objects. Identifying which aspects of the
trajectories are relevant is of primary interest in terms both of prediction and testing
specific theories. In this work, we propose an innovative approach based on data mining
feature selection algorithms. The approach is in two steps. We start by automatically
extracting several properties of the sequences. Using a life course approach, we focus here
on features related to three key aspects of the life course: sequencing, timing and duration
of life events. Then, in a second step, we use feature selection algorithms to identify the
most relevant properties associated with the outcome. We discuss the use of two features
selection approaches a random forest approach (Boruta) and a LASSO method (Stability
Selection). We also discuss the inclusion of control variable such as socio-demographic
characteristics of the respondent in this selection process. The proposed approach is
illustrated through a study of the effects of family and work trajectories between age 20
and 40 on health and income conditions in midlife.
1 Introduction
Many social sciences research questions are interested in the link between a previous
trajectory and a given later-life outcome. The life course paradigm (e.g. Elder et al.,
1
2003) stresses the importance to situate any outcome within its individual temporal
dynamics, and therefore to relate it with the previous trajectory. Moreover, life
domains might be intertwined. Decisions and events happened in one domain may
have a strong influence into another life domain. Therefore, one should take into
account the unfolding of previous trajectories in other life domains as well.
Several studies have focused on the consequence of the life course on later life
outcomes (Fasang, 2012). Social stratification researches have studied how and if
previous life course patterns result in late life inequalities (Fasang, 2012; Gabriel
et al., 2014). Social epidemiologists studied how previous professional, family
or even housing tenure life-course shape aging, wellbeing and health condition
at old age (e.g. McMunn et al., 2015; Sabbath et al., 2015; Hoven et al., 2017;
Vanhoutte et al., 2017). Demographers analyzed how cohabitation trajectories
influence leaving home (Rossignon et al., 2016). Criminologists studied how family
formation events influence criminal careers (Zoutewelle-Terovan et al., 2012).
Two different strategies are generally used to study the link between previous
trajectories and a given outcome (Rossignon et al., 2016). First, some authors
include indicators of these trajectories in their models, such as whether the respon-
dent experienced a given key event or the time spent in a given state (Gabriel
et al., 2014; Blossfeld et al., 2007). However, the approach is limited, as previous
trajectories might be complex resulting in the need to take many indicators into
account, which might be relevant... or not (Rossignon et al., 2016). Furthermore,
the current situation might result from complex interaction between past events
that are difficult to identify. As a result, several authors use sequence analysis
to operationalize the concept of past trajectories. This typically involves coding
the trajectories as sequences, computing dissimilarities between them, creating a
typology using cluster analysis and include it in their own analysis (e.g. Rossignon
et al., 2016; Fasang, 2012; Gabriel et al., 2014; Hoven et al., 2017). The procedure
has the advantage to take into account the timing, duration and sequencing of these
trajectories, three key dimensions according to the life-course paradigm (Studer
and Ritschard, 2016). Finally, the use of sequence analysis avoid making an a priori
choice about the relevant dimensions of the previous trajectories probably resulting
in the identification of more complex dynamics (Rossignon et al., 2016).
The use of sequence analysis raises several issues. First, the creation of a single
typology aims to capture the timing, duration and sequencing of the previous
trajectory altogether. However, it is often interesting to understand the more
specific impact of each aspect taken separately (Studer and Ritschard, 2016).
Indeed, each of these aspects might be linked to different sociological theories and
models when studying the relationship between previous trajectory and outcome.
For instance, the time spent in a state might capture the relationships between
“time exposure” to unemployment and future health. The timing of a situation
2
or an event is thought to influence later life outcomes within the “critical period”
model. According to this model, a situation such as unemployment only affect
the outcome when it occurs within some specific age range, for instance after
50 years old. Finally, a link between specific sequencing and an outcome might
reveal “mandatory steps” or the impact of specific “dynamics”. For instance, many
back-and-forth movements between employment and unemployment might have
specific effect on future health.
Second, it might be difficult to give a clear interpretation of the association
between previous trajectory typology and later outcome. Sequence analysis ty-
pologies are generally defined by an implicit set of rules (Studer, 2018). In other
words, the typology works by grouping together similar trajectories. However, the
exact properties of the sequences that distinguish the types are unknown. It might
therefore be difficult to have an in-depth understanding of the links between the
trajectory and a subsequent outcome.
Finally, any clustering technique works by simplifying the information. The aim
is to ignore small variation in the sequences to build a few types. However, such
simplification is made without taking the outcome into account. Therefore, the key
variations in the sequences that explain the variation in the outcomes might very
well have been “simplified” or ignored in the process. The procedure can therefore
hinder the identification of some key properties of the sequences, therefore leading
to wrong conclusions.
In this article, we propose an innovative methodology combining feature selection
algorithms and sequence analysis tools to study the link between previous trajectory
and later-life outcome. This combination aims to automatically identify the specific
key dimensions of a previous trajectory that are linked with the outcome of interest.
As such, it overcomes the indicator approach by taking many aspects of the previous
trajectory into account, while providing information about the specific impact of
timing, duration and sequencing.
The proposed methodology starts with the creation of a very broad and auto-
matically defined set of indicators of the previous trajectory aiming to capture its
timing, sequencing and duration dynamics. In a second step, it relies on machine
learning feature selection algorithms to select the most relevant indicators, and
therefore, the most important aspects of the trajectory for the domain under investi-
gation. Two feature selection approaches are discussed: Boruta and LASSO. Boruta
aims to select all relevant features associated with the outcome and can capture
non-linear association and interaction effects between aspects of interest. LASSO
aims to identify a parsimonious subset of features that best predict the outcome.
We also discuss the inclusion of control variable such as age in this selection process
to avoid the selection of confounders. The proposed methodology is illustrated
through a study of the effect of previous family and professional trajectory on
3
self-rated health and income in midlife in Switzerland using longitudinal data from
the Swiss Household Panel.
The remainder of this paper is organized as follows. In Section 2 we present
the motivating examples. Data are discussed in Section 2.3. Section 3 discusses
how to extract features from sequential data. In Section 4, we introduce our
methodological approach discussing two ways of selecting the relevant features. In
Section 5 we analyze the association between family and work histories, and health
and wealth. The software used is briefly discussed in Section 6. Finally, Section 7
reports concluding remarks and discussion on future extensions.
2 Motivating Example: The Effect of Family and
Working Life on Health and Income
In order to illustrate the proposed method and its usefulness, we rely on two sample
applications. We focus on the consequences of professional and family trajectories
before forty years old on self-rated health and household income in middle life.
These two examples allow us to illustrate the use of the proposed methodology to
study categorical (health in our example) as well as numerical (income) outcomes.
2.1 Working and Family Trajectories and Health
Drawing on life course and cumulative disadvantage literature (O’Rand, 2002;
Dannesfer, 2003), many studies have focused on the link between late-life health
inequalities and previous professional or family life trajectories.
The negative association between unemployment and physical and mental health
has been explored in several studies (e.g. Paul and Moser, 2009; Falba et al., 2005;
Laitinen et al., 2005; Strully, 2009). Most of them focused on the short-term effect
of unemployment. Few exceptions include Clark et al. (2001) and Daly and Delaney
(2013) that studied its long-term effects. Aside from the overall time spent in
unemployment, having unstable employment histories is often though as a potential
source of distress over a working life which might have a direct impact on health.
Family life trajectories can also play a role (Henretta, 2007; Kravadal et al., 2012;
Williams et al., 2011; O
´
Flaherty et al., 2016). Studies have found an association
between parity and different types of later-life health conditions (e.g., Doblhammer,
2000; Hurt et al., 2006; Grundy and Kravdal, 2007; Grundy and Holt, 2000). The
timing of childbearing is also relevant. Women who had children earlier in life had
increased both metabolic (McMunn et al., 2015) and inflammatory markers (Lacey
et al., 2015). Sironi (2018) found a non-linear effect of age at first child on health.
Having a teenage pregnancy or having the first child after age 35 is associated with
4
an increased risk of chronic conditions later in life.
The marriage and its duration has also been found to be positively associated
with health (Grundy and Holt, 2000; Henretta, 2010). The positive effect of
marriage might be due to accumulation of social, emotional and economic support
over a long period of time. However, early marriage might have negative effect on
health (Grundy and Holt, 2000) probably due to higher risk of divorce later on and
being disruptive for educational attainment.
Taking a broader perspective, Sabbath et al. (2015) and McKetta et al. (2018)
studied the relationship between the unfolding of professional and family social
roles over the life course and health outcomes. In these studies, role accumulation
was linked with better health, while weak labor market relationship was generally
associated with lower health in the US and UK.
2.2 Working Career, Family Life and Income
Losing a job is not only directly associated with income loss, but it might lead to
deterioration of future labor market prospects (e.g. Heckman and Borjas, 1980;
Arulampalam, 2008) and earnings due to a depreciation of human capital (Becker,
1993), underutilization and underdevelopment of skills. Social and working related
stigma might be (see e.g. Gibbons and Katz, 1991). Employers might interpret
career interruptions as a lack of commitment and ability with negative effects in
terms of future wages and occupational advancements.
Due to different roles in the society expected for men and women, the effect of
career interruption on earning differs by gender. Despite women experienced more
often transitions in and out employment for childbearing, caring for children or
older parents, the wage penalty for unemployment over the career is lower among
women (Spivey, 2005).
The magnitude of the effects depends on the timing and duration of career
interruption during the working life (Albrecht and Vromon, 1999; Beblo and Wolf,
2002; Spivey, 2005). Unsurprisingly, experiencing a longer period of unemployment
(duration) is associated with lower re-employment’s chances (Eriksson and Rooth,
2014). Youth unemployment has been found to have a scarring effect on later
working life in terms of wages, working conditions and careers (Gregg, 2001).
According to Pohlig (2019), most studies have focused on the timing and
duration of unemployment because of the use of simple indicators. He argues that
we need a more detailed understanding of the consequence unemployment patterns
and sequencing. This is one of the aims of the methodology proposed in this paper.
