An Assessment of the Adverse Effects of Teenage
Fatherhood on Later Life Outcomes
T.G. (Ties) Busschers
1June 7, 2018
MSc. Thesis
Abstract
This thesis studies the causal impact of teenage fatherhood on later life outcomes in 13 European countries. The adverse effects of teenage fatherhood are compared and contrasted with the adverse effects of teenage motherhood. The SHARE and SHARELIFE dataset enables the assessment of the adverse effects of teenage
parenthood at various stages of one’s life. Employing Propensity Score Matching and a parametric approach, this thesis finds that the adverse effects of teenage fatherhood are of a similar order and magnitude as the detrimental effects attached to teenage motherhood. Both teenage fatherhood and teenage motherhood substantially reduces educational attainment, the likelihood of finding your lifetime partner before the age of 46 and late-life physical and mental health. Moreover, lower educational attainment, caused by teenage parenthood, drives a substantial part of the adverse effects of teenage parenthood on late-life physical health. Lastly, the results of this thesis imply that (supra)-national policies and programs should be developed to provide adequate aid for teenage fathers. These programs are currently lacking.
Keywords: Teenage Fatherhood, Teenage Motherhood, Semi-Parametric
Identification.
JEL Classification: I12, J12, J14, J16
Supervisors: dr. J.O. (Jochen) Mierau & R.D. (Roel) Freriks, MSc. Course Code: EBM877A20
1Student Number: S2556324
1
Preface
This thesis is part of my final project for the Master Economics at the University of
Groningen. Officially, this project has started in February, and today, the beginning of June, this project is completed. In reality, this project has started in November 2017 when I took an interest in an Economics discipline unfamiliar to me at that point of time. In the course Health Economics and Policy I was confronted with the Economics behind an ageing population. This course convinced me that Health Economics, and the economists Health Economics brings forth, are imperative to solve the issues society faces with an ageing population.
Economics as a science is often misunderstood by the layman: how can an Economics student graduate with a thesis on the causal impact of teenage father- and motherhood on later life outcomes? For me, Economics is characterized by incentives and trade-offs, these can be found in everyday decisions. Such as, the decision to either take the bus or to go by bike to the university library, or in life-altering decisions, such as the decision to become a teenage parent.
I profoundly enjoyed most parts of this project, this is largely due to the inclusion of Propensity Score Matching techniques and other statistical approaches not taught in the standard Economics curriculum at the University of Groningen. This allowed me to indulge in new material to study and investigate, which provided me with great satisfaction.
Of course, this thesis would not have existed in its current form without the aid of my supervisors, and other interested individuals. I would like to thank Roel Freriks for his guidance throughout this thesis and for being the most approachable supervisor one can imagine. Moreover, I thank dr. Mierau for his excellent help and critical perspective on this thesis. I am grateful for the helpful comments by some of my classmates. Lastly, I would like to thank my girlfriend for encouraging me to take a course in Health Economics, which ultimately led to this thesis.
June 7, 2018.
2
1. Introduction
In 1968 Nobel laureate Gary Becker extended economic science with his seminal contribution to topics previously considered as the matter of interest for sociologists and medical
researchers. In 1971, the academic literature was enriched with economic theory applied to a trade-off between the quantity and quality of children. This opened the gates for economists to investigate what tickled their interest. In merely 5 years after this, 1976, economic theory and practices were applied to teenage and adolescent pregnancy and childbearing by Baldwin (1976).
The economic literature on teenage parenthood has become synonymous with the research on teenage motherhood. As very little resources have been devoted to the study of teenage fatherhood. This phenomenon is exemplified by Morrison, Samulon and Zellman (1981). Morrison et al. conducted a study called ‘Teenage Parenthood: A review of risks and
consequences’, in which not a single word is dedicated to teenage fatherhood and its risks and consequences. Teenage fatherhood is either unrightfully so deemed as uninteresting or its adverse effects were assumed to be negligible.
The adverse effects of teenage motherhood are conspicuous. Pregnancy, parturition and lactation present physiological challenges not endured by men. The adverse effects of teenage fatherhood, however, are less apparent but not less substantial. Although teenage fatherhood is on the decline in Europe, the individual effects are of a considerable size (Treffers, 2003). Moreover, the adverse effects of teenage father- and motherhood can be long-lasting, teenage parents experience lower late-life health. For teenage fathers adequate aid and programs are lacking (Gesell and Van Dijk, 2010). This thesis assesses the adverse effects of teenage fatherhood, the results imply that teenage fathers could benefit from adequate programs, which are currently lacking.
3
For this reason, teenage fathers build lower levels of human capital. This can have long-lasting effects (Venti and Wise, 2015).
A comparison of the later life outcomes of teenage parents with a generic group of non-teenage parents does not yield the causal effect of non-teenage parenthood on later life outcomes. Teenage parenthood is associated with socioeconomic status and childhood characteristics. Teenage parents are more likely to be born into families with a low socioeconomic status than non-teenage parents. For this reason, a comparison of teenage parents with a generic group of individuals (who did not experience teenage parenthood) does not capture the causal
component. Prior to the seminal paper of Geronimus and Korenman (1992), researchers did not take the low socioeconomic status of teenage parents into account. Therefore, the literature prior to Geronimus and Korenman found excessive detrimental effects of teenage parenthood on later life outcomes. These effects were driven by the lower socioeconomic status component, as well as the causal impact component. For this reason, the greater part of this thesis concerns itself with the causal identification of the adverse effects of teenage parenthood.
The strategy for the isolation of the causal effect is two-fold. On one hand, the semi-parametric approach Propensity Score Matching (PSM) is employed. PSM constructs a synthetic control group that is as similar as statistically possible. Subsequently, later life outcomes of teenage parents with the synthetic control are contrasted. On the other hand, a parametric approach, Logit model, is employed with the inclusion of a rich set of control variables. These control variables mitigate the issue of confounding factors hampering the causal identification. Subsequently, consistent lower bound marginal effects are derived following the procedure of Oster (2015).
The empirical sections employ the SHARE and SHARELIFE dataset, which allows this thesis to infer the causal effect of teenage parenthood in several phases of an individual’s life. This thesis finds that in both the short-, medium- and long-term the adverse effects of teenage fatherhood is of a considerable size. These adverse effects do not shrink into insignificance when contrasted to the adverse effects of teenage motherhood. Thus, the results indicate that the relative attention devoted to teenage motherhood, in comparison to teenage fatherhood, cannot be justified by the order and magnitude of the adverse effects.
4
long-term effects of teenage parenthood is rare and tends to focus on teenage motherhood. Second, the combination of the SHARE and SHARELIFE dataset enables this thesis to follow teenage parents in several phases of their life. Thereby, it is possible to infer the causal effects of teenage parenthood in the short-, medium- and long-term. Third, this thesis is the first to explicitly compare and contrast the adverse effects of teenage parenthood for both genders. Thereby, this thesis is able to conclude that the results for both genders are of a similar order and magnitude.
This thesis is organized as follows. Section 2 discusses the relevant literature. Section 3 describes the data. Section 4 discusses the empirical strategy and results. Section 5 investigates potential transmission channels that lead to lower late-life health of teenage parents. Section 6 infers the robustness of the PSM results, section 7 discusses and concludes this thesis.
2. Literature Review
2.1 Teenage Motherhood
Teenage mothers are disproportionately born into families with a low socioeconomic status. That is, those experiencing teenage motherhood are not a random subset of society. A low socioeconomic status leads to lower late-life outcomes (Landbergis et al., 2003), teenage motherhood leads to lower late-life outcomes as well (Angelini and Mierau, 2016). For this reason, variables related to socioeconomic status should be taken into account. Lundberg and Plotnick (1995) develop a model of economic incentives, in the context of the decision to become pregnant, and consequently whether to bear a child or not. Lundberg and Plotnick thus show that the decision to take children is non-random. These (confounding) factors should be controlled for, if one wants to capture causal effects. Hamburg (1986) and Burton (1990) show the non-random characteristics of (teen) childbearing as well.
