Couples in context: Understanding family dynamics in Norway

(1)

Janna Bergsvik

Couples in context:

Understanding Iamily dynamics in Norway

Department of Sociology and Human Geography University of Oslo

2020

(2)

© Janna Bergsvik, 2021

Series of dissertations submitted to the Faculty of Social Sciences, University of Oslo No. 841

ISSN 1564-3991

reproduced or transmitted, in any form or by any means, without permission.

Cover: Hanne Baadsgaard Utigard.

Print production: Reprosentralen, University of Oslo.

(3)

Acknowledgements

This thesis is written as part of the project Changing families and the gender revolution, funded by the Research Council of Norway. The project has been led by Trude Lappegård, who has also served as my supervisor. I am forever grateful to her and to Øystein Kravdal, my co-supervisor, for scholarly and emotional support during my dissertational work.

The demographic research environment has been inviting and generous from start on and I am grateful to everyone I crossed on my way.

I am grateful to my excellent co-authors for generously sharing their knowledge with me, and for teaching me a lot about style, content and methods. A special thanks to Sara Cools for being thorough, inspiring and pragmatic when needed, and a special thanks to Hege Kitterød for sharing incredible knowledge on Norwegian policies, gender and (un)paid work with me and for inviting me to collaborate.

Special thanks to Rannveig K. Hart and Kenneth A. Wiik, who have been engaging coauthors and important mentors to me through my whole dissertational process.

Thanks to my wonderful and bright colleagues at the unit for demography and social research at Statistics Norway. Fruitful discussions with colleagues at Statistics Norway have considerably improved my research. Special thanks to Astri Syse, Rosanna Johed and Lene Sandvik for always being great and caring.

Thanks to Andreas Kotsadam for an inspiring seminar on causal analysis.

Gratitude is also extended to the Research Council of Norway for sponsoring my visiting scholarship abroad, and to the Department of Sociology at the University of California, Los Angeles, for hosting me. Special thanks to Megan Sweeney for kindly inviting me to Los Angeles. Thanks also to Judith Seltzer and everybody at the California Center for Population Research for including me in your daily scholarly activities. Special thanks to Heeju Sohn for sharing the office and L.A. experiences. Thanks also to Patrick Heuveline and fellow PhD students at UCLA for discussing social demography and reminding me not to take access to affordable childcare, parental leave and abortion or even contraception as a given across societies. In Norway we are lucky.

Last but not least, thanks to my friends and family for providing a richness of anecdotes and real-life experiences on the pitfalls and joys of modern family life.

Oslo, October 15, 2020 Janna Bergsvik

(4)

(5)

Summary

This thesis studies contemporary family behavior. Part A comprises an introduction and discussion of the overall contribution of the thesis, whereas Part B comprises five independent papers: one based on a systematic international literature review and four based on empirical analyses of data from Norwegian administrative registers. Together with coauthors, I analyze how fertility, couple specialization and union stability are interrelated and linked to wider social contexts using a combination of descriptive and quasi-experimental techniques. The papers concern three main areas. First, Norwegian family policies are placed in an international context by means of a systematic literature review on the effects of policies on fertility (Paper 1). Next, couple specialization and family dynamics in Norway are empirically examined, recognizing the historically unique shifts towards increasing gender equality and cohabitation (Papers 2 and 3). Finally, an empirical assessment of how Norwegian couples’ embeddedness at the very local level is interrelated with their childbearing behavior is provided (Papers 4 and 5).

There exists a large body of work examining the potential that family policies may have in boosting fertility. The first paperof the thesis is a systematic review of this literature, wherein our focus is on evaluating the evidence of the effectiveness of policies on fertility in major Western countries since the 1970s. Our bird’s eye view of quasi-experimental results corroborates observations long made by demographers: Family-friendly policies docontribute to higher fertility. Comparing fertility effects across types of policies such as parental leave, public childcare, health services, universal child transfers and welfare measures, we find the strongest evidence for an effect for public childcare and cash transfers. Countries with limited support for families should reflect on this solid base of evidence if their aim is to boost fertility.

With the advance of the gender revolution in many Western societies, income dynamics in couples are changing. Nonetheless, parenthood still promotes specialized gender roles. The second paperof this thesis investigates possible changes in the associations between parenthood and within-couple inequality in earnings in Norway in recent years. Using interactions and fixed effects models, we compare the development of within-couple gender gaps in earnings over time between childless couples and couples with children of different ages, as well as changes within couples before and after childbirth. Despite Norway’s generous family policies and gender egalitarian ambitions, the results show that the gender gap in earnings in couples increases with the number of children and is most distinct among couples with children below school age. However, the association between parenthood and within-couple inequality in

(6)

earnings is reduced across the study period, a development partly driven by a decreasing fatherhood premium evident from 2009 onwards. Our findings thus indicate marked changes in how men and women prioritize paid labor after childbirth.

Gender specialization and union dynamics are also in focus in the third paper. It is a widely found phenomenon that couples with female main earners have higher divorce risks.

However, there is inconsistent statistical evidence on the underlying reasons, which are likely to be complex, while even less is known about cohabiting couples. Using regression discontinuity methods, this study explores whether union dynamics differ between couples when the woman contributes slightly more to the couples’ total earned income, as compared to couples where the woman’s income is below the man’s income. We find that union dissolutions are higher among couples where the woman earns more than the man, a result mainly driven by married couples, as no significant discontinuity is found in the union dynamics of cohabiting couples. As such, gendered expectations on income seem important for union dynamics even in Norway, a supposedly gender egalitarian society.

Lastly, the focus shifts to explore how couples’ embeddedness at the very local level is interrelated with their childbearing behavior. Geographical variations in fertility and the diffusion of fertility across space and social networks are central topics in demographic research. Less is known, however, about the role of neighborhoods and neighbors for geographical variations in fertility. The fourth paperof this thesis investigates spatial variations in family size by analyzing third births in a neighborhood context. Using unique longitudinal geo-data to construct ego-centered neighborhoods, this paper introduces a new geographical dimension of fertility variation. Data on housing and fixed effects for statistical tracts are used to account for sorting into housing and urban versus rural districts. The analysis shows that the likelihood of two-child couples having another child increases with the share of families in the neighborhood that have three or more children. Consequently, this study indicates that transitions to third births may be linked to social interaction effects among neighbors, in addition to well-known processes of selective residential sorting.

