The independent woman—locus of control and female labor force participation

Hennecke, Juliane Article — Published Version The independent woman—locus of control and female labor force participation Review of Economics of the Household Provided in Cooperation with: Springer Nature Suggested Citation: Hennecke, Juliane (2023) : The independent woman—locus of control and female labor force participation, Review of Economics of the Household, ISSN 1573-7152, Springer US, New York, NY, Vol. 22, Iss. 1, pp. 329-357, https://doi.org/10.1007/s11150-023-09650-0 This Version is available at: https://hdl.handle.net/10419/309908 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by/4.0/ Rev Econ Household (2024) 22:329–357 https://doi.org/10.1007/s11150-023-09650-0 The independent woman—locus of control and female labor force participation Juliane Hennecke 1 Received: 29 April 2022 / Accepted: 20 February 2023 / Published online: 6 March 2023 © The Author(s) 2023 Abstract This paper contributes to the research on heterogeneity in labor force participation decisions between women. This is done by discussing the role of the personality trait locus of control (LOC), a measure of an individual’s belief about the causal relationship between behavior and life outcomes, for differences in participation probabilities. The association between LOC and participation decisions is tested using German survey data, finding that internal women are on average 13 percent more likely to participate in the labor force. These findings are also found to translate into higher employment probabilities at the extensive and intensive margin as well as in a lifetime perspective. Additional analyses identify a strong heterogeneity of the relationship with respect to underlying monetary constraints and social working norms. In line with the existing literature, an important role of LOC for independence preferences as well as subjective beliefs about returns to investments are proposed as theoretical explanations for the findings. JEL classification D13 ●J22 ●J16 Keywords Locus of control ●Labor supply ●Female labor force participation ● Personality ●Social norms ●Women and work 1 Introduction Research on female labor force participation has a long tradition. Triggered by the growing labor supply of women in the second half of the last century, a large strand of theoretical and empirical research on this issue has developed over the past *Juliane Hennecke [email protected] 1 Otto von Guericke University Magdeburg, Universitätsplatz 2, 39106 Magdeburg, Germany Supplementary information The online version contains supplementary material available at https://doi. org/10.1007/s11150-023-09650-0. 1234567890();,: decades. Especially the gender differences in participation and its potential explanations have been the center of attention in the economic literature (see e.g., Angrist, 2002, Blau & Kahn, 2007,2017, Goldin, 1990, Goldin & Katz, 2002, Juhn & Murphy, 1997, Mincer, 1962). The most recent research has put much focus on nonmonetary incentives and disincentives such as in specific social norms in order to explain gender gaps (see e.g., Bertrand et al., 2015, Charles et al., 2018, Fortin, 2015, Gay et al., 2017, Goldin, 2006, Knabe et al., 2016). Based on this literature, we already know a lot about why women keep on having elower participation rates and why these variables started converging in recent decades. However, between-women heterogeneity in participation probabilities can only be explained within this framework to a limited extend. Traditional economic models largely attribute these unexplained differences in decision outcomes to idiosyncratic shocks or unobserved constraints and opportunities. As opposed to this, modern behavioral economic and applied microeconomic approaches started investigating these differences with respect to unobserved, inherent beliefs and preferences. A growing literature is thus interested in investigating the psychological black box behind female labor supply. Especially the most recent literature investigates the role of inherent personal attributes for female decision making on the labor market (see e.g Wichert & Pohlmeier, 2010). This paper contributes to the literature by investigating the role of a specific personality trait, an individual’s perception of control, also called locus of control (LOC), for women’s labor supply decisions. LOC can be characterized as a "generalized attitude, belief, or expectancy regarding the nature of the causal relationship between one’s own behavior and its consequences”(Rotter, 1966) and describes whether individuals believe in the effects of their own efforts and abilities on their life outcomes. While individuals with an internal LOC (internals) believe that their own efforts and abilities will be rewarded in their future, individuals with an external LOC (externals) attribute life outcomes mainly to luck, chance, fate or other people. LOC has already been shown to have an important effect on economic behavior and decision making on the labor market in such areas as educational attainment (Coleman & DeLeire, 2003, Mendolia & Walker, 2015), job search effort (Caliendo et al., 2015, McGee & McGee, 2016), occupational attainment (Cobb-Clark and Tan, 2011, Heywood et al., 2017), entrepreneurial activity (Caliendo et al., 2014, Hansemark, 2003) and labor market mobility (Caliendo et al., 2019) and, as an outcome of them, wages (Osborne Groves, 2005, Schnitzlein & Stephani, 2016, Semykina & Linz, 2007). 1 Nevertheless, literature that directly relates female labor force participation to LOC is scarce. Most prominently, Heckman et al. (2006)find a significant positive effect of a combined measure of LOC and self-esteem on the individual probability of being employed at age 30 for the sample of young individuals from the NLSY79. They show that this relationship is much more pronounced for females. Using Australian data from the HILDA, Xue et al. (2020)find that these results on higher employment probabilities for internal individuals can be replicated using twin fixed effects in a sample of non-identical twins while they vanish if identical twins are used. They argue that this indicates a high importance of genetic factors in the 1 See Cobb-Clark (2015) for detailed discussion of the concept as well as an overview of the literature on LOC in labor economics. 330 J. Hennecke formation of LOC. Berger and Haywood (2016) analyze the effect of LOC on mother’s return to employment after parental leave. Using German survey data, they find that women with an internal LOC return to employment more quickly. Based on a heterogeneity analysis with respect to the underlying flexibility in the women’s occupations, they conclude that the effect is mainly driven by different subjective expectations about future career costs of maternity leave. That study is most closely related to the paper at hand. Nevertheless, Berger and Haywood (2016) concentrate on a very specific group of women in a rather exceptional stage of life and the study thus lacks external validity for the decision making for other women. The study at hand is intended to draw a much more general picture and shed light on the important interplay between constraints and preferences in female decision making on the labor market. In addition, this study especially contributes to the discussion of the individual decision making processes underlying the participation of women on the labor market. While the existing economic literature on LOC was mainly concentrated on the role of LOC for subjective expectations about future payoffs, this study also considers the