robust measures of intergenerational poverty …the impact of poverty during childhood on...
TRANSCRIPT
Robust measures of intergenerational poverty transmission
for Europe: the role of education∗
Luna Bellani§
University of Konstanz
Michela Bia‡
LISER
This Version: June 2015
(This draft compiled: June 8, 2015)
Preliminary version. Please do not quote without permission
Abstract
This paper examines the causal channels through which growing up poor affects the individual’s economic
outcomes as an adult. We contribute to the growing literature on intergenerational transmission, on the one
end providing a quantitative measurement of the causal effects of poverty in childhood, applying a potential
outcome approach to this question and implementing a series of robustness checks. On the other end, analyzing
the impact of the interplay between growing up in poverty and individual human capital accumulation on the
children outcomes later in life. The analysis is based on the module on intergenerational transmission of 2011
of the EU-SILC data, where retrospective questions about parental characteristics (such as education, age,
occupation) were asked. We find that even after controlling for possible unobserved confounders, e.g. child
ability, being poor in childhood significantly decreases the level of income in adulthood, increasing the average
probability of being poor. Moreover, our results reveal a significant role of human capital accumulation in this
intergenerational transmission.
Keywords: poverty, intergenerational transmission, potential outcome, dynamic treatment effects, education
JEL classification codes: D31, I32, J62
∗This work has been supported by the second Network for the analysis of EU-SILC (Net-SILC2), funded by Eurostat.The European Commission bears no responsibility for the analyses and conclusions, which are solely those of the authors.In addition, Bellani acknowledges financial support from an AFR grant (PDR 2011-1) from the Luxembourg Fonds Nationalde la Recherche cofunded under the Marie Curie Actions of the European Commission (FP7-COFUND). Usual disclaimersapply.§E-mail:[email protected]‡E-mail:[email protected]
1
1 Introduction
The impact of poverty during childhood on individuals economic outcomes later in life is a topic of
active research and a major policy concern in many developed as well as developing countries.
The economic literature on intergenerational transmission focused typically on the estimates of the
intergenerational elasticity in income or earnings of parents and their offspring. In the United States,
the first studies on this topic revealed the elasticity coefficients of children’ earnings with respect to
parent’s earnings to be of around 0.2 (Behrman and Taubman, 1985; Becker and Tomes, 1986). More
recently Chadwick and Solon (2002) estimate elasticities in the range 0.35 to 0.49 for family income,
and Altonji and Dunn (2000) in the range 0.35 to 0.41, for hourly wages. A substantial literature
followed, estimating this mobility for many other countries, focusing on Europe in particular, among
others for Scandinavian Countries (Bjorklund and Jantti, 1997), for Britain (Blanden et al., 2007), for
France (Lefranc and Trannoy, 2005) and OECD (2010) for European OECD countries.
Many of these estimates are based on predicted fathers’ earnings, rather than the actual ones, using
fathers’ characteristics (father’s education and occupation in Bjorklund and Jantti (1997); only educa-
tional levels in Grawe (2004); different levels of education, occupational groups and age in Lefranc and
Trannoy (2005); detailed occupation sectors in Leigh (2007)).
The general message coming form these contributions suggests that the United States and the United
Kingdom have higher rates of intergenerational persistence than other countries, while the Scandinavian
countries have the lowest, with respect to the other countries instead the ordering varies.
Fewer studies have instead been focusing on poverty persistence across generations (see among oth-
ers Ermisch et al. (2004); Mayer (1997); Shea (2000); Acemoglu and Pischke (2001)). These papers
find significant impact of parental income or parental financial difficulties on children human capital
accumulation and later labor market outcomes, in the range of 5% decrease in education given parental
joblessness for Britain, and a 1.4 percentage point increase in the probability of attending college for
an increase of income of 10% for the United States.
Blanden et al. (2007) analyze in detail the association between childhood family income and later adult
earnings, among sons, exploring the role of education, ability, non-cognitive skills and labor market
experience in generating intergenerational persistence in the UK. They do so by decomposing the es-
timated mobility coefficient conditional on those mediating variables. They show that inequalities in
achievements at age 16 and in post-compulsory education by family background are extremely impor-
tant in determining the level of intergenerational mobility. In particular they find a dominant role of
education in generating persistence. Cognitive and non-cognitive skills both work indirectly through
2
influencing the level of education obtained, with the cognitive variables accounting for 20% of inter-
generational persistence and non-cognitive variables accounting for 10%.
As we have already argued in a previous contribution (Bellani and Bia, 2015), although those studies
agree that growing up in a poor family raises the probability of falling below the poverty threshold in
adulthood, the key contentious question for policy here is whether this association is truly causal in the
sense that poverty in childhood per se influences later outcomes or whether it is driven by other factors
correlated with both childhood poverty and later outcomes, such as parenting styles, family structure,
neighborhood influences, genetic transmissions, etc. Moreover, as suggested also by Blanden et al.
(2007), it is relevant for policy to examine a plausible causal channel through which being born poor
affects the individual’s economic and social status as an adult. In this paper we thus not only provide a
series of robustness checks on the impact of growing up in poverty that is provided in Bellani and Bia
(2015), but we also analyze this process, introducing individual human capital as intermediate variable.
