\citep*{Feingold1994}\citep{Heckman_2006}

Results

Decision

For our analysis on the decision we restricted our sample to individuals finishing studys at the age of 35 or younger. We aülüllrgue that otherwise, the condition to move out before graduating, becomes otherwise irrelevant.
A very intuitive and obvious result is the age of completing education. The later young adults complete tertiary education the higher propensity they show to not live at home at the time of graduation.
Also we observe trends as they were described in the mentioned literature about patterns appearing across all countries. Throughout all sample sizes and after controlling for all possible covariates we find a significant impact of the birth year for people born after 1979. For children born in the nineties this impact gets even larger, stays highly significant and is completely in line with trends in recent history, as described in \cite{Holdsworth2000}.   
The gender variable has throughout all regressions a positive effect on the probability. This is in line with the findings of \citep{Billari2001} across all European countries.  The insignificance is due to the correlation of gender with personality traits, which will be discussed later.
For the second regression, we include the social background of each individual. This data is missing only for 78 individuals, so we only face a minor reduction. The variables of the first regressions stay, while the only one of significant effect added is the education of the mother. It stays significant except of one regression, which we not gonna grant any further interpretation. 
As Locus of Control, Willingness to take Risk and the Big Five were first measured in the 2000s we exclude the seventies dummy variable for these regressions, because none of the students at time of measuring was born before 1970.  Otherwise we would our analysis would suffer a dummy variable trap. Furthermore our sample consists after including the Big Five of less than half of the original sample, precisely 546 . Including locus of control and willingness to take risk we end up with minimum sample size of 485 individuals. Due to this restriction our further interpretations have to be considered with caution. But on the other hand, this reduced sample possibly leaves us with very constraints possibilities to find significant effects. Due to the questioning, the treatment and the control group only consist of individuals finishing tertiary education. As non-cognitive skills, like the Big Five and Locus of Control play a major role in education decisions \citep{Heckman_2006}, therefore our individuals have a small variance in these skills. Also Willingness to take Risk is strongly correlated to social background and cognitive skills \citep[see][]{Dohmen_2010}. Therefore to find significant parameters for the decision, the regression needs to a have the smallest omitted variable bias as possible. We see this for the gender variable pretty clear in our regression, as the gender variable gets insignificant depending on the control variables we use. This is in line with the literature stating a correlation between gender and the Big Five \citep{Weisberg2011}, Locus of Control \cite{Feingold1994} and Willingness to take Risk \citep{Dohmen2010}.  The Education of the mother also gets significant depending on the variables included. As before we will only interpret the significance level in our last regression. For household income we can interpret this finding even further. As mentioned before it shouldn't play a major role in the leaving decision and it only gets significant once we control for all personal and personal heterogeneities.  It is then positively significant, which is in line with \citep*{Laferrere2004}.  For the migration background, we can again approve the hypotheses of the strong German welfare system, as it never gets significant and therefore doesn't seem to play a major role in the decision to leave.
The only facet of the Big Five driving significant the decision to move out is Agreeableness. Again this only becomes significant, after including all explanatory variables, which are able for us to control for. This result seems to be intuitive, as a high score indicates a social compatibility, which is key to adapt in new living arrangements. Also we would assume further heterogeneiteis driving the decision, but due to the small varieties in these measure due to the selection of college students, we cannot identify further factors. 

Wage Regression

We run several cross-section and panel regressions in order to see if leaving your parental household during tertiary education influences an individual's earning .  In both settings we start with a standard Mincer-regression in which we include our decision dummy \cite{jacob1974,Mincer_1958}. Subsequently we augment the standard regression with measures for cognitive and non cognitive skills. 