role of personality profiles for preferences in line with the psychological literature. Additionally, the explicit consideration of the interplay between personality and underlying preferences with monetary and non-monetary constraints is new to the literature on female labor force participation as well as the literature on behavioral effects of personality traits. The theoretical considerations mainly discuss two potential mechanisms of the association of LOC and labor force participation, which are (1) difference in the direct marginal utility from participating as well as (2) differences in the subjective beliefs about returns to investment. Firstly, based on the psychological literature on the connection between LOC and independence considerations, the direct nonmonetary gain from participation is expected to be higher for internal women. Internals put greater weight on the status of being active l. They not only derive utility from the consumption level as an outcome of participation, but also from the fact that they themselves had control over generating it. Secondly, in line with earlier literature (Caliendo et al., 2015), LOC can be assumed to have an effect on beliefs about positive future returns to individual efforts, such as parental investments, job search and the investments into future career advancements. Therefore, in the empirical part of the paper, the conditional association between LOC and current labor force participation of a woman is estimated in a reduced form approach. The estimations are conducted using the extensive information available from the Socio-Economic Panel (SOEP, 2020), a large representative longitudinal household panel from Germany. Using this data, the average marginal effects of a woman’s LOC on her probability of participating in the labor force is estimated using a binary logit estimation conditional on standard socio-economic determinants of participation. In this context, labor force participation is defined as a general availability for market production in line with the definition of the International Labor Organization (2018) and thus concentrates on the behavioral implications of LOC on labor supply decisions. The analysis finds a significant positive relationship between having an internal LOC and being available to the labor market. Internal women are on average 13 percent less likely to stay at home. Nevertheless, the relationship is found to be non-linear with especially very external women having significantly lower participation rates. Additional analysis reveals that these effects also translate The independent woman—locus of control and female labor force participation 331 into significantly higher actual labor force activity at the extensive and intensive margin, also in a lifetime perspective. Furthermore, a subgroup analysis reveals that while a strong relationship can be observed for cohabiting women and mothers, the effect for childless women is lower or even zero, depending on family status. This indicates a crucial heterogeneity with respect to underlying monetary incentives to work. In addition, a second heterogeneity analysis shows that the estimated effects are also sensitive to the underlying social norms of working as measured by regional as well as cohort differences. The outline of the paper is as follows. Section 2gives a brief overview the theoretical considerations behind the empirical analysis. Section 3describes the data as well as the empirical strategy and Section 4presents the main estimation results. Section 5gives an overview over various heterogeneity analyses. The paper concludes in Section 6. 2 Theoretical considerations Based on the underlying definition of LOC, multiple hypotheses can be formed about the relationship between LOC and female labor force participation which will guide the empirical analysis. All further considerations will assume a static decision situation which abstracts from the formation process of LOC. This is based on the assumption of relative stability of LOC during adulthood, which will be discussed in further detail in Section 3. Additionally, as the focus of this paper is to analyze the behavioral aspects of labor supply, it concentrates on labor force participation as opposed to actual employment. This reduces the risk of biased results due to omitted returns in employment probability in the empirical section. In line with the ILO definition of “labor force”, a woman is assumed to participate on the labor market if she is either already employed or self-employed or if she is unemployed and intends to participate by indicating that she is searching for a job (see International Labor Organization, 2018). Thus, labor force participation also equals one if the woman does not work but is available to the market through job searching. In this simplification, given a certain expected market wage, no assumptions on labor market conditions and frictions are necessary. The link between a woman’s labor force participation decision and her individual characteristics only depends on her individual preferences and expectations and not on demand-side responses to her characteristics, i.e., a higher or lower employment probability based on e.g., her LOC. In line with this simplification, labor force participation is thus reduced to a binary decision, i.e., for or against participation, at the extensive margin and abstracts from the continuous decision at the intensive margin, i.e., working hours or employment years over the life cycle. Within this framework, the decision making of the woman can be formalized using a basic neoclassical model of labor-leisure choice in which the woman maximizes her utility function Ui¼fC i;Li ðÞ ð1Þ given the limitations imposed by her budget constraint Ci¼~ wTLi ðÞþVi:ð2Þ 332 J. Hennecke with Cbeing the woman’s consumption level, Lbeing leisure or more general all time not spend on labor force participation. Thus leisure equals the total time allocated to the woman Tminus the number of hours she decides to supply given the return she expects from each unit of labor supply ~ w.~ wthus refers to the expected net hourly wages (which depends on the gross wage itself (Mincer, 1962) the tax regime (Eissa & Liebman, 1996, Fuenmayor et al. 2016, James, 1992) but also child care costs (Morrissey, 2016) as well as the probability of receiving this wage (i.e., the joboffer arrival rate) (Caliendo et al., 2015). Vcorresponds to her non-labor income, which includes e.g., partners income (Devereux, 2004, Lundberg, 1988) or social transfers. The purpose of this formal depiction is to create a theoretical framework for the empirical findings discussed in the later sections and not to build the most realistic and detail theoretical model of the decision making of women between paid work, unpaid work and leisure. This is why the framework explicitly abstracts from a formal differentiation of unpaid work and leisure time. The empirical focus of the paper is on the decision whether the woman decides to work or not to work, independent from how the time not at paid work is spent. Thus, the time not spend for paid work is condensed under the label “leisure”although differences between leisure and unpaid work will be considered in the formation of hypothesis on the role of LOC. Based on this framework, three major mechanisms for an association between LOC and female labor force participation are proposed: 1) differences in the marginal utility from participation driven by latent preferences, 2) differences in the subjective expected returns to efforts such as e.g., job search or parental and workplace investments and 3) direct