Our analysis is based on the module on intergenerational transmission of 2011 of the EU-SILC data,
where retrospective questions about parental characteristics (such as education, age, occupation) were
asked. We find that even after controlling for possible unobserved confounders, e.g. child ability,
being poor in childhood significantly decreases the level of income in adulthood, increasing the average
probability of being poor. Moreover, our results reveal a significant role of human capital accumulation
in this intergenerational transmission.
The remainder of the paper is organized as follows. Section 2 introduces the estimation strategy and in
Section 3 the data used through the whole paper are described. In Section 4 we analyze both the average
and the distributional impact of growing up poor, while in Section 5 we focus on a possible channel of
this impact, analyzing the role of attaining higher education. Section 6 concludes.
2 Estimation strategy
Experiencing financial difficulties growing up is not the only determinant of outcomes later in life,
because of the complexity of the process, different statistical techniques have been used, each of which
relies on a different set of assumptions, as reviewed in our previous contribution on the topic (Bellani
and Bia, 2015).
In this paper we follow the framework of Potential Outcomes approach for causal inference (Rubin,
1974, 1978), which considers a randomized experiment where (a) subjects are randomly selected from
the target population; (b) a binary treatment is randomly allocated to the subjects; (c) there are no
hidden versions of the treatment and there is no interference between units (Stable Unit Treatment
3
Value Assumption - SUTVA) as the golden standard for estimating causal effects. In observational
studies, the existence of other difficulties in addition to the critical problem of non-random treatment
assignment implies that additional assumptions have to be made in order to estimate the causal effects
of the treatment.
In our study, the critical problem of non-random treatment assignment (assumption (b) above) implies
that additional assumptions have to be made in order to estimate the causal effects of the treatment.
An important identifying assumption is the selection on observables (unconfoundedness) (Rosenbaum
and Rubin, 1983). For a review of the statistical and econometric work focusing on estimating average
treatment effects under this assumption, see Imbens (2004)).)
Let us consider a set of N individuals, and denote each of them by subscript i: i = 1, . . . , N . Let Ti
indicate whether a child was growing up in a poor household, Ti = 1 (treated), or not, Ti = 0 (control).
For each individual, we observe a vector of pre-treatment variables, Xi and the value of the outcome
variable associated with the treatment, Yi(1) for being a poor child, Yi(0) for not being a poor child.
The central assumption of our approach is that the “assignment to treatment” is unconfounded given
the set of observable variables: Yi(0), Yi(1) ⊥ Ti|Xi.. Let p(X) be the probability of growing in a
poor household given covariates: p(X) = Pr(T = 1|X = x) = E[T |X = x]. Rosenbaum and Rubin
(1983) show that if the potential outcome Yi(0) is independent of treatment assignment conditional on
X, it is also independent conditional on p(X). Following Rosenbaum and Rubin (1983), we adjust
for the propensity score, that is, treatment and potential outcomes are independent also conditional on
p(X): Yi(0), Yi(1) ⊥ Ti|p(X), thus removing all biases associated with differences in the covariates.
For a given propensity score, exposure to treatment can be considered as random and thus poor and
non poor children should be on average observationally identical. Therefore, we apply a propensity
score matching method to select a control group of non-treated individuals (in this case non poor as
a child) who are very similar to treated individuals conditional on a set of observable characteristics
(parental characteristics, family composition, and other features fixed in childhood, such as the number
of siblings or birth order) (unconfoundedness). The matched samples of poor and non-poor children
will then be used to assess impacts on adulthood outcomes, primarily equivalent income and probability
of being poor.
Formally, given the population of units i, if we know the propensity score p(Xi), then the average effect
of being poor on the population (ATE) is estimated as follows:
τ = E[Y1i − Y0i] = E[E[Y1i − Y0i|p(Xi)]] = E[E[Y1i|Ti = 1, p(Xi)]]− E[E[Y0i|Ti = 0, p(Xi)]]
and the average effect of being poor on those exposed to poverty (ATT) as follows:
τt = E[Y1i − Y0i|Ti = 1] = Ep(Xi)|Ti=1[E[Y1i − Y0i|Ti = 1, p(Xi)]] =
4
Ep(Xi)|Ti=1[E[Y1i|Ti = 1, p(Xi)]]− Ep(Xi)|Ti=1[E[Y0i|Ti = 0, p(Xi)]]
Finally in Section 5, we introduce a dynamic treatment approach based on sequential conditional-
independence assumptions to assess if there is any impact of growing up in poverty on children out-
comes later in life after controlling for individual human capital. Statistical analyses concerning the role
of an intermediate variable between a particular treatment and an outcome are important for understand-
ing issues concerning mechanism. Several ways to conceptualize the mediatory role of an intermediate
variable in the treatment - outcome relationship have been proposed in the causal inference literature
(Joffe et al., 2007).
We consider dynamic matching models that extend the static Neyman-Robin model based on selec-
tion on observable (see Gill and Robins (2001) and Lok (2008), and for application to economics
Fitzenberger et al. (2008) and Lechner and Miquel (2010)). In particular we consider a discrete-time
finite-horizon model. Under the assumption that the actual treatment assignment S is sequentially ran-
domized (SR), we can sequentially identify the causal effects of treatment and construct the distribution
of the potential outcomes Y (s) for any treatment sequence s in the support of S. Moreover, the no-
anticipation condition (NA) ensures that potential outcomes for a treatment sequence s equal actual
(under treatment T0) outcomes up to time t− 1 for agents with treatment history st−1 up to time t− 1.
Formally, Y t−1(s) = Y t−1 on the set {St−1 = st−1}.