Cross Section

With exception of the first two columns we include several control variables like the state of residence, social background information and  information regarding the field in which the individual is employed.
As expected the standard mincer factors, years of education, experience and squared experience have sigvaficant effects throughout every of our regressions. This is conform with the impressive importance of the mincer equation in the human capital literature \citep{Heckman_2003} .
Even our decision dummy shows positive significance on the 10% level in the first expansion of the standard mincer regression framework. Unfortunately this effect vanishes as soon as we include the control variables. We suppose that two main reasons drive this results.
First we assume there is a selection bias, because the decision to leave the parental household is associated with a following up process by the GSOEP.  This makes it potentially more unlikely for an  individual to stay in the GSOEP, if she left home compared to co-residence during tertiary education. Additional our identification is based on the exact reporting of the residence status. As in Germany forming a new household is associated with administration costs, like broadcast contribution or other possible monetary disadvantages, our sample could be exposed to a kind of Social Desirability Bias. That is an incentive to not report, if they already moved out, because it would contradict their official deceleration. Following these arguments our sample could exist of a smaller share of people leaving the household.
Second there is an identification bias. As written in chapter 4, we could only identify for about 10 percent of  all individuals, who obtained tertiary education, whether they left their parental household or not. This is due to the fragmentary data from the BIOEDU data set, which left us creating the dummy variable only for individuals followed through their whole adolescence.  Therefore it seems to be likely that our treatment variable doesn't reflect all the individuals left home before graduating. 
The results regarding the personality traits summarized by the big fives are ambiguous. Only the coefficients for agreeableness and neuroticism are significant on the 1% level through all regressions in which the big fives are included. We find a negative impact on wages for both factors.  An increase of one standard deviation in Agreeableness and Neuroticism leads to an decrease in wages of  3%, respectively 2.8% (see Table columns). This seems to be in line with the previous literature's results.  So the reverse of Neuroticism, emotionally stability has a positive impact on wages in   \citealt{Heineck2011,Judge1999,Mueller_2006,Nyhus2005} and  \citealt{Boudreau2001}., whereas agreeableness is associated with negative impact on wages in  \citealp{Mueller_2006,Boudreau_2001} and \citealt{Heineck2010}.  Additionally Conscientiousness is significant at the 10% level in the regression where we expand the mincer framework with the big five factors, even when we add the control variables. However Conscientiousness becomes non significant as soon as we integrate more personality related variables like locus of control or willingness to take risk to our regressions. We argue that the mixed results for the impact of Openness on wages in the economic literature  allows to  more or less ignore the non significance of Openness ( \citealt{Mueller_2006,O_Connell_2011,Heineck2008,Heineck2010,Heineck2011}).  Especially since ( ) state that the strong correlation between Openness and IQ leads to a upward bias. Nonetheless our results for Conscientiousness and Extraversion don't coincide with the significant effects found for both factors in \citealt*{Heineck2010} and \citealt{Judge1999} . Only \citealt*{Nyhus2005} and \citealt{Heineck2011} state non significance for Conscientiousness respectively  Extraversion. 
In all our regressions that include Locus of Control as a dependent variable it shows as highly significant on the 1% level. An increase of one unit in Locus of Control leads to an increase in wage by 6.44% ( see table column).  This seems to be conform with  findings by \citealt{Cebi_2007} and  \citealt*{Heineck_2010}.  
In contrast to Locus of control, the effect of the self reported Willingness to take Risk shows only significantly different from zero as long as the big five are not included in the regression.
Last but not least we  included last grades in school in math, German and foreign language as a proxy for cognitive ability. Note that the implementation of this proxy for cognitive skills shrinks the sample size dramatically. Nonetheless we find that only the last grades in math have a significantly different effect from zero. Because of the German grading system that goes from 1 to 6 where 1 indicates the best achievable grade the coefficient is negative. A decline of one level , indicated by an increase in the related  variable by one unit, decreases wages on average by 4.35%. 

Panel Regression

Unfortunately it would be to restrictive to construct a balanced panel from the 2005, 2010 and 2015 wave because that would lead to a immense decrease in the number of individuals in our group of interest.  Additionally an unbalanced panel poses the risk of an sample selection bias, because the response rate could be dependent on an individuals characteristics. Therefore we have to abstain from a fixed effects estimation approach.  We try to compensate for this by including random effects to account for unobservable heterogeneity and time fixed effects in our regressions.  Note that the random effects approach relies on the assumption that the individual specific effect is independent from all independent variables. We argue that the variety of variables that we include in our regression represent a considerably amount of an individual's characteristics. Following this argument we assume that the random effects estimator is consistent. 
The results from the panel regressions confirm the cross section results. As before all Mincer coefficients are significantly different from zero on the 1% level throughout all regressions.  In the pooled OLS regressions in which we excluded the control variables our decision dummy shows significant on the 1% level.  Nonetheless the effect disappears when adding family background, state of residence and  fields of industry as control variables.  As mentioned before this could be associated with the different biases we were facing during the identification process.
However when we include all control variables, the locus of control,  own willingness to take risk and the big five in a random effects regression the coefficient for our treatment dummy shows significantly different from zero on the 1% level.  This means that an individuals  that are part of our treatment group earn on average a 7.06% higher wage (see Table 3 column 7).
The results for the big five seem to be accord with the cross section analysis. As before the coefficients of Agreeableness and Neuroticism are significant on the 1% level throughout every regression. 
A increase of one standard deviation in Agreeableness  and Neuroticism leads to an decrease in wages of 2.75% respectively 2.04%.  As mentioned before this seems to  be in line with previous literatur's results \cite{Boudreau2001,Mueller2006,Judge1999,Heineck2011}.  Unlike in the cross section regressions,  Conscientiousness has a positive significant effect  on the 5% respectively the 10% level even if locus of control and own willingness to take risk is included.  It becomes insignificant when we include last grades in school as a proxy for cognitive skills.
In contrast to the cross section regressions, own willingness to take risk has also an significant effect on the 1% level even if the big five are included. An increase of one level in own willingness to take risk increases an individual's  wage on average by 0.7%.
Including last grades in math, German and foreign language decreases the number of observations from 15519 to 8102.  Nonetheless the patterns look similar to the previous regressions. Last grades in math are still significant on the 1% level, grades in German are significant on the 10% level but only in the pooled OLS regression whereas foreign language skills doesn't show significantly different from zero at all.