5
matching techniques. In turn, these empirical approaches, in the context of teenage motherhood, will be discussed below.
The sibling effects approach controls for socioeconomic background by comparing sisters or even twins. Although the validity of the siblings approach is disputed by some academics, as for example voiced by Griliches (1979), it is widely used in Economics. Geronimus and Korenman employ the sibling approach to account for parental background. Geronimus and Korenman find that early childbearing has a negative effect on educational attainment. Holmlund (2005) adds to the sibling effects approach by controlling for heterogeneity within families. Holmlund finds that controlling for sister heterogeneity substantially reduces, but not nullifies, the detrimental effect of teen childbearing on educational attainment. The adverse effects of teenage motherhood, however, are not limited to educational attainment. Teenage motherhood reduces (late-life) health, as found in Günes (2016) and Martin, Visscher and Webbink (2008). Additionally, Günes reports that lower later life health are partially caused by lower labor force participation and matching with a lower quality spouse. Fletcher and Wolfe (2009) find negligible effects of teen childbearing on adverse health behavior, both in the short- and long-term, by employing a sibling analysis. In short, the sibling effects
approach literature finds that teenage motherhood lowers educational attainment and (later life) health.
A different identification strategy to infer the causal relationship of early childbearing on later life outcomes, is the Instrumental Variables (IV) approach. The task of IV analysis is to find relevant and valid instruments to battle endogeneity problems. Klepinger, Lundberg and Plotnick (1999) use the age of first menarche as an instrument of fertility. Subsequently, the short-term effects of teen childbearing on human capital accumulation is investigated. The outcomes of Klepinger et al. suggest that teenage childbearing lowers educational attainment. Identical results using different instruments are found by Marini (1984) and Ashcraft,
Fernández-Val and Lang (2013).
6
and Painter (2003), Lee (2010) and Kane et al. (2013) are aligned with those of Chevalier and Viitanen, using PSM techniques. Lee (2010) adds to the literature by employing PSM with a sensitivity analysis using Rosenbaum bounds. The Rosenbaum bounds method suggests that the residual effect of selection bias on unobserved covariates is not sufficiently large to alter the PSM estimates. Angelini and Mierau stand out as the first to investigate the causal
relationship of teenage motherhood on late-life health, using PSM. Angelini and Mierau show that women experiencing teen childbearing are 5.1 percentage points less likely to report self-assessed very good or excellent health, and 5.5 percentage points less likely to lack depressive symptoms. Furthermore, evidence is presented that education, experiencing divorce and income are potential transmission channels.
To conclude, all three identification strategies reveal that teenage motherhood lowers educational attainment and (late-life) health.
2.2 Teenage Fatherhood
Much less resources have been devoted to the adverse effects of teenage fatherhood. Teenage fatherhood is a rarity, compared to teenage motherhood. This may be due to the fact that mothers are younger than fathers when their first child is born, and for fathers it is much easier to not acknowledge their child than for mothers 2. Although, teenage fatherhood is much rarer than teenage motherhood, the individual adverse effects can be of a substantial size, as found in this thesis. Teenage fatherhood is associated with lower socioeconomic status, a phenomenon reported in the teenage motherhood literature as well (Cairns, Cairns and Xie, 2001). Therefore, the same statistical approaches as for teenage motherhood are required to establish causal interference.
In contrast to teenage mothers, teenage fathers can withdraw themselves from the process of pregnancy and childbearing. Experiencing teenage parenthood for women constitutes to biological transformations and developments, whereas men can experience teenage
parenthood without being aware of it. Therefore, limited effect of teenage fatherhood on later life outcomes is expected if fathers are not involved in the pregnancy and the years following the birth of their child. As noted in Klerman and Jekel (1973) and Robinson (1988) a large majority of the fathers support the teenage mother during the pregnancy and thereafter. 81% of the teenage fathers date the pregnant women and 85% support her either financially or
7
through other means (Miller, Smolen and Vaz, 1983). Similar conclusions are drawn in Marsiglio (1987). Moreover, teenage fathers work and earn more following shortly after the birth than their peers. Thus, although the male body is not burdened by biological changes during a pregnancy, teenage fatherhood can require a drastic change in money and time allocation.
Card and Wise (1978) find that both women and men are affected by teenage parenthood. Teenage parents are more likely to surpass their ideal number of children at age 29 than their generational counterparts. Moreover, teenage fathers attain less education than their
counterparts, this effect is more pronounced for women. Nock (1998), and Fletcher and Wolfe (2012) find similar results as Card and Wise. An interesting paper is the one of Sigle-Rushton (2005). Sigle-Rushton employs a PSM approach in the context of teenage fatherhood and later life outcomes. She finds that teenage fathers are more likely to receive means-tested benefits, publicly subsidized housing, or experiencing life dissatisfaction. No effect on voting behavior is found. In contrast, Covington et al. (2017) find that teenage fatherhood reduces the
likelihood of voting.
In short, teenage fatherhood has adverse effects on educational attainment and late-life outcomes.
2.3 Gender Differences
In contrast to teenage motherhood, the adverse effects of teenage fatherhood are less apparent. Pregnancy, parturition and lactation present physiological challenges not endured by men. Women advancing to the latest weeks of their pregnancy are therefore forced to temporarily cease their education, or take a break from their jobs. Subsequently, returning to their prior education seems to be problematic, since the literature finds that teenagers experiencing childbearing have a lower educational attainment, relative to their peers.
8
rarity, as most teenage fathers support the mother (Miller et al.). For teenage fathers that support the mother with financial means, an economic rational choice could be to abandon educational ambitions, in order to finance the costs of their child. Therefore, it is expected that teenage fathers build a lower level of human capital, in comparison to their peers. In
conclusion, women, in contrast to men, experience biological developments and
transformations. However, both genders experience adverse effects of teenage parenthood. This suggests that there are differences in effects and mechanisms for men and women, driving these adverse effects of teenage parenthood. Four of these mechanisms are discussed below.
2.3.1 Education
Above, it has been established that teenage parenthood leads to a lower level of educational attainment in the short-term. Nonetheless, the adverse effects of teenage parenthood do not vanish in the medium- and long-term. The reason for this, is that the profits reaped from education are long‐lived. Education affects life outcomes through lifetime earnings. This in turn, affects socioeconomic status. The socioeconomic gradient in health3 is well embedded within the health economics literature. Moreover, Conti, Heckman and Urzua (2010) report that a lower educational attainment leads to adverse health behavior later in life, such as smoking and excessive eating. The causal effect is more pronounced for men than for women.
2.3.2 Entry into the Labor Market and Job Stability
Sigle-Rushton reports that men, not abstaining from any involvement in the upbringing of their child, experience lower paid and less stable jobs. Moreover, Sigle-Rushton suggests that these effects are more pronounced for men than for women. In a similar vein, Card and Wise find that virtually all teenage fathers enter the labor market at an earlier age than their peers. This is due to the financial obligations attached to young children. For women a reverse pattern is observed, women experiencing teenage childbearing are less likely to enter the labor market at an earlier age than their peers. For men these effects tend to peter out. Thus, men not experiencing teenage parenthood do not face the same financial urgency to drop out of school. Consequently, these men finalize school and subsequently enter the labor market at a later age. Again, a reverse pattern is observed for women.