Consider a two-period version of the model in which agents take either treatment (1) or control (0)
in each period. Then, S(1) and S(2) have values in S = {0, 1}. The potential outcomes in period t
are Y (t, (0, 0)), Y (t, (0, 1)), Y (t, (1, 0)) and Y (t, (1, 1)). For example, Y (2, (0, 0)) is the outcome in
period 2 in the case that the agent is assigned to the control group in each of the two periods.
In a nutshell, in this paper we consider two treatments (poverty, education) and two periods. In period
1, a member of the population can be observed in exactly one of two treatments (Poor, Non poor). In
period 2, she faces the subsequent treatment (Tertiary Education), thus she participates in one of four
treatment sequences (0, 0), (1, 0), (0, 1), (1, 1), depending on what happened in period 1.
Therefore, a comparison of causal effects of T (growing up in poverty) on Y (children outcomes later
in life) for individual now matched also on the value of the intermediary outcome (tertiary education in
our specific case) provides information on the extent to which a causal effect of growing up in poverty
on later outcomes occurs together with a causal effect of growing up poor on the intermediate outcome
human capital status.
5
3 Data
The analysis is based on data from the the European Union Statistics on Income and Living Conditions
(EU-SILC), which provides comparable, cross-sectional and longitudinal multidimensional data on in-
come, poverty, social exclusion and living conditions in the European Union.1 For the specific purpose
of this project we will use the module on intergenerational transmission of 2011, where retrospective
questions about parental characteristics (such as education, age, occupation) referring to the period in
which the interviewee was a young teenager (between the age of 14 and 16) were asked to each house-
hold member aged over 24 and less than 66. We treat missing data as an additional category, keeping
cases that would otherwise be dropped. “This is particularly appropriate when the unobserved value
simply does not exist. For instance, individual may have questions on mother’s and father’s education,
but the father or mother was unknown or never part of the family”.(Allison, 2001)
We further restrict our sample to the individual in working age between 35 to 55 to maintain a degree
of homogeneity in the period of the life cycle in which the outcomes of interests are measured.
More in details, our treatment variable is constructed based on the financial situation of the household
(vary bad or bad), while the variables we used as pre-treatment are the country of residence and of
birth, the gender, the year and quarter of birth, the n. of adults in the household, n. of persons in the
household in work, number of siblings, family composition, year and country of birth, highest level of
education, main activity, main occupation and citizenship of the father and of the mother and tenancy
status.
The outcome in adulthood that we are interested in are the log of the equivalized income,2 and the
probability of being poor (defined as having an equivalized income lower than 60% of the median in
his/her country in that year).
As already mentioned at the end of Section 2, as intermediate outcome that are after treatment and have
an impact on the final outcome of interest we are analyzing the probability of having a post-secondary
education.
The descriptive statistics and the T-test on the means of the variables used in the following analysis can
be found in the Appendix A.3
1Refer to chapter 2 in Atkinson and Marlier (2015) for a description of this database.2We use equivalized disposable income that is the total income of a household, after tax and other deductions, that is
available for spending or saving, divided by the number of household members converted into equalized adults; householdmembers are equalized using the so-called modified OECD equivalence scale.
3We are aware that individuals may suffer from retrospective recollection bias. The retrospective technique has beenextensively applied in studies of intergenerational occupational mobility (Atkinson et al., 1983). In this specific case, webelieve that the type of question asked is less affected by this problem compared, for example, with a direct question on thelevel of income in the household.
6
4 Results
4.1 Propensity score based methods
A preliminary analysis of these data can be found in Bellani and Bia (2015), where a propensity score
base method has been applied for both 2005 and 2011 data. Here, as in there, we firstly estimate
by means of probit model, each individual’s propensity score, i.e. his/her probability to be poor in
childhood given the observed characteristics introduced in the previous section.4
As already explained in Section 2, the propensity score is a balancing score (Rosenbaum and Rubin,
1983), that is, within strata with the same value of p(X) the probability that T ={0, 1} (being poor or
not) does not depend on the value of X. This balancing property, combined with the unconfoundedness
assumption, implies that, for a given propensity score, exposure to a treatment status is random and
therefore treated and control units should be on average “similar” conditional on observables character-
istics. As a result, to be effective, propensity score based methods should balance characteristics across
treatment groups. The extent to which this has been achieved can be explored by comparing balance in
the pre-treatment covariates before and after adjusting for the estimated propensity score (PS). Figures
1 provides the standardized bias (in percentage) for unmatched and matched units,5 showing a huge
improvement in the balancing property when adjusting for the PS, with a bias always around 0.
Another important requirement for identification is given by the common support, as shown in figure 2,
which ensures that for each treated unit there are control units with the same observables. In the matched
sample, the comparison of baseline covariates may be complemented by comparing the distribution of
the estimated PS between treated and controls. The common area of support serves as a first evaluation
4For the results of this first step of the estimation strategy please refer to table 25.3 in Bellani and Bia (2015) and therelative comments.
5The reduction of bias due to matching is computed as:
BR = 100(1− BM
B0)
whereBM
is the standardized bias after matching
BM =100(xMC − xMT )√
S2MC
+S2MT
2
andB0
is the standardized bias before matching
B0 =100(x0C − x0T )√
S20C
+S20T
2
,
where subscript M denotes after matching, 0 denotes before matching.