Having a child is often linked with career interruptions and lowers a woman’s
lifetime earnings (the so-called motherhood penalty) (Budig and England, 2001;
Correll et al., 2007). The timing of career interruption due to childbearing matters.
Studies have shown that motherhood gap is less important when childbearing is
5
postponed (Miller, 2011; Leung et al., 2016). These results are consistent with the
human capital depreciation theory, where late career interruptions are thought to be
less disruptive, as workers have already acquired the most important skills. Studies
did not find a ”fatherhood gap” but rather a ”fatherhood premium” (Glauber,
2008).
Family unions have positive social and economic consequences (Ribar, 2004).
Marriage has a positive effect on economic condition among men (marriage premium)
(Hill, 1979; de Linde Leonard and Stanley, 2015). Its effect on women’s earnings is,
however, more complex to distinguish. Traditionally, childbearing follows marriage
in subsequent years (typical sequencing of events). For these reasons, the potential
positive effect of marriage on earning is often counterbalanced by the motherhood
penalty (Budig and England, 2001).
2.3 Data
We use information on family and working life from the “Life History Calendar”
questionnaire of the Swiss Household Panel (SHP) (Tillmann et al., 2016). This
calendar collected detailed retrospective data on life trajectories in different life
domains. In this work, we consider family history and working life between 20
and 40 years old on a sample of around 3,000 individuals (1626 women and 1378
men) aged 40 to 65 in 2013. For each individual we constructed yearly parallel
trajectories of family and professional trajectories.
In order to compare the proposed methodology with the sequence analysis, we
created a typology of family and professional trajectories taken separately. We
used optimal matching with constant substitution costs to measure the distance
between each pair of sequences. Then, we applied the partitioning around medoids
(PAM) algorithm and retained the solution with the highest average silhouette
width (Studer, 2013).
The professional trajectory allows distinguishing between: 1) Full-Time Work,
2) Part-Time Work, 3) Inactive or Unemployed (Non-Working), 4) Education. We
add a fifth state in case of missing data. Missing information is quite rare with
around 2% of missing states. The typology allows identifying seven groups of
working trajectories (Figure 1) and five typologies of family history (Figure 2).
From previous studies, we know that professional careers are highly gendered in
Switzerland, resulting in different kinds of dynamics for women while men tend
to follow a full-time breadwinner pattern (Levy et al., 2006; Widmer et al., 2003).
This is confirmed in our sample, where 75% of men follow the full-time pattern
while only 23% of women do so.
The cohabitation trajectories code the living arrangement until forty years old
using five states: single, single with child(ren), with a partner, with a partner
and child(ren) and other situations. We anticipate less structured cohabitation
6
Figure 1: Typology of Professional Trajectories
FT
Freq. (n=1407)
y20 y25 y30 y35
0.0 0.6
Late FT−NW
Freq. (n=290)
y20 y25 y30 y35
0.0 0.6
FT−NW
Freq. (n=247)
y20 y25 y30 y35
0.0 0.6
PT
Freq. (n=347)
y20 y25 y30 y35
0.0 0.6
Late FT−PT
Freq. (n=316)
y20 y25 y30 y35
0.0 0.6
Edu−FT
Freq. (n=328)
y20 y25 y30 y35
0.0 0.6
NA
Freq. (n=69)
y20 y25 y30 y35
0.0 0.6
Education (Edu)
Full−time (FT)
Missing (NA)
Non working (NW
Part−time (PT)
Figure 2: Typology of Cohabitation Trajectories
Single
Freq. (n=489)
y20 y25 y30 y35
0.0 0.6
Early parenthood
Freq. (n=851)
y20 y25 y30 y35
0.0 0.6
With partner
Freq. (n=398)
y20 y25 y30 y35
0.0 0.6
Other paths
Freq. (n=167)
y20 y25 y30 y35
0.0 0.6
Late parenthood
Freq. (n=1099)
y20 y25 y30 y35
0.0 0.6
Single with child(ren)
Other situations
With partner and child(ren)
With partner
Single
trajectories, but still centered around a few types (Levy et al., 2006). Figure 2
presents the five-group typology built using the same methodology as for professional
trajectories. The most common pattern are early and late parenthood.
Using the prospective nature of the SHP data, we aim to analyze the effect of
family life and working career on later-life self-rated health and later-life economic
condition (income level). Self-rated health is measured as a binary variable,
reporting “Very Well” or “Well” health versus other answers at the time of the
7
interview.
1
Income is measured using equalized disposable household income, since
it takes family composition (household size) into account directly.
3 Trajectories’ Features Extraction
The proposed methodology follows a three-step approach to identify relevant
sequence properties. In the first step, we automatically extract several features
of the trajectories. In the second step, we select the most relevant ones using
data-mining feature selection algorithms. Finally, the remaining features are used
as covariates in a traditional regression model to estimate the magnitude and the
direction of their effects on the variable of interest.
3.1 Automatic Sequences Feature Extraction
The aim of the feature extraction is to build a set of indicators measuring the
different aspect of the trajectories. In this article, we propose to build this set
of indicators based on the three key aspects of life-course research identified by
Studer and Ritschard (2016), namely sequencing, timing, and duration. We further
propose some extensions of it.
The notion of sequencing relates to the order of the events or states within the
trajectories. It regroups the quantum, i.e. which states or events occur, and the
dynamics of the sequences coded through the succession of states (Billari et al.,
2000). Both concepts can be thought to have an impact on later life outcomes. For
instance, the succession and accumulation of roles is thought to have an impact on
health in later life (McMunn et al., 2015). Traditional studies on the effect of life
histories on later-life economic and health conditions might not be able to capture
the importance of the order of events on the outcome since often rely on simple
indicators such as experiencing the event or not (e.g., having or not the child),
duration in a condition (e.g. duration of unemployment) and/or the time of events
(e.g. age at first child).
Following Studer (2018), we rely on frequent subsequence mining to extract
indicators of the sequencing of the states. This method aims to uncover frequent
patterns of states out of a set of sequences (Studer et al., 2010; Agrawal and Srikant,
1995; Zaki, 2001).
A subsequence
s
is defined as a subsequence of
x
if all the states of
s
occur
in the same order as in
x
. For instance, the sequence
A−C
is a subsequence of
sequence
A−B−C
because
A
and
C
occurs in the same order. In order to focus
1
The self-rated health conditions are defined by the question “How do you feel right now?”.
Five possible answers were proposed: “not well at all”, “not very well”, “so, so” , “well”, “very
well”.
8
on sequencing we look at the subsequence of the sequences of distinct successive
states (i.e., without repeating a state to account for duration). A subsequence
therefore captures either direct transition (two successive states) or medium term
ordering. These two kinds of dynamics are interesting in many social sciences
applications. For instance, when studying social mobility, direct transition can
capture promotions, while medium term ordering should be able to measure job
change, potentially with transitional unemployment spell, that results in upward
mobility.
A subsequence might be of length one. Hence
A
and
B
are subsequences of
A−B
. These subsequences capturing the quantum of each situation along the
trajectories. The notion of a subsequence is therefore able to capture the two
aforementioned aspects of sequencing.
In this framework, a subsequence is said to be frequent if it is found in more
than a predefined percentage of sequences, called the minimum support, that
we typically set at 5%. Aside from a list of frequent subsequence, the analysis
generates one variable per identified subsequence storing the presence or absence
of the subsequence in each trajectory. It can therefore be used in any subsequent
analysis, such as feature selection or a regression.
We extracted (see Table 1) 25 subsequences for the professional trajectories and
32 for the family one using in both cases a minimal support (i.e. frequency) of 5%.
For instance, the pattern [Single
→
With partner
→
With partner and Kid] was
followed by a little more than half of our sample. Aside from its frequency, we also
have a dummy indicator for each individual whether this pattern occurs within his
or her trajectory. Frequent subsequences are not mutually exclusive. Therefore,
any individual following the [Single
→
With partner
→
With partner and Kid]
also follows the [With partner →With partner and Kid] pattern as well.
The notion of timing relates to when a situation is experienced. The timing
of events is crucial in many life-course researches. In the critical period model of
life course epidemiology, events are often thought to have different consequences
depending on when they occurred. Losing a job at the beginning or at the end of a
professional career might have very different applications, but the same apply to
the age at childbirth Lesnard (2010); Leung et al. (2016); Sironi (2018, e.g.).
Studer (2018) uses the situation at each age to capture the timing of the process.
We take here a more parsimonious approach by computing the time spent in each
state for each 5
−
years period, i.e. 20-24, 25-29, 30-34 and 35-39 years old. Table 2
presents the average of these indicators. This allows to quickly identify the general
trend of the timing of these trajectories as well. Unsurprisingly, education mostly
occurs at the beginning of the sequence (until age of 30), while living with kids
and a partner occurs more frequently towards the end of the considered family life
span when respondents had 30/35 years old.
9
Table 1: Frequent subsequences in family and professional trajectories and their
associated frequencies.
Work trajectories Family trajectories
Subsequence Perc. Subsequence Perc.
FT 91.4% Sin 84.8%
PT 45% Part 82.5%
FT →PT 32.7% Part/Kid 76%
NW 28.4% Sin →Part 69.8%
Edu 26.2% Part →Part/Kid 65.2%
Edu →FT 22.8% Sin →Part/Kid 62.3%
FT →NW 22.1% Sin →Part →Part/Kid 54.8%
NA 20% Oth 26.2%
FT →FT 16.9% Oth →Part 16.8%
PT →FT 15.4% Oth →Part/Kid 16.4%
FT →NA 13.8% Sin →Oth 16.3%
NA →FT 13.2% Sin →Sin 13.7%
Edu →PT 12.5% Oth →Part →Part/Kid 12.2%
NW →PT 12% Part →Sin 11.3%
PT →NW 10.7% Oth →Sin 10.9%
PT →PT 9.5% Sin →Oth →Part 9.6%
FT →NW →PT 9.3% Sin →Oth →Part/Kid 9.4%
FT →NA →FT 8.4% Sin →Part →Sin 8.9%
NA →PT 8.2% Sin/Kid 8.5%
FT →PT →FT 7.9% Sin →Sin →Part 8.4%
Edu →FT →PT 7.7% Part →Part 8.2%
FT →PT →NW 6.7% Oth →Sin →Part 7.2%
Edu →NW 6.4% Part →Sin →Part 6.8%
NW →FT 5.3% Sin →Oth →Part →Part/Kid 6.8%
FT →NA →PT 5.1% Sin →Part →Part 6.4%
Sin →Oth →Sin 6.1%
Part/Kid →Sin/Kid 6.1%
Oth →Sin →Part/Kid 5.9%
Sin →Sin →Part/Kid 5.4%
Sin →Sin/Kid 5.3%
Part →Sin/Kid 5.3%
Sin →Part →Sin →Part 5.3%
Note: The states are abbreviated as follows for professional trajectories: FT (full time), PT
(Part-time), NW (non-working), Edu (education), NA (missing value). For family trajectories: Sin
(Single), Part (with partner), Part/Kid (with partner and kids), Oth (other situations), Sin/Kid
(single with kids).
