Data from the UK Household Longitudinal Study, or “Understanding Society”, (UKHLS) is used. This is a large household panel survey of around 40,000 households in the United Kingdom (2013). The sample is restricted to respondents who answered in the first two waves (2009–2010 and 2010–2011) as some mediators were only asked in the first wave, and all outcomes were measured in the second wave. The initial response rate was 81.8% and further attrition at wave 2 was 22.7% (Lynn et al. 2012). All outcomes are weighted to account for this attrition. The sample consists of 3,965 respondents. They were 16–30 years old in the first wave and not in full-time education in the second wave. Literature on scarring shows that the labour market experiences during this early period can have long-lasting effects on the later career (Gregg and Tominey 2005; Mavromaras et al. 2013).
The UKHLS has not yet been used for the study of the effects of paternal worklessness as most British studies on this topic used the cohort studies. The UKHLS is used here because it contains information on variables concerning mental health and well-being, which allow the study of this mediating mechanism. It also allows richer information on the type of employment through prospective questions about job characteristics, as compared to retrospective employment histories. This study complements work using the cohort studies, such as Mavromaras et al. (2013).
In addition to the UKHLS, the quarterly Labour Force Survey (LFS) for the United Kingdom is used from 2002 through to 2010. The LFS is a nationally representative sample in the UK with about 60,000 households, maintained by the Office for National Statistics (ONS). It is used to calculate median hourly wages for different groups.
The father’s working status when the child is 14 is measured through three categories depending on whether the father worked and in what 3-digit occupation he worked at age 14 if he was employed. These occupations are divided in two groups. Based on the LFS, the weighted median hourly wage by occupation is calculated and ranked quarterly. These rankings are averaged between 2002 and 2010. Occupations with an average rank below the first quartile are classified as low-paying since the median wage in that occupation was among the lowest 25%. Of the 3,965 respondents, 472 (12%) reported having a father who did not work at age 14 and 949 (24%) had a father who was working in an occupation within the lowest quartile of median wage. The results are found to be robust to changes in the threshold of a low wage occupation from the lowest quartile to the lowest half and the lowest decile.
In order to estimate the effect of having a father who did not work at age 14, the counterfactual of their outcomes had their father worked is needed. The difference between the observed outcome and these potential outcomes for children of workless fathers is the average effect of treatment for the treated (ATT) (Schafer and Kang 2008). These potential outcomes can be estimated by predicting the outcome for children whose fathers did not work using the prediction equation from an appropriate regression model for the relevant outcome in the control group (Rubin 1979; Schafer and Kang 2008). This prediction is the estimate of the potential outcome. The difference between the observed outcome for the treated group and their estimated potential outcomes is tested using a paired sample T-test. Equation 1 presents the calculation of the average treatment effect, with T indicating treatment and Y indicating the outcome for individual ‘i’. Ŷ
0 indicates the predicted outcome for the treated group based on the equation estimated in the control group (T = 0). The control group are those whose fathers worked in lower paying occupations. Restricting the control group means that the conditions while growing up of those whose fathers did not work are not as big as when comparing them to everyone.
This method allows for a non-additive treatment effect and for possibly different returns to other characteristics, whereas including an indicator variable for paternal worklessness in a linear regression might not capture the full effect.
3.1 Labour market outcomes
The effect of paternal worklessness is estimated on eight different outcomes. The sample varies depending on whether employment probability or job characteristics are the outcome. The final sample for employment probability consists of 472 children whose fathers did not work and a control group of 856 young adults whose fathers worked in a lower paying occupation. When studying job characteristics, the sample consists of 259 young adults whose fathers did not work and a control group of 622 young adults.