7
Figure 1: Balancing property
Figure 2: Common support
of whether covariates means included in the PS model are similar between the two groups (Ho et al.,
2007). We con notice here that, as expected given the topic of interest, children with a probability of
being poor higher than 0.7 cannot be included in our study as it is not possible to find children with
these levels of pre-treatment variables belonging to both groups. This results n giving us a sort of lower
bound for the real effect which could be sensibly be considered even higher if this part of the population
was included.
8
We implement the propensity score procedure applying a single nearest-neighbor matching to remove
bias associated with differences in covariates and estimate the effects of being poor on adulthood out-
come, primarily equivalent income and probability of being poor later in life.
Table 1: Propensity Score Matching
(1) (2) (3)Main Outputs Intermediate Outputs
Income Poverty Education
ATE -0.02 0.03 -0.06(0.007) (0.006) (0.008)
ATT -0.04 0.04 -0.04(0.006) (0.006) (0.006)
N. 133647 133818 132789
Results are presented in table 1. The results on income show a substantial decrease in the equivalized
income in adulthood due to exposure to poverty in childhood. This decrease is on average close to 3
percentage points.
With respect to our intermediate outcome of interest, our results show an average of 5% decrease in the
probability of reaching a higher level of education for the children experiencing financial difficulties
while growing up.
This significant decrease in the probability of getting higher education could be considered as a cause
of the strongly significant difference in income in adulthood between these two groups of children. In
the last part of the paper (section 5) we address this research question in more details.
4.2 Robustness checks
4.2.1 Combining Matching and Regression: Doubly-Robust Estimation
In this section we combine both matching on the propensity score and regression, doing a doubly-robust
estimation. By doing the matching we trim from the data all the observations that are not “similar” in
their probability of belonging to any of the treatment groups, with respect to the pre-treatment observ-
ables. In practice, we eliminate units that are not on common support, including in our sample only the
control cases that are on average “identical” to the treatment cases and discarding the remaining data.
In our analysis the psmatch2 Stata package6 gives us the information we need via the weights. We know
which individuals we should discard and which should be counted more than once7. Hence, over the
matched units, we run a regression with all the same covariates used in the propensity score estimation.
6See Leuven and Sianesi (2003) for more details on the package.7More detailed results are available upon request from the authors.
9
The marginal effect of the treatment is our doubly-robust estimate of the average treatment effect, as
reported in table 2.8 The average treatment effect estimate, relative to each outcome considered in our
study, is overall very close to the findings derived from the ATT and ATE estimation when using the PS
based technique, with slightly different results only for income and education at the population level.
Table 2: Doubly-Robust Estimation
(1) (2) (3)Main Outputs Intermediate Outputs
Income Poverty Education
ATE -0.03 0.03 -0.04(0.002) (0.003) (0.004)
ATT -0.03 0.04 -0.04(0.003) (0.004) (0.004)
N. 133647 133818 132782
4.2.2 Sensitivity analysis for poverty risk
One of the central assumptions of our analysis is that being poor in childhood can be considered as ran-
dom, given a set of precedent covariates X and thus the outcome of non-poor children can be used to
estimate the counterfactual outcome of the poor children if they were not experiencing poverty in child-
hood. The plausibility of this assumption heavily relies on the quality and amount of information con-
tained inX .9 The validity of this assumption is not directly testable, since the data are completely unin-
formative about the distribution of the potential outcomes, but its credibility can be supported/rejected
by additional sensitivity analysis. Our analysis would be in fact biased if we were to believe that even
conditional on all the covariates we can observe (parental education and occupation, family situation,
child own age, sex, year and country of birth and number of siblings, etc.), being poor in childhood
would be linked to some unobserved parental genetic ability which would not only influence the proba-
bility of the parents of falling into poverty (being treated) but also the child potential outcome as a result
of the genetic transmission of ability. In this setting, it is assumed that the conditional independence
assumption holds given X and the unobserved variable A: Yi(0) ⊥ Ti|Xi, Ai10 and knowing A would
be sufficient to consistently estimate the ATT: E[Y0i |Ti = 1, Xi, Ai] = E[Y0i |Ti = 0, Xi, Ai].
In order to asses the credibility of our assumption, we follow the approach suggested by Ichino et al.
8 The results presented in this table are the average marginal effect of the treatment variable, given by the OLS coefficientestimation for the continuous variable log of Income, and the average of the marginal effects calculated from the probitestimates for the probability of being poor in adulthood and to attain tertiary education.
9Refer to Section2 for a more detailed description of the assumptions made.10If the average treatment effect of interest is the ATT, the CIA assumption is then reduced to: Yi(0) ⊥ Ti|Xi, where, within
each cell defined by X , treatment assignment is random, and the outcome of controls are used to estimate the counterfactualoutcome of treated in case of no treatment.
10
(2008) and assume that the unobserved ability variable A can be expressed as a binary variable taking
value H=high, L=low. In addition, A is assumed to be i.i.d. distributed in the cells represented by
the Cartesian product of the treatment and outcome values. The distribution of the binary confounding
factor A can be fully characterized by the choice of four parameters: pij ≡ Pr(A = 1|T = i, Y =
j) = Pr(A = 1|T = i, Y = j,X) with i, j ∈ 0, 1, which give the probability that A = 1(high) in each
of the four groups defined by the treatment status (poor as a child) and the outcome value (poverty in
adulthood)11. Given arbitrary values of the parameters pij , a value of A is attributed to each individual,
according to her/his belonging to one of the four groups defined by their poverty status in childhood
and adulthood.