9 2.3.3 Marital Status
The findings of Card and Wise and Nock suggest that teenage fathers are more likely to be single and less likely to marry than teenage mothers. Whereas, teenage motherhood is associated with a lower quality spouse, compared to their peers and teenage fathers.
Presumably, because shortly after experiencing teenage childbearing, women (by lack of other providers) are in need of financial support and therefore more willing to quickly marry a lower quality spouse (Günes). It is well embedded in the literature that marriage on one hand and happiness and late life health on the other hand are correlated. (Frey and Stutzer, 2005; Goldsteen, Mirowsky and Ross, 1990; Gallagher and Waite, 2000; Ermisch and Pevalin, 2005).
2.3.4 Allostatic Load
Another potential mechanism associated with early entry into teenage parenthood is allostatic load. Allostatic load has been conceptualized as “a multisystem physical dysregulation resulting from cumulative effects of responding to multiple stressors and operating through the functioning of several regulatory systems” (Arevalo et al., 2013). Young parents are less resilient against physical, emotional and financial stress, involved with raising children than their peers (Barban, 2013). Early entry into parenthood is associated with a higher allostatic load for both men and women, this effect is marginally larger for women (Grundy and Read, 2015). These stressors materialize in lower mental and physical health later in life.
3. Data
10
European countries 4. This allows this thesis to infer the causal effects of teenage parenthood in the short-, medium- and long-term. Individuals are disregarded if the variable of interest is missing. The number of observations is contingent on the dependent variable of interest. For the dependent variable for the medium-term, marital status, the number of observations is around 7,000. For the three remaining dependent variables the number of observations are approximately 10,000 for each gender.
3.1 Main Independent Variables 3.1.1 Teenage Fatherhood
The most vital independent variable is teenage fatherhood. A teenage father is classified as such, if the male of interest has a child before the age of 21. This is in contrast to common practice, which classifies a male as a teenage parent if the male has a child before the age of 20. This deviation of common practice is conducted, because it raises the number of teenage fathers substantially, almost doubles, from 113 to 216. Around 1.1 and 2.2 percent of the males in the dataset identify as a pure teenage father and an individual that has a child before his 21st birthday, respectively. It is acknowledged, indeed, that this thesis mildly bends the
definition of a teenage father. However, given a uniform distribution of birthdays over the year, 75 percent of the fathers at age 20 made their partner pregnant before the age of 20. Therefore this trade-off is deemed as beneficial 56. To validate this assumption, it is assessed whether those that experience fatherhood at age 20 are inherently different to those that experience it before the age of 20. This analysis finds that there are no statistical differences between those that experience fatherhood at age 20 and those before 20.
An important issue for consideration, is that this dataset merely includes teenage fathers acknowledging that they experienced teenage fatherhood. The adverse effects of teenage fatherhood are negligible for those that fail to acknowledge and support their child. This is, because these teenage fathers are not forced to temporarily cease their educational ambitions,
4 The countries are: Austria, Belgium, The Czech Republic, Denmark, France, Italy,
Germany, Greece, The Netherlands, Poland, Spain, Sweden and Switzerland.
5 The results do not change significantly for pure teenage fathers and mothers.
6 The average age of a father when the first child is born is 28, for females this is 25. The
11
in order to support the mother and child. Therefore, the adverse results of teenage fatherhood on later life outcomes in section 4 and onwards overstate the adverse effects for all teenage fathers. However, teenage fathers that fail to acknowledge and support their children are not part of the focus of this thesis, because they are not subject to (the same) adverse effects of teenage fatherhood.
3.1.2 Teenage Motherhood
In a similar vein as teenage fathers, teenage mothers are defined as females that have had a child before the age of 21. Similarly as for men, there are no statistical differences for women that experience childbearing at age 20 and those that experience childbearing before the age of 20. Although, there are sufficient teenage mothers in the dataset, this choice is made to remain comparability with their male counterparts. Of the 12196 women in the dataset 684 and 1208 classify as pure teenage mothers and women that became a mother before the age of 21, respectively. This is approximately 5.6 and 9.9 percent of the women in the dataset. There is a considerable larger amount of teenage mothers than their male counterparts.A reason for this may be that fathers are on average older than mothers when their first child is born.
3.2 Dependent Variables 3.2.1 Short-Term
Researchers tend to find that due to teenage childbearing women follow on average one year education less than their peers, this effect is less pronounced for men. To assess the effect of teenage parenthood on education, the International Standard Classification of Education (ISCED) of 1997 is employed. ISCED 1997 is a harmonized statistical framework that ensures comparability of education levels across the globe. To control for educational differences between countries ISCED 1997 is chosen in favor of the number of years of schooling. ISCED 1997 is a 7-point scale, ranging from zero to six, where higher values indicate higher levels of education. The ISCED 1997 score corresponding to an individual is inferred by the interviewer, and indicates the individuals highest educational degree.
Subsequently, a binary variable is created which equals one if an individual has at least a post-secondary education, and zero in all other cases 7. Thus, a one indicates an educational degree after the decision whether to become a teenage parent is made. Education has been chosen, since education is often perceived as the pathway to success in the Western World. Moreover,
12
education and its effects on leading a more financially unconstrained life are well documented within the economic literature (Houthakker, 1959).
3.2.2 Medium-Term
For the medium-term, this thesis focuses on marital status in the middle of the average life cycle. A binary variable is constructed that equals one if the person is married before or at the age of 45, and is still together with this person at the time of the interview. The binary
variable equals zero if the person is never married, married after the age of 45, or has experienced a divorce. Thus, this variable equals one if the person has found his or her lifetime partner before the individual is of age 46 and has never experienced a divorce, in all other cases this variable equals zero. Lifetime partner is defined as a long-term relationship that has lasted until at least when the survey took place. It is to be expected that those
experiencing teenage parenthood are a less attractive spouse. This decreases the likelihood of finding their lifetime partner before the age of 46. This specific age is chosen for the reason that it approximately is in the middle of an average person his or her life. Moreover, the long-term variables start at age 50 (as the data consists of those aged at least 50). Thus, in order to distinguish different stages in an individual’s life, the age of 46 is decided upon 8.
3.3 Long-Term
3.3.1 Self-Assessed Health, (Physical Health)
The dependent variable of interest for late life outcomes are self-assessed health and (a lack of) depressive symptoms. Thereby, this thesis takes an identical approach as Angelini and Mierau. However, this thesis complements their paper by comparing and contrasting the effects of both genders, rather than solely investigating mothers experiencing teenage
childbearing. As found by Conti, Heckman and Urzua, lower educational attainment leads to more adverse health behavior later in one’s life. Favorable health is a significant driver of life satisfaction (Caligari, 2015), this effect is observed for individuals at old age as well (Lebo, 1953). Self-assessed health is measured by asking the individuals to rate their health from best to worse as follows: excellent, very good, good, fair and poor. A one represents excellent health whereas a five represents poor health. Subsequently, a binary variable is constructed that equals one if the individual has excellent, very good or good health and zero otherwise.
8 The results presented in section 4 and onwards are all robust to different sets of ages
13
3.3.2 (Lack of) Depressive Symptoms, (Mental Health)
Mental health is operationalized by the EURO-D depression scale. The EURO-D depression scale is constructed based on twelve questions in which the individual has to self-assess their emotional well-being (e.g. tearfulness, life-enjoyment among other items). The EURO-D depression scale is a thirteen point scale, ranging from a zero (no depressive symptoms at all) to twelve (highly depressed). A score above three indicates that some depressive symptoms are present within the individual (Prince et al., 1999a, 1999b). For this reason, an indicator variable is created which equals one when the individual lacks depressive symptoms, and zero if there are depressive symptoms present within the individual. Depression is the leading cause of an increase in disability adjusted life years worldwide, however effective treatment is often lacking (Institute for Health Metrics and Evaluation, 2010; Alexopoulos, 2005).