monetary returns to differences in LOC. Preferences The first potential channel suggests that LOC affects a woman’s preferences for the different components of her utility function and thus the marginal utility she derives from participation. In line with Almlund et al. (2011), we can assume that a woman’s marginal gains from leisure and consumption depend on a vector of individual attributes and preferences which are i.e., shaped by personality traits such as e.g., locus of control loci. Ui¼fC i;Li;loci ðÞ ð3Þ Based on the basic idea of the model, we can assume that non-monetary benefits from working, i.e., the “joy of working”is captured by the woman’s preference for leisure. This component incorporate, besides others, known concepts like identity and purpose (Akerlof & Kranton, 2000, Jahoda, 1981, Knabe et al., 2016) but also financial and economic independence. Similar to the argumentation in Cobb-Clark et al. (2014) about the effect of LOC on healthy behavior, internal women (i.e., women with a high LOC) are likely to have a stronger preference for being active on the labor market than external women because they prefer to directly affect their life outcomes and thus be independent of external forces. Thus, they derive more direct utility from participation than externals do, i.e., less utility from leisure time: ∂2Ui ∂Li∂loci <0:ð4Þ The independent woman—locus of control and female labor force participation 333 Holding everything else constant, an internal woman is on average more likely to participate than an external woman as her marginal rate of substitution (∂Ui=∂Li ∂Ui=∂Ci)is lower, i.e., she is willing to give up more leisure hours in exchange for an additional unit of consumption. Thus, consumption which is generated from self-earned income is valued higher than consumption generated from external income such as partner’s earnings or social transfers. Based on these theoretical considerations, internal women are ex-ante expected to be more likely to participate. These considerations are in line with findings from earlier economic and psychological literature. We, for example, already know from the existing literature that income autonomy is an important driver of labor division of paid and unpaid work in couples (Görges, 2014). The role of independence and autonomy considerations for LOC has already been discussed especially in the context of early childhood skill formation in the psychological literature (see e.g., Hill, 2011, Wichern & Nowicki, 1976). As opposed to this, in the presence of children in the household, internal women might consider the effect of own actions on their children more carefully than external women. This is in line with the findings by Lekfuangfu et al. (2018) on the strong effect of maternal LOC on attitudes towards parental style as well as actual parental time investments. Thus internal mothers might have stronger preferences for home production. If we assume that home production in this very simple model is captured by leisure Li(i.e., the opposite of participation), internal women might derive higher utility from every unit of Li:∂2Ui ∂Li∂loci>0 and might thus on average be less likely to participate. Subjective Expectations The second channel proposes that LOC directly affects a woman’s subjective expectations about the returns to own efforts and investments. The expected monetary returns to participation are higher for internal individuals as they believe in the direct causality between their own efforts and life outcomes. Internal women have higher subjective job-offer arrival rates and higher subjective future income paths (Berger & Haywood, 2016, Caliendo et al., 2015). Hence, they expect higher (current and future) returns to participation and have a steeper budget curve: ∂~wi ∂loci>0. Internal woman thus expect higher utility from availability for market production as their budget constraints allows for higher returns to participation in expected consumption levels. They are, on average, more likely to participate. Monetary returns to LOC Besides these two main mechanisms, the raw difference between internal and external women could also be driven by differences in the objective monetary returns to LOC and thus, indirectly, via different constraints. One potential explanation for this may be positive demand-side responses to an internal LOC, i.e., higher realized wage rates (Heineck & Anger, 2010) which are correctly anticipated by women and thus incorporated into the decision-making independent from the subjective beliefs. Additionally, internal women have been found to select occupations that are less open for flexible employment paths (Cobb- Clark & Tan, 2011). These occupations are likely to be associated with higher future career costs of non-participation and thus higher disincentives for home production. Additionally, LOC might also be correlated with the partners’earnings, captured by Viand thus family income driven by assortative mating or mating probabilities in general (see e.g., Lundberg, 2012). 334 J. Hennecke Nevertheless, as opposed to the first two hypothesized channels, this heterogeneity can be captured by a number of control variables and is discussed in Section 4.2 in more detail. 3 Data and empirical identification Based on these theoretical considerations, the goal of this paper is to empirically analyze the role of LOC in explaining women’s current labor force participation. This is done by using data from the German Socio-Economic Panel (SOEP, 2020). The SOEP is an annual representative household panel that follows a general-purpose approach. It has been studying about 22,000 individuals living in 12,000 households in Germany since 1984. Personal questionnaires are completed by all individuals aged 18 or older. For more information on the SOEP see Goebel et al. (2019). The SOEP contains a measurement of LOC over multiple waves, rich information on current labor-market outcomes and family status as well as the opportunity to connect women to their partners’characteristics if they are surveyed in the same household. Sample Restriction The sample restriction process is intended to create a relatively homogeneous sample of women who could potentially participate in the workforce. Thus, the sample only consists of women in the traditional working age, which is defined as 25 to 65 years, as well as only women who are not in school, academic or vocational education, not already in (early) retirement or in military service. Additionally, only women who live in single-adult or in couple households with or without children are kept. All women in multi-generation households or other unknown household combinations are dropped in order to enable a more straightforward argumentation about intrahousehold decision making. 2 Finally, only women for whom it is possible to observe all the relevant socio-economic control variables are kept. This leaves 70,662 observations for 11,013 women over the period 2000 to 2018 in which women are observed for, on average, 10.35 years. Column (1) in Table A.1 in the Appendix gives an overview of the descriptive statistics of the sample. 3.1 Locus of Control LOC is surveyed within the SOEP in the years 1999, 2005, 2010 and 2015. Respondents are asked how closely a series of 10 statements characterizes their views about the extent to which they influence what happens in life. A four-point Likert scale ranging from 1 (‘applies fully’)to4(‘does not apply’) was used in 1999, while in 2005, 2010 and 2015, responses were measured on a seven-point Likert scale ranging from 1 (‘disagree completely’)to7(‘agree completely’). A list of the items can be found in Table A.2 in the Appendix. 