The simulated A is then treated as any other observed covariate and is included in the set of matching
variables used to estimate the propensity score and to compute a simulated ATT estimate, derived as an
average of the ATTs over the distribution of A.
We can thus control for the conditional association of A with Y0 and T by measuring how each config-
uration of pij leads to an impact of A on Y0 and T .
In order to do so, we run the sensatt program developed by Nannicini (2007) and estimate a logit
model of Pr(Y = 1|T = 0, A,X) at each iteration, reporting the average odds ratio of A as the
“outcome effect” (Γ) and “selection effect” (∆) of the simulated confounder:
Γ =
Pr(Y=1|T=0,A=1,X)Pr(Y=0|T=0,A=1,X)
Pr(Y=1|T=0,A=0,X)Pr(Y=0|T=0,A=0,X)
,
i.e. the effect of parental ability on the outcome of non-poor children, controlling for the observable
covariates (X),
∆ =
Pr(T=1|A=1,X)Pr(T=0|A=1,X)
Pr(T=1|A=0,X)Pr(T=0|A=0,X)
,
i.e. the effect of parental ability on the probability of experiencing poverty (T=1), controlling for the
observable covariates (X).
In this section, we perform two simulation exercises. In the first one, the pij are set so as to let our
simulated parental abilityAmimic the behavior of parental education variables, as their strong, although
11Note that, in order to perform the simulation, in Nannicini (2007) two assumptions are made: i) binary confounder ii)conditional independence of A given X.
11
not perfect, positive correlation is well known in the literature (see among others Anger and Heineck
(2010); Bjorklund et al. (2010)).
In the second one, a set of different pij is built in order to capture the characteristics of this potential
confounder that would drive the ATT estimates to zero. (Killer confounder)
In tables 3 and 4 the results of these sensitivity checks are presented. The significance of the impact
is robust to the introduction of the calibrated unobserved ability, although this considerably reduces its
size. These results show that both the outcome and the selection effect need to be very strong (2.2 and
3.9, respectively) in order to kill the ATT, i.e., to explain almost entirely the positive baseline estimate
of the ATT.
Table 3: ATT estimation
Baseline Father’s Educ-Calibrated Mother’s Educ-Calibrated Killer
0.04 0.01 0.02 0.00(0.005) (0.006) (0.007) (0.007)
Table 4: Sensitivity Analysis: pij values and odds ratio
Father’s Educ-Calibrated Mother’s Educ-Calibrated Killer
p11 0.1 0.1 0.5p10 0.2 0.2 0.3p01 0.3 0.2 0.2p00 0.4 0.4 0.1
Outcome Effect 0.49 0.52 2.25Selection Effect 0.31 0.29 3.89
4.3 Distributional effect
Following the results presented in column 1 of Table 1 we know that experiencing financial difficulties
while growing up lead to a decrease of income in adulthood of on average 3%, in this subsection we
are interested in exploring the distributional effect of this impact. In order to do so we present here the
cumulative densities of the distributions of income in adulthood of the children belonging to the treated
and control group together with some measures of inequalities and poverty in both samples.
At first, if we consider the whole sample and compare the distribution of individuals’ income in both
group without controlling for their probability of experiencing poverty in childhood, we can notice that
the distribution of the non-poor children first order stochastic dominates the other, implying thus higher
social welfare in an hypothetical society in which no one experience poverty as a child than one in
which childhood poverty is common (see part (a) and (c) in figure 3). Interestingly, when we control
12
for the propensity score and we look at the matched sample (see part (b) and (d) in figure 3) this result
no longer holds, suggesting that when we can control for the characteristics which are associated with
experiencing poverty in the first place, the impact of the parental financial difficulties, although being on
average significant, does not clearly predict such a lower welfare achievement. It is worth recall here,
that we excluded from our sample the children with observable characteristics which were leading a
probability of experiencing poverty higher than 70%. This part of the population was clearly notably
decreasing the social welfare when included in the sample of the poor.12 This could also be noticed
from table 5, where we report some measure of inequality and poverty by ”treatment” group in both the
whole (part (a) and (b)) and the matched (part(c) and (d) )sample.
(a) Cumulative density-all sample (b) Cumulative density-matched sample
(c) Generalized Lorenz Curves-all sample (d) Generalized Lorenz Curves-matched sample
Figure 3: Distributional Graphs-Equivalized Income
To conclude this part, we also analyzed the distribution of the highest level of education attained and
we can notice the same patter that we have described for income, driven by the higher average level
12In Appendix B the bottom of the distribution is plotted for both cases, which clearly shows dominance in one case andnot in the other.
13
Table 5: Equivalized Income Inequality
(a) (b) (c) (d)Not Matched Matched
Non-Poor Poor Non-Poor Poor
Mean 16660 13379 14349 14202Median 13027 10322 10329 8875p90/p10 11.23 11.97 11.11 12.05p75/p25 4.01 3.93 4.12 4.42Gini 0.44 0.44 0.45 0.48Theil Index 0.32 0.33 0.34 0.39Atkinson Index (α = 2) 0.78 0.71 0.62 0.62
of education of the non-poor sample (see Figure 4 part (a) and (b)).13 Once we exclude from the the
sample the extreme cases based on their probability of being poor as children (i.e. the ones who do not
belong to the common support of the propensity score as explained in Section 2), we are again not able
to assess which distribution is better from a social welfare prospective.