The first outcome is the young adult’s employment probability. The counterfactual is estimated through a binary logistic regression. Respondents work if they had paid work in the last week or if they had a paid job despite not working in the last week. All other cases are classified as out of work, and respondents in full-time education are not included. 3,019 (76.1%) of the respondents in the full sample were working. This includes the self-employed. Since a father and child can share many characteristics that make them both more likely to be employed or not, the following control variables are included. First of all, the respondent’s gender, age and highest obtained educational qualification are controlled for. Whether the respondent is white or non-white and whether the respondent speaks English as a native language are also taken into account. Variables indicating whether the respondent is cohabitating or married and whether (s)he has children are included as this may influence labour supply. Having poor health was controlled for as health is related to the transmission of socio-economic status (Bianchi et al. 2005; Smith 2004). Variables measuring how often the respondent sees their father and whether the child lived with the father at age 16 are included to capture the strength of the relationship between father and child. To account for the general employment situation when the child was 14 years old, the UK-wide unemployment rate in the year the child was 14 years old, acquired from the OECD (http://stats.oecd.org, accessed on 31/04/2013), was included as this could influence the children’s attitude towards unemployment (Ochsen and Welsch 2011). Father’s current age and whether the father and mother had a higher educational degree are included since parental education may influence the child’s labour market outcomes (Andersen 2011).
Hours spent working, on average, each week is another dependent variable capturing labour supply. This outcome is measured through ordinary least squares (OLS) with the same control variables as when estimating the probability of being employed.
The final indicator of labour supply is a dummy variable for working part-time. This is estimated using a binary logistic regression with the same control variables as when estimating the probability of employment.
Working on a fixed-term contract may increase job insecurity and is therefore an important part of the quality of the job. The potential outcome is estimated using a binary logistic regression on a dummy variable using all control variables as above, with the exception of whether the respondent cohabitates or has children. Socio-demographic background, family background and relation to father as well as unemployment rate at age 14 are still included.
In order to assess the quality of employment, a dummy indicating whether the respondent’s earnings are lower than those of his/her peers is used. This relative wage can be an important indicator of job quality. It compares the individual wage to an appropriate peer group which the person might take as a comparison as well. The median gross hourly wage is calculated by age category (16–19; 20–25; 26–30), gender and highest educational qualification. This is calculated in the LFS from 2009 to 2010, weighted appropriately. A dummy variable indicates that the respondent’s gross hourly wage, calculated from the UKHLS, lies below the nationally representative median hourly wage for people of similar age, gender and educational qualifications. The model is estimated through a binary logistic regression on all controls except cohabitating and being a parent. These two demographic variables are not thought to be important in determining the relative state of hourly wage. Working on a fixed-term contract and working part-time are also included as controls as someone’s position in the wage distribution may depend on their type of contract.
The respondent’s relative wage position was also calculated with regards to the hourly wage by age category, gender and 3-digit occupation, again using the LFS. This indicates that the respondent has a wage in the lower half of earnings compared to people of the same age and gender who work in the same occupation. The counterfactual is estimated similarly to the wage position given age, gender and education.
The logarithmic transformation of the child’s monthly labour market outcomes serves as a straight-forward measure of labour market success. It relates directly to the financial dimension of job quality (Kalleberg 1977). The counterfactual labour market income for children whose fathers did not work is estimated through an OLS regression of the logarithm of gross monthly labour market income on all control variables that are used in estimating whether someone works on a fixed-term contract, with the addition of the average hours worked per week.
The final labour market outcome is the self-reported job satisfaction of working that respondents indicate about how they experience their work (Kalleberg 1977). Respondents in the UKHLS are asked how satisfied they are with their job and can respond from 1, completely dissatisfied, to 7, completely satisfied. If respondents reported to be somewhat dissatisfied (3) or less, this is classified as being dissatisfied on a dummy variable. The counterfactual job satisfaction is estimated using a binary logistic regression using all controls used for monthly wage with the inclusion of all other labour market outcomes. This variable complements the more objective job characteristics by adding the respondent’s evaluation of the job. It may indicate different expectations of a job and therefore different evaluations of the available work conditions on average.
Ideally the analyses should be carried out separately for men and women. This strongly reduces the sample however, which is why gender is only controlled for. As a robustness test, the analyses were separated by gender, showing small differences.
Paternal worklessness is expected to affect a young adult’s probability of employment at least partly through some other mediating variables. In this analysis, only employment probability is estimated as an outcome. Mental health, wellbeing and attitudes were tested for their role in the transmission of worklessness.