The fifth paperseeks to unpick the contribution of these two causal mechanisms using the sex composition of the two firstborn children and twin births as instrumental variables (IVs) for having a third child. We measure effects of the third child on three separate outcomes:

parents’propensity to move, characteristics of their final neighborhood, and the fertility of their neighbors. We identify selective moves as one plausible causal driver of the spatial correlation

(7)

the spatial correlation in fertility is also driven by factors that we effectively control for in our design – most importantly self-selection based on preferences for family size and a family- oriented environment and life style. The second mechanism examined relates to social interaction effects among neighbors. Because of the difficulty of measuring social interaction effects among neighbors we are reluctant to say that they do not exist, even though we do not identify them. Nevertheless, we contribute to the understanding of fertility and relocation, but also to the literature on social interaction effects in fertility by testing the relevance of yet another network, i.e. that of neighbors.

Following the downward trend in fertility rates since 2009, fertility has received political attention in Norway. Our systematic review of policy effects on fertility showed that the generous Norwegian family policies might have contributed in raising fertility levels. At the same time, however, findings from this thesis reveal that children still come with labor market costs for women, and increasingly also for young men. Gender roles persist in Norway and continue to influence family behavior. If one is looking to increase fertility, housing policies and/or policies relating to family-friendly infrastructures might be a new direction to explore, especially given the evidence for self-selection into certain areas and neighborhoods. Such topics have received little attention in family policies so far but have been shown to be important in the empirical research presented in this thesis.

(8)

(9)

PART A

1 Introduction

¹

Societal change and shifting family behavior go hand in hand. Individuals’ decisions on partnerships, childbearing and organizing family life are not made in a vacuum but are linked to each other and to broader societal dynamics. The ambitious aim of research in family demography is to understand these links and social mechanisms, thus furthering insights into how individuals and societies may fare in the future (Seltzer et al. 2005).

Family dynamics impact individuals through their whole life course. How family life looks, will influence children as they grow up (Kravdal and Grundy 2019; Härkönen, Bernardi, and Boertien 2017; McLanahan, Tach, and Schneider 2013). The acceptance of alternative concepts of living together, such as unmarried cohabitation or same-sex marriages, shapes choices of young adults when forming unions (Cherlin 2004). How couples organize their work and family life impacts men’s and women’s life prospects (Anxo et al. 2011; Drange and Rege 2013) and not least the everyday life of all family members (Flood, Meier, and Musick 2020;

Musick and Meier 2012). Union dissolutions, childlessness and life-long singlehood are increasingly socially accepted but may have detrimental consequences if social bonds to close family are lacking, especially at older ages (Kravdal et al. 2012; Schoen et al. 1997). Welfare states are built around these changing preferences and needs, usually with the intention to provide safety nets for the population at all life stages, but also with the purpose of enabling individuals to act according to their preferences and thus facilitate well-being. Consequently, knowledge about demographic trends and processes that inform decision-making, policies and priorities, is a premise for a for a well-functioning welfare state (Kravdal 2010).

Family dynamics are affected by the economy, which in turn pave the way for further changes in both working life and the economy (Ruggles 2015). The shift from a society of agricultural work and family businesses to paid labor heralded the rise of the male breadwinner family, wherein young men were provided the opportunity to build their own family independent of generational dependence (Ruggles 2015). Women’s increased opportunities have been emphasized in the more recent shift from the male-breadwinner to the dual-earner family. Since the 1970s, the gender revolution has radically changed family life in many parts

1 I am grateful to Rebecca Gleditsch, Trude Lappegård, Astri Syse, Michael Thomas and Kenneth A. Wiik for comments to (parts of) previous drafts of this Introductory chapter. I am also grateful to Rannveig K. Hart for useful discussions.

(14)

of the world (Goldscheider, Bernhardt, and Lappegård 2015). Not only did the rise in women’s employment challenge the gendered division of labor in homes and in work places, the educational success of women also brought about new mating patterns and new standards for the ordering of life events. In several countries, couples increasingly consist of partners with equally high levels of education (Kravdal and Rindfuss 2008; Klesment and Van Bavel 2017) with young men and women increasingly finishing higher education and embarking on stable employment trajectories before they have their first child (Hart 2015). In Norway this ideal is maintained by family policies favoring employed parents.

Notions about how to start, organize and maintain family life vary by social context.

Understanding family dynamics means to situate family behavior within the wider social context both historically, geographically, and politically. For good reasons, the nation state with its laws and social policies is one context in which analyses of family behavior often are embedded (Cooke and Baxter 2010). Still, demographic transitions have usually transformed whole regions and resulted in differential subnational dynamics (Bongaarts and Watkins 1996;

Goldstein and Klüsener 2014). Hence, analyzing family dynamics also along different geographical dimensions is desirable (Boyle 2003; Voss 2007; Weeks 2004). This thesis explores how family decisions interrelate with three social structures in which family life is embedded: national policies, gender ideology and the residential context.

This thesis consists of two parts: Part A comprises an introduction and a broader discussion of the overall contribution of this thesis, whereas Part B comprises five independent papers examining various subtopics within the broader framework outlined above. The papers concern three main areas: Paper 1 situates Norwegian family policies within an international context by means of a systematic review of the literature of policy effects on fertility. Next, couple specialization and family dynamics in Norway are empirically examined in light of the historically unique situation with an increasing share of gender-egalitarian and cohabiting couples (Papers 2 and 3). Finally, the focus shifts to an empirical assessment of how Norwegian couples’ embeddedness at the very local level is interrelated with their childbearing behavior (Papers 4 and 5).

Part A of this thesis continues with an introduction to the Norwegian context in Section 1.1. The research questions and overall aim of this thesis are elaborated on in Section 1.2. The theoretical frameworks that have inspired this thesis are outlined in Section 2. A short introduction to the methods used in the empirical investigations in the individual papers in part B, as well as a more general discussion of identification of causality, are given in Section 3,

(15)

3 before the data that have been used are described and discussed in Section 4. Section 5 provides a short summary of the individual papers. Finally, Section 6 discusses the findings and overall contribution of this thesis, before offering some concluding thoughts.

1.1 The Norwegian context

Together with the other Nordic countries, Norway is known for its family-friendly working life, its generous universalistic welfare state and high gender equality. A large public sector with opportunities for part-time work and a short normal working week (37-40 hours) have contributed to elevate female employment rates in Norway to levels very close to that of men.

About 70 percent of men and 65 percent of women between ages 15-74 were employed in 2019.² The universalistic welfare state supports life-long full-time work for men and women alike, and the combination of work and childcare is further enabled by family policies such as a long paid parental leave of about one year³ and widely available and affordable public childcare for all children above one year of age. Public acceptance for childcare is high, and 92 percent of all children between ages 1-5 were enrolled in childcare in 2019.4

Since the 1970s, the main goal of Norwegian family policies has been to support a symmetrical family model wherein women and men share domestic duties and paid work equally between them (Ellingsæter and Leira 2006). To promote gender equality in family- related tasks, and in the labor market, certain weeks of the parental leave allowance have been earmarked for fathers since 1993.⁵ These policy measures have been expanded and strengthened over the last decades. In line with this, the dual-earner model is widespread among contemporary Norwegian families.