(a) Generalized Lorenz Curves-all sample (b) Generalized Lorenz Curves-matched sample
(c) Lorenz Curves-all sample (d) Lorenz Curves
Figure 4: Distributional Graphs-Highest Level of Education
13Around 20% more individual with post secondary education in this group.
14
5 Results: Poverty in childhood-Education-Poverty in adulthood
As introduced at the end of section 2, in this last part of our analysis we implement a dynamic treatment
approach to our research question in order to uncover the role of human capital accumulation in the
intergenerational poverty transmission.
Table 6 reports the results if we insert education in the set of covariates we use for the matching. The
highest level of education attained is used in the matching procedure as pre-treatment characteristic.
It provides information on how experiencing the (psychological and) physical dimension of poverty
as a child gives less labor market opportunities to young generations via the highest educational level
achieved by the individual. Any progress and advancement at school plays indeed a substantial role
both in terms of intermediate outcome (column 3, Table 1) and control variable (intermediate variable)
to define the causal effect estimation of poverty on labor market outcomes later in life (columns 1 and
2, Table 6). Being poor as a child will more likely translate into an exclusion from tertiary education
and, as a result, more likely into lower income levels as an adult.
Table 6: Propensity Score Matching
(1) (2)Income Poverty
ATE -0.01 0.02(0.008) (0.005)
ATT -0.03 0.02(0.006) (0.007)
N. 130413 130577
Finally, as a robustness check on our measure of human capital accumulation, we present in table 7 the
results using instead of the constructed dummy for post secondary education, the categorical variable
Highest ISCED level of education attained.14
Table 7: Propensity Score Matching
(1) (2) (3)Education Income Poverty
ATE -0.07 -0.01 0.02(0.018) (0.005) (0.006)
ATT -0.12 -0.02 0.02(0.014) (0.005) (0.007)
N. 132789 132620 132789
146 categories: Less than primary education, Primary education, Lower secondary education, Upper secondaryeducation,Post-secondary non -tertiary education, Tertiary education.
15
6 Concluding remarks
This paper examines the causal channels through which growing up poor affects the individual’s eco-
nomic outcomes as an adult. We provide a quantitative assessment of the causal effect of poverty in
childhood at the European level, performing a series of robustness checks on the potential outcome
approach chosen. Moreover, we analyze the impact of the interplay between growing up in poverty and
individual human capital accumulation on the children outcomes later in life. This analysis reinforce
the need for policies devoted to eliminate the source of these increased risk, reducing child poverty.
Moreover, after having showed that childhood poverty has indeed a relevant detrimental impact on eco-
nomic, and potentially also social and behavioural, outcomes in adulthood, a fundamental question left
to answer to be able to make more effective policy recommendation regards the channels through which
being raised in a poor family affects the individual’s economic and social status as an adult. In order
to do so, as a first step in this direction, this paper provides evidence of the important role that human
capital has in this process.
References
Acemoglu, D. and Pischke, J.-S. (2001). Changes in the wage structure, family income and childrens
education. European Economic Review, 45:890904.
Allison, P. D. (2001). Missing Data. Quantitativa Applications in the Social Sciences, Sage Publica-
tions.
Altonji, J. and Dunn, T. (2000). An intergenerational model of wages, hours, and earnings. Journal of
Human Resources, 35:221–258.
Anger, S. and Heineck, G. (2010). Do smart parents raise smart children? the intergeneretional trans-
mission of cognitive abilities. Journal of Population Economics, 23:1255–1282.
Atkinson, A., Maynard, A., and Trinder, C. (1983). Parents and Children: Incomes in Two Generations.
London: Heinemann.
Atkinson, T. and Marlier, E., editors (2015). Income, work and deprivation in Europe. Eurostat Statis-
tical.
Becker, G. S. and Tomes, N. (1986). Human capital and the rise and fall of families. Journal of Labor
Economics, 4:S1–S39.
16
Behrman, J. and Taubman, P. (1985). Intergenerational earnings mobility in the united states: Some
estimates and a test of beckers intergenerational endowment model. Review of Economics and Statis-
tics, 67:144–151.
Bellani, L. and Bia, M. (2015). The impact of growing up poor in europe. In Atkinson, T. and Marlier,
E., editors, Income, work and deprivation in Europe. Eurostat Statistical.
Bjorklund, A., Eriksson, K. H., and Jantti, M. (2010). Iq and family background: Are associations
strong or weak? The B.E. Journal of Economic Analysis & Policy, 10.
Bjorklund, A. and Jantti, M. (1997). Intergenerational income mobility in sweden compared to the
united states. American Economic Review, 87:1009–1018.
Blanden, J., Gregg, P., and Macmillan, L. (2007). Accounting for intergenerational income persistence:
Noncognitive skills, ability and education. The Economic Journal, 117:C43C60.
Chadwick, L. and Solon, G. (2002). Intergenerational income mobility among daughters. American
Economic Review, 92:335–344.
Ermisch, J., Francesconi, M., and Pevalin, D. J. (2004). Parental partnership and joblessness in child-
hood and their influence on young peoples outcomes. Journal of the Royal Statistical Society,
167:69101.
Fitzenberger, B., Osikominu, A., and Volter, R. (2008). Get training or wait? long-run employment
effects of training programs for the unemployed in west germany. Annales d’Economie et de Statis-
tique, (91/92):pp. 321–355.