2 LOC might have a direct effect on the household type a woman is living in if e.g., internal women are more prone to leaving the parental household and thus less likely to live in multi-generation households. This might lead to sample selection issues if these household are excluded from the estimation sample. Table S.3 in the Supplementary Material analyzes this association and results in column (1) and (2) indeed show a weak negative connection between the continuous measure of LOC and the probability to live in a multi-generation household. Nevertheless, the estimation results are robust against the inclusion of multigeneration households into the estimation sample (columns (3) and (4)). The independent woman—locus of control and female labor force participation 335 In order to harmonize the scales, the responses from 1999 are reversed and “stretched”. 3 Afterwards, an exploratory factor analysis is conducted jointly for all years in order to investigate the way these items load onto latent factors. The factor analysis clearly indicates one underlying factor with an eigenvalue of 1.78 (with the eigenvalue of a second factor being 0.55). The rotated factor loadings indicate that items 2, 3, 5, 7, 8 and 10 have a positive loading and item 1 has a negative loading on this factor. 4 Items 4, 6 and 9 have relatively low factor loadings and are excluded from the factor prediction. 5 Excluding these three items improves the internal consistency and scale reliability of the resulting factor, as Cronbach’s alpha (Cronbach, 1951) increases from 0.63 to 0.69, which is in line with the findings from Specht et al. (2013). Nevertheless, the estimation results are robust against the inclusion of item 4,6 and 9. Results of the sensitivity check can be found in Section 4.4. In line with the previous literature (see e.g., Piatek & Pinger, 2016), a two-step procedure is used in order to create a continuous and uni-dimensional LOC factor. First, the scores for items 2, 3, 5, 7, 8 and 10 are reversed such that all seven items are increasing in internality. Second, confirmatory factor analysis is used to extract a single factor. This has the advantage that it avoids simply weighting each item equally, as averaging would do, and instead allows the data to determine how each item is weighted in the overall index. Simple averaging of items would risk measurement error and attenuation bias (Piatek & Pinger, 2016). 6 The resulting factor is increasing in internal LOC and its distribution is shown in Fig. 1. In order to fill the observation gaps, LOC is imputed forwards lagged by at least one year, i.e., the LOC observed in 1999 is imputed into the years 2000 to 2005 and so forth. On the basis of the generated and imputed continuous LOC factor variable, a categorical variable is created that splits the continuous LOC in three terciles, in order to identify non-linear relationships. 3.2 Labor force participation Labor force participation (LF) is measured as a binary indicator that indicates a woman’s availability to the labor market. The decision at the extensive margin still is the most prominently discussed decision situation in female labor force participation especially with respect to intrinsic factors such as personality traits. Decisions about participation at the intensive margin are often much more strongly determined from the demand side with, for example, working hours being commonly restricted to specific full-time and part-time options. An additional analysis (in Section 4.3) 3 In line with Specht et al. (2013), this process preserves the relative differences between individuals. The process results in values of 1, 3, 5 or 7 such that a ‘1’on the 1999 four-point scale, for example, becomes a ‘7’on the 2005–2015 seven-point scales. The robustness of results with respect to this is checked in Section 4.4. 4 A scatterplot of the loadings can be found in Fig. S.1 in the Supplementary Material. 5 The exclusion of items 4 and 9 is in line with the literature and supported by Specht et al. (2013). The exclusion of item 6 is specific to this paper as the near-zero loading seems to be driven by the sample of women while earlier studies have found a low but negative loading (see e.g., Caliendo et al., 2019, Preuss and Hennecke, 2018). 6 Sensitivity checks include a re-estimation of the results using this simple index. The results are found to be robust against this variation and can be found in Section 4.4. 336 J. Hennecke new baseline in the following, in order to eradicate parts of the bad control problem. In columns 5 and 6 of Table A.3 potentially omitted information on the industry and occupation type of women in their current or last job, as well as net labor income of the last observed working spell, are added as controls to the model. Although the effect size does go down, especially when wage is controlled for, the effects remain significantly positive. Hence, an effect of LOC on participation probabilities via occupational selection and differences in the expected future costs of nonparticipation can largely be rejected while we do observe a demand-side response to LOC via higher expected wages which leads to higher participation probabilities. Lastly, information on a woman’s partner has to be controlled for in order to rule out assortative mating as a cause for the observed relationship between LOC and labor force participation. Fortunately, the SOEP makes it possible to merge cohabiting women with their partners. Thus, column 8 present the results of the estimation in which the partner’s current net labor income is included as additional control variables for cohabiting women. The results do not change if partner’s net income and LOC are included as control variables, indicating that the results of the main estimation are not driven by assortative mating. 4.3 Labor force activity, working hours and lifetime participation The behavioral implications of LOC on labor force availability have been the center of attention in the theoretical considerations as well as the main part of the empirical analysis. Nevertheless, it is interesting to investigate whether those static behavioral effects actually translate into higher employment probabilities, higher participation on the intensive margin as well as higher average lifetime participation, as these are the variable with the desired positive macro- and microeconomic consequences in the long run. If a higher probability of being available to the market for internal women does not translate into higher employment probabilities, the positive economic implications of LOC are limited by other unobserved factors such as market conditions and frictions. In order to assess the generalizability of the results with respect to the choices made about the participation indicator as described in Section 3.2, three major components of the dependent variable are investigated: (1) the concentration on labor force availability instead of labor force activity, (2) the restriction to the extensive margin as well as (3) the focus on a one-period discrete choice rather than a intertemporal lifetime perspective on labor force participation. The results of these additional estimations can be found in Table 3. Labor Force Activity and Working Hours As a first step, the dependent variable is adjusted such that it only captures labor force activity (“working”) instead of availability. Thus, the indicator is one if a woman is actually employed or selfemployed and zero if she is unemployed or not-working, independent of her intention to work. This alternative definition was neglected in the main