These concepts mediate the effect of father’s worklessness if they are influenced by father’s worklessness and in turn affect the child’s employment probability, when controlling for father’s worklessness (Mackinnon and Dwyer 1993). The total effect of father’s worklessness on the probability of employment is decomposed into a direct effect and an indirect effect. This indirect effect is the part that is accounted for by the mediator. This decomposition is not straight-forward when it involves binary outcomes (Breen et al. 2013).
In this paper, an extension of the method explained above using counterfactuals is applied. First of all, the effect of paternal worklessness on the mediating concept is estimated as the difference between the average value of that variable in the treatment group and the counterfactual based on a prediction equation in the control group. This difference indicates the extent to which growing up with a workless father affects that mediator.
We are interested in whether these mediators explain some of the effect of growing up with a workless father on the employment probability. To do this the average proportion of employment in the treatment group is compared to two counterfactuals. The first counterfactual is estimated without accounting for the value of the mediator. This indicates the total effect. A second counterfactual is based on a prediction equation in the control group where the mediator is included. The effect of paternal worklessness is then calculated based on the two counterfactuals. The difference between the effect of paternal worklessness when taking the mediating concept on board and ignoring it indicates the role played by that concept. The relative difference between the two indicates the percentage of the total effect due to an indirect effect through the mediator.
The first mediator is the respondent’s psychological wellbeing. It is measured through two dummy variables (Warr 1990). The first one indicates the respondent scores in the top quartile of the general health questionnaire (GHQ). This is a validated scale for mental health status, where a higher score indicates higher probability of mental problems (Goldberg et al. 1997). Another dummy indicates that the respondent felt completely, mostly or somewhat dissatisfied with life in general. The correlation between the two dummies is 0.31 and significant at p < 0.05, which is not too high.
The second mediator consists of attitudes or non-cognitive skills. Seven indicators measure this. One is a factor built from seven items that indicate a positive outlook on life and self-confidence. These items are: ‘feeling optimistic about the future’; ‘feeling useful’; ‘feeling relaxed’; ‘dealing with problems well’; ‘thinking clearly’; ‘feeling close to others’ and ‘able to make up own mind’. The scale has a Cronbach’s alpha of 0.86, and a higher score indicates a more positive outlook. Sense of control is measured through three dummy variables that indicate that someone feels moderate or strong powerlessness regarding life or occurrences at home and whether the respondent feels overwhelmed with demands (Armstrong 2012; Groves 2005). These dummies correlate at 0.38 at most, which is not too high. Attitudes towards risk and trusting people can be influenced by parental experiences and influence economic outcomes (Dohmen et al. 2012). A dummy indicating the respondent does not believe most people can be trusted and two variables ranging from 0 to 10 capturing whether the respondent is prepared to trust strangers and prepared to take risks in general are included. These three variables correlate at 0.38 at most.
To test whether respondents whose fathers did not work experience worklessness differently than their peers whose fathers worked, the association of being out of work with overall wellbeing is studied. If young adults whose fathers did not work experience worklessness as less negative, a weaker association between being out of work and dissatisfaction with overall life is expected for these respondents.
Being dissatisfied with life is logistically regressed on the control variables used for the employment equation. The respondent’s employment status in both waves, the father’s employment status when the father was 14 years old, and the interaction of the respondent’s employment status in the second wave and the father’s employment status are also included. If being out of work is experienced differently in terms of dissatisfaction with life by respondents depending on their father’s employment status while growing up, the interaction term will be significant.
3.3 Missing observations
There are many missing observations among these variables, which is problematic as the sample is quite small. To deal with the missing data, 50 multiple imputations through chained equations are used (Royston and White 2011). All control variables as well as the mediators and labour market outcomes in waves 1 and 2 are used in the imputation model. Multiple imputation assumes that the data are missing at random, conditional on all variables that are used in the imputation model. This is superior to a complete cases analyses if responses are not missing completely at random (Enders 2010).