Finally, the notion of duration refers to the overall time spent in each state
(Studer and Ritschard, 2016). It can be linked to the concept of “time exposure” to a
given situation over the life course. As already discussed, duration of unemployment
over the life course, length of marriage among others have strong effects on wellbeing.
The overall time spent in each state in our sample is presented in Table 2 (last
row). As observed also in Figure 1, respondents spent the majority of time being in
a Full-Time working conditions (on average 12 years). For the family trajectories,
the most common condition is having a partner and kids (Part/Kid state).
In some applications, it might be useful to distinguish further between the overall
10
Table 2: Average of Timing’s and Duration’s Indicators.
Work trajectories Family trajectories
Age group Edu FT NA NW PT Sin/Kid Oth Part/Kid Part Sin
20-24 0.9 3.2 0.3 0.2 0.5 0.0 0.5 0.3 1.0 3.1
25-29 0.1 3.3 0.2 0.6 0.7 0.1 0.4 1.4 1.6 1.5
30-34 0.0 3.0 0.2 0.8 1.0 0.1 0.2 2.7 1.2 0.8
35-39 0.0 2.8 0.1 0.7 1.3 0.2 0.2 3.4 0.7 0.5
Overall 1.0 12.3 0.8 2.3 3.5 0.4 1.3 7.8 4.5 5.9
Note: The states are abbreviated as follows for professional trajectories: Edu (education), FT
(full time), PT (Part-time), NW (non-working),NA (missing value). For family trajectories: Sin
(Single), Part (with partner), Part/Kid (with partner and kids), Oth (other situations), Sin/Kid
(single with kids).
time spent and typical spell duration. This can be included by adding the number of
spells in each state for instance. When studying professional integration trajectories,
this might be interesting to make the distinction between long unemployment spells
and multiple short employment spells for instance. In our application, multiple
spells of unemployment are rare and in general trajectories are quite stable over
time with most individuals experience only one spell in each state during our
observational period. We have therefore not included it in our analysis.
Aside from the key dimensions discussed above (Studer and Ritschard, 2016),
we also added the complexity index (Gabadinho et al., 2010). It aims to capture
the unpredictability and instability of the trajectories. The exact value of this
index is not interpretable, but a higher value can be linked to a higher complexity
of the trajectories.
3.2 Designing and Adapt the Set of Properties
Building on previous experiences, we argue that four key points should be taken into
account while considering the features to extract. First, the properties should have
a meaningful interpretation for the specific applications. Ideally, these indicators
should be linked with relevant theoretical assumptions. This ensures that the
results are interpretable. The features we have extracted here might not be relevant
in all applications. We strongly encourage its adaptation to fit the research issue as
much as possible. Some indicators could be removed, for instance if complexity is
not of central interest, and others added. For instance, when studying professional
integration trajectories it might be useful to add a specific indicator for the age of
the individual at the time he/she lost his/her job.
Second, in order to have meaningful theoretical interpretation, the set of
indicators should be as complete as possible. It should cover all the aspects of
the trajectories that are relevant for the domain under study. In the end, we will
11
conclude on the effect of the previous trajectories taken as a whole, we therefore
need to ensure that all relevant aspects were taken into account.
Third, each indicator should be interpretable on its own, even when the other
indicators are not selected. Practically, a single concept should therefore not be
coded using two (or more) variables. The procedure might end up selecting only
one of the two making the interpretation difficult. For instance, the age at first
birth should not be coded in two features like having the first baby and then, for
those who experienced this event, the age at which is happened. If only one feature
ends up selected, the remaining one will not represent the concept of timing of the
transition to parenthood.
Finally, we should not have several indicators of the same underlying concept.
Our set should be parsimonious for two reasons. It allows mitigating multiple
testing issues and this is important even if most feature selection algorithms take
it into account. It also typically leads to more stable results as feature selection
methods might be affected by highly correlated features.
Parsimoniosity can be further achieved by removing duplicate and constant
indicators. A constant indicator will never help in distinguishing the effect of the
previous trajectory on a given outcome. For instance, in our example looking at
trajectories up to age 40, nobody is in education in the last two age groups (30-35,
35-39) and the related indicators always equal 0. We therefore directly removed
them from the list.
Some indicators might also be exact duplicates. In this case, it is better to
remove them to avoid multicolinearity issues. These indicators can be identified
by looking at the correlation matrix between the indicators. Several methods are
available to do it automatically (e.g. Orestes Cerdeira et al., 2018; Jolliffe, 2002),
even to exclude highly but not perfectly correlated items (Cadima and Jolliffe,
2001). We did not use these methods here, but they might be useful when a large
number of features are automatically extracted. However, we removed the last age
group of our timing features, as it can be directly derived using a linear combination
of the other age groups and the overall time spent in each state.
Summarizing, we need to build the smallest but exhaustive possible set of
theoretically relevant indicators, which are interpretable on their own.
In our empirical example, we extracted 45 indicators for the professional trajec-
tory and 53 for the family one. We now turn to the selection of the relevant ones
in studying the association between family and working histories, and income and
health later in life.
In the rest of the paper, we will refer to all these indicators more generally as
sequence features.
12
4 Feature Selection
The second step in our proposed approach is to identify relevant sequence features
using data mining feature selection’s methods. The aim of such methods is to
automatically select a set of features that is statistically linked with the outcome
variable.
There are two different kinds of feature selection methods. First, the “all-
relevant” approach tries to identify all the features that are significantly linked with
the outcome variable. Specialized methods are required to do so in order to take
multiple testing into account. These methods are useful when we are interested in
looking at an entire set of meaningful features (Degenhardt et al., 2017). However,
as we will show, these methods often select a high number of possibly redundant
features, making the exact interpretation of the effect of each feature difficult.
The second approach, often called “predictive”, tries to directly select a relatively
small set of features with the best predictive power, typically in a regression-like
setting. However, by doing so, it discards relevant features that correlate with one
or several of the selected features.
In this paper, we explore the use of both approaches and compare their results.
For the “all-relevant” approach, we use Boruta (Kursa and Rudnicki, 2010), an
extension of random forest, which was shown to feature among the best methods in
this approach (Degenhardt et al., 2017). We rely on the well-known stability selec-
tion and LASSO regression (Meinshausen and B¨uhlmann, 2010) in the “predictive”
approach. Before presenting these approaches, we discuss the issue of controlling
for confounders in the selection process, as it applies to both methods.
4.1 Controlling for Confounders
When we study a quantitative relationship, we usually want to account for possible
confounders. The same applies when studying the relationships between trajectories
and income or self-rated health. We control for socio-demographic characteristics
that might be correlated with health and economic condition. More specifically, we
control for migration status (Swiss born or not), level of education (having at least
tertiary education) and age to account for age difference at the time of the interview.
Since we have cross-sectional data, age is also equivalent to the cohort, and the
cohorts are strongly linked to life trajectories in Switzerland (e.g. Levy et al.,
2006). We therefore want to control for age/cohorts to avoid selecting trajectories
features because they measure in fact an age/cohort effect. Since professional and
family histories are strongly gendered, we will perform the selection of the features
stratified by gender.
In order to control for possible confounders, we use a two-step approach.
13
•
Step 1. We start by estimating a regression model with the set of control
variables on the outcome. We use a linear regression for income and binomial
regression model for self-rated health.
•
Step 2. We use the residuals of the model estimated in step 1 in the feature
selection algorithm.
By doing so we ensure that our outcome is orthogonal to any confounder.
Therefore, any selected feature should not be linked with a possible confounder
but directly with the outcome. For the analysis of self-rated health, we used the
residuals on the “deviance” scale resulting in a quantitative variable that can be
analyzed with feature selection algorithms in linear models.
4.2 Boruta
The aim of Boruta is to identify all relevant features that are linked to an outcome
variable (Kursa and Rudnicki, 2010). This is particularly interesting if we aim to
understand in detail the relationship between trajectories measured through their
features and the outcome of interest. Boruta featured among the best methods
in a recent review of all-relevant feature selection methods, particularly in low-
dimensional spaces, i.e. when there are fewer features than observations, as in our
case (Degenhardt et al., 2017).
The all-relevant features approach faces two main challenges. First, one should
take multiple testing into account. If we keep the usual 5%-significance threshold
and have 100 independent features, we would end up with 5 false positives, i.e.
features flagged as relevant even if there are independent. Second, as explained by
Kursa and Rudnicki (2010), the lack of a direct relationship with the outcome does
not mean that a feature is not important when considering other features. For this
reason, we need to rely on so-called wrapper methods, that take multiple features
into account at the same time.
Boruta is a wrapper method built around the random forest (RF) approach (for
technical details please refer to Breiman, 2001). RF is commonly used in machine
learning, and only recently it has been started using in social sciences too (Perry,
2013; Berk et al., 2016; Bruno et al., 2018), to rank variables according to their
“importance”. It requires relatively small model tuning is not too computational
intensive and provides reliable performances.
Based on regression tree approach, RF runs the analysis over many sub-datasets
made by randomly selecting features (ensemble algorithm). The outcome is a sort
of model averaging over (possible thousands) regression trees. The advantages of
RF are to reduce overfitting issues and mitigate instability of regression trees. RF
can be used with a large number of variables, even if they are highly correlated. It
identifies interactions and non-linear associations, and any type of variable (numeric,
14
categorical) can be used as input. The typical output of RF is a ranking of the
importance of each variable included in the analysis.