part of the empirical analysis as it captures unobserved returns to LOC with respect to employment probabilities and therefore does not concentrate on the behavioral aspects of labor force participation. Column 1 of Table 3give the results of this new indicator while still concentrating on the extensive margin. The results indicate that the behavioral changes are fully The independent woman—locus of control and female labor force participation 343 translated into higher employment probabilities. The effects are considerably stronger than in the main estimations. This is likely due to unobserved returns to LOC in employment probabilities as observed in the negative correlation between LOC and unemployment as seen in Table 1. Having a medium (high) LOC thus on average increases the probability of working by 3.7 (3.9) percentage points. In addition to this, columns 2 and 3 give the estimated marginal effects of LOC on participation indicators at the intensive margin. For the sub-sample of all women who are employed (LF =1 in column 1), the outcome variable in column 2 indicates the actual working hours (contracted hours plus overtime) of a woman whereas the outcome variable in column 3 indicates whether a woman is full-time employed (FT), defined by at least 35 contracted working hours per week. As the goal of the main estimation model was to capture behavioral changes in participation decisions instead of actual labor force activity, which is strongly influences by demand side restrictions such as fixed full- or part-time options for working hours, this continuous measure of participation was neglected in all previous analysis. While a medium LOC positively affects labor force availability as well as participation at the extensive margin, no effects can be identified at the intensive margin. Table 3 Additional Results (Marginal Effects): Outcome Variable - Labor Force Activity and Aggregated Participation Sample: All All Employed Cross-Section—Women 65+a Outcome: Work (E/SE) Working Hours FT Years in LF 25- 65y Years Employed 25- 65y Years % Years % (1) (2) (3) (4) (5) (6) (7) LOC Factor (cont.) 0.021*** 0.188* 0.008* 0.379* 0.009 0.497** 0.012** (0.003) (0.107) (0.004) (0.200) (0.005) (0.198) (0.005) Observations 70,662 50,007 50,007 3810 3810 3810 3810 Full Controls ✓✓ ✓✓✓✓✓ Locus of Control Terciles (Ref.: [LOCmin,LOCP33]) (LOCP33,LOCP66] 0.037*** 0.147 0.010 0.083 −0.000 0.376 0.007 (0.006) (0.226) (0.009) (0.441) (0.012) (0.437) (0.012) (LOCP66,LOCmax] 0.039*** 0.428* 0.016* 1.197** 0.030** 1.455*** 0.036*** (0.007) (0.239) (0.010) (0.476) (0.013) (0.472) (0.013) Observations 70,662 50,007 50,007 3810 3810 3810 3810 Full Controls ✓✓ ✓✓✓✓✓ SOEP, waves 2000–2018, version 36, https://doi.org/10.5684/soep.v35, own calculations Standard Errors in parentheses. *p< 0.1; **p< 0.05; ***p< 0.01. The estimation sample for column (1) is the main estimation sample in line with Table 2, the estimation sample in columns (2) and (3) is a subsample of all women who are employed at time t and the estimation sample. Estimations in columns (4) to (7) refer to a new cross-sectional estimation sample of the first available observation for all women in the age of 65 or older. Panels I and II of the table refer to separate estimation models with (I) including the continuous measure of LOC and (II) including the discrete measure of LOC as the main explanatory variable aLOC is calculated based on the average over all available LOC observations 344 J. Hennecke As opposed to this, having a high LOC does on average increase the amount worked by 0.43 hours and the probability of being full-time employed by 1.6 percentage points. Lifetime Participation Additionally, the lifetime perspective should be considered in order to understand whether this static relationship actually translates into differences for the whole working life due to the potentially important role of path and state dependencies in women’s employment biographies. Thus, in the additional results presented in columns 4 to 7 of Table 3, the accumulated years in the labor force as well as in employment between the age of 25 and 65 are the outcome variables of interest. Using the detailed biographical information available for every SOEP participant, the aggregated time in the labor force is calculated by adding the years a woman spent in employment or registered unemployment during those 40 years. 12 As no biographical information is available on the job-search behavior, the analysis relies on the reported labor force status in order to identify LF. As job-search is likely to be an important determinant of true willingness to participate, it has to be taken into account that this is, therefore, only a rough measure of participation. The cross-sectional estimation sample consists of the first available observation for women in the age of 65 or older. The explanatory variable is a measure for the average LOC over all available observations. The effects are estimated using a linear regression model. Columns 4 and 6 present the results for absolute count of years while columns 5 and 7 present the share of years in LF or employment of the total years labor force statuses are observed for the woman during this time-span. The results indicate a significant positive effect of a high LOC on lifetime labor force availability and activity. Women with a high LOC spent on average approximately 1.2 more years in the labor force (3.0%) and 1.5 more years in employment (3.6%) during this time. 4.4 Robustness checks Omitted Variable Bias As has already been discussed in detail in Section 3.3,a drawback of the study at hand is that it does not rule out endogeneity in LOC due to omitted variables bias. In order to access the impact of additional unobserved factors on the estimation results, the approach proposed by Oster (2019) is applied. This method exploits the assumption that the bias from observed factors provides information about this unobserved bias, as it assumes a certain amount of proportionality between both biases and assesses the movements in coefficients and R-squared. Table 4contains the results of this sensitivity analysis. Comparing Columns (1) and (2) in Table 4reveals that the estimated effect of the continuous LOC Factor on LF decreases from 0.028 in a linear probability model without any control variables, except time and region fixed effects, to 0.011 in the full specification, which includes all sets of control variables except the endogenous variables discussed in Section 4.2. Guided by the rule of thumb provided in Oster (2019), the maximum R2(i.e., the R-squared from a hypothetical regression of the 12 A women is assumed to spend a full year in a certain labor force status if she only reports one spell during a certain year. If she reports multiple spells during one year, she is assumed to have spent an equal share of the year in either spell and consequently the value (1\number of spells) is added to the counter. The independent woman—locus of control and female labor force participation 345 outcome on the treatment and both observed and all unobserved controls) is set to 1.3 times the R2in the fully-controlled model for each of the estimations. The method is based on assumptions about the relative degree of selection on observed and unobserved variables (δ). δ=1 would imply that observed and unobserved factors are equally important in explaining the outcome, while δ> 1 implies a larger impact of unobserved than observed factors. Column (3) contains the identified set of coefficients at δ=1 which is [0.005; 0.011] for the continuous LOC factor and would thus still be positive even if we consider