fase 1: Presentación de la historieta y motivación a la propuesta

The question I ask in this paper is whether being resettled from the slums at an early age can have an impact on the education, integration and aspiration outcomes of young Roma children. The following regression presents my empirical specification:

yij =α+βaij +γ0xij +ρ0zj+uj+uij, (i= 1, ..., N)

(j = 1, ..., J) (3.1) whereiandj index individual and family, respectively. The dependent variable yij corresponds to the outcomes I want to study, e.g., Not repeated grade (the child has not repeated grade at school), which takes values 1 (has not repeated grade) or 0 (has repeated grade). The exogenous variableaij captures the age at which childiwas resettled from the slum, and I define it as a dummy variable that takes value 1 if childi

32_{In cases when the children responded to the survey, their replies were taken into account if confirmed}

in familyjwas resettled (left the slum) before age 6 and value 0 otherwise. The vectors xijandzjinclude the confounding factors that I control for in the model:xijis ak1×1 vector of exogenous individual-specific variables (age, sex) and zj is a k2 ×1vector of family-specific variables that I assume constant across all children in the family: a dummy for parents’ literacy (that takes value 1 if one or both parents can read and write), a set of dummies for slum of origin, and a set of dummies for ancestors’ origin, where ancestors’ origin is classified in the following categories: Castellanos33, Gallegos, Extreme˜nos, Portuguese, or of mixed origin. This classification is based on a broadly accepted classification of the Spanish Roma (San Rom´an, 1997). The error term has two components: a family-specific error term uj that includes omitted characteristics that are constant across family members, and a random error termuij, that is independently, identically distributed overiandj, with mean 0 and varianceσ_u2.

Estimating equation (3.1) using ordinary least squares (OLS) yields the impact of being resettled before age 6 conditioned on the covariates specified in the model. However, this estimate ofaij is biased becauseaij could be correlated with unobserved family characteristics (uj); for example, the preferences of the parents for their chil- dren’s education and welfare may be correlated with the ages that their children had at the time they left the slum –one cannot rule out that the parents that care the most about their children’s education may have made a larger effort to move to slums that they suspected could be dismantled when their children were very young, to ensure their children would live most of their childhood and life in a flat.

To eliminate the problem of unobserved family-specific variables, I estimate my regression using family fixed effects. In particular, given the assumption I impose in equation (3.1) on family-specific unobservables,uj, staying constant for all children in a given family but exhibiting variation across families, we can take the deviation from the mean across individuals in a given family and obtain

yij−y¯j =β(aij −¯aj) +γ0(xij−x¯j) + (uij −u¯j), (i= 1, ..., N) (j = 1, ..., J) (3.2) wherey¯j = (1/N) PN i=1yij, ¯aj = (1/N) PN i=1aij, x¯j = (1/N) PN i=1xij,

andu¯j = (1/N)

PN

i=1uij. Taking deviations from the mean as in equation (3.2) elim- inates the unobserved family effects. The error term ij = (uij −u¯j) is now inde- pendently, identically distributed over i and j, with mean 0 and variance σ2, and the least-squares regression of equation (3.2) provides unbiased and consistent estimates of βandγ, which are the family fixed effects estimators of equation (3.1).

A key assumption I make in equation (3.1) is the functional form chosen for the dummy variable aij, which takes value 1 if the child was resettled (left the slum) be- fore age 6. The decision to not treat age at resettlement as a continuous variable and to choose this cut-off point in particular is based on the early childhood development literature. Almond and Currie (2011) survey this literature; they define early childhood as starting at birth and ending at age five and claim that the work they review con- clusively shows that events before five years old can have large long term impacts on adult outcomes; for example, Cunha and Heckman (2007) and Heckman (2007) argue that childhood environment is the principal factor affecting children outcomes and their future adult outcomes, and Dobbie and Fryer Jr (2011) provide strong evidence sup- porting the claim. A number of studies in the immigration literature have used similar cut-off points to the one I choose to evaluate how an immigrant child’s age at arrival to a new country affects his life outcomes; for example, Ohinata and van Ours (2012) use a cut-off age of 5 to evaluate test scores of maths and science of immigrant children aged 9 and 10 in The Netherlands and conclude that children that arrive at age 5 or older have lower scores than those who arrived at an earlier age; Bratsberg, Raaum, and Red (2011) conclude that immigrant children that arrive in Norway at a later age have a smaller chance of completing secondary school, especially if they arrive after age 7. I thus choose the cut-off point at age 6. However, to alleviate the concern that this choice could be arbitrary, I provide my results for different cut-off points for the variable that captures whether the child was resettled at an early age, and for the variable age at reset- tlement expressed in continuous form (and continuous and squared). The results from these robustness checks are displayed and discussed in section 3.7.

The sample used to estimate my theoretical model in equation (3.1) by family fixed effects is smaller than the sample one can use to estimate the model using least squares. The transformations performed in order to estimate the model using family fixed effects can only be done in families that have more than one child (I >1), i.e., the sample of children aged 7 to 14 is reduced to those whose families have more than one

Table 3.7: Families with children aged 7 to 14. Sample size by family type

Number of children Number of families

Panel A. Families with children aged 7 to 14

Total 784 464

Panel B. Families with only 1 child aged 7 to 14

Total 229 229

Panel C. Families with more than 1 child aged 7 to 14

Families with no variation in key dimension 365 156 Families with variation in key dimension 190 79

Notes: The key dimension variable inPanel Cis the variableResettled before age 6, which takes value 1 (variation in key dimension) if the family has children resettled before age 6 and children resettled when they were 6 years old or older, and takes value 0 othewise (no variation in key di- mension). Total sample size is 784 children (464 families). I have excluded from the sample the 13 children (10 families) for which there is no information on the key dimension variableResettled before age 6.

child in this group age. I show the number of families that fulfil this criteria in Panel C in Table 3.7. Column (2) in the table presents the total number of children aged 7 to 14 in my sample (Panel A), the total number of children aged 7 to 14 in the sample that belong to families that have only one child aged 7 to 14 (Panel B) and the total number of children aged 7 to 14 in families with more than one child in the age range 7 to 14 (Panel C). Column (3) in the table shows the number of families that correspond to each of these subsamples.

My theoretical model, equation (3.1), imposes one further restriction on the sam- ple because I define age at resettlement as a dummy variable that takes value 1 if the child was resettled before age 6 or 0 if the child was resettled at or after age 6. This divides the families with more than one child aged 7 to 14 into two groups: Those that present variation in this key dimension (from all children aged 7 to 14 in the family, at least one was resettled before age 6 and one was resettled at or after age 6), and those families that do not present this variation. The children belonging to families that present variation in this key dimension are those that form part of my sample for estimating equation (3.1) by family fixed effects. The number of children and fami- lies that correspond to this sample are shown in Panel C in Table 3.7: In total, I have a sample of 190 children that correspond to 79 different families. Table 3.8 presents this information by type of outcome variable. The sample size for estimating equation

Table 3.8: Number of children by variable outcome and family type. Families with children aged 7 to 14

Family type

All Only 1 child More than 1 child No variation Variation

Panel A. Education outcomes