The random forest algorithm does not take any decision about the selection of
the features. It just ranks them. Boruta solves these issues by using a resampling
technique to determine the statistical importance of the features. It distinguishes
those who are “important” from those who are “unimportant”.
The algorithm is based on “shadow features” made by shuffling the values of
the original features. A RF approach is then applied and the importance of both
the shuffled and the original features are computed. Boruta then compares the
importance of the original and the shadow features and marks the original feature
as important if its importance is constantly higher. Repeating this procedure
iteratively and eliminating progressively irrelevant features, we end up with a
robust set of features that have been frequently identified as important. This
procedure allows Boruta to find all the relevant features, including those that are
weakly but significantly associated with the outcome. For more details on Boruta,
please refers to Kursa and Rudnicki (2010).
As mentioned above, Boruta relies on random forest and therefore on decision
trees. For this reason, it is able to capture non-linear relationships between features
and outcomes. It might also detect interaction effects, i.e. a feature that would
have a different effect depending on the value of another feature. For instance,
we might think that the effect of non-working on income at later age depends on
whether a woman live with a partner or not. These properties and Boruta’s ability
to capture weak relationships tend to result in large set of important features.
Table 3 presents the features selected by Boruta for the income and the self-
rated health focusing on women. The results for men are available in the appendix.
The column “Imp.” shows the standardized importance of each feature. A higher
value indicates a more important feature. The column “Prop.” measures the
proportion of resampling in which the feature had a higher importance than the
most important “shadow” feature. We highlighted in blue the feature that reached
this importance at least 95% of the time as this threshold fully account for multiple
testing. Boruta, as other feature selection methods, does not estimate the direction
of the association between features and the outcome but it only shows their relative
importance.
Focusing on income, no sequencing features were identified as important in the
family life trajectory. The order of family life events has no direct or indirect effect
on household income in midlife. However, the timing and the time spent in different
living arrangement conditions (duration) matter. More specifically, the timing and
duration of living with a partner, in union with kids (lower relatively importance),
or single as the most relevant aspect of family trajectories (respectively named in the
table as Part, Part/Kid and Sin) were identified as important. Surprisingly, living
15
Table 3: Features selected by Boruta for income and self-rated health for women.
Family Trajectories.
Feature Income Health
Imp. Prop. Imp. Prop.
Sequencing
Part/Kid 3.45 0.69
Part →Part/Kid 6.75 0.99
Sin →Part/Kid 4.61 0.92
Sin →Part 3.56 0.71
Sin →Part →Part/Kid 4.62 0.91
Timing
20-24 Part/Kid 6.95 0.99
20-24 Part 4.34 0.96 5.84 0.96
20-24 Sin 3.06 0.73 5.62 0.97
25-29 Part/Kid 3.76 0.90 6.85 0.99
25-29 Part 5.12 0.98 4.99 0.93
25-29 Sin 3.68 0.89 4.96 0.92
30-34 Sin/Kid 3.34 0.64
30-34 Part/Kid 2.83 0.67 5.49 0.97
30-34 Sin 3.23 0.78 3.34 0.63
Duration
Overall Sin/Kid 5.37 0.96
Overall Oth 4.36 0.85
Overall Part/Kid 5.15 0.99 9.46 1.00
Overall Part 7.65 1.00 5.77 0.96
Overall Sin 5.76 0.99 6.74 0.99
Complexity
C-index 7.66 1.00
Work Trajectories.
Feature Income Health
Imp. Prop. Imp. Prop.
Sequencing
Edu →FT 3.07 0.75
Edu →FT →PT 3.39 0.86
Edu →PT 2.74 0.64
FT 3.43 0.83
FT →PT 4.16 0.95
NA 3.29 0.64
NW →PT 3.34 0.64
Timing
20-24 Edu 4.52 0.97
20-24 FT 5.55 0.99
20-24 NW 3.63 0.87 7.03 0.99
20-24 PT 3.17 0.77
25-29 FT 4.24 0.95 4.61 0.89
25-29 NW 5.27 0.99 5.09 0.95
25-29 PT 3.72 0.88 3.26 0.61
30-34 FT 3.87 0.92 5.65 0.98
30-34 NW 2.89 0.68 4.13 0.84
30-34 PT 3.80 0.93 4.53 0.88
Duration
Overall Edu 4.30 0.96
Overall FT 6.08 0.99 6.77 0.99
Overall NW 4.87 0.97 6.02 0.97
Overall PT 6.40 1.00 6.58 0.99
Note: Imp: standardized importance. Prop. proportion of resampling in which the feature had a higher importance than the
most important. For the abbreviation of the states, please refer to Table 1.
single with children is not selected. Focusing on work trajectories, the time spent
in the different types of employment status (education, full-time, non-working or
part-time) were found important. Some sequencing features capturing education or
the back-and-forth movements in the labor market were also selected as important,
such as going back to part-time after being in a full-time position. These sequencing
aspects are also found in the timing indicators. Despite the fact we control for
education attainment in the first step of our selection procedure, the time spent in
education at the beginning of the trajectory (i.e., staying in education after age of
20) has been selected as relevant.
Many features selected for income are also selected for health. For the latter,
the timing and sequencing of the state living with a partner with kids seem to be of
key importance, confirming the importance of the timing of child birth (McMunn
et al., 2015). The timing and the overall time spent in full-time, part-time and
non-working are also important.
The sequencing, timing and duration of working states seem to be less relevant
for men than women with much fewer features selected (see Appendix A). This can
be explained by the stability of male working trajectories that are characterized by
long periods of full-time work. Several family features were selected. Interestingly,
living with a partner and children is also associated with health, but at a later age.
Separation (the subsequence Part
→
Sin) was been found to be associated with
income and health among men but not women.
The strengths of Boruta are close to its limits. On the one hand, Boruta selects
16
a high number of features, including weak association, non-linear relationships
and interaction effects among them. It therefore provides a “global” overview of
sequence’s features that are associated with the outcome of the interest. On the
other hand, the high number of selected features makes the interpretation of the
results less straightforward. For this reason, the method is particularly relevant if
one aims to gain a global understanding of the features playing an important role
in the relationship between previous trajectory and later life outcome.
4.3 LASSO and Stability Selection
LASSO regression model is a well-known approach for variable selection (Tibshirani,
1996; Hastie et al., 2015). Differently from Boruta, LASSO aims to find an efficient
subset of features to predict the outcome variable, i.e. income or health in our
case. LASSO uses a penalization method of non-zero coefficients in its estimation
procedure. As a result, when the weight given to penalization is high, many
coefficients are set to zero and are therefore unselected by the model. For further
details on LASSO method, please refers to the seminal work of Tibshirani (1996).
Stability selection (SS) (Meinshausen and B¨uhlmann, 2010) extends LASSO by
using resampling approach to overcome two of its common issues. First, LASSO
does not provide any guidance on the number of features to keep. SS offers such
guidance by looking at risk error, as in usual statistical reasoning. Second, when
correlated variables are used in LASSO, the procedure tends to keep only one
of them, and shrink toward zero the coefficient of the others, even if they were
significant. SS milds this effect by using resampling techniques. This aspect is
crucial while selection features from sequential data since indicators tend to be
highly correlated.
In this article, we use an improvement of Stability Selection proposed by Shah
and Samworth (2012) and further discussed by Hofner et al. (2015). It works by
applying LASSO on random subsample of the data and look at “stable” features,
i.e. feature that are often selected by the model.
When using SS, two parameters need to be set. Hofner et al. (2015) discuss
in detail the choice of these parameters. First, we need to specify the minimum
percentage of subsamples in which a variable was selected in order to be considered
as stable. According to Hofner et al. (2015), any value between 0
.
5 and 1 can be
used and should not have a relevant impact on the result. Following Meinshausen
and B¨uhlmann (2010), we set to cut-off point at 0
.
6. Second, the per-family error
rate (PFER) measures the expected number of false positive. It should be set at
least to the significance threshold (typically
α
= 0
.
05) to fully account for multiple
testing, i.e. when each feature measures a completely different concept. Hofner
et al. (2015) recommend taking a value in the range [
α
;
m·α
] where
m
is the
number of features. Setting the PFER to a value of
m·α
would lead not to take
17
Table 4: Features selected by stability selection for income and self-rated health
for women and men.
Women Men
Income Health Income Health
1 Family: Overall Part* Family: Part →Part/Kid* Work: NW Work: 30-34 NW*
2 Work: FT* Family: 20-24 Part/Kid* Work: Edu →FT
3 Work: FT →PT* Work: Overall Edu**
4 Work: NW Work: 30-34 FT*
5 Work: Edu →FT*
6 Work: 20-24 Edu*
Note: Features selected with P F ER = 1 are highlighted in blue. For the abbreviation of the states, please refer to Table 1. A
star indicates a feature selected by Boruta as well.
multiple testing into account.
Table 4 presents the features selected by the SS algorithm with
P F ER
=
m·α
.
As we expected, the set of features selected by stability selection is much smaller
than the one estimated by Boruta. The number of features selected ranged from
1 to 6 according to the outcome and gender considered. With
P F ER
=
α
, no
features were selected since the procedure would be too conservative as our features
are highly correlated due to the stability of sequences. The features highlighted in
blue were selected with P F ER = 1, the default value in the R package “stabs”.
According to SS and as expected, family trajectory features are mostly important
for women. The overall time spent with a partner is linked with income, while the
pattern and timing of children matter for self-rated health. No features from family
history have been selected among men.
Apart from education, which matters for men and women, income among
women is mostly linked to back-and-forth movement on the labor market as mostly
indicators of sequencing have been kept. This includes the occurrence of a full time,
or non-working spells as well as the full-time
→
part-time pattern. Among men, in
addition to the time spent in education after age 20, the full-time employment after
leaving education and having experienced a period of unemployment (NW) matter
for income level in midlife. Not working between age of 30 and 34 is associated
with later-life health as well.