a set of potential unobserved factors which has an equal importance as the already very rich set of control variables in our full specification. Reassuringly, the identified set of coefficients only includes zero if ~ δexceeds 1.73, meaning the unobserved factors would have to be nearly twice as important as the already included observed factors. Similar robustness can also be identified for the binary indicators in Panel II of Table 4. Especially the estimated effect of a medium LOC is very robust with the Table 4 Relative Degree of Selection Test (Oster (2019)) (OLS, Outcome—Moderate or Regular Drinking) Baseline effect (SE), [R2] Controlled effect (~ β) (SE), [R2] Identified Set ~ β;β  ~ δfor β=0 given Rmax (1) (2) (3) (4) LOC Factor (cont.) 0.028*** 0.011*** [0.005, 0.011] 1.73 (0.003) (0.003) [0.023] [0.235] Observations 70,662 70,662 Rmax 0.306 Locus of Control Terciles (Ref.: [LOCmin,LOCP33]) (LOCP33,LOCP66] 0.043*** 0.023*** [0.017, 0.023] 3.57 (0.007) (0.006) (LOCP66,LOCmax] 0.056*** 0.020*** [0.006, 0.021] 1.42 (0.007) (0.006) [0.022] [0.235] Observations 70,662 70,662 Rmax 0.306 Time &Region ✓✓ Socio-Demographics ✘✓ Family ✘✓ Personality ✘✓ SOEP, waves 2000–2018, version 36, https://doi.org/10.5684/soep.v35, own calculations Clustered standard errors in parentheses. *p< 0.1; **p< 0.05; ***p< 0.01. Columns (1) and (2) report coefficients of linear regressions using the continous LOC Factor (in Panel I) or the binary LOC indicators (in Panel II). While column (1) contains the baseline model, which only controls for time and region fixed effects, column (2) includes all controls of the full estimation model in line with Eq. 5. As opposed to the main estimation model, the coefficients in column (1) and (2) report the coefficients of a linear probability model instead of the marginal effects of a logit model. Rmax is set to 1:3~ Rand reported in the bottom row of each panel. Column (3) reports the identified set, which is bounded below by β*at δ=1atRmax and above by ~ β. Column (4) shows the value of ~ δthat would produce β=0 given the values of Rmax 346 J. Hennecke identified set of coefficients only includes zero if the set of potential unobserved factors is over thrice as important as the already included observed factors. Locus of Control As a second important set of sensitivity checks, the construction and imputation of LOC as explanatory variable is tested. Panel 1 of Table A.4 presents the results of four alternative forms of construction of the LOC factor, which have alreadybeenintroducedinSection3. The construction of the factor is robust against variations in the items considered in the factor predictions (columns (3) to (8)) as well as against the use of a simple index instead of the prediction based on the loadings from factor analysis (columns (9) and (10)). Secondly, the timing of the LOC measurement and thus the imputation approach is tested for its impact on the robustness of the estimated results. As Preuss and Hennecke (2018) pointed out, there is a considerable risk of reverse causality or attenuation bias due to temporary measurement errors in the LOC. Using the same data from the SOEP, they found a significant negative short-run effect of exogenous job-loss on LOC for individuals who are still unemployed during the LOC interview. They conclude that this is likely to be driven by temporary state-dependent reporting in the LOC for unemployed individuals even though LOC can be assumed to be stable in the long-run. Due to the fact that employed and non-employed individuals are pooled in the present estimation sample, there might be a risk of biased results due to a measurement bias in LOC, which would, by definition, be greater in the group of non-participating women due to a higher share of non-employed individuals in this group 13 .Inorderto circumvent this measurement problem, two alternative approaches are implemented. Firstly, instead of the forward imputed LOC, a variable which averages all available LOC observations of an individual between 1999 and 2015 is used as the explanatory variable. This approach is likely to reduce the attenuation bias in the LOC due to temporary measurement errors to a minimum. The results of this alternative estimation are presented in columns (1) and (2) in panel II of Table A.4. The estimated effects increase in magnitude and remain statistically significant. It has to be noted that these alternative estimates nevertheless are again at risk of being biased by reverse causality, which is why the forward imputation is still the preferred imputation method. Thus, as a second robustness check, LOC is imputed from the first available LOC observation of an individual in the SOEP. In this alternative imputation LOC from 1999 if available, if not from 2005 and so on in order to go as far back in time as possible. Results in columns (3) and (4) show show that results are also robust against this change in the imputation method. Using the average or first LOC nevertheless does not solve problems with reverse causality if the measurement error is selective, as women who are not employed in t have a higher probability to also be not employed in the periods before and after t. These women thus always report a lower LOC due to their non-activity on the labor market. Therefore, additionally the LOC observation during the closest employment or self-employment spell to tis used. The two conditions for imputing the LOC 13 While in the group of participating women potentially only some of the women, i.e., those who are unemployed, might have a state-bias in their observed LOC, the share is expected to be greater in the group of non-participating women as 100% of women in this sample might be affected by such a state-bias. The independent woman—locus of control and female labor force participation 347 observation from a period t+xor t−xinto tare that (a) LOC has to be observed in that year and (b) the woman is observed to be employed or self-employed in that year. 14 Nevertheless, this approach has one main caveat: by imputing from the closest employment spell, all women who are never observed in (self-)employment are lost. Never being observed in (self-)employment is highly endogenous to the model in line with the argumentation in Section 4.2. Columns (5) and (6) in panel II of Table A.4 thus check the effect of the LOC variable in the baseline model, using only the sample of women for which the LOC variable from the closest employment is observed. As expected, although still positive and significant, the estimated effect is now considerably smaller, indicating a problem with endogeneity in the observability of employment spells. Based on this reduced sample, columns (7) and (8) present the results for the alternative approach of imputation for the LOC factor. When using the reduced sample, the alternative LOC variable slightly increases the estimated effects. Thus, if the main estimations are at risk of being biased, this is likely to be a bias towards zero by attenuation. Lastly, in addition to concerns with respect to measurement problems which might cause attenuation, the effect of the scale changes between 1999 and 2005 on the estimation results on the estimation results have been analyzed. This has been done by restricting the estimation sample to the period 2006–2018 and thus avoiding the use of any imputed values of LOC from 1999. The results in columns (9) and (10) show that estimation results are robust against this change too. 5 Effect heterogeneity The influence of personality on participation can be assumed to crucially depend on the overall size of underlying participation