3.4 Pre-Pregnancy Control Variables
In the literature review it has been stressed that teenage parents are not randomly distributed over each socioeconomic class in society. This hampers the causal relationship of teenage parenthood on later life outcomes. Therefore, a rich set of control variables are employed to mitigate the problem of confounding factors. These control variables are pre-pregnancy, as this thesis aims to find the causal effect of teenage pregnancy of later life outcomes. Therefore, one must control for pre-pregnancy circumstances. The control variables must possibly correlate with socioeconomic status, to account for the lower socioeconomic status of teenage parents. The control variables can be divided into two groups. In the first group are all the variables that can be attributed to individual characteristics. These variables are age, the square of age, assessed relative math and language performance at the age of ten, self-assessed health at the age of ten and country of residence. The square of age is included to account for any non-linear effects of age. In the second group all characteristics that can be attributed to the parents or legal guardian are included. This includes rooms per person; number of people living in house; with whom they lived at home; which facilities the house had; the number of books in house (ranging from one if there are no books present, to five if there are more than 200); whether the parents engaged in adverse health behavior; whether the individual grew up in a rural or urban area (ranging from one if the individual lived in a large city to five if the individual lived in a rural area or village); occupation of the main
14
individual is not vaccinated and one if the person is vaccinated). Vaccination is strongly correlated with socioeconomic status (Galarce, Minsky and Viswanath, 2011), for this reason childhood vaccination is included. All non-descriptive (e.g. ‘Refusal’ and ‘Do not know’) answers by individuals are dropped. Lastly, all correlations among the explanatory variables have been assessed, no issues of multicollinearity are arising.
Table 1 and 2 provide some valuable insights into the life of teenage parents before their child is born, their later life outcomes are presented as well. Note that Table 1 and 2 do not provide evidence that teenage parenthood leads to a lower socioeconomic status, but it rather shows that teenage parents are born into families with a lower socioeconomic status. The causal relationship will be established in section 4. The characteristics of the household the teenage parents grew up in, are associated with lower socioeconomic status. Teenage mothers disproportionately come from one-parent households with fewer rooms and books in house and no running water. The parents of teenage mothers are more likely to consume a
substantial amount of alcohol. Moreover, the main breadwinner has had an occupation associated with a lower socioeconomic status. The teenage mothers tend to score lower at language and mathematics tests at the age of ten. Furthermore, teenage mothers are less likely to obtain post-secondary school degrees, to find their lifetime partner before the age of 46 and report lower self-assessed health and absence of depression. The differences between teenage fathers and their peers are less pronounced and observable. This implies, that teenage
15 Table 1
Descriptive Statistics Males (N=9957)
Name Min. Max.
Mean (All) Mean (TF=0) Mean (TF=1) Individual Characteristics Teenage Fathers 0 1 0.022 0 1*** Age 50 97 64.205 64.262 61.676***
Relative Math performance 1 5 2.625 2.623 2.700
Relative Language performance 1 5 2.772 2.769 2.907**
Childhood Health Status 1 6 1.982 1.985 1.856*
Parental and Chilldhood Characteristics
Rooms in Household 0 41 3.734 3.733 3.764
Number of People in Household 0 45 5.583 5.596 5.019***
Biological Mother in Household 0 1 0.967 0.967 0.954
Biological Father in Household 0 1 0.914 0.916 0.852***
No Running Water in Household 0 1 0.278 0.279 0.227*
Number of Books in Household 1 5 2.083 2.081 2.185
Parents Smokes 0 1 0.653 0.653 0.662
Parent Drinks 0 1 0.082 0.081 0.097
Area of Residence 1 5 3.714 3.717 3.611
Occupation of Main Breadwinner 1 11 6.236 6.232 6.412
Vaccinated during Childhood 0 1 0.924 0.923 0.968**
Dependent variables
Short-term: Education 0 1 0.269 0.271 0.199**
Medium-term: Lifetime Partner (N=3446) 0 1 0.544 0.549 0.383***
Long-term: (Very) Good or Excellent Health 0 1 0.710 0.711 0.657*
Long-term: (Lack of) Depressive Symptoms 0 1 0.853 0.855 0.796**
Note: */**/*** indicate significant differences at the 10%/5%/1% level based on the mean differences of TF=0 and TF=1, assessed with a t-test.
16 Table 2
Descriptive Statistics Females (N=12196)
Name Min. Max.
Mean (All) Mean (TM=0) Mean (TM=1) Individual Characteristics Teenage Mothers 0 1 0.099 0 1*** Age 50 99 64.045 64.120 62.707***
Relative Math performance 1 5 2.775 2.769 2.831**
Relative Language performance 1 5 2.598 2.592 2.654**
Childhood Health Status 1 6 2.103 2.105 2.084
Parental and Chilldhood Characteristics
Rooms in Household 0 40 3.680 3.737 3.163***
Number of People in Household 0 50 5.580 5.591 5.4768
Biological Mother in Household 0 1 0.961 0.963 0.945***
Biological Father in Household 0 1 0.910 0.917 0.844***
No Running Water in Household 0 1 0.278 0.271 0.334***
Number of Books in Household 1 5 2.123 2.138 1.981***
Parents Smokes 0 1 0.611 0.609 0.626
Parent Drinks 0 1 0.081 0.077 0.110***
Area of Residence 1 5 3.726 3.725 3.740
Occupation of Main Breadwinner 1 11 6.245 6.200 6.654***
Vaccinated during Childhood 0 1 0.929 0.928 0.934
Dependent variables
Short-term: Education 0 1 0.207 0.218 0.105***
Medium-term: Lifetime Partner (N=3297) 0 1 0.499 0.500 0.492
Long-term: (Very) Good or Excellent Health 0 1 0.654 0.666 0.549***
Long-term: (Lack of) Depressive Symptoms 0 1 0.704 0.710 0.642***
Note: */**/*** indicate significant differences at the 10%/5%/1% level based on the mean differences of TM=0 and TM=1, assessed with a t-test.
4. Statistical Identification and Empirical Results
The aim of this thesis is to infer, contrast and compare the causal effects of teenage
17
The advantages attached to the (semi-parametric) PSM approach to parametric approaches such as Ordinary Least Squares or a Logit model are twofold. Firstly, PSM allows the construction of a counterfactual with observational data. Thereby, PSM considers only comparable individuals. Other statistical practices do not provide this edge with non-experimental data. Secondly, linear and logistic models rely on assumptions regarding the functional form, this obstacle is partially surrounded by using PSM. The non-parametric characteristics of PSM ensures that PSM is less susceptible to issues of misspecification. The methodology of Altonji et al., Bellows and Miguel, Oster and Gonzalez and Miguel is employed to contrast and compare the (semi-parametric) PSM results to. Vastly different methodologies, yielding similar results reinforces our believe in the validity and robustness of the results.