Norway is consistently ranked one of the most gender-equal countries in the world.⁶ Besides a high female labor force participation, women now outnumber men in higher education and are better educated than their partner in an increasing number of couples. In the cohorts born 1940 to 1964, 38 percent of first-time mothers had the same education level as the fathers, and the share of parental couples in which the mother had the highest level of education

2 https://www.ssb.no/arbeid-og-lonn/faktaside/arbeid

3 In 2020 it was 49/59 weeks with 100/80 percent wage compensation up to a cap at six times the social security base income (G), e.g. in total around 600,000 NOK (approximately 55,000 Euros). This cap is above the average income of women but below that of men.

4 https://www.ssb.no/en/utdanning/statistikker/barnehager

5 The fathers’ quota has been adjusted several times. In 1993 it was 4 weeks, while it is 15/19 weeks in 2020, depending on the compensation grade (100/80 percent).

6 See, for instance, the United Nations Gender Inequality Index (http://hdr.undp.org/en/content/gender- inequality-index-gii) and the Global Gender Gap Report by the World Economic Forum

(http://reports.weforum.org/).

(16)

increased from 19 percent for the earliest cohorts to 30 percent for the most recent ones (Kravdal and Rindfuss 2008). In 2019, 60 percent of students enrolled in higher education were women.

Nonetheless, the fields of study are gender-segregated: Relatively few women have top positions in industry and commerce (Bertrand et al. 2019), a gender wage gap persists, and in 2019 about 36 percent of Norwegian women worked part-time. Among the newest cohorts of mothers with preschool-aged children the part-time work rate is, however, lower (Ellingsæter and Jensen 2019).

Norwegian family policies favor employed parents and several studies indicate that couples nowadays choose to postpone childbearing until both partners have established themselves in the labor market (Hart 2015). Along with women’s higher educational attainment, this has led to rising ages at first birth in Norway. In 2019, the average age at first birth was 29.8 years for women and 32.0 years for men.⁷ This postponement of parenthood is one of two reasons behind a decline in the total fertility rate in Norway after 2009. Over the last ten years, the total fertility rate has declined substantially in all Nordic countries, with a reduction in Norway from 1.98 (2009) to 1.53 (2019) children per woman. This fertility decline has weakened the Nordic countries’ exceptional position as countries with both high gender equality and relatively high fertility. Even though completed cohort fertility remained at 1.96 for Norwegian women who were 45 years old in 2019, the low period fertility has resulted in a historically new political focus being placed on fertility and potential ways to raise its levels in Norway. Consequently, pronatalist perspectives are becoming more widespread on the political agenda.

An additional reason for Norway’s declining fertility is that fewer women are having three or more children. Women with many children have previously been shown to be overrepresented among housewives (Kitterød and Rønsen 2013) and among couples with weaker connections to the labor market (Hart, Rønsen, and Syse 2015). The share of women with three or more children has declined in all cohorts of women born after 1960. Having two children is thus currently most common. The most recent cohort of women that have reached age 45, and thus for all practical purposes have finished their fertile careers, was born in 1974.

In this cohort, 15 percent have one child, 42 percent have two children, while 29 percent have three or more children. Only 14 percent are childless (see Table 5.2, Syse, Thomas, and Gleditsch 2020). The link between the declining number of families with more than two children and the declining total fertility rate has traditionally been considered important in

7 https://www.ssb.no/en/befolkning/statistikker/fodte

(17)

5 Norway (Kravdal 1992), and while the number of large families is falling, almost half of Norwegian survey respondents in 2012 continue to regard three or more children as the ideal number for a family (ISSP Research Group 2016).

Family sizes are spread unevenly across the country. For several decades, the share of women with at least three children has been highest in the southwestern parts of Norway, also referred to as the Norwegian Bible Belt. As in most countries, women have more children in rural areas and suburbs than in urban areas (Kulu, Boyle, and Andersson 2009; Kulu, Vikat, and Andersson 2007). Besides these regional patterns, fertility rates differ between neighborhoods. For example, within the capital city of Oslo, the difference in total fertility rates between urban districts stood at 0.8 children in 2015 (TFR of 2.08 in Bjerke versus 1.29 in St.

Hanshaugen) (Syse, Hart, and Aase 2016). In contrast, the difference between counties was 0.3 children (Syse, Hart, and Aase 2016). While the association between sociodemographic characteristics and third births in Norway is relatively well-studied (Kravdal and Rindfuss 2008; Hart, Rønsen, and Syse 2015), the origin of the uneven spatial distribution of large families has received less attention.

In an international context, the Norwegian society is seen as highly secularized and individualized and was early to adopt diverse union forms (Lesthaeghe 2010). Currently, most co-residential partnerships in Norway begin as unmarried cohabitations and more than 50 percent of first births occur within cohabiting unions (Noack, Bernhardt and Wiik 2014). At the same time, the attitudes, behaviors, and socioeconomic characteristics of cohabitors differ from those of married persons (Wiik, Keizer, and Lappegård 2012; Kravdal 1999), implying that such union types are qualitatively different from marriages, and often a result of self-selection of those who disapprove of, or face barriers to, marriage. Cohabitors are also more prone to dissolve their unions than those who are married, even when couples have common children (Hart, Lyngstad, and Vinberg 2017).

Union dissolutions have in general increased considerably since the 1960s. Today, around 40 percent of Norwegian marriages are estimated to end in divorce, compared with only 10 percent of those that were initiated in 1960 (Dommermuth et al. 2015). Recently, however, divorce rates have stabilized and even decreased, particularly among those aged 40 and younger. One plausible explanation is an increased selection of only the most committed cohabitors into marriage.

(18)

1.2 Research questions and aim of thesis

The gender egalitarian social-democratic Norwegian welfare state (Esping-Andersen 1990) of the last two decades (2000-2018) serves as main empirical case for this thesis and is examined within an international, national and local perspective. Important changes in family behavior are addressed, touching upon three core trends: the declining number of children, changing gender roles in families, and increasing union instability. The five separate papers that comprise Part B of this thesis each examine various topics yet build upon and complement each other.

While the overarching research aim concerns how couples’ decisions regarding family behavior are linked to their wider social (i.e. policy, gender, residential) context, each paper deals with its own specific research question.

International: 1) Do policies affect childbearing?

National: 2) Does childbearing affect gender specialization in income?

3)Do gendered expectations on income affect union stability?

Local: 4)Is neighbors’ fertility behavior linked to each other?

5)How can we explain the residential clustering of fertility?

