Our empirical analysis focuses on the following two dependent variables: the probability to save and net worth. For the first model, we compare households that are saving with those that are not saving. Because a binary variable has been observed in (almost) all panel waves, we can exploit the variance within households over time with a fixed effects (FE) regression. The main advantage of this method is that it excludes any unobserved heterogeneity bias. Even though we have a dichotomous dependent variable, we estimate a linear probability model.Footnote 4 As main independent variables, we include the number of dependent children living in the household by age groups (0–4 years, 5–9 years, 10–14 years or 15–24 years). Children older than 15 years are considered as dependent if they are in education and do not work full-time. In addition, we include a binary variable for planning/wanting a child to test whether anticipating a child increases motivation for saving. This variable has not been collected in the German SOEP, however. To test the mediating effect of income on savings, we include household income (yearly earnings adjusted to household size)Footnote 5 and working hours (mean working hours for couples). We control for the following variables that have been revealed to be important in previous studies (Finke and Pierce 2006; Pericoli and Ventura 2011; Vespa and Painter II 2011): age and its squared term (of the household head), civil status, years of cohabitation of the couples and home ownership. Despite its possible endogeneity, we also include an indicator about home ownership because repayment of a mortgage might not be considered as savings and because home ownership might increase the money available for non-housing consumption.
The main advantage of the FE models is that they capture the causal effect of dependent children on the probability to save. As these models have two important limitations, we complement the analysis with an OLS regression on net worth. First, FE models can only analyse changes that occur within the duration of the panel (e.g., maximal until children are 25 in the SOEP). Accordingly, FE models measure the impact of having dependent children in the household compared to the situation of no dependent children in the household. Second, the dependent variable indicates the presence of savings but not the amount saved (this information is available only for Germany). The OLS regression can capture long-term effects of children (once children left the household) on wealth accumulation and will reveal the size of the effect. Net worth is defined as the sum of all assets minus the level of accumulated debts. For couples, we split the amount of household wealth in half. As each country is analysed separately, we use national currencies and adjust for inflation. As a method of analysis, we use (pooled) linear regressions.Footnote 6 To address individuals without wealth or debts and to limit the influence of extremely high values, we apply an inverse hyperbolic sine transformation (hereafter IHS) on total net worth (see Friedline et al. 2015 for details). Because we are interested in the effect of children on wealth accumulation in the long term, we need to consider children irrespective of their age and of whether they live in the household. Following Dockery and Bawa (2015), we compute a variable that we call child-years. This variable multiplies the number of children by their age with a maximum of 18 years per child. Lacking more precise information, the maximum of 18 is set as the average age for independency. This maximum considers that children pursuing professional training might become independent before and that children enrolled in university might finish their education later. As in the FE model, we estimate a separate model controlling for income and labour supply. Income refers here to permanent household income, which we define as the average of all available previous earnings and pensions. In addition to the variables included in the FE model (age, civil status, years of cohabitation and home ownership), we control for variables’ stable characteristics. The educational level (three levels with the highest educational level of the couple) is included as a proxy for wiser choices in saving and investing behaviours. Living in a city centre is included because it might imply not only higher living costs but also higher property values. We also take into account the number of siblings (in Switzerland the presence of siblings) and a measure for parental socio-economic status to capture possible effects of inheritances on the accumulation of wealth over the life-course. Other control variables are country specific. In Australia, we include a binary variable for the English mother tongue because non-native speakers find difficulties in terms of integration, whereas we identify foreign-born individuals in Switzerland and in Germany. In Switzerland, we include a variable for the linguistic region, and in Germany we distinguish the Western from the Eastern part. For age, nationality, siblings and parental socio-economic status of couple households, we consider the information provided by the main earner.
When comparing models of different countries and different surveys, we have to pay attention to different sample sizes and differences in the definition of the variables. We commented on these differences whenever necessary.