4.1.1 Semi-Parametric Identification: Propensity Score Matching
In order to infer a causal effect of teenage parenthood on later life outcomes, one needs a counterfactual. PSM provides this counterfactual in the form of a synthetic control group. PSM necessitates a four-step procedure to infer the causal effect. In the context of this thesis, the causal effect of teenage parenthood on later life outcomes. Firstly, a logistic or probit prediction model is used to identify the factors that leads teenagers to be a parent. This yields a propensity score of becoming a teenage parent for each individual, irrespective of them actually being a teenage parent. Secondly, a synthetic control group is constructed based on the similarity of propensity scores, this control group functions as the counterfactual. Thus, a teenage parent has a propensity score and is matched with a peer that has a similar propensity score who has not experienced teenage parenthood. This control group, if the procedure is done properly, is as similar as possible to the treated group with the exception that the individuals in this control group are not a teenage parent. In order to be an accurate
counterfactual, the reason of these individuals to postpone or suspend their decision to take children should be exogenous. An example of this exogeneity is infertility. Thus, that infertile individuals, that would have been a teenage parent without this infertility, are in the control group 9. Thirdly, the quality of the match needs to be assessed. Lastly, the causal impact of teenage parenthood on later life outcomes is deduced as the differences in later life outcomes of the teenage parents and their probabilistic peers. This is a causal effect if the entire
procedure is done properly, since these individuals are as alike as statistically possible. The
18
similarity of propensity scores can be evaluated in various ways, for the matching in this thesis full Mahalanobis matching is employed 10.
A logistic model is employed for the estimation of the probability of one to become a teenage parent 11. This is done with the aid of all pre-pregnancy variables in table 1 and 2.
Subsequently, a teenage parent with a propensity score of 𝛼𝑖∈ [0,1] is matched with his or
her nearest non-teenage parent neighbor in terms of the propensity score. The nearest neighbor is identified by minimalizing the Mahalanobis distance. Moreover, a non-treated individual (i.e. a non-teenage parent) cannot be matched with multiple teenage parents, this is the so-called no-replacement method.
4.1.2 Matching Quality
The matching procedure is conducted two times for each gender. On one hand, the procedure is done for the short- and long-term variables. On the other hand, for the medium-term variable. This method is decided upon, because the number of observations for the medium-term is considerably lower. There are no statistical differences between both samples for both genders. Therefore, the medium-term can be considered as a random subsample of the full dataset.
Figure 1 and 2 below provides some preliminary evidence of the quality of the match for both genders. Figure 1 and 2 do not constitute definite statistical evidence in favor of a proper execution of the matching procedure. Nonetheless, figure 1 and 2 yield the valuable insight that the quality of matches are satisfactory, as a considerable portion of the densities overlap, i.e. the groups are sufficiently similar to be compared.
19
Figure 1: Predicted probabilities of teenage parenthood for each gender, short- and long-term sample.
20
are below the 5 percent level. For the short- and long-term variables the median biases are 2.3 and 1.6 percent, for men and women respectively. For the smaller subsample of the medium-term the values are 4.9 and 3.3 percent, for men and women respectively.
Another method of assessing the matching quality is by a simple two-sample t-test, and inspect if there are significant differences in the covariate means for the treated and control group. This method shows that for none of the covariates the means of the treated and control group differ significantly at any conventional level. The lack of significant differences
between the means of the covariates for the treated and non-treated group implies a high matching quality (i.e. the differences between the treated and non-treated disappeared and therefore the control group can be used for comparison). The results of a F-test are aligned with the two sample t-tests.
Finally, Sianesi suggests that the matching quality can be assessed by performing the matching procedure with a subsample of only the treated and non-treated individuals and analyzing the new pseudo 𝑅2. This pseudo𝑅2 indicates how well the regressors explain the
probability of becoming a teenage parent. After the initial matching procedure is conducted, no systematic differences between the treated and non-treated should be observed. Therefore a second matching procedure with this (non-)treated subsample should yield a fairly low pseudo 𝑅2. This procedure yields a fairly low pseudo 𝑅2 of 0.016 and 0.004 for men and women,
respectively, for the short- and long-term variables. The subsample for medium-term yield pseudo 𝑅2 values of 0.037 and 0.009 for men and women, respectively.
All in all, the conclusions of both the formal and informal tests of the qualities of the matches are aligned. Each of the four tests imply that the matches are correctly executed. Now, all necessary prerequisites have been established to interpret the results of the PSM procedure.
4.1.3 PSM Results
Table 3 below reports the average treatment effects for educational attainment, whether the person is married to his or her lifetime partner before the age of 46, self-assessed health and depressive symptoms. Teenage fatherhood decreases the likelihood of reporting a
21
females. This finding aligns with the literature that finds that women experiencing teenage childbearing marry a spouse of lower quality, whereas men marry at a later stage in life. Furthermore, individuals experiencing teenage fatherhood and motherhood are respectively 13.0 and 10.3 percentage points less likely to report good, very good or excellent health. Additionally, individuals experiencing teenage fatherhood and motherhood are respectively 7.4 and 5.9 percentage points less likely to lack depressive symptoms. Interestingly, the adverse effects are more pronounced for men than for women. These results confirm that the relative attention devoted to teenage motherhood in contrast to teenage parenthood is
unjustified. For the reason, that effects of teenage fatherhood on later life outcomes are not smaller than for teenage motherhood. All in all, teenage fatherhood lowers later life outcomes substantially.
Table 3
Average Treatment Effects Nearest Neighbor Matching
Education LifeTime
Partner Health Depression
Teenage Father -0.106*** -0.080* -0.130*** -0.074**
(0.021) (0.048) (0.042) (0.036)
Teenage Mother -0.096*** -0.048* -0.103*** -0.059***
(0.015) (0.026) (0.020) (0.019)
Notes: */**/*** indicate statistical significances at the 10%/5%/1% level, brackets indicate standard errors.
4.2.1 Parametric Identification: Logistic Model
An alternative identification strategy to the PSM methodology is a Logit model. The strategy here is to include a rich set of control variables, to mitigate the problem of factors driving both our main independent variable of interest (i.e. teenage parenthood), as well as the dependent variables. Subsequently, the magnitude and the sign of the potential bias arising from unobservables is assessed with the aid of the Altonji ratio and the Oster lower bound formula.
22
depressive symptoms. Again, the medium-term sample is considerably smaller than the short- and long-term sample. For this reason, the inference below consists of a large sample and a smaller subsample. The subsample, however, does not statistically differ from the larger sample.
The results of column 1 do not represent a causal relationship, since these simple regressions suffer from confounding factors hampering the causal relationship. This problem can be mitigated by extending column 1 in two ways. Firstly, by adding individual characteristics prior to the individual experiencing teenage parenthood, as showcased in column 2 of tables 4 to 7. Secondly, The Logit model is sophisticated by adding parental and individual
characteristics, as presented in column 3 of table 4 to 7. In order to remain comparability with the semi-parametric PSM approach, identical covariates is decided upon.