Gill, R. D. and Robins, J. M. (2001). Causal inference for complex longitudinal data: The continuous
case. Ann. Statist., 29(6):1785–1811.
Grawe, N. (2004). Generational Income Mobility in North America and Europe,, chapter Intergenera-
tional mobility for whom? The experience of high and low-earning sons in international perspective.
Cambridge University Press, Cambridge.
Ho, D., Imai, K., King, G., and Stuart, E. (2007). Matching as nonparametric preprocessing for reducing
model dependence in parametric causal inference. Political Analysis, 15.
Ichino, A., Mealli, F., and Nannicini, T. (2008). From temporary help jobs to permanent employment:
what can we learn from matching estimators and their sensitivity? Journal of Applied Econometrics,
23(3):305–327.
17
Imbens, G. (2004). Non parametric estimation of average treatment effects under exogeneity: a review.
Review of Economics and Statisitcs, 86:4–29.
Joffe, M., Small, D., and Hsu, C.-Y. (2007). Defining and estimating intervention effects for groups
that will develop an auxiliary outcome. Statistical Science, 22:74–97.
Lechner, M. and Miquel, R. (2010). Identification of the effects of dynamic treatments by sequential
conditional independence assumptions. Empirical Economics, 39(1):111–137.
Lefranc, A. and Trannoy, A. (2005). Intergenerational earnings mobility in france: is france more
mobile than the us? Annales d’Economie et de Statistique, 78:57–78.
Leigh, A. (2007). Intergenerational mobility in australia. The B.E. Journal of Economic Analysis &
Policy, 7.
Leuven, E. and Sianesi, B. (2003). Psmatch2: Stata module to perform full mahalanobis and propensity
score matching, common support graphing, and covariate imbalance testing. Statistical Software
Components, Boston College Department of Economics.
Lok, J. J. (2008). Statistical modeling of causal effects in continuous time. Ann. Statist., 36(3):1464–
1507.
Mayer, S. E. (1997). What Money Cant Buy. Family Income and Childrens Life Chances. Harvard
University Press, Cambridge MA.
Nannicini, T. (2007). Simulation-based sensitivity analysis for matching estimators. The Stata Journal,
7(3):334–350.
OECD (2010). Economic policy reforms-going for growth. Technical Report chapter 5.
Rosenbaum, P. R. and Rubin, D. B. (1983). The central role of the propensity score in observational
studies for causal effects. Biometrika, 70:41–55.
Rubin, D. (1974). Estimating causal effects of treatments in randomized and nonrandomized studies.
Journal of Education Psychology, 66.
Rubin, D. (1978). Estimating causal effects of treatments in randomized and nonrandomized studies.
Annals of Statistics, 6.
Shea, J. (2000). Does parents money matter? Journal of Public Economics, 77:155–184.
18
A Descriptive statistics
Table 8: Descriptive Statistics
mean sd min maxquarter of birth 2.45 1.11 1 4year of birth 1965.6 5.95 1956 1976sex 1.52 0.50 1 2n. of adult in hh 2.49 1.21 0 67n. of children in hh 2.39 1.45 0 41n. of person in work 1.84 1.11 0 30year of birth of father 1935.4 9.35 1880 1981year of birth of mother 1938.5 8.86 1841 1988Poor as child 0.13 0.34 0 1Income 4.08 0.42 -0.87 5.14Poor as adult 0.14 0.35 0 1country==AT 0.025 0.16 0 1country==BE 0.023 0.15 0 1country==BG 0.027 0.16 0 1country==CH 0.030 0.17 0 1country==CY 0.018 0.13 0 1country==CZ 0.031 0.17 0 1country==DE 0.049 0.22 0 1country==DK 0.024 0.15 0 1country==EE 0.021 0.14 0 1country==EL 0.024 0.15 0 1country==ES 0.061 0.24 0 1country==FI 0.037 0.19 0 1country==FR 0.043 0.20 0 1country==HR 0.026 0.16 0 1country==HU 0.049 0.22 0 1country==IE 0.017 0.13 0 1country==IS 0.015 0.12 0 1country==IT 0.085 0.28 0 1country==LT 0.021 0.14 0 1country==LU 0.026 0.16 0 1country==LV 0.025 0.16 0 1country==MT 0.017 0.13 0 1country==NL 0.045 0.21 0 1country==NO 0.020 0.14 0 1country==PL 0.057 0.23 0 1country==PT 0.023 0.15 0 1country==RO 0.030 0.17 0 1country==SE 0.026 0.16 0 1country==SI 0.050 0.22 0 1country==SK 0.026 0.16 0 1country==UK 0.029 0.17 0 1country of birth==EU 0.037 0.19 0 1country of birth==LOC 0.90 0.30 0 1country of birth==OTH 0.064 0.24 0 1family composition==Lived with both parents 0.75 0.43 0 1family composition==Lived with single father (single parent family) 0.015 0.12 0 1family composition==Lived with single mother (single parent family) 0.075 0.26 0 1family composition==Lived in another private household, foster -home 0.011 0.10 0 1Observations 179060
19
Table 9: Descriptive Statistics(cnt.)