This thesis begins by placing Norwegian family policies in an international context.

Inspired by the microeconomic theory of fertility, social scientists have produced a large volume of literature on how policies affect fertility decisions. Results from (quasi)experimental analyses are an important piece of that puzzle, but no systematic review of the full body of such studies exists to date. This thesis’ first paper “Can policies stall the fertility fall? A systematic review of the (quasi)experimental literature” aims to fill this gap and discusses the results of a systematic review of the literature on the effects of policy on fertility since 1970 in Europe, the US, Canada and Australia.

The Nordic countries with their well-developed policies supporting the combination of employment and childcare for men and women are often seen as ideal templates for promoting a gender-equal dual-earner/dual-caregiver family model (Esping-Andersen 2009). Thus, we next focus on the implications of an increasing share of gender-egalitarian couples. Based on this, the second paper of this thesis “Parenthood and couples’ relative earnings in Norway”

analyzes the role children play for gender specialization in income. This question is certainly

(19)

7 not new, but our focus on whether the association between parenthood and within-couple inequality in income has changed in recent years, contributes to the existing literature in this field by assessing whether such changes are linked to a reduced impact of children on women’s income and/or a stronger impact on men’s income.

Gender specialization and union dynamics are also of interest to the third paper

“Women’s relative income and the dynamics of marital and cohabiting unions”. In this analysis, the aim is to investigate whether gendered expectations on income provision in couples influence the relationship functioning of couples where the woman earns more than the man, and whether such possible patterns differ by union type. This provides insights into gendered expectations in modern couples and may give an indication of whether cohabiting couples are qualitatively different from married couples in this respect.

Figure 1 Overview of research topics.

Lastly, the focus shifts to explore how couples’ embeddedness at the very local level is interrelated with their childbearing behavior. Geographical variations in fertility and the diffusion of fertility across space and social networks are central topics in demographic research. Less is known, however, about the role of neighbors and neighborhoods with regard to geographical variations in fertility. The fourth paper “Linking neighbors’ fertility: Third births in Norwegian neighborhoods” contributes to the literature by widening spatial and

Policies P1 + P2

Gendered practices

P2 + P3

Local residential

context P4 + P5

Ideas about family life and couple decisions

- fertility (P1, P2, P4, P5) - couple specialization (P2, P3) - union dissolution and marriage

among cohabitors (P3)

NATIONAL DYNAMICS

SUBNATIONAL DYNAMICS

(20)

network perspectives in fertility research to also encompass neighbors and neighborhoods. The fifth paper “Explaining residential clustering of fertility” contributes to the literature by using exogenous variation in fertility behavior to identify a possible causal effect of an additional child on residential moves and neighbors’ fertility.

The research questions all have their roots in the microeconomic theory of the family (Becker 1991), as well as the two most prominent contemporary demographic frameworks: The second demographic transition (Lesthaeghe 2010; van de Kaa 2001) and the gender revolution (Esping-Andersen and Billari 2015; Goldscheider, Bernhardt, and Lappegård 2015). These theories are outlined in more detail in Section 2.

1.3 Delimitations

The papers in this thesis are necessarily limited in scope as they mostly focus on contemporary family dynamics of different-sex couples and quantitative research within social demography from countries with comparably advanced economies.

Fertility and family behavior are understood as relatively free choices, enabled by law and subsidized by the welfare state. In the contemporary Norwegian context, regulation costs are low. There is broad access to contraception and early medical abortion. Unintended pregnancies are therefore assumed to be a minor issue and are not discussed further. In the study period (2000-2014), about every fifth pregnancy in Norway was terminated.⁸ All measures used in this thesis refer to live-born children, and thus this thesis makes no strict distinction between the desired or intended versus actual number of children, because desires and intentions are interrelated and subsequently revised to match possibilities and constraints (Iacovou and Tavares 2011).

Same-sex couples are excluded from the studies of gendered practices to reduce complexity, although admittedly much interesting insight emerges from studying the division of labor in same-sex couples (Evertsson and Boye 2018; Andresen and Nix 2019). Including family complexity, as for instance stepfamilies, only as a ‘control variable’ is not giving the topic the attention that it deserves (Sweeney 2010). As such, a thorough examination of same- sex couples, as well as other more recent partnering patterns and dynamics, for instance stepfamilies, increasing educational homogamy and long-term singlehood (Wiik and Dommermuth 2014) falls outside of the scope of this thesis.

8Yearly between 14,000 to 16,000 pregnancies were terminated, compared to 55,000 to 60,000 births (see, http://statistikkbank.fhi.no/abort/ and http://statistikkbank.fhi.no/mfr/).

(21)

9 All empirical studies are based on register data, which also necessarily limits the scope of the analyses. For instance, it is not possible to include direct measures of norms, values or preferences using register data. Such issues would require other sources of data such as those collected via household surveys or from qualitative studies. Some implications of this are discussed in Section 6.2.

(22)

2 Theoretical framework and previous literature

This thesis draws upon three theoretical frameworks which are widely used to describe, explain and predict ongoing social demographic processes: the microeconomic theory of the family (Becker 1960; Becker 1991; Becker, Landes, and Michael 1977), the second demographic transition (Lesthaeghe 2010; van de Kaa 2001) and the gender revolution (Esping-Andersen and Billari 2015; Goldscheider, Bernhardt, and Lappegård 2015). In this Section, each of the theories will be briefly outlined. Next, the theories will be linked to the research questions of this thesis.

2.1 The Microeconomic Theory of the Family

The microeconomic theory of the family is a rational choice theory based on individual and family utility maximization. Family behavior is explained by rational weighting of the costs and benefits of having children (Becker 1960), the division of (un)paid labor within couples (Becker 1991) and of being partnered (Becker, Landes, and Michael 1977).

Raising children takes time and money, and in its simplest form, the economic theory of the family postulates that the number of children couples have depends on the couples’

amount of time and money, as well as their preferences for spending that time and money on children or other purposes (Becker 1991). If children are a ‘normal good’ (i.e. a good for which consumption increases with income), increased incomes or decreased costs will translate into larger family sizes. However, several mechanisms make the expected relationship between resources and fertility more complicated.

First, having more resources could make parents invest more in each child, e.g. provide better housing or schooling. This would in turn increase the cost of raising a child and could reduce the demand for children. Such a quality-quantity trade-off (and the preference for quality above quantity) can lead to (counterintuitive) negative income effects (Becker 1991). The expected relationship between resources and fertility is further complicated by the fact that (at least one parent in) most families earn the bulk of their income in the labor market. With increasing wages and stronger attachment to the labor market, the losses from taking time off work to care for children (the opportunity cost) increase, too. This substitution effect complicates a precise understanding of the fertility effect of household income. Time costs have been disproportionally taken by mothers, and if this has dampened fertility (after the gender revolution), shifting the costs to fathers may have pronatalist effects (Goldscheider, Bernhardt, and Lappegård 2015). However, such effects will in a rational choice framework emerge only

(23)

11 if fathers’ increased costs do not negatively impact fertility more than the positive effects from mothers’ reduced burden.