23 Table 4
The Effect of Teenage Parenthood on Educational Attainment for Both Genders
Logit Model
Dependent Variable: Education
Males Females Individual Characteristics (1) (2) (3) (1) (2) (3) Teenage Parenthood -0.672*** -0.576*** -0.553*** -1.004*** -1.008*** -0.859*** (0.174) (0.185) (0.188) (0.100) (0.103) (0.108) Age 0.004 0.011 0.69* -0.031 -0.029 -0.014 (0.035) (0.037) (0.039) (0.038) (0.040) (0.042) Age2 -0.000 -0.000 -0.001** -0.000 -0.000 -0.000 (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Relative Math performance -0.439*** -0.433*** -0.342*** -0.3556
(0.033) (0.034) (0.033) (0.034)
Relative Language performance -0.584*** -0.455*** -0.550*** -0.385***
(0.034) (0.035) (0.035) (0.036)
Childhood Health Status 0.025 -0.024 -0.021 -0.008
(0.026) (0.027) (0.025) (0.026)
Parental and Childhood Characteristics
Rooms in Household 0.063*** 0.084
(0.016) (0.016)
Number of People in Household -0.071*** -0.090***
(0.017) (0.015)
Biological Mother in Household 0.178 0.033
(0.169) (0.151)
Biological Father in Household 0.122 -0.200*
(0.104) (0.106)
No Running Water in Household -0.314*** -0.489***
(0.079) (0.085)
Number of Books in Household 0.375*** 0.406***
(0.026) (0.026) Parents Smokes -0.085 -0.058 (0.058) (0.058) Parent Drinks -0.276** -0.007 (0.108) (0.099) Area of Residence -0.133*** -0.082*** (0.019) (0.019)
Occupation of Main Breadwinner -0.090*** -0.116***
(0.013) (0.012)
Vaccinated during Childhood -0.124 0.094
(0.100) (0.109)
Country Dummies Yes Yes Yes Yes Yes Yes
Pseudo R2
0.063 0.144 0.215 0.117 0.175 0.254
Number of Observations 9957 9957 9957 12196 12196 12196
24 Table 5
The Effect of Teenage Parenthood on Finding a Lifetime Partner before the age of 46 for Both Genders
Logit Model
Dependent Variable Marital Status
Males Females (1) (2) (3) (1) (2) (3) Teenage Parenthood -0.473** -0.484** -0.471** -0.204* -0.196* -0.222** (0.224) (0.230) (0.227) (0.120) (0.120) (0.121) Age -0.264*** -0.263*** -0.257*** 0.002 0.006 -0.008 (0.061) (0.062) (0.062) (0.058) (0.059) (0.060) Age2 0.002*** 0.002*** 0.002*** -0.000 -0.000 -0.000 (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Relative Math performance -0.071 -0.079 -0.068 -0.075
(0.048) (0.048) (0.050) (0.051)
Relative Language performance 0.104** 0.073*** -0.002 -0.045
(0.049) (0.051) (0.050) (0.051)
Childhood Health Status -0.128*** -0.122*** -0.094** -0.071**
(0.040) (0.040) (0.038) (0.038)
Parental and Childhood Characteristics
Rooms in Household 0.007 0.018
(0.022) (0.027)
Number of People in Household 0.026 0.041*
(0.020) (0.022)
Biological Mother in Household 0.164 0.033
(0.242) (0.212)
Biological Father in Household 0.073 0.165
(0.150) (0.152)
No Running Water in Household 0.023 0.112
(0.113) (0.119)
Number of Books in Household -0.059 -0.099**
(0.039) (0.040) Parents Smokes 0.179** -0.028 (0.085) (0.088) Parent Drinks -0.185 -0.261* (0.127) (0.140) Area of Residence 0.067** 0.116*** (0.028) (0.030)
Occupation of Main Breadwinner 0.005 0.002
(0.019) (0.019)
Vaccinated during Childhood 0.075 0.075
(0.141) (0.165)
Country Dummies Yes Yes Yes Yes Yes Yes
Pseudo R2
0.139 0.142 0.147 0.174 0.176 0.188
Number of Observations 3446 3446 3446 3297 3297 3297
25 Table 6
The Effect of Teenage Parenthood on Self-Assessed Health for Both Genders
Logit Model
Dependent Variable Self-Assessed Health
Males Females (1) (2) (3) (1) (2) (3) Teenage Parenthood -0.476*** -0.476*** -0.463*** -0.554*** -0.556*** -0.480*** (0.156) (0.160) (0.160) (0.067) (0.067) (0.068) Age -0.065** -0.060* -0.043 -0.115*** -0.103*** -0.083*** (0.033) (0.033) (0.034) (0.027) (0.027) (0.028) Age2 0.000 0.000 -0.000 -0.000** 0.000** 0.000 (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Relative Math performance -0.126*** -0.107*** -0.083*** -0.067**
(0.030) (0.031) (0.027) (0.028)
Relative Language performance -0.133*** -0.080*** -0.179*** -0.130***
(0.032) (0.033) (0.028) (0.029)
Childhood Health Status -0.197*** -0.199*** -0.290*** -0.283***
(0.024) (0.024) (0.021) (0.021)
Parental and Childhood Characteristics
Rooms in Household 0.033** 0.062***
(0.015) (0.015)
Number of People in Household -0.020** -0.014
(0.011) (0.011)
Biological Mother in Household 0.120 0.136
(0.136) (0.111)
Biological Father in Household -0.013 -0.121
(0.092) (0.077)
No Running Water in Household -0.134** -0.132**
(0.061) (0.054)
Number of Books in Household 0.095*** 0.081***
(0.027) (0.023) Parents Smokes -0.048 -0.035 (0.052) (0.044) Parent Drinks -0.214** -0.288*** (0.085) (0.076) Area of Residence -0.010 0.007 (0.018) (0.016)
Occupation of Main Breadwinner -0.042*** -0.037***
(0.012) (0.010)
Vaccinated during Childhood 0.089 0.031
(0.075) (0.068)
Country Dummies Yes Yes Yes Yes Yes Yes
Pseudo R2
0.078 0.090 0.098 0.088 0.108 0.117
Number of Observations 9957 9957 9957 12196 12196 12216
26 Table 7
The Effect of Teenage Parenthood on Depressive Symptoms for Both Genders
Logit Model
Dependent Variable (Lack of) Depressive Symptoms
Males Females (1) (2) (3) (1) (2) (3) Teenage Parenthood -0.610*** -0.614*** -0.586*** -0.331*** -0.325*** -0.265*** (0.178) (0.180) (0.182) (0.067) (0.067) (0.067) Age 0.200*** 0.208*** 0.218*** 0.129*** 0.143*** 0.150*** (0.039) (0.039) (0.039) (0.026) (0.027) (0.027) Age2 -0.002*** -0.002*** -0.002*** -0.001*** -0.001*** -0.001*** (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Relative Math performance -0.122*** -0.109*** -0.185*** -0.172***
(0.039) (0.039) (0.027) (0.028)
Relative Language performance -0.097** -0.057 -0.009 0.021
(0.041) (0.043) (0.028) (0.029)
Childhood Health Status -0.172*** -0.168*** -0.205*** -0.195***
(0.029) (0.029) (0.020) (0.021)
Parental and Childhood Characteristics
Rooms in Household -0.029 0.038**
(0.019) (0.015)
Number of People in Household -0.007 -0.021**
(0.013) (0.011)
Biological Mother in Household 0.104 0.053
(0.168) (0.114)
Biological Father in Household 0.087 0.121
(0.110) (0.079)
No Running Water in Household -0.017 -0.075
(0.077) (0.055)
Number of Books in Household 0.079** 0.049**
(0.033) (0.023) Parents Smokes 0.123** 0.118*** (0.064) (0.045) Parent Drinks -0.391*** -0.465*** (0.098) (0.074) Area of Residence 0.035 0.039** (0.022) (0.016)
Occupation of Main Breadwinner -0.022 -0.014
(0.015) (0.011)
Vaccinated during Childhood 0.223** 0.149**
(0.087) (0.068)
Country Dummies Yes Yes Yes Yes Yes Yes
Pseudo R2
0.051 0.059 0.065 0.053 0.065 0.072
Number of Observations 9957 9957 9957 12196 12196 12196
27 Table 8
Average Marginal Effects of Teenage Parenthood
Male Female Education -0.082*** -0.103*** (0.028) (0.013) Lifetime Partner -0.095** -0.042** (0.046) (0.023) Self-Assessed Health -0.084*** -0.093*** (0.029) (0.013)
(Lack of) Depressive Symptoms -0.069*** -0.050***
(0.021) (0.013)
Notes: Marginal effects are based on column 3 of table 4 to X7, */**/*** indicate significance at the 10%/5%/1% standard errors are provided within the brackets.