incentives. If monetary and non-monetary incentives for either market or home production are very strong, the power of personality to affect participation probabilities may be comparably low. Thus, the estimated effects are expected to be highly heterogeneous with respect to, for example, overall size of the available household income or the monetary and nonmonetary utility from home production as well as with respect to underlying differences in social norms of working. Consequently, in this section, the heterogeneity of the estimated effects with respect to the family status (i.e., existence of a partner and children in the household) as well as with respect to underlying differences in social norms of working (represented by region of living and cohort indicators) is considered. Nevertheless, a major drawback of this heterogeneity analysis is the endogeneity of these variables and the high likelihood of them being correlated with other potentially important unobserved factors such as for example latent values and norms. The heterogeneity analysis at hand, thus, only provides additional evidence and has to be interpreted with caution. Since not only β2, i.e., the marginal effect of loc, is regarded to be heterogeneous, the heterogeneity is examined using fully separated models for the different 14 Backwards imputation is allowed to avoid problems with sample size. This is based on the assumption that, besides measurement bias in LOC through non-employment, non-employment has no long-term effect on LOC based on the findings in Preuss and Hennecke (2018). 348 J. Hennecke subgroups SGit: PLF it ¼1jSGit ðÞ¼Pβ1þβ2locitnþβ3Xit þβ4Fit þβ5PiþϵitjSGit ðÞ:ð6Þ In order to reduce problems with selection into these sub-groups depending on LOC, LOC is standardized and cut into terciles for each sub-group separately such that women are always compared to women in the same sub-group 15 . Family Status and Children Table 5presents the results for the sub-samples based on family status and existence of biological children under the age of 16. These subgroup analyses correspond to the supposed heterogeneity of the effect of LOC on participation probabilities with respect to underlying monetary and non-monetary incentives and disincentives to work, driven by the existence of partners and children in the household. Participation shares in the sub-samples are reported in the bottom row of the table and already indicate the different levels of incentives for the different groups with incentives for working being especially high in the groups of women without partners in the household (columns 4 and 5), likely due to the absence of a unconditional baseline household income provided by the partner. But also the absence of children in the household increases participation probabilities, likely due to lower monetary and non-monetary incentives for home production such as childcare costs or direct non-monetary utility from spending time with your children. Looking at the estimated average marginal effects for the separate groups, we can see that the effect is, in large part, driven by cohabiting women. Cohabiting women with a medium (high) LOC are, on average, ceteris paribus 2.2–3.5 (2.3–2.8) percentage points more likely to be in the labor force than cohabiting women with a low LOC, depending on whether they have children under 16 in the household (column 6 and 7). The effects differ only marginally between cohabiting women with and without children, with the non-linearity of the LOC effect being stronger for women without children. For non-cohabiting women, the effect is insignificant and close to zero if no children are present in the household. However, in the subgroup of noncohabiting women with children under 16, i.e., single mothers, the effect of a medium LOC is positive and significant on the 10% level. Single mothers with a medium LOC are, on average, ceteris paribus 2.1 percentage points more likely to be in the labor force than single mothers with a low LOC. Nevertheless, a high LOC does not significantly increase the probability of being in the labor force for single mothers. 16 All these results support the theoretical idea that the effect of LOC on participation probabilities strongly interacts with underlying incentives and disincentives to work. If the monetary incentives for market production, such as in the case of single women 15 Nevertheless, the use of within group standardization of LOC (group specific tertile groups), might again restrict absolute comparability of results across groups, especially if distributions differ across groups: Women reporting the same LOC may be in the bottom of one distribution, but in the middle of another. The sensitivity of the results to the within-group definition of LOC has been checked and results do not change if instead an across-group distribution is used. 16 Table S.4 in the Supplementary Material also provides analog estimation results for mothers depending on the age of their children. Women with pre-school and young school children (until the age of 12) exhibit the largest effects (columns 1 to 4). The independent woman—locus of control and female labor force participation 349 without children, 17 already considerably exceed the decision threshold, personality and preferences have no power to affect the participation decision. 18 There is no clear evidence for the theoretical idea that an internal LOC might be associated with a lower participation probability for mothers due to considerations about their own influence on children’s outcomes, but this consideration might be reflected in the non-linearity of the effects. Social Norms of Working In additional to budget constraints, a woman’s decision making might also be constrained by prevailing non-monetary utility from participation such as social norms of working. If, for example, one group of women is exposed to strong social norms for working and another group is exposed to weak social norms of working, even women in the first group who individually gain lower marginal utility from participation (i.e., external women) still have a high probability of participating as the marginal utility from participation is already considerably high. This is also in line Table 5 Heterogeneity Analysis (Marginal Effects): Family Status and Children Baseline All Non-Cohabiting Cohabiting Children under 16 Children under 16 Children under 16 No Yes No Yes No Yes (1) (2) (3) (4) (5) (6) (7) LOC Factor (cont.) 0.012*** 0.011*** 0.014*** 0.004 0.008 0.014*** 0.015*** (0.003) (0.004) (0.003) (0.004) (0.005) (0.004) (0.004) Observations 70,662 36,076 34,586 9170 5382 26,906 29,204 Full Controls ✓✓✓✓✓✓✓ Locus of Control Terciles (Ref.: [LOCmin,LOCP33]) (LOCP33,LOCP66] 0.024*** 0.025*** 0.025*** 0.002 0.021* 0.035*** 0.022** (0.006) (0.007) (0.008) (0.009) (0.012) (0.009) (0.009) (LOCP66,LOCmax] 0.022*** 0.019** 0.025*** 0.012 0.009 0.023** 0.028*** (0.006) (0.009) (0.008) (0.010) (0.014) (0.011) (0.009) Observations 70,662 36,076 34,586 9170 5382 26,906 29,204 Full Controls ✓✓✓✓✓✓✓ LF =1 82.08% 85.59% 78.41% 93.03% 86.88% 83.06% 76.85% SOEP, waves 2000–2018, version 36, https://doi.org/10.5684/soep.v35, own calculations Standard Errors in parentheses. *p< 0.1; **p< 0.05; ***p< 0.01. Panels I and II of the table refer to separate estimation models with (I) including the continuous measure of LOC and (II) including the discrete measure of LOC as the main explanatory variable 17 Consideration about monetary constraints do not fully apply for single mothers with young children. In German law, employment is, among others, not “reasonable”if this employment would, for example, endanger the upbringing of children. As is regulated in §10 SGB II, this applies to children under the age of 3. Hence, these single mothers do have the opportunity to choose home production and receive social transfers as an equivalent to partners income. 