Employment, Income and Divorce – What Can We Learn from Causal Analysis?

Daniel Brüggmann, Hertie School of Governance

In 2002 the male breadwinner model is still the predominant model in West Germany. A large share of married women is either non-employed or at most marginal or part-time employed and therefore dependent on the household income. Consequently, these women experience a dramatic short-term drop in household income when their marriage ends. This paper analyses the consequences of divorce by modeling a quasi-experimental design and estimates [1] the extend to what women adjust their labor market participation and [2] to what extend this monetarizes in earned income. We analyze the effect for the first divorce of all West German women with at least five years of marriage and who are aged between 25 to 55.

Results, for all West German women in the sample, show that a divorce significantly encourages women to participate in the labor market. The preferred type of employment is thereby regular employment (employment with social security contribution) as opposed to marginal employment. The increase in employment rates leads to an increase in the average income following divorce. Nevertheless, once we control for labor market participation, as is necessary since income is only observed for those in the labor market, we do not find evidence that divorced women work or earn more than employed married women.

We are also able to identify strong heterogeneous effects among divorced women. West German women with no labor market participation in the five years prior union dissolution respond stronger to divorce than women with labor market participation prior dissolution.

Our results are derived from weighted Difference-in-Difference estimations and were scrutinized by robustness checks for variations in covariates and sensitivity analyses for deviations from underlying model assumptions.

Divorce rates haveincreased in most industrialized societies since the 1960s. In the EU-25, forexample, the crude divorce rate per 1000 was 0.6 in 1960. This figure hadincreased to 1.5 in 1980 and to 1.9 in 2000 (Eurostat, 2006/07). Consequently,a large body has amassed that examined the impact of divorce and separation onearnings and labor market behavior.

Latest research hasshown that women in Germany experience a dramatic short-term drop in equalizedhousehold income of approximately 26% in the year following divorce. Governmenttaxes and transfers reduce this drop to -17% (Hauser et al. 2016). Theequalized household income of men instead increases by 4% before taxes andtransfers and it drops to -4% once they are taken into account.

In this paper, we addto the previous literature by using longitudinal register data to examine theconsequence of divorce on labor income and employment participation in Germanyfor the year 2002. We go beyond prior research by providing estimates for aspecific year and investigate how divorce affects employment participation inregular employment (employment with social security contribution) and marginalemployment, net of any labor market conditions, seasonal aspects and policychanges. Research so far was mostly limited by low case numbers and in order tocircumvent such constraints scholars often combined observed divorces formultiple years. Hence, these estimated consequences are averaged economicconsequences over a specific time window (mostly for some years up to adecade). The shortcoming of such studies is that results are often severelyinfluenced by reforms simultaneous taken place as well as different conditionsin the labor market throughout the research window. Such reforms might affectthe labor market behavior of divorced and married women in different ways andthus, the economic consequences. Economic consequences for a divorce in aspecific year instead come much closer to the true divorce effect.

Data for thisinvestigation comes from the Statutory German Pension System. For our purpose we combine therecords of the Sample of Active Pension Accounts(Versicherungskontenstichprobe) with the records of the Pension RightsAdjustments Statistic (Versorgungsausgleichsstatistik). This is a novel becauseit combines for the first time longitudinal data about the individualemployment history with the individual data about marriages, marriage duration,divorces and the individual financial compensation due to pension rightsadjustments. We concentrate only on the first divorce and neglect higher ordermarriages and divorces (because these women have a greater labor marketattachment). Due to data limitations it is not possible to control forre-partnering. Re-partnering is an effective coping strategy and research hasshown that women living with a new partner dispose of higher incomes (Regt,Mortelmans, Marynissen 2012) because a new partner enhances the economicsituation and substitutes for the income of the former partner.

The analysis is doneunder the framework of matching at the propensity score. With non-experimentaldata this requires special care because we cannot rule out differences onunobservables that are simultaneously related to divorce and the outcomemeasure. Matching deals with the selection process by constructing a comparisongroup of individuals with observable characteristics similar to the treated.The challenge of matching is to ensure that the “correct” set of observables isbeing used so that the observations of non-treated are what the observations oftreated would be had they not divorced. For that purpose, matching requiresabundant good quality data to be at all meaningful. (Blundell, Dias 2002)

We complement thepropensity score matching method by “Difference-in-Differences”. Combining bothapproaches allows us to control for selection on observable covariates as wellas controlling for selection on time-constant unobservable confounders. Bothapproaches combined will significantly improve the quality of non-experimentalevaluation results and will increase the validity of our estimated treatmenteffects. Difference-in-Differences is usually implemented by comparing thedifference in outcomes before and after the treatment for the eligible groupwith the before and after contrast for the comparison group. In the absence ofa randomized experiment and under certain conditions, this approach can be usedto recover the average effect of the treatment on those individuals enteredinto the treatment. It does this by removing time-constant unobservable individualeffects and common macro or policy effects. The weakness is if the macro orpolicy effect has a differential impact across the two groups (Blundell, Dias2002).

All estimates are alsoscrutinized by extensive robustness checks and sensitivity analysis. Robustnesschecks are important in order to check estimates against variations in chosencovariates. We select covariates which are either derived from machine learningor data driven procedures. Since robustness checks do not address the problemof deviations from underlying model assumptions we conduct also varioussensitivity analyses. Sensitivity analyses explicitly model the presence ofunobserved covariates and thus, deviations from the “no hidden bias”assumption.

Preliminary Results:





Presented in Session 1233: Posters