Differentiated gender roles have long been considered functional and necessary for family stability (Parsons 1949) and the mutual dependence of spouses arising out of their specialized functions is seen as the major gain to marriage in the microeconomic theory of the family (Becker 1991). Although women’s employment and income increase the total resources of the family and hence the family economic utility, women’s economic independence is expected to increase the risk for union dissolution (Becker, Landes, and Michael 1977). A specialized division of labor within the household will, according to the specialization hypothesis, result in a better financial situation and reduce time pressure.

In principle, it could be of minor importance whether it is the man or the woman who specializes more strongly in (un)paid work. However, small but nonetheless significant biological sex differences, which predispose women to domestic production and reproduction and men to paid labor, build the underlying rationale for a gender-typical couple specialization, which is further supported by men’s historically higher earning potential, statistical discrimination against women in the labor market and the persisting gender pay gap. A gender- typical division of labor between household work and paid labor is, in other words, maximizing the family utility and is thus a rational response to structural and biological sex differences (Becker 1991). Such gender role specialization might be strengthened at the arrival of children as sex differences become more pronounced in the process of childbearing and child-rearing.

Also, women’s and men’s earnings potential might influence their bargaining power over time use, which often is especially tight after the arrival of children (Angelov, Johansson, and Lindahl 2016).

A high degree of specialization in the household may at the same time be risky, and historically primarily for women. The loss of one of the specialized partners, due to for instance a break-up, implies a reduction in the economic well-being of the other partner and/or their child(ren) as the exchange of specific skills is no longer desired or possible to the same degree.

Consequently, critics state that the family utility may be highest when both partners contribute economically to the household (Oppenheimer 1997, 1994).

Because family behavior happens at the intersection of emotional and rational choice, the early version of the microeconomic theory of the family has been heavily criticized, and several modifications exist. The lack of acknowledgment of ideals and norms as well as the

(24)

minor weight placed on the role of social networks within which individuals are embedded have been especially criticized. Easterlin (1975: 55) argues that “It is through tastes or subjective preferences that attitudinal considerations stressed by sociologists operate, such as norms regarding family size and the "quality" of children (standards of child care and rearing)”. Hence, according to him, research on the formation of tastes could bridge sociological and economic research on family behavior and should be given high priority.

2.2 The Second Demographic Transition

Since Becker’s influential economic analysis of fertility (1960), family behavior has changed considerably in most parts of the industrialized world. Trends such as the increased spread of cohabitation, postponement of childbearing, childbearing outside of marriage, increased union dissolution and re-partnering have become widely known as phenomena of ‘the second demographic transition’ (Lesthaeghe 2010; Lesthaeghe and Surkyn 1988; van de Kaa 2001). In essence, the theory of the second demographic transition highlights an underlying ideational change towards ‘less family’ in the traditional sense, stemming from the rise of higher order needs such as self-realization and individualization (van de Kaa 2001).

The theory of a second demographic transition builds on a concept of ‘bounded rational choice’, thereby allowing for autonomous preference drifts (Lesthaeghe 2014). Changing preferences are given a more pronounced role than they were in earlier, pure versions of a rational choice-based theory of the family. The perspective of shifting preferences in modern societies builds on Maslow’s popular concept of the hierarchy of needs (Maslow 1954). In short, with increasing standards of living individuals and societies are expected to move from physiological and safety needs to psychological needs of belonging and self-esteem, before self-fulfillment needs become the dominant driver of human behavior. As such, the basic ideas of the theory are related to theories about ‘post-materialism’ (Inglehart 1990), ideas about postmodern families (Beck and Beck-Gernsheim 1995) and general social theories on increased individualization (Beck and Beck-Gernsheim 2002).

The theory of the second demographic transition is closely related to previous demographic transition theories and more general theories of fertility diffusion (Casterline 2001), although its emphasis on ideational shifts and the cultural component stands out.

Whether it is through structural forces or ideational factors, the ideas of social diffusion of new family behavior through space and time are common themes of the demographic transition theories. This has contributed to the analysis of larger worldwide trends, but also given room

(25)

13 for important regional differences and dynamics. In its outmost consequence, predictions and empirical investigations based on the theory of the second demographic transition have highlighted the importance of population composition, local context, and social interaction for understanding variations in family behavior through time and space.

In the theory of the second demographic transition, little emphasis has been placed on gender roles, which is a major critique repeatedly raised against the validity and relevance of the theory (Zaidi and Morgan 2017; Vitali, Aassve, and Lappegård 2015).

2.3 The Gender Revolution

In many countries, gender roles have changed during the last decades, although this development has been described also as uneven and stalled (England 2010). Societal expectations towards the behavior of men and women have traditionally differed the most when couples have children. The ‘doing gender’ perspective argues that gender is constructed through social interactions (West and Zimmerman 1987). That is, individuals’ behavior is evaluated in light of gendered expectations, and deviations are discouraged through social pressures.

Traditionally, masculinity has been strongly linked to breadwinning and femininity to mothering and housework (Johnston and Swanson 2006). Several theories based on changing gender roles offer alternative interpretations and predictions of the changes in contemporary family behavior that are highlighted in the second demographic transition (e.g. McDonald 2000;

Esping-Andersen and Billari 2015; Goldscheider et al. 2015).

In most countries, women’s educational attainment and labor market participation have increased considerably in parallel with the decline in fertility and the increase in couple instability. Questioning whether the decline of traditional norms of family life indeed reflects individualization, Esping-Andersen and Billari (2015) describe the ongoing process as containing multiple equilibria (see Figure 1). Comparisons between countries reveal that family life and couple specialization is either marked by traditional gender roles, egalitarian practices or may be defined to be in an unstable phase of transition. Of those phases, only the transitional phase is characterized by individualization and ‘less family’. Rather than witnessing normatively unanchored family forms in gender egalitarian societies, a new dominant model emerges: ‘[O]ne premised on broad conformity with principles of gender equity and symmetry’

(Esping-Andersen et al. 2013: 1280).

According to Goldscheider, Bernhardt, and Lappegård (2015), the gender revolution has two stages: In the first stage, that began around the 1960s and 1970s in Europe, women entered

(26)

the public sphere, i.e. the labor market and the political arena. The resulting growth in female labor market participation and women’s higher educational attainment challenged traditional family roles and led to a mismatch between gender equality in the public and the private sphere.