4.2.2 Lower Bound Calculation
An extensive list of control variables has been included in section 4.2.1 to relieve the model from omitted variable bias. However, the results could still suffer from omitted variable bias. The relative importance of omitted variable bias can be gauged by investigating the
robustness of the coefficients of teenage parenthood to the inclusion of control variables. If the effect of teenage parenthood is robust to the addition of a rich set of (observable) control variables, then we can conclude that unobservables are unlikely to alter the order of
magnitude of the results, or even render the results insignificant. On the other hand, if the effect of teenage parenthood is largely subject to the inclusion of the observable control variables, we can safely assume that the unobservables are likely to do the same. This
intuition is formalized in the following ratio, popularized in Altonji et al., henceforth referred to as the Altonji ratio:
𝛽 ̂𝑅𝑒𝑠
28
Catholic school attendance on their outcome variables. Below the Altonji ratios are provided, which are derived from column 1 and 3 in table 4 to 7.
Table 9 Altonji Ratios Male Female Education 5.647 6.924 Lifetime Partner 236.500 -11.333 Self-Assessed Health 36.615 7.486
(Lack of) Depressive Symptoms 25.417 5.015
Notes: See formula 1 for the ratio underlying these numbers, inputs are derived from the teenage parenthood coefficients in column 1 and 3 in table 4 to 7.
The Altonji ratios in table 9 indicate that unobservables are unlikely to render the causal effects insignificant. The smallest positive Altonji ratio is substantially larger than the 3.55 benchmark in Altonji et al., it is unlikely that the impact of unobservables is five times as large as the impact of observables. The negative Altonji ratios warrants some extra caution, however negative Altonji ratios are rare but not alarming. Negative Altonji ratios are caused by covariates that enhance the effect of interest, rather than diminish, this pattern is observed for the medium-term variable for females. Negative Altonji ratios imply that the causal effect is in fact larger than those presented in table 5 12.
An issue of the Altonji ratio yet unaccounted for, is that the Altonji methodology assumes that the unobservables exhibit similar covariance properties as the observables. This
unconfirmable assumption can be avoided with the Oster lower bound calculation, these lower bound marginal effects of teenage parenthood can be gauged with formula 2 13.
𝛽̂𝐿𝐵= 𝛽̂𝐹𝑢𝑙𝑙 − (𝛽̂𝑅𝑒𝑠− 𝛽̂𝐹𝑢𝑙𝑙) ∗𝑅𝑀𝑎𝑥2 −𝑅𝐹𝑢𝑙𝑙2
𝑅𝐹𝑢𝑙𝑙2 −𝑅𝑅𝑒𝑠2 (2)
𝛽̂𝑅𝑒𝑠 and 𝛽̂𝐹𝑢𝑙𝑙 have the same definition as in equation (1), additionally 𝛽̂𝐿𝐵 is the lower bound
of the teenage parenthood coefficient, 𝑅𝑀𝑎𝑥2 represents the 𝑅2 for the unobservable Logit model, in which the observables and the unobservables are included, 𝑅𝐹𝑢𝑙𝑙2 and 𝑅𝑅𝑒𝑠2 represent the 𝑅2 for the full and restricted model, respectively. All inputs of formula 2, but 𝑅
𝑀𝑎𝑥2 , are
12 papers finding negative Altonji ratios all reported that although negative Altonji ratios are
unusual, they do not constitute an issue, among these papers are Beck et al. (2013) and Satyanath, Voigtländer and Voth, (2013).
29
uncomplicated to derive. 𝑅𝑀𝑎𝑥2 however, requires some additional assumptions as the true 𝑅𝑀𝑎𝑥2 cannot be observed. Bellows and Miguel suggest that 𝑅̂𝑀𝑎𝑥2 = 𝑅𝐹𝑢𝑙𝑙2 + (𝑅𝐹𝑢𝑙𝑙2 − 𝑅𝑅𝑒𝑠2 ), whereas Oster suggests 𝑅̂𝑀𝑎𝑥2 = 𝑀𝑖𝑛(1.3𝑅𝐹𝑢𝑙𝑙2 , 1) 14. The
literature does not provide a decisive answer which 𝑅̂𝑀𝑎𝑥2 calculation is preferred, the slightly more conservative Oster calculation is chosen 15. The marginal effects attached to the Oster lower bound values are reported below.
Table 10
Average Marginal Effect of Teenage Parenthood with Oster Lower Bound Values
Male Female
Education -0.075 -0.093
Lifetime Partner -0.093 -0.056
Self-Assessed Health -0.081 -0.076
(Lack of) Depressive Symptoms -0.065 -0.036
All Oster lower bound average marginal effects are well below zero, this implies that omitted variable bias does not nullify the results. The adverse life effects of teenage parenthood are all of a considerable size, and economically and statistically different from zero. The Oster lower bound average marginal effects of the medium-term variable, whether the individual found his or her lifetime partner before the age of 46, are for women larger than the non-lower bound values. The reason for this, is that the addition of covariates enhances the adverse effect of teenage parenthood on the medium-term variable, instead of diminish. Following the logic of equation 1 and 2, this indicates that the causal effect is larger than those presented in table 8, which is corrected for in table 10.
The Oster correction formula is formally for linear specifications, however the method above has been applied to a Logit model. As a check, the procedure of this section is repeated with a linear probability model with robust standard errors, to account for the heteroscedasticity involved in linear probability models. These results are presented in the Appendix. The Oster corrected values derived from the linear probability models are presented in table 11. Table 11 presents slightly lower values, but similar conclusions are drawn from table 11 as table 10.
14 The 1.3 multiplication factor has not been chosen haphazardly, it is verified across many
top economic journal from 2008 to 2013.
30 Table 11
Oster Lower Bound Adverse Effects of Teenage Parenthood on Later Life Outcomes, Linear Probability Model
Male Female
Education -0.075 -0.064
Lifetime Partner -0.078 -0.068
Self-Assessed Health -0.055 -0.072
(Lack of) Depressive Symptoms -0.065 -0.022
4.2.3 Pooled Logistic Model
The adverse effects of teenage motherhood are well embedded in the literature. Therefore, it is of no surprise that the results in section 4 are aligned with this notion. In contrast, the literature has devoted much less time and effort to the potential adverse effects of teenage fatherhood. The results of this thesis, however, imply that the adverse effects of teenage fatherhood are not inferior to those of teenage motherhood. This notion can be formally asserted by reiterating the procedure of section 4.2.1 and 4.2.2 for a pooled sample, with the inclusion of an interaction term of gender and teenage parenthood. Insignificancy of these interaction terms signify that the adverse effects of teenage parenthood are not greater for females than for men. Table 12 presents the Logit coefficients of these interaction terms. The adverse effects of teenage parenthood on education and depressive symptoms are larger for women than for men. In distinction to education and depressive symptoms, there is no
discernable difference between men and women for the adverse effects of teenage parenthood for finding your lifetime partner before the age of 46 and self-assessed health 16. In
conclusion, the adverse effects of teenage parenthood on education and depressive symptoms are more pronounced for women than for men, but there are no statistical differences for marital status and self-assessed health. The results of the Pooled Logistic model disclose that the adverse effects of teenage fatherhood are of a considerable size, and should be valued as such 17.
16 Transforming the coefficients to Oster lower bound values does not render the interaction
term insignificant, thereby confirming that indeed the adverse effects of teenage parenthood on these variables are more pronounced for women than for men.
17 Note that the conclusions drawn in this section are based on the Logit model and not the
31 Table 12
Coefficients of Interaction Term
Education Lifetime
Partner Health Depression
Teenage Parent -0.539** 0.161 -0.195 -0.536***
(0.216) (0.245) (0.169) (0.190)
Notes: */**/*** indicate significance at the 10%/5%/1% standard errors are provided within the brackets.