18 These findings are supported by an additional heterogeneity analysis with respect to available family income presented in Table S.5 in the Supplementary Material. Family income is approximated by subtracting the reported labor net income as well as individual unemployment insurance payments from the reported net household income. The variable is thus assumed to capture all earnings which are not generated through own labor force participation. 350 J. Hennecke with the idea that, for example, for men the social norms of “being the breadwinner”are expected to be very strong in general and thus independent from their LOC (see e.g., Bertrand et al., 2015, Charles et al., 2018, Killingsworth & Heckman, 1986, Knabe et al., 2016). 19 The same might be true for groups of women who are subject to very strong social norms of working. For them, the harm from staying at home exceeds the gains from participation independent of their personality. As prevailing social norms of working are unobserved and no direct measure for them exists in the data at hand, the analysis relies on three different discontinuities of social working norms already observed in the earlier literature: East and West Germany, urban and rural regions and age cohorts. Table 6pre- sents the results of this heterogeneity analysis. The estimation results again are presented for fully separated models and LOC is predicted and standardized separately within each group. Firstly, heterogeneity can be expected with respect to differences between the eastern and western parts of Germany. Due to the long-term socialist political influence in the former GDR, the east of Germany has a longer tradition of women’s participation in the labor force. 20 As we would nevertheless also expect the rurality of the region to play a role for the prevalence of social working norms, we distinguish between 4 types of regions: West urban, West rural, East urban and East rural 21 . As the direct marginal utility from participation is expected to be higher for eastern German women as well as women in urban areas, the absolute effect of LOC on participation probabilities is expected to be lower. The observation numbers (columns 1 to 4 in bottom panel of Table 6) support this assumption especially with respect to East versus West Germany. The participation probability is with around 80% distinctly lower in the west of Germany than in the east of Germany (approx. 88%) but largely independent from whether the woman lives in a rural or urban area. In line with the assumption the results reveal that the significant positive marginal effect of a medium and a high LOC is mainly observable for women in the rural areas in the west of Germany. The effect for urban regions in the West is still significant but distinctly lower. While the effect for rural regions in the East is, as expected, essentially zero, we do see some effects for a high LOC in urban regions in the East. Nevertheless, these effects are driven by women in Berlin who make up 36% of the women in the East German, urban regions sample 22 . A potentially strongly connected explanation 19 In an additional analysis, presented in Table S.6 in the Supplementary Material, we replicate the results in our main analysis Table 2for the sample of men and find significant positive but distinctly lower effects. 20 The socialist system was characterized by a strong emphasis on the dual-earner/state-carer system of family labor supply, i.e., an extremely high levels of female labor force participation in combination with an extensive system-level organization of family-support structures and child care (see e.g., Braun et al., 1994, Rosenfeld, Trappe & Gornick, 2004). 21 The indicator for rural vs. urban settlement is based on spatial categorization provided by the Federal Institute for Research on Building, Urban Affairs and Spatial Development (BBSR, 2020). Urban settlements include all counties which have at least a population density of 150 inhabitants per km2or include at least one large city with 100.000 inhabitants or more. 22 A more detailed sensitivity analysis with respect to women in Berlin as well as the main estimation results when Berlin is excluded from the sample can be found in Table S.7 in the Supplementary Material. Estimation results for the main analysis are robust to this exclusion. The independent woman—locus of control and female labor force participation 351 for the regional differences could be underlying differences in child care availability in East and West as well as rural and urban regions, which could also explain the specific case of Berlin. Non-monetary and monetary factors can, thus, not be cleanly separated. Additionally, based on the continuous decrease in the importance of traditional gender roles over time in almost all modern Western societies (see e.g., Goldin, 2006), women in later cohorts are assumed to be more affected by a generalized social pressure to be economically independent from external forces than women of earlier cohorts (Heim, 2007). For the former, the marginal utility from participation can be assumed to be higher than for the latter. They might therefore have a higher participation probability independent from LOC. Thus, columns 5 to 7 of Table 6present the results of the estimations. The cutoffs for the manifestations of the birth cohort indicator “early”, “middle”and “late”were generated based on the terciles of year of birth in the full estimation sample, i.e., P(33) =1958 and P(66) =1968, in order to obtain groups of approximately similar size. The results indicate a strong heterogeneity of the effect with respect to cohort. The distinct marginal effects of a medium and high LOC on participation probabilities can mainly be observed for women from the early cohorts, i.e., born before 1958 (column 5). The effect is distinctly lower for both the women in the medium as well as in the latest cohorts. Table 6 Heterogeneity Analysis: Social Working Norms (Sample: All) Region Cohorta West East Urban Rural Urban Rural Early < ′58 Middle ′58–′68 Late > ′68 (1) (2) (3) (4) (5) (6) (7) LOC Factor (cont.) 0.009** 0.027*** 0.014** 0.004 0.015*** 0.013*** 0.006 (0.004) (0.007) (0.006) (0.005) (0.005) (0.004) (0.004) Observations 38,882 13,228 7550 11,002 23,786 26,654 20,217 Full Controls ✓✓✓✓✓ ✓ ✓ Locus of Control Terciles (Ref.: [LOCmin,LOCP33]) (LOCP33,LOCP66] 0.021*** 0.038*** 0.010 0.010 0.035*** 0.023*** 0.010 (0.008) (0.014) (0.012) (0.010) (0.011) (0.008) (0.008) (LOCP66,LOCmax] 0.016* 0.050*** 0.030** 0.000 0.028** 0.018* 0.014* (0.009) (0.015) (0.013) (0.012) (0.013) (0.009) (0.008) Observations 38,882 13,228 7550 11,002 23,786 26,654 20,217 Full Controls ✓✓✓✓✓ ✓ ✓ LF =1 80.10% 79.83% 87.55% 88.01% 78.68% 86.58% 80.15% SOEP, waves 2000–2018, version 36, https://doi.org/10.5684/soep.v35, own calculations Standard Errors in parentheses. *p< 0.1; **p< 0.05; ***p< 0.01. Panels I and II of the table refer to separate estimation models with (I) including the continuous measure of LOC and (II) including the discrete measure of LOC as the main explanatory variable. In order to achieve convergence for all the small samples, all estimation models do not include federal state indicators as well as interview characteristics aCohort Cutoffs based on terciles of distribution in birth years: Early—born before 1958, Middle—born 1958–1968, Late—born after 1968 352 J. Hennecke