McDonald (2000) has earlier made the distinction between gender equity in individual-oriented institutions such as education and labor markets and gender equity in the family. Slow improvements in gender equality within the family and domestic tasks entail a work-family conflict for women, which might lead them to withdraw both from marriage and childbearing.

Hence, according to Goldscheider, Bernhardt, and Lappegård (2015), the second stage of the gender revolution involves a change in men’s roles, including a more equitable division of household labor and parenting.

Figure 2 The idea of multiple family equilibria (Esping-Andersen and Billari 2015).

Source: Esping-Andersen and Billari, 2015: 9.

The gender revolution is described as an ongoing process, and no country has yet completed the second stage with men and women participating equally in both the labor market and in family life. However, many countries are described as being in the transition between the two stages (Esping-Andersen et al. 2013). During such a transition, societies may face uncertainties about gender roles and young men and women may have unclear prospects about what to expect from their partner. Studies repeatedly show that gender norms may be quite sticky (Endendijk, Derks, and Mesman 2018; Bertrand, Kamenica, and Pan 2015). However, in

(27)

15 line with the propositions of the theories relating to the gender revolution, a certain level of gender equality in couples may have become normative and desired by both men and women in Nordic countries, as exemplified by the fact that most Swedish and Norwegian fathers use their reserved part of the parental leave, and in some cases more (Lappegård et al. 2020). The value of the theories for countries outside the relatively rich and progressive Nordic context and for less privileged individuals, i.e. men and women without higher education, has been questioned and the theories have been criticized for lacking a perspective on social inequality (Cherlin 2016).

2.4 Family life between utility maximization and gender roles

Since policies that affect fertility will typically work by affecting the time and money available to parents, Paper 1 on the effect of policies on fertility builds primarily on the microeconomic theory of fertility (Becker 1960). Although gender roles are clearly relevant for many family policies, the perspective is not directly utilized in Paper 1. Papers 2 and 3 combine the perspectives found in the microeconomic theory of the family with perspectives found in theories relating to the gender revolution. More specifically, we draw on Becker’s ideas of (economic) couple specialization (Becker 1991; Becker, Landes, and Michael 1977), whereas we consider possible changes over time in light of the gender revolution (Esping-Andersen and Billari 2015; Goldscheider et al. 2015). In this thesis, the perspective put forward in the second demographic transition are implicitly drawn upon, since ideational change is a concept that is hard to grasp using register data. However, Papers 4 and 5 have been inspired by the demographic transition theories acknowledgement of variations in family behavior through time and space.

(28)

3 Methodology

This thesis consists of five empirical papers. Except for the systematic review, all rely on different, although related, statistical techniques. All papers include detailed descriptive parts on variables of interests and context. Some papers also include novel methodological attempts to complement the classical approaches. The idea behind the analytical strategy of each paper is critically discussed in the papers themselves, including reflections on selection, as well as the reliability of each strategy. In combination, the papers provide new insights for topics that have to now been less explored, while others draw on novel techniques and/or apply a different focus to uncover new insights in more established areas of research. This chapter aims to give a summarized overview and provide a discussion of the techniques and methods for the reader who might be less familiar with them.

Most papers in this thesis use a combination of methods. Since quasi-experimental research designs are used, I discuss identification of causality in general in section 3.1. Then, event history analysis, applied in papers 3 and 4, is presented in Section 3.2. Linear probability models as an alternative to logistic regression models are discussed in Section 3.3. Extensions using fixed effects (applied in papers 2 and 4) are presented in Section 3.4. Although some of these techniques may provide estimates with a causal interpretation, I suggest causality only for limited parts of paper 2. Three papers (paper 1, 3 and 5) include quasi-experimental techniques.

Yet, the causal interpretation of their estimates is also here not presupposed but critically discussed for each case in the papers themselves. In section 3.5 I give an overall introduction to quasi-experimental methods, before I briefly characterize each of the applied approaches:

regression discontinuity and instrumental variable analysis. This thesis also includes a systematic literature review (paper 1), which is discussed last in Section 3.7. The systematic review includes quasi-experimental studies only.

3.1 A note on causality

Regression estimates might or might not permit causal interpretation. Causal research designs in the econometric sense give estimates that, under certain identifying assumptions, can be given a causal interpretation in the form of a one unit increase in X giving an increase in Y (Angrist and Pischke 2009). For such estimates the terms ‘effect estimates’ or ‘causal estimates’

are used in this thesis. In other cases, estimates are referred to as a representation of an

‘association’ or ‘correlation’ between X and Y.

(29)

17 This understanding of causality stems from the counterfactual model of causality (Gangl 2010). In the counterfactual model, the effect of X on Y is the difference in Y between a situation where X has happened compared to a counterfactual situation where everything else is the same, but X did not happen. The counterfactual approach to causality is useful when the interest lies in the effect of ‘intervention’ X on outcome Y (Angrist and Pischke 2009).

Identifying this causal effect requires research designs that overcome selection problems and confounding from unmeasured factors (omitted-variable bias). Quantitative designs used to identify causal effects are also called identification strategies. Several non-experimental strategies to tackle identification challenges can allow for causal inference, such as advanced panel data models (Section 3.4) or analytical designs exploiting natural experiments or exogenous reforms for a counterfactual analysis (Section 3.5).

Importantly, the counterfactual model is not necessarily a causal model (e.g. Elster 1998).

It should rather be perceived as a tool to quantify the strength of a causal mechanism.

Conversely, theory testing is not dependent on causal estimates (Berk 2010). Doubtlessly, empirical studies whose estimates do not have a causal interpretation still picture a social reality generated by causal processes. Whether a counterfactual analysis is possible depends on the phenomenon one is interested in. Many demographic determinants, such as sex and age, cannot be randomized. Hence, (quasi)experimental analyses are not adequate for all research questions that are important when trying to understand family behavior. This thesis relies on an understanding of causation as a generative process (Goldthorpe 2001) and argues that well- informed descriptions obtained from population wide longitudinal data and causal estimates from quasi-experiments in combination provides us with valuable information and knowledge about societal dynamics (Moffitt 2005).

3.2 Event history analysis

Event history analysis, also termed ‘hazard regression’ or ‘survival analysis’ (Allison 2010), refers to a group of statistical models used with longitudinal data to study the occurrence and timing of events. Event history analysis is frequently used in quantitative analysis of life course transitions such as becoming a parent, having another child or leaving a partner. The dependent variable in event history analysis is the ‘hazard’ of a given event occurring at time t, conditional on the event not having occurred before t. Event history analysis is particularly well suited for handling censored data and models with many time-dependent covariates (Allison 2010).