4.3 Comparison of Parametric and Semi-Parametric Results
The effects of teenage parenthood on later life outcomes has been assessed both
parametrically and semi-parametrically, by means of a Logit model and PSM. A comparison of both results can be made by concentrating on table 3, which shows the PSM results, and table 8, which presents the Logit model results. The results of the PSM are slightly larger than those of the Logit model. However, the results are well aligned with regards to the message they convey. Both tables illustrate that the adverse life effects of teenage fatherhood are of a considerable size relative to teenage motherhood. That is, the adverse effects of teenage fatherhood on later life outcomes do not shrink into insignificance compared to the adverse effects of teenage motherhood. The adverse effects of teenage motherhood on later life outcomes are easier imaginable and more apparent, relative to teenage fatherhood. However, this thesis does not find evidence that the disproportional attention devoted to teenage motherhood, compared to teenage fatherhood, is justified, as the adverse effects of teenage fatherhood are of a similar order and magnitude.
5. Transmission Channels
Above. it has been established that teenage parenthood affects life outcomes in the short-term, those experiencing teenage parenthood are less likely to report a post-secondary school
32
However, it has not been assessed, yet, whether the long-term effects can partially be explained by the adverse effects observed in the short- and medium-term. This section assesses whether the short- and medium-term variables operate as transmission channels for the adverse health effects of teenage parenthood on late-life outcomes. Additionally, an income indicator is employed to infer the potential transmission channels. Income is added, because lower levels of education particularly materializes in lower wages later in life.
Income is Inverse Hyperbolic Sine18 transformed, as suggested in Burbidge, Magee and Robb
(1988). The three pathways that potentially reduce late-life outcomes are added as covariates. A lower marginal effect or average treatment effect of teenage parenthood on late-life health, due to the addition of these (potential) pathways, reveal that these (potential) pathways partially explain late life health 19. Angelini and Mierau and Baron and Kenny (1986) note that a potential pathway can only be classified as such, if teenage father- and motherhood are statistically significant when the potential pathways are regressed on teenage father- and motherhood, respectively. These results are presented below in table 13. To diminish omitted variable bias and remain comparability, the full set of covariates are employed as in section 4.2 with a Logit model.
An issue for consideration, is that the results in this section can be deluded by endogeneity of the pathways, as noted in Angelini and Mierau. Therefore, the results and conclusions
presented below should be interpreted as potential pathways, rather than undisputed causal transmission channels.
Table 13
Average Marginal Effects of Teenage Parenthood on Potential Pathways
Education Income Marriage
Teenage Father -0.100*** -0.010 -0.076
(0.029) (0.008) (0.046)
Teenage Mother -0.098*** 0.003 -0.046*
(0.014) (0.007) (0.025)
Notes: */**/*** indicate statistical significances at the 10%/5%/1% level, brackets indicate standard errors.
18 Inverse Hyperbolic Sine transformation: Log(X+(X+1)0.5), this method is superior to simply
33
The results in table 13 illustrate that when education, marital status and income are regressed on teenage fatherhood, only education is statistically significant. This in turn, following the procedure of Angelini and Mierau and Baron and Kenny, implies that merely education is appropriate to be added as a covariate. This covariate assesses whether it is a potential pathway of adverse health effects, caused by teenage fatherhood. The same procedure for women indicate that education and marital status are potential candidates as transmission channels. The prerequisites of pathway analysis have now been established.
The empirical strategy of section 4 is two-fold, the causal impact of teenage parenthood on later life outcomes has been assessed parametrically and semi-parametrically, the Logit model and PSM, respectively. In a similar vein, the potential pathways can be gauged by the Logit model and PSM. Both methods are employed, yet again, because if these vastly different approaches yield approximately similar results, our believe in the validity of these results is reinforced 20. The results of both methods are presented below. A pathway is identified as
such if it reduces the marginal or the average treatment effects. For the average marginal effects there is an additional requirement that the pathway is statistically significant. Therefore, the results in table 14 are to be compared to table 8. For males, the effect of self-assessed health and depressive symptoms is substantially lower, in absolute value, from 8.4 to 4.5 percentage points and from 6.9 to 5.0 percentage points, respectively. The average
treatment effects in table 15 are to be compared to those in table 3. The average treatment effect is smaller, in absolute value, for self-assessed health from 13.0 to 0.9 percentage points and larger for depressive symptoms, from 7.4 to 9.3 percentage points. For females, for the Logit model, the negative effect for self-assessed health is decreased from 9.3 to 3.0
percentage points, while the negative effect of depressive symptoms is decreased from 5.0 to 4.0 percentage points. For the semi-parametric approach, PSM, for females the negative self-assessed health effect is decreased from 10.3 to 6.1 percentage points, and the negative depression effect has decreased from 5.9 to 4.1 percentage points.
20 Post-treatment matching as done in this section is not an undisputed method, because it
raises the probability of a bad control problem as noted in Angrist and Pischke (2008).
34 Table 14
Pathway Analysis Logit Model
Male Female
Dependent
Variable: Self-Assessed Health Self-Assessed Health
Teenage Parenthood -0.045 -0.030 (0.044) (0.025) Education 0.040* 0.113*** (0.022) (0.026) Income Lifetime Partner 0.072*** (0.022) (Lack of) Depressive
Symptoms
(Lack of) Depressive Symptoms Teenage Parenthood -0.050 -0.040 (0.0367 (0.025) Education -0.029 0.027 (0.018) (0.026) Income Lifetime Partner 0.006 (0.022)
Notes: */**/** indicate significance at the 10%/5%/1% level, standard errors are within the brackets, numbers represent marginal effects.
Table 15 Pathway Analysis PSM Self-Assessed Health (Lack of) Depressive Symptoms Teenage Father -0.009** -0.093*** (0.046) (0.035) Teenage Mother -0.061* -0.041 (0.033) (0.032)
Notes: */**/*** indicate significance at the 10%/5%/1% level, standard errors are within the brackets, numbers represent average treatment effects.
35
late-life. The pathway is insignificant in table 14 and the value is increased, in absolute value, in table 15. The average treatment and marginal effects are not nullified, this signifies that educational attainment does not capture all pathways. In other words, not all variation in self-assessed health and depressive symptoms for those experiencing teenage parenthood can be explained by variation in educational attainment.
The results of table 14 and 15 illustrate that a lower educational attainment and the medium-term variable substantially lower the adverse effect of teenage motherhood on self-assessed health. The results of depressive symptoms for women are inconclusive. In table 14 the pathways are insignificant, indicating that education and the medium-term variable are no pathways. Whereas table 15 shows a reduced average treatment effect, compared to table 3. Again, the pathways, although the adverse effects of teenage parenthood is substantially decreased, do not nullify the results. The inability of the pathways to nullify the results, reflects that other pathways are in place that cause the adverse long-term effects of teenage motherhood.
To conclude this section, the results above imply that there is some evidence that lower educational attainment, caused by teenage fatherhood, is an important transmission channel of adverse long-term physical health effects. The analysis in this section, illustrates that a lower educational attainment and not finding your lifetime partner before the age of 46 are
transmission channels of adverse physical health effects in the long-term for women. That is, teenage mothers experience adverse physical health effects later in life, partially, because of their lower education.
6. Robustness Checks
6.1 Propensity Score matching
The propensity score matching procedure consists of several steps, these steps are outlined in section 4.1.1. Step three comprises of matching the treated (i.e. teenage parent) with
propensity score 𝛼𝑖∈ [0,1] to a non-treated (non-teenage parent) individual with a similar