4.4 Comparing Boruta and Stability Selections Features
In this section, we briefly compare the features selected by Boruta and SS. First
of all, SS identifies a much smaller number of features that were also selected by
Boruta, except the occurrence of not-working spell when studying income. However,
timing and duration features of not-working were flagged as important for women.
Features selected by both methods are indicated with a start in Table 4.
To further compare the results obtained by Boruta and Stability Selection
methods, we first estimated the direct linear relationship between each Boruta’s
18
Table 5: Comparison Features selected by Boruta and Stability Selection for income
and self-rated health for women.
Income
Feature
Sequencing
Work Trajectories
Edu →FT *
Edu →FT →PT A
Edu →PT A
FT *
FT →PT *
Timing
Work Trajectories
20-24 Edu *
20-24 FT
20-24 NW A
20-24 PT
25-29 FT
25-29 NW S
25-29 PT
30-34 FT A
30-34 NW A
30-34 PT
Family Trajectories
20-24 Part
20-24 Sin
25-29 Part/Kid A
25-29 Part A
25-29 Sin
30-34 Part/Kid A
30-34 Sin
Duration
Work Trajectories
Overall Edu S
Overall FT
Overall NW A
Overall PT
Family Trajectories
Overall Part/Kid A
Overall Part *
Overall Sin
Health
Feature
Sequencing
Work Trajectories
NA
NW →PT
Family Trajectories
Part/Kid
Part →Part/Kid *
Sin →Part/Kid A
Sin →Part A
Sin →Part →Part/Kid A
Timing
Work Trajectories
20-24 NW A
25-29 FT
25-28 NW A
25-29 PT
30-34 FT
30-34 NW
30-34 PT
Family Trajectories
20-24 Part/Kid *
20-24 Part
20-24 Sin A
25-29 Part/Kid S
25-29 Part A
25-29 Sin
30-34 Sin/Kid A
30-34 Part/Kid
30-34 Sin
Duration
Work Trajectories
Overall FT
Overall NW
Overall PT
Family Trajectories
Overall Sin/Kid A
Overall Oth
Overall Part/Kid
Overall Part A
Overall Sin
Complexity
Family Trajectories
C-index A
Note: For the abbreviation of the state, please refer to Table 1.
* = features selected by both methods; A = direct linear relationship is accounted for when controlling; S = direct linear
relationship remains significant after controlling; empty = direct linear relationship is not significant in any case.
features and the outcome. We then estimated whether this direct linear relationship
19
is still significant when controlling for the SS features.
These results are presented in Table 5 for women and Table 8 for men as follows:
“*” for features selected by both methods, “A” if the direct linear relationship
identified for Boruta feature is Accounted for when controlling for SS features, “S” if
it remains Significant after controlling, and nothing if the direct linear relationship
is not significant.
The results allow us to illustrate the added value of each approach. First of all,
several Boruta features have no direct linear relationship with the outcome (empty
sign: 13 out of 29 for income and 18 out of 32 for health. For women). This can
be explained by the ability of Boruta to capture non-linear and interaction effects.
These interactions and non-linear relationships are neglected by linear approaches
such as stability selection.
Most of the time, direct linear relationships found by Boruta are accounted
for when controlling for SS features (“A” in the table).
2
In other words, only a
few Boruta’s features are still correlated with the outcome after controlling for SS
features (“S” in the table).
3
The comparisons of the results for men yield to a very
similar conclusion (see Appendix B).
Different reasons can drive this result. On the one side, SS aims to identify
a small set of features highly correlated with the outcome using a linear model.
So, it is strongly focused on finding the best predictive linear model. It is not
surprising then that after controlling for SS features, some correlation found for
Boruta features disappear. On the other side, it also means that the SS’s features
“represents” other sequences features that are correlated with the selected one.
Multiple features might bring a similar information and SS will usually select only
one of them. In other words, SS features might hinder other correlated features.
We should therefore take that into account in our interpretation of the results.
5 Regression Models
Feature selection algorithms allow us to identify the most relevant properties, but
not to understand their effect. As a final step, we run a regression model including
the SS features to interpret the relationships. We do not consider Boruta features,
as their number and correlation would lead to multicolinearity issues.
Table 6 presents several logistic regression on health for women. The base
model only includes our control variables. As expected, the chances to have a
good self-rated health decrease with age and tends to be higher for those born in
Switzerland. In the second model, the sequence analysis typologies of professional
210 features out of 16 for income and 11 out of 14 for self-rated health.
3
Namely, for income: not-working between ages 25-29; overall duration of education after age
20; for health: having a partner and children at ages 25-29.
20
Table 6: Logistic regressions on health for women.
Base Seq. An. Stab. Sel. Complete
Intercept 2.74∗∗∗ 2.33∗∗∗ 2.32∗∗∗ 2.19∗∗∗
(0.54) (0.59) (0.55) (0.59)
Age −0.03∗∗ −0.02∗−0.02∗−0.02
(0.01) (0.01) (0.01) (0.01)
Tertiary educ. −0.02 −0.04 −0.09 −0.09
(0.14) (0.14) (0.14) (0.15)
Born in CH 0.31∗0.34∗0.30∗0.34∗
(0.14) (0.14) (0.14) (0.15)
Family Life
Part →Part/Kid 0.49∗∗∗ 0.61∗∗
(0.14) (0.20)
20-24 Part/Kid −0.20∗∗∗ −0.15∗
(0.05) (0.06)
Professional Cluster (ref=Full-time)
Late FT-NW −0.08 −0.26
(0.24) (0.25)
FT-NW −0.20 −0.34
(0.25) (0.26)
PT −0.01 −0.11
(0.24) (0.24)
Late FT-PT −0.36 −0.51∗
(0.23) (0.23)
Edu-FT −0.07 −0.15
(0.32) (0.32)
NA 0.24 0.20
(0.44) (0.44)
Family life Cluster (ref=Single)
Early parenthood −0.07 −0.15
(0.23) (0.29)
With partner 0.22 0.22
(0.27) (0.27)
Other paths 0.32 0.32
(0.41) (0.41)
Late parenthood 0.60∗0.25
(0.24) (0.28)
AIC 1328.57 1330.39 1303.28 1311.84
BIC 1349.63 1404.11 1334.87 1396.09
Log Likelihood −660.28 −651.20 −645.64 −639.92
∆χ2(vs Base) 18.18 29.29∗∗∗ 40.73∗∗∗
∆χ2(vs Complete) 5.94∗∗∗ 22.55∗∗∗ 11.44
∗∗∗p < 0.001, ∗∗ p < 0.01, ∗p < 0.05
Note: For the abbreviation of the state, please refer to Table 1. Cluster of Sequence Analysis are rep orted in Figure 1 and 2.
and family life trajectories were added. Consistently with the literature, “late
parenthood” type is associated with higher health than those in the “single” or “early
parenthood” clusters. No significant effect is found for working career typology.
More globally, a likelihood ratio test against the base model shows no significant
overall improvement.
The third model includes the two SS features and shows a significant improve-
ment over the base model (∆
χ2
= 29
.
29;
p <
0
.
001). The sequencing feature
“Part
→
Part/Kid” (living with children and partner after a living with a partner
episode) is associated with higher health. The feature “20-24 Part/Kid” (time
spent in living with partner and children between 20 and 24 years old) is associated
with lower self-rated health. In order to correctly interpret the proper effect of
each property, the reference it should be clearly identified. In our case, it is a
woman who did not have a child early in life, and not lived with a partner and
a child after living with a partner. In our sample, this is mostly women without
21
children. Summarizing, we found the lowest self-rated health for early mothers,
followed by women without children and finally late parenthood. All these results
are consistent with previous studies on the link between family trajectories and
midlife health conditions showing the negative effect of having children early in life
(McMunn et al., 2015).
The last model includes the typologies and the SS features to provide a bench-
mark. It shows that adding SA typologies on top of SS features does not significantly
improve the model (likelihood ratio test ∆
χ2
= 11
.
44). In this empirical example,
the SS features capture all the relevant information provided by the SA typologies.
On the contrary, comparing the last and the SA model, we can conclude that
adding SS features to the SA model significantly increases the model fit (likelihood
ratio test ∆
χ2
= 22
.
55
∗∗∗
). SS features therefore provide a more precise information
as these features might be shared by individuals classified in different types. It also
provides a much clearer interpretation.
Table 7 presents linear regression models of household income following the
same modeling strategy as for health. The SA model (second column) highlights
two significant and positive effects on income: living with a partner without
child(ren) and the role of the time spent in education (typology labeled “Edu —
Full-time” type). Unsurprisingly, couple without children is linked with higher
household equivalized income as well as further education. The SS model provides
a more precise interpretation. Aside from living with a partner and education,
it shows the importance of working full-time, the full-time
→
part-time pattern
(both positively associated with later-life income), experiencing nonworking spells
(negative impact on income in midlife). The effect of the Edu
→
FT pattern is
not significant, probably because of multicolinearity (
V IF
= 2
.
9). Here again, this
model significantly improves the base model.
As before, a proper interpretation of these relationships between work-related
features and household income requires us to identify the reference individual: a
woman who did not experience any full-time or non-working spells and without
time spent in education between 20 and 24 years old. These are predominantly
women experiencing a full part-time trajectory.
We then observe a significant effect of experiencing a full-time spell and a
negative effect of non-working spells. These results were expected. Surprisingly,
the “FT
→
PT” pattern (full-time followed by a part-time spell) is associated with
higher household income. This could be explained by a reverse causality, female
workers who decided to reduce their labor force participation are also those who
live already in a wealthy family and can afford it.
Here again, the complete model shows us that all the effects identified by SA
typologies are accounted for by the SS features (i.e. they are not significantly asso-
ciated with the outcome when including SS features). No significant improvement
22
Table 7: Linear Regression on Equivalized Household Income for women.
Base Seq. An. Stab. Sel. Complete
Intercept 33595.51∗∗∗ 23152.00∗∗ 9352.45 7511.66
(8143.71) (8785.52) (8909.74) (9650.40)
Age 472.33∗∗ 613.10∗∗∗ 683.49∗∗∗ 675.52∗∗∗
(149.75) (151.23) (147.59) (149.77)
Tertiary educ. 14295.51∗∗∗ 12412.24∗∗∗ 10511.73∗∗∗ 10687.09∗∗∗
(2148.27) (2162.95) (2158.73) (2169.23)
Born in CH 4023.59 4567.69∗5220.70∗5185.78∗
(2166.80) (2157.51) (2142.27) (2164.30)
Family Life
Overall Part 959.41∗∗∗ 922.78∗
(232.40) (399.32)
Working Life
FT 7522.80∗8091.90
(3666.16) (4234.45)
FT →PT 5977.57∗∗ 7003.72∗
(2280.57) (2774.28)
NW −6214.28∗∗ −7008.99∗
(2143.38) (3121.32)
Edu →FT 1385.17 1130.34
(4151.16) (4201.92)
20-24 Edu 3972.86∗∗∗ 3678.14∗∗
(1066.36) (1161.58)
Professional Cluster (ref=Full-time)
Late FT-NW −2188.74 −346.42
(3562.81) (4501.34)
FT-NW −7193.93 125.31
(3840.47) (4760.45)
PT 2390.33 1559.38
(3494.07) (4051.02)
Late FT-PT 2229.19 −3952.89
(3482.17) (4001.61)
Edu-FT 18376.90∗∗∗ 3450.20
(4722.45) (5862.58)
NA −305.77 −956.49
(6194.77) (6164.14)
Family life Cluster (ref=Single)
Early parenthood 979.11 3163.78
(3701.11) (3716.98)
With partner 12117.93∗∗ 2602.80
(4159.02) (5977.35)
Other paths −3917.41 −3079.26
(5721.13) (5677.09)
Late parenthood 4078.11 2862.18
(3640.47) (3698.68)
R20.03 0.06 0.09 0.09
∆F(vs Base) 4.72∗∗∗ 15.06∗∗∗ 5.94∗∗∗
∆F(vs Complete) 5.94∗∗∗ 7.74∗∗∗ 0.49
BIC 35969.09 35995.23 35924.61 35992.73
∗∗∗p < 0.001, ∗∗ p < 0.01, ∗p < 0.05
Note: For the abbreviation of the state, please refer to Table 1. Cluster of Sequence Analysis are rep orted in Figure 1 and 2.
is found by adding the SA typologies to the SS model. Unsurprisingly, here again
the SS approach provides the most parsimonious model.
We do not discuss in detail the model for men as they lead to similar conclusion.
In all cases, SS features were able to detect the association found with the SA
typologies. Furthermore, it also makes it possible to interpret more precisely the
effect of the previous trajectory on the outcome of interest. Finally, in some models,
additional interpretations were made possible by identifying features that were not
taken into account in the typology.
23
6 Running the Analysis in R
We briefly present here how to run the proposed methodology in
R
. Nevertheless,
feature selection approaches are available in other software too.
The automatic extraction of features is available using the
seqpropclust
function available in the
WeightedCluster
package (Studer, 2013, 2018). Aside
from the state sequence object
myseq
, one needs to specify the
properties
to be
extracted, and the
prop.only=TRUE
arguments to retrieve the list of extracted
features. The function can extract other properties as well, see Studer (2018) for
more information.
library(WeightedCluster)
## Extracting features
features <- seqpropclust(myseq,
properties=c("pattern","agerange","duration"),
prop.only = TRUE)
## Controlling for confounders
regconfond <- glm(outcome˜confounders)
## Running Boruta
library(Boruta)
Boruta(residuals(regconfond)˜features)
## Running stability selection
library(stabs)
stabsel(features, residuals(regconfond), cutoff=0.6,PFER=1)
Once the features are extracted, the Boruta can be run using the
Boruta
function
from the package of the same name (Kursa and Rudnicki, 2010). The
outcome
variable and the sequence
features
are specified through the usual
R
formula
interface. Additional parameters of interest include
maxRuns
and
num.trees
which
can be both increased to get more stable results.
Stability selection using LASSO regressions is available in the
stabs
package
(Hofner and Hothorn, 2017) and the
stabsel
function. The
features
and
outcome
should be specified. Threshold for stability can be set using the
cutoff
argument,
and the acceptable error through the PFER argument.
24
7 Conclusion
In this article, we proposed a new methodological approach to study the link
between a previous trajectory and a given outcome. It starts by automatically
extracting properties of trajectories coded as sequences. We proposed indicators
to capture meaningful aspects of trajectories in a life-course perspective, namely
duration, complexity, sequencing, and timing. Although we discussed an automatic
feature extraction procedure, we also emphasized the need to adapt it for specific
applications, by removing or adding indicators.
In a second step, relevant features are selected among the previously defined set
of properties. We discussed two approaches. First, Boruta is an all relevant feature
selection procedure able to capture non-linear and interaction effects. Second,
stability selection using LASSO regression identifies a relevant subset of features to
be included in subsequent analysis. These two approaches are complementary, as
the first one allows to take a broader look on relevant trajectories’ properties, while
the second allows identifying a relevant subset for future uses. While presenting
this framework, we also discussed how to consider possible confounders based on
residuals.
Finally, the effects of relevant features can be interpreted using regression
modeling. In our sample applications, we saw that the approaches provided more
precise results and interpretation than the usual sequence analysis approaches. Our
illustrative examples on the effect of family and working lives on later-life economic
and health condition further showed that the proposed approach provides more
precise results and interpretation than commonly used methods such as sequence
analysis typologies or user-defined indicators.
The proposed approach is complementary to the more traditional SA typology.
Both approaches aim to summarize the complexity of life trajectories, but they have
different aims. The goal of sequence typologies is to provide a descriptive “global”
view on trajectories, by identifying similar patterns. The feature selection approach
proposed here aims to identify the specific characteristics of the trajectories that
are significantly linked to a later-life outcome. Depending on the research question,
one or the other approaches could be more useful.
While the proposed approach is promising for many applications, we see rooms
for future developments. First, Boruta results showed us interaction and non-
linear effects between trajectory indicators that should be taken into account. The
resulting set of selected features is relatively large. Further work is needed to
include them in the regression model afterward. This would allow us to study the
combined effect of the intertwinement of different life domains, one of the core
principles of the life-course paradigm.
Many feature selection algorithms were developed in the data-mining literature.
We used here two approaches that are well recognized. However, other approaches
25
might be useful in life-course research too. Among others, the group-LASSO
approach, a method that takes prior information on the grouping of the features
into account (Yuan and Lin, 2006), might be interesting to consider life-domain
interactions.
In this empirical example, the professional and family trajectories were con-
sidered separately, without accounting directly for the interaction between them,
even if Boruta might have identified some of them. Future studies might look
more specifically on life domains interactions using, for instance, multichannel
techniques.
Finally, future works should focus on missing data to better understand how to
properly treat them. In this work, missing states was infrequent and was mostly
made of short episode. The results are therefore probably not strongly affected by
missing data.
The proposed approach can also be used as starting point for further complex
analysis. For instance, for causal inference, the features selected by Boruta could
be used as matching variables in a propensity score matching framework if we want
to match individuals according to their previous trajectories. Further studies are
needed to compare this to existing approaches that use the (dis)similarity between
previous trajectories as matching criteria (e.g. Barban et al., 2017).
References
Agrawal, Rakesh and Ramakrishnan Srikant. 1995. “Mining Sequential Patterns.”
In Proceedings of the International Conference on Data Engeneering (ICDE),
Taipei, Taiwan, edited by Philip S. Yu and Arbee L. P. Chen, pp. 487–499. IEEE
Computer Society.
Albrecht, James W and Susan Vromon. 1999. “Unemployment Finance Companse-
tion and Efficiency Wages.” Journal of Labor Economics 17:141–167.
Arulampalam, Wiji. 2008. “Is Unemployment Really Scarring? Effects of Unem-
ployment Experiences on Wages.” The Economic Journal 111:586–606.
Barban, Nicola, Xavier de Luna, E Lundholm, I Svensson, and Francesco C. Billari.
2017. “Causal Effects of the Timing of Life-course Events: Age at Retirement
and Subsequent Health.” Sociological Methods & Research pp. 1–34.
Beblo, Miriam and Elke Wolf. 2002. “How much does a year off cost? Estimating
the wage effects of employment breaks and part-time periods.” Brussels Economic
Review 45:191–217.
26
Becker, Gary S. 1993. Human Capital: A theoretical and Empirical Analysis with
Special Reference in Education. The University of Chicago Press. 3rd Edition.
Berk, Richard A, Susan B Sorenson, and Geoffrey Barnes. 2016. “Forecasting
domestic violence: a machine learning approach to help inform arraignment
decisions.” J. Empir. Legal Stud 13:94–115.
Billari, Francesco C., Johannes F¨urnkranz, and Alexia Prskawetz. 2000. “Timing,
sequencing, and quantum of life course events: a machine learning approach.”
Working Paper 010, Max Planck Institute for Demographic Research, Rostock.
Blossfeld, Hans-Peter, Katrin Golsch, and G¨otz Rohwer. 2007. Event History
Analysis with Stata. Mahwah NJ: Lawrence Erlbaum.
Breiman, Leo. 2001. “Random Forests.” Machine Learning 45:5–32.
Bruno, Arpino, Marco Le Moglie, and Letizia Mencarini. 2018. “Machine-Learning
techniques for family demography: An application of random forests to the
analysis of divorce determinants in Germany.” RECSM Working Paper 56.
Budig, Michelle J and Paula England. 2001. “The Wage Penalty for Motherhood.”
American Sociological Review 66:204–225.
Cadima, Jorge F. C. L. and Ian T. Jolliffe. 2001. “Variable Selection and the
Interpretation of Principal Subspaces.” Journal of Agricultural, Biological, and
Environmental Statistics 6:62–79.
Clark, Andrew, Yannis Georgellis, and Peter Sanfey. 2001. “Scarring: The Psycho-
logical Impact of Past Unemployment.” Economica 68:221–241.
Correll, Shelley J, Stephen Bernard, and In Paik. 2007. “Getting a Job: Is There a
Motherhood Penalty?” American Journal of Sociology 122:1297–1339.
Daly, Michael J. and Liam Delaney. 2013. “The scarring effect of unemployment
throughout adulthood on psychological distress at age 50: Estimates controlling
for early adulthood distress and childhood psychological factors.” Social Science
and Medicine 80:19–23.
Dannesfer, Dale. 2003. “Cumulative Advantage Disadvantage and the Life Course:
Cross-Fertilizing Age and Social Science Theory.” The Journals of Gerontology:
Series B 58:327–337.
de Linde Leonard, Megan and T. Stanley. 2015. “Married with children: What
remains when observable biases are removed from the reported male marriage
wage premium.” Labour Economics 33:72–80.
27
Degenhardt, Frauke, Stephan Seifert, and Silke Szymczak. 2017. “Evaluation of
variable selection methods for random forests and omics data sets.” Briefings in
Bioinformatics 20:492–503.
Doblhammer, Gabriele. 2000. “Reproductive history and mortality later in life:
A comparative study of England and Wales and Austria.” Population Studies
54:169–176.
Elder, Glen H., Monica Kirkpatrick Johnson, and Robert Crosnoe. 2003. “The
Emergence and Development of Life Course Theory.” In Handbook of the Life
Course, edited by JeylanT. Mortimer and MichaelJ. Shanahan, Handbooks of
Sociology and Social Research, pp. 3–19. Springer US.
Eriksson, Stefan and Dan-Olof Rooth. 2014. “Do Employers Use Unemployment as
a Sorting Criterion When Hiring? Evidence from a Field Experiment.” American
Economic Review 104:1014–39.
Falba, Tracy, Hsun-Mei Teng, Jody L Sindelar, and William T Gallo. 2005. “The
effect of involuntary job loss on smoking intensity and relapse.” Addiction
100:1330–1339.
Fasang, Anette Eva. 2012. “Retirement Patterns and Income Inequality.” Social
Forces 90:685–711.
Gabadinho, Alexis, Gilbert Ritschard, Matthias Studer, and Nicolas S. M¨uller. 2010.
“Indice de complexit´e pour le tri et la comparaison de s´equences cat´egorielles.”
Revue des nouvelles technologies de l’information RNTI E-19:61–66.
Gabriel, Rainer, Michel Oris, Matthias Studer, and Marie Baeriswyl. 2014. “The
persistence of social stratification? A life course perspective on old-age poverty
in Switzerland.” Swiss Journal of Sociology (forthcoming).
Gibbons, Robert and Lawrence F Katz. 1991. “Layoffs and lemons.” Journal of
Labor Economics 9:351–380.
Glauber, Rebecca. 2008. “Race and Gender in Families and at Work: The Father-
hood Wage Premium.” Gender & Society 22:8–30.
Gregg, Paul. 2001. “The Impact of Youth Unemployment on Adult Unemployment
in the NCDS.” The Economic Journal 111:626–653.
Grundy, Emily and Gemma Holt. 2000. “Adult life experiences and health in early
old age in Great Britain.” Social Science and Medicine 51:1061–1074.
28
Grundy, Emily and Oystein Kravdal. 2007. “Reproductive history and mortality
in late middle age among Norwegian men and women.” American Journal of
Epidemiology 167:271–279.
Hastie, Trevor, Robert Tibshirani, and Martin Wainwright. 2015. Statistical
Learning with Sparsity: The Lasso and Generalizations. Chapman & Hall/CRC.
Heckman, James and George Borjas. 1980. “Does Unemployment Cause Future
Unemployment? Definitions, Questions and Answers from a Continuous Time
Model of Heterogeneity and State Dependence.” Economica 47:247–83.
Henretta, John C. 2007. “Early childbearing, marital status and women’s health
and mortality after age 50.” Journal of Health and Social Behavior 48:254–266.
Henretta, John C. 2010. “Lifetime marital history and mortality after age 50.”
Journal of Aging and Health 22:1198–1212.
Hill, Martha. 1979. “The Wage Effects of Marital Status and Children.” Journal of
Human Resources 14:579–594.
Hofner, Benjamin, Luigi Boccuto, and Markus G¨oker. 2015. “Controlling false
discoveries in high-dimensional situations: boosting with stability selection.” BMC
Bioinformatics 16.
Hofner, Benjamin and Torsten Hothorn. 2017. stabs: Stability Selection with Error
Control. R package version 0.6-3.
Hoven, Hanno, Nico Dragano, David Blane, and Morten Wahrendorf. 2017. “Early
Adversity and Late Life Employment History—A Sequence Analysis Based on
SHARE.” Work, Aging and Retirement 4:238–250.
Hurt, Lisa S, C Ronsmans, and S L Thomas. 2006. “The effect of number of births
on women’s mortality: Systematic review of the evidence for women who have
completed their childbearing.” Population Studies 60:55–71.
Jolliffe, Ian T. 2002. Principal Component Analysis. Springer Series in Statistics.,
second edition edition.
Kravadal, Oystein, Emily Grundy, T Lyngstad, and K Wiik. 2012. “Family life
history and late mid-life mortality in Norway.” Population and Development
Review 38:237–257.
Kursa, Miron B. and Witold R. Rudnicki. 2010. “Feature Selection with the Boruta
Package.” Journal of Statistical Software 36.
29
Laitinen, Jaana, Ellen Ek, and Ulla Sovio. 2005. “Stress-related eating and drinking
behavior and body mass index and predictors of this behavior.” Preventive
Medicine 34:29–39.
Lesnard, Laurent. 2010. “Setting Cost in Optimal Matching to Uncover Con-
temporaneous Socio-Temporal Patterns.” Sociological Methods and Research
38:389–419.
Leung, Man Yee Mallory, Fane Groes, and Raul Santaeulalia-Llopis. 2016. “he Re-
lationship between Age at First Birth and Mother’s Lifetime Earnings: Evidence
from Danish Data.” PloS one 11.
Levy, Ren´e, Jacques-Antoine Gauthier, and Eric Widmer. 2006. “Entre contraintes
institutionnelle et domestique : les parcours de vie masculins et f´eminins en
Suisse.” Cahiers canadiens de sociologie 31:461–489.
McKetta, Sarah, Seth J. Prins, Jonathan Platt, Lisa M. Bates, and Katherine
Keyes. 2018. “Social sequencing to determine patterns in health and work-family
trajectories for U.S. women, 1968–2013.” SSM - Population Health 6:301–308.
McMunn, Anne, Rebecca E Lacey, Meena Kumari, Diana Worts, Peggy McDonough,
and Amanda Sacker. 2015. “Work-family life courses and metabolic markers in
mid-life: evidence from the British National Child Development Study.” Journal
of Epidemiology and Community Health 70:481–487.
Meinshausen, Nicolai and Peter B¨uhlmann. 2010. “Stability selection.” Journal of
the Royal Statistical Society: Series B (Statistical Methodology) 72:417–473.
Miller, Amalia. 2011. “The effects of motherhood timing on career path.” Journal
of Population Economics 24:1071–1100.
O
´
Flaherty, Martin, Janeen Baxter, Michele Haynes, and G Turrell. 2016. “The
family life course and health: partnership, fertility histories, and later-life physical
health trajectories in Australia.” Demography 53:1–28.
O’Rand, Angela. 2002. “Cumulative Advantage Theory in Life Course Research.”
Annual review of gerontology & geriatrics 22:14–30.
Orestes Cerdeira, Jorge, Pedro Duarte Silva, Jorge Cadima, and Manuel Minhoto.
2018. subselect: Selecting Variable Subsets. R package version 0.14.
Paul, Karsten I and Klaus Moser. 2009. “Unemployment Impairs Mental Health:
Meta-Analyses.” Journal of Vocational Behaviour 74:264–282.
30
Perry, Chris. 2013. “Machine learning and conflict prediction: a use case.” Stab.
Int. J. Secur. Dev. 2.
Pohlig, Matthias. 2019. “Unemployment sequences and the risk of poverty: from
counting duration to contextualizing sequences.” Socio-Economic Review .
Ribar, David. 2004. “What Do Social Scientists Know About the Benefits of
Marriage? A Review of Quantitative Methodologies.” IZA Discussion Paper 998.
Rossignon, Florence, Matthias Studer, Jacques-Antoine Gauthier, and Jean-
Marie Le Goff. 2016. “Childhood family structure and home-leaving: A combina-
tion of survival and sequence analyses.” In LaCOSA II .
Sabbath, Erika L., Ivan Mej´ıa Guevara, M. Maria Glymour, and Lisa F. Berkman.
2015. “Use of Life Course Work–Family Profiles to Predict Mortality Risk Among
US Women.” American Journal of Public Health 105:e96–e102.
Shah, Rajen D. and Richard J. Samworth. 2012. “Variable selection with error
control: another look at stability selection.” Journal of the Royal Statistical
Society: Series B (Statistical Methodology) 75:55–80.
Sironi, Maria. 2018. “Fertility histories and chronic conditions later in life in
Europe.” European Journal of Ageing (forthcoming).
Spivey, Christy. 2005. “Time Off at What Price? The Effects of Career Interruptions
on Earnings.” ILR Review 59:119–140.
Strully, Kate W. 2009. “Job loss and health in the U.S. labor market.” Demography
46:221–246.
Studer, Matthias. 2013. “WeightedCluster Library Manual: A practical guide to
creating typologies of trajectories in the social sciences with R.” LIVES Working
Papers 24, NCCR LIVES, Switzerland.
Studer, Matthias. 2018. “Divisive Property-Based and Fuzzy Clustering for Sequence
Analysis.” In Sequence Analysis and Related Approaches: Innovative Methods
and Applications, edited by Gilbert Ritschard and Matthias Studer, volume 10
of Life Course Research and Social Policies, chapter 13. Springer.
Studer, Matthias, Nicolas S. M¨uller, Gilbert Ritschard, and Alexis Gabadinho.
2010. “Classer, discriminer et visualiser des s´equences d’´ev´enements.” Revue des
nouvelles technologies de l’information RNTI E-19:37–48.
31
Studer, Matthias and Gilbert Ritschard. 2016. “What Matters in Differences
between Life Trajectories: A Comparative Review of Sequence Dissimilarity
Measures.” Journal of the Royal Statistical Society, Series A 179:481–511.