(30)

In this thesis, (variations of) discrete time hazard regression models are partially applied in papers 3 and 4 using data organized in a long format with logistic regression software (SAS and Stata). In these applications, the probability of an event occurring in a given time interval is used as an approximation of the hazard of the event occurring. Couples are followed from the time they are under risk of experiencing the studied event (or from when we can observe them for the first time) until the event happens or they are right-censored (because they are no longer at risk or not observable anymore).

Although discrete-time event history analysis is used as a starting point for these papers, none of the applications strictly adheres to a classic event history study for several reasons. The study of within-couple income inequality and union dynamics in the third paper is extended to a regression discontinuity design. The logistic regression for the study of neighbors’ fertility and a couple’s risk of having another child in the fourth paper is converted to a linear probability model for practical reasons (see next Section 3.3).

3.3 Linear probability models

In sociological and demographic research logistic regression seems to be the default option when analyzing binary dependent variables. In economics, however, the linear probability model has become an increasingly popular alternative (Angrist and Pischke 2009). This chapter briefly outlines some pros and cons of both practices.

The most pressing and often-overlooked issue in applications of logistic regression analyses emerges from logit estimates being affected by the degree of unobserved heterogeneity in the model (Mood 2010). Comparing the frequently reported unintuitive (Williams 2012) odds ratios across models or samples is therefore problematic.⁹ Mood (2010: 78f.) discusses the issue of comparability in logistic regression and presents the use of linear probability models as valid solution. The main reservations against using linear regression with binary dependent variables, i.e. getting predicted probabilities out of range, heteroscedastic and non-normal residuals resulting in invalid standard errors and a mis-specified functional form, have been demonstrated to have little practical relevance (Breen, Karlson, and Holm 2018; Mood 2010).

Papers 3, 4 and 5 implement linear probability models with robust standard errors that adjust for potential heteroscedasticity due to the binary dependent variable and the correlation

9 In a logistic model, the parameter estimates not only reflect the correlations between x and y but also the degree of unobserved heterogeneity. Logit transformations rely on the assumption that the error term has a standard logistic distribution with a fixed variance. Hence, any changes in the explained variance, for example due to the inclusion of additional covariates, also change the variance of the dependent variable (see Mood 2010).

(31)

19 of observations over time or within units, where needed (Mood 2010; Snijders and Bosker 2012:

197). Hence, neither heteroscedastic residuals nor predicted probabilities out of range appear to be a problem in the analyses conducted here. Because all main predictors are categorical, the functional form is of minor relevance, too.¹⁰ In such cases, no continuous probability function is modeled, but discrete probabilities associated with each category of the predictor are estimated. Coefficients from linear probability models are closely related to the often-used average marginal effects from logit models (Breen, Karlson and Holm 2018: 50).¹¹ This is also what I found in my work with this thesis (see Paper 4 for an example).

3.4 Extending panel regressions with fixed effects

Fixed effects can be a powerful tool to control for unobserved confounders when working with panel data. If they are applied at the lowest level of analysis, i.e. the individual, couple or household level, changes within individuals over time are in focus and the individual serves as its own control (Allison 2009). In such a case, the model measures effects of within-individual change in the predictor on the individual outcome, while differences between individuals are discarded. If the fixed effects are at a higher level than the outcome, e.g. the neighborhood, they are in a strict sense not much different from other dummy variables included in a regression model, i.e. they may be considered absorbed dummies.

In papers 2 and 4 different forms of fixed effects are used to reduce the problem of omitted-variable-bias. While these models give estimates with less bias caused by unmeasured characteristics, fixed effects may always also absorb interesting information. Therefore, in both papers fixed effects are used only as extensions.

3.4.1 Couple fixed effects

To obtain estimates of the effect of parenthood on income gaps within couples, we ran couple- level fixed effects models in paper 2. Couple-level fixed effects models were added for the subsample of couples who had a child during the observation period, comparing earning inequalities within couples before and after they had a(nother) child. Couple-level fixed effects capture all stable couple characteristics. Hence, these models give the net change in relative earnings by shifting from one family status to another and eliminate baseline differences in

10 In the applied cases of this thesis, nearest neighbors’ fertility, as well as women’s relative income and the third birth is measured in categories.

11 Breen, Karlson and Holm (2018: 50) state “LPM coefficients are closely related to average marginal effects derived from logit or probit models. In models including only a binary predictor, the two will be identical. In models including multiple predictors or continuous covariates, they will differ, but often not by very much (…)”.

(32)

relative income due to differential selection into parenthood.Obviously, for such models all observations without such a change (in this case: a birth) are omitted from the analysis. The estimation is done using variations of the following equation:

Yi,t = Įi + ȕ 1Xi,t + ei,t

Y^i,t denotes the outcome (couples’ income gap) for couple i in period t ; Įⁱis the couple fixed effect, which is absorbed by the within-estimator; Xi,t is a dummy variable for the event of interest, here a childbirth set to 1 for the year of birth until the year after. The pre-event period is the reference period and is in our applied case limited to the preceding two years for all couples. ei,t is the error termcapturing the remaining, unobserved stochastic disturbance.

3.4.2 Neighborhood fixed effects

The papers of this thesis that link residential contexts to family size use individualized and very small-scale neighborhood definitions. This approach was enabled by having access to geocoded addresses for the whole population of Norway. Administrative data offers great opportunities for including many important traits as covariates in regression models. Still, in neighborhood studies, shared unmeasured confounders among neighbors are likely to be present. Therefore, models with spatial unit fixed effects based on administrative neighborhoods are used in the last model of paper 4. Including such neighborhood fixed effects reduces the risk that the main estimates capture unmeasured neighborhood characteristics which correlate with family size.

In the applied case, the chosen administrative unit was statistical tracts, representing a level between the smallest statistical unit and municipalities. Fixed effects take account of time- constant characteristics of these statistical tracts, which may be the built environment, childcare facilities, and other opportunity structures for families that were shared at this or a higher geographical level and that remained constant over the observation period, including relatively time-stable values or norms. The model is represented by the following equation:

Yit = ȕ0 + ȕNbors XNbors,it + ȕTime ZTime,it + ȕTimeSq (ZTime,it x ZTime,it) + ȕControls ZControls,it + σ^௦௧_௦௧ୀଵ ȕ^StatTractZStatTract,it + İ^it

Y^itis the outcome and the subscripts denote the i^thcouple in the t^th quarter of year. X^Nbors,it represents the predictor of interest, which is a measure based on individualized neighborhoods.

ZTime,it is a continuous counter variable (process time) where the first couple-quarter for each

couple is coded as 0, and each subsequent quarter of year incremented by 1. ZControls,it represents other couple characteristics, which serve only as controls. ZStatTract,it are dummies for the about 1,550 statistical tracts in Norway (statistical tract fixed effects), for which effects are absorbed.