mean sd min maxfather country of birth==country of residence 0.72 0.45 0 1father country of birth==another EU-27 0.059 0.24 0 1father country of birth==Another European 0.017 0.13 0 1father country of birth==outside Europe 0.025 0.16 0 1father citizenship==country of residence 0.72 0.45 0 1father citizenship==another EU-27 0.045 0.21 0 1father citizenship==Another European 0.015 0.12 0 1father citizenship==outside Europe 0.021 0.14 0 1mother country of birth==country of residence 0.74 0.44 0 1mother country of birth==another EU-27 0.062 0.24 0 1mother country of birth==Another European 0.017 0.13 0 1mother country of birth==outside Europe 0.025 0.16 0 1mother citizenship==country of residence 0.74 0.44 0 1mother citizenship==another EU-27 0.046 0.21 0 1mother citizenship==Another European 0.015 0.12 0 1mother citizenship==outside Europe 0.021 0.14 0 1Education father==Less than primary education 0.024 0.15 0 1Education father==Primary education 0.46 0.50 0 1Education father==Lower secondary education 0.21 0.41 0 1Education father==Upper secondary education 0.083 0.28 0 1Education mother==Less than primary education 0.035 0.18 0 1Education mother==Primary education 0.52 0.50 0 1Education mother==Lower secondary education 0.20 0.40 0 1Education mother==Upper secondary education 0.052 0.22 0 1Main activity of father==Employee 0.61 0.49 0 1Main activity of father==Self-employed 0.15 0.36 0 1Main activity of father==Unemployed 0.0049 0.070 0 1Main activity of father==Retired 0.0081 0.090 0 1Main activity of mother==Employee 0.39 0.49 0 1Main activity of mother==Self-employed 0.085 0.28 0 1Main activity of mother==Unemployed 0.0046 0.067 0 1Main activity of mother==Retired 0.0037 0.060 0 1Tenancy status== Owner 0.70 0.46 0 1Tenancy status==Tenant 0.24 0.43 0 1Main occupation father==Armed forces 0.0085 0.092 0 1Main occupation father==Managers 0.042 0.20 0 1Main occupation father==Professionals 0.054 0.23 0 1Main occupation father==Technicians and associate professionals 0.065 0.25 0 1Main occupation father==Clerks 0.033 0.18 0 1Main occupation father==Service and sales workers 0.047 0.21 0 1Main occupation father==Skilled agricultural, forestry and fishery workers 0.099 0.30 0 1Main occupation father==Craft and related trades workers 0.18 0.39 0 1Main occupation father==Plant and machine operators 0.11 0.31 0 1Main occupation father==Elementary occupations 0.085 0.28 0 1Main occupation mother==Armed forces 0.00052 0.023 0 1Main occupation mother==Managers 0.012 0.11 0 1Main occupation mother==Professionals 0.045 0.21 0 1Main occupation mother==Technicians and associate professionals 0.038 0.19 0 1Main occupation mother==Clerks 0.053 0.22 0 1Main occupation mother==Service and sales workers 0.078 0.27 0 1Main occupation mother==Skilled agricultural, forestry and fishery workers 0.069 0.25 0 1Main occupation mother==Craft and related trades workers 0.040 0.20 0 1Main occupation mother==Plant and machine operators 0.028 0.16 0 1Main occupation mother==Elementary occupations 0.092 0.29 0 1Financial situation==don’t know 0.015 0.12 0 1Financial situation==very bad 0.040 0.20 0 1Financial situation==bad 0.089 0.28 0 1Financial situation==Moderately bad 0.17 0.38 0 1Financial situation==moderatley good 0.39 0.49 0 1Financial situation==good 0.25 0.43 0 1Tertiary Education 0.31 0.46 0 1health 0.93 0.25 0 1Observations 179060
20
Table 10: T-test on the pre-treatment variables means by treatment status
quarter of birth 0.0198∗
(2.13)
year of birth 1.233∗∗∗
(27.00)
sex 0.000970(0.25)
n. of adult in hh -0.146∗∗∗
(-15.74)
n. of children in hh -0.700∗∗∗
(-63.59)
n. of person in work 0.0963∗∗∗
(11.28)
year of birth of father 2.021∗∗∗
(25.80)
year of birth of mother 1.880∗∗∗
(26.51)
country -0.396∗∗∗
(-6.09)
country of birth -0.0260∗∗∗
(-10.75)
family composition -0.314∗∗∗
(-62.79)
Highest ISCED level of education attained by father 0.556∗∗∗
(69.14)
Highest ISCED level of education attained by mother 0.393∗∗∗
(60.93)
Main activity status of father 0.0702∗∗∗
(9.07)
Main activity status of mother -0.415∗∗∗
(-26.84)
countrty of birth of father 0.0411∗∗∗
(6.36)
citizenship of father 0.0730∗∗∗
(11.06)
country of birth of mother -0.0714∗∗∗
(-13.38)
citizenship of mother -0.0451∗∗∗
(-8.09)
Tenancy status -0.150∗∗∗
(-31.82)
Main occupation father 0.263∗∗∗
(10.16)
Main occupation of mother 0.191∗∗∗
(6.64)
Financial situation 2.416∗∗∗
(316.84)t statistics in parentheses∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001
21
Table 11: T-test on the post-treatment variables means by treatment status
Income 0.101∗∗∗
(31.24)
Poor as adult -0.0956∗∗∗
(-35.00)
Tertiary Education 0.166∗∗∗
(46.87)t statistics in parentheses∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001
22
B Cumulative density at the bottom of the distribution.
(a) Without Matching (b) Matched
Figure 5: Cumulative density-2011
23