Tibshirani, Robert. 1996. “Regression Shrinkage and Selection via the Lasso.”
Journal of the Royal Statistical Society: Series B (Statistical Methodology) 58:267–
288.
Tillmann, Robin, Marieke Voorpostel, E Antal, Ursina Kuhn, Florence Lebert,
Val´erie-Anne Ryser, Oliver Lipps, and Wernli Boris. 2016. “The Swiss Household
Panel Study: Observing social change since 1999.” Longitudinal and Life Course
Studies 7:64–78.
Vanhoutte, Bram, Morten Wahrendorf, and Jacques Yzet Nazroo. 2017. “Duration,
timing and order: How housing histories relate to later life wellbeing.” Longitudinal
and Life course Studies 8.
Widmer, Eric D., Ren´e Levy, Alexandre Pollien, Rapha¨el Hammer, and Jacques-
Antoine Gauthier. 2003. “Between Standardisation, Individualisation and Gen-
dering: An Analysis of Personal Life Courses in Switzerland.” Swiss Journal of
Sociology 29:35–65.
Williams, Krist, Sharon Sassler, Arianne Frech, Fenaba Addo, and Elizabeth
Cooksey. 2011. “Nonmarital childbearing, union history, and women’s health at
midlife.” American Sociological Review 76:465–486.
Yuan, Ming and Yi Lin. 2006. “Model selection and estimation in regression with
grouped variables.” Journal of the Royal Statistical Society: Series B (Statistical
Methodology) 68:49–67.
Zaki, Mohammed Javeed. 2001. “SPADE: An Efficient Algorithm for Mining
Frequent Sequences.” Machine Learning 42:31–60.
Zoutewelle-Terovan, Mioara, Victor van der Geest, Aart Liefbroer, and Catrien
Bijleveld. 2012. “Criminality and Family Formation.” Crime & Delinquency
60:1209–1234.
32
A Boruta results for men
Table 8: Features selected by Boruta for income and self-rated health for men.
Family Trajectories.
Feature Income Health
Imp. Prop. Imp. Prop.
Sequencing
Oth →Part/Kid 3.82 0.70
Part/Kid 2.83 0.65 4.62 0.86
Part 3.99 0.75
Part →Part 3.48 0.80
Part →Sin 3.09 0.73 5.19 0.91
Part →Sin →Part 3.72 0.67
Sin →Oth 3.79 0.70
Sin →Oth →Part/Kid 4.02 0.75
Sin →Part/Kid 4.03 0.75
Sin →Part 4.20 0.81
Sin →Part →Part 3.87 0.88
Sin →Part →Sin 3.58 0.81 5.69 0.94
Sin →Sin 4.63 0.87
Timing
20-24 Oth 4.26 0.78
20-24 Part 3.64 0.83
20-24 Sin 2.98 0.67 5.11 0.89
25-29 Oth 5.43 0.92
25-29 Part/Kid 5.30 0.98 7.61 1.00
25-29 Part 3.10 0.72 4.82 0.85
25-29 Sin 5.05 0.90
30-34 Oth 5.82 0.95
30-34 Part/Kid 3.11 0.67 7.25 1.00
30-34 Part 3.66 0.85 6.35 0.98
30-34 Sin 4.20 0.77
Duration
Overall Oth 7.41 0.98
Overall Part/Kid 5.54 0.98 10.30 1.00
Overall Part 5.21 0.98 8.94 1.00
Overall Sin 3.82 0.86 8.81 1.00
Complexity
Complexity Index 4.02 0.90 9.63 1.00
Work Trajectories.
Feature Income Health
Imp. Prop. Imp. Prop.
Sequencing
Edu →FT →PT 3.71 0.83
FT →NW 5.85 0.95
NW 5.45 0.92
NW →FT 3.44 0.59
Timing
20-24 Edu 4.76 0.95
25-29 Edu 3.96 0.86
30-34 FT 4.04 0.88
30-34 NW 7.12 0.99
Duration
Overall Edu 5.24 0.97
Overall FT 3.24 0.72
Overall NW 6.21 0.97
Overall PT 3.69 0.85
33
B Comparison Boruta and Feature Selection re-
sults for men
Table 9: Features selected by Boruta for income and self-rated health for men.
Income
Feature
Sequencing
Work Trajectories
Edu →FT →PT
Family Trajectories
Part/Kid S
Part →Part S
Part →Sin
Sin →Part →Part
Sin →Part →Sin
Timing
Work Trajectories
20-24 Edu A
25-29 Edu A
30-34 FT *
Family Trajectories
20-24 Part
20-24 Sin
25-29 Part/Kid
25-29 Part
30-34 Part/Kid
30-34 Part
Duration
Work Trajectories
Overall Edu *
Overall FT
Overall PT
Family Trajectories
Overall Part/Kid A
Overall Part S
Overall Sin
Complexity
Work Trajectories
Complexity Index
Health
Feature
Sequencing
Work Trajectories
FT →NW S
NW S
NW →FT A
Family Trajectories
Oth →Part/Kid
Part/Kid
Part A
Part →Sin
Part →Sin →Part
Sin →Oth
Sin →Oth →Part/Kid
Sin →Part/Kid
Sin →Part
Sin →Part →Sin
Sin →Sin
Timing
Work Trajectories
30-34 NW *
Family Trajectories
20-24 Oth
20-24 Sin
25-29 Oth
25-29 Part/Kid S
25-29 Part S
25-29 Sin
30-34 Oth A
30-34 Part/Kid
30-34 Part S
30-34 Sin
Duration
Work Trajectories
Overall NW A
Family Trajectories
Overall Oth
Overall Part/Kid
Overall Part S
Overall Sin
Complexity
Family Trajectories
Complexity Index
Note: For the abbreviation of the state, please refer to Table 1.
* = features selected by both methods; A = direct linear relationship is accounted for when controlling; S = direct linear
relationship remains significant after controlling; empty = direct linear relationship is not significant in any case.
34
C Regression results for men
Table 10: Logistic regressions on health for men.
Base Seq. An. Stab. Sel. Complete
Intercept 2.85∗∗∗ 2.91∗∗∗ 3.21∗∗∗ 3.14∗∗∗
(0.66) (0.70) (0.68) (0.71)
Age −0.03∗−0.03∗−0.03∗∗ −0.03∗
(0.01) (0.01) (0.01) (0.01)
Tertiary educ. 0.47∗∗ 0.50∗∗ 0.47∗∗ 0.48∗∗
(0.18) (0.18) (0.18) (0.18)
Born in CH 0.49∗∗ 0.47∗∗ 0.46∗∗ 0.44∗
(0.17) (0.18) (0.18) (0.18)
30-34 NW −0.71∗∗∗ −0.79∗∗
(0.18) (0.28)
Professional Cluster (ref=Full-time)
Late FT-NW −1.93∗∗ −0.38
(0.74) (0.98)
FT-NW −2.19 1.72
(1.48) (2.01)
PT −0.18 −0.15
(0.40) (0.40)
Late FT-PT −0.26 −0.27
(0.56) (0.56)
Edu-FT −0.03 −0.05
(0.26) (0.26)
NA −0.30 −0.34
(0.79) (0.79)
Cohabitation Cluster (ref=Single)
Early parenthood −0.20 −0.20
(0.26) (0.27)
With partner 0.40 0.41
(0.33) (0.34)
Other paths −0.42 −0.47
(0.35) (0.35)
Late parenthood 0.06 0.06
(0.24) (0.24)
AIC 910.76 916.34 897.23 909.75
BIC 931.06 987.42 922.61 985.90
Log Likelihood −451.38 −444.17 −443.62 −439.88
∆χ2(vs Base) 14.41 15.53∗∗∗ 23.00∗
∆χ2(vs Complete) 23.00∗8.59∗∗ 7.48
∗∗∗p < 0.001, ∗∗ p < 0.01, ∗p < 0.05
35
Table 11: Linear regression on standardized household income for men.
Base Seq. An. Stab. Sel. Complete
Intercept 23828.06∗23978.86∗15736.06 19533.85
(10141.48) (10415.81) (10743.23) (13866.84)
Age 617.02∗∗∗ 652.04∗∗∗ 550.45∗∗ 584.60∗∗
(185.16) (184.70) (182.77) (185.26)
Tertiary educ. 21089.93∗∗∗ 17798.61∗∗∗ 17133.64∗∗∗ 17115.24∗∗∗
(2815.74) (2827.91) (2811.20) (2823.09)
Born in CH 5931.39∗7695.88∗∗ 8718.30∗∗ 8369.80∗∗
(2729.46) (2722.93) (2732.32) (2754.07)
Overall Edu 3593.51∗∗∗ 2973.84∗
(1024.47) (1311.61)
30-34 FT 1796.30 1427.80
(979.36) (1882.32)
NW −15769.75∗−14866.84∗
(6300.30) (6879.35)
Edu →FT 5144.14 3467.91
(5113.73) (5255.09)
Professional Cluster (ref=Full-time)
Late FT-NW −9976.19 8183.00
(17784.21) (19254.94)
FT-NW −45351.73 −31127.89
(27235.83) (29160.91)
PT −5591.29 −1827.54
(5842.43) (10307.74)
Late FT-PT −1364.23 1731.20
(8728.32) (10225.14)
Edu-FT 23910.12∗∗∗ 6176.80
(3766.40) (6981.74)
NA −11621.80 −5341.30
(13117.06) (15004.19)
Cohabitation Cluster (ref=Single)
Early parenthood −6541.92 −6148.36
(4153.91) (4137.29)
With partner −244.63 484.98
(4529.64) (4531.78)
Other paths −6388.10 −5814.24
(5901.85) (5890.23)
Late parenthood −5995.59 −5088.41
(3606.73) (3604.00)
R20.05 0.09 0.10 0.10
∆F(vs Base) 5.50∗∗∗ 15.96∗∗∗ 5.05∗∗∗
∆F(vs Complete) 5.05∗∗∗ 3.80∗∗∗ 0.70
BIC 30996.03 31013.07 30961.95 31026.34
∗∗∗p < 0.001, ∗∗ p < 0.01, ∗p < 0.05
36