(33)

21 To the extent that associations are found, these capture how individual neighborhoods (XNbors,it) deviate from the statistical tract (ZStatTract,it ) wherein the couple lived in a certain quarter of year.

3.5 Quasi-experimental research designs

Methods for causal inference have developed in parallel in statistics (Rubin 1974; Holland 1986), econometrics (Angrist, Imbens, and Rubin 1996; Heckman 1978) and epidemiology (Greenland 1990), to some extent with diverging terminology and specific techniques. The techniques referred to in this thesis (Papers 1, 3 and 5) are mainly inspired by quasi- experimental strategies as found in Angrist and Pischke (2009). The quasi-experimental literature is growing also within family demography and is an important complement to studies of variations in family dynamics across time and space.

Quasi experiments are analytical designs that exploit ‘naturally occurring’ experiments for a counterfactual analysis of the phenomenon of interest. Quasi-experimental studies have high internal validity and are unlikely to interpret other societal changes as effects of the predictor they evaluate. However, effects may depend on the larger societal structure and vary over time and space. Demographers have tended to emphasize external validity, allowing some compromises with respect to internal validity, i.e. potential endogeneity issues (see e.g.

Bhrolcháin and Dyson 2007). In a discussion of causal analysis in population studies, Moffitt (2003: 457) suggests that the limitations of external validity should be met with careful

“weighing of evidence from different studies with different strengths and weaknesses”. For such weighing to take place, structured summaries of available, sound studies are crucial. This thesis contributes with such a systematic review of quasi(experimental) studies of policy effects on fertility (Paper 1).

This thesis also includes empirical applications of two different quasi-experimental strategies: regression discontinuity designs and instrumental variable analysis. Before the strategies themselves are described in Section 3.5.1 and 3.5.2, it makes sense to reflect on some general characteristics of quasi-experimental study designs. In several quasi-experimental study designs, effects must be interpreted as average effects of the treatment on individuals who are treated only because of their treatment status (compliers), also termed the local average treatment effect (LATE). Individuals with strong preferences for being treated (always takers) or untreated (never takers) might in many cases either experience different effects of a policy (Paper 1), or of having a specific income arrangement (Paper 3) or of having a third child (Paper 5). In many cases, estimated effects of an intervention cannot simply be translated to represent

(34)

the total effect of that intervention, and vice versa. Social interaction effects, nonlinearities and the importance of different thresholds may entail that effects evaluated at a specific local margin underestimate the total importance of the intervention(s), (see also Olivetti and Petrongolo 2017).

In general, more conservative identification strategies should be less likely to yield biased results. Findings from this thesis (Papers 1 and 5) demonstrate that there is a trade-off between Type I and Type II errors. Typically, bias is away from zero (Type I error) and, as found in Papers 1 and 5, more rigorous designs more frequently yield no effect. More conservative designs are at the same time more prone to Type II errors, i.e. not identifying existing effects (see discussion in Paper 1). Many quasi-experimental studies analyze relatively marginal changes, which one cannot reasonably expect to lead to large changes in outcomes. Even when based on full population data, these designs may be underpowered for detecting the small effects that one could reasonably expect.

The quasi-experimental methodological framework is more extensively discussed in excellent overviews targeting demographers (Moffitt 2003; Bhrolcháin and Dyson 2007;

Engelhardt, Kohler, and Prskawetz 2009) or social sciences in general (Angrist and Pischke 2009; Morgan and Winship 2015).

3.5.1 Instrumental variable analysis

Instrumental Variable designs (IV) use third variables as instruments to obtain exogenous (random) variation in an otherwise endogenous regressor. Endogeneity is addressed by searching for an instrumental variable (Z) that impacts the explanatory variable of interest (X) but which in itself is uncorrelated to the outcome (Y). Such instrumental variables are hard to find, but if available, they are a powerful tool for handling selection and confounding factors.

Valid instruments must be relevant, i.e. the instrument (Z) must significantly predict the instrumented variable (X), and affect the outcome through the instrumented variable only, i.e.

the exclusion restriction (Angrist and Pischke 2009). These criteria are discussed more thoroughly in Paper 5 of this thesis, where two much-used instrumental variables to obtain exogenous variation in family size are applied: twin births (Rosenzweig and Wolpin 1980) and the children’s sex composition (Angrist and Evans 1998).

Twin births represent an unplanned immediate increase in family size and a permanent increase in family size for women who would otherwise not have had more children. To the extent that having twins is conditionally random (i.e. if mothers of twins are no different from

(35)

23 mothers of singletons after observable characteristics are netted out), it is potentially valid as an instrumental variable for family size. The sex-composition instrument relies on the fact that many couples have a preference for having one child of each sex (Andersson et al. 2006), so that they will have a third child, if and only if the two first born are of the same sex. As child sex is random in most developed countries, so are increases in family size induced by child sex composition. Tests for (conditional) randomness on observable characteristics are presented in Paper 5, and in the case of twin births not fully satisfying.

IV estimation is done in a two-stage least squares (2SLS) regression and consists of two steps. First, the effect of the instruments on the predictor of interest are given as first stage estimates. In the second stage of the analysis, the outcome is regressed on the part of the variation in the predictor of interest which is tied to the instruments. That is, the endogenous regressor is replaced with the predictions from the first step. The IV estimate thus measures the average effect of the predictor of interest among those moved by the instrument (LATE, see previous Section 3.5).

3.5.2 Regression discontinuity methods

Regression Discontinuity designs (RD) use naturally occurring random variation in a treatment defined by certain thresholds. They are suitable when arbitrary thresholds or ‘cut-offs’ – rather than individual choice – define who is affected by an intervention. In such a case, those being

‘just’ affected should be similar to those being ‘just’ unaffected and therefore constitute a good comparison group.

In paper 3 of this thesis a local-linear regression discontinuity (RD) approach is used to analyze whether the association between women’s relative income and union dynamics is distinct at the (threshold) point where she earns just more than him. The use of such a RD design is inspired by previous studies that have used this threshold of women’s share of the household income to analyze the existence of gendered expectations on income in couples (Bertrand, Kamenica, and Pan 2015; Wieber and Holst 2015; Eriksson and Stenberg 2015; Pierce, Dahl, and Nielsen 2013).

The RD approach exploits the fact that couples with relatively gender-equal income arrangements should, on average, be very similar, apart from a small difference in the relative income of each partner. In each year, there is some randomness in whether he or she earns more among couples with relatively equal incomes. However, if there is a gendered expectation that women should not earn more than their partners, it is violated at the point where her earnings

Couples in context: Understanding family dynamics in Norway

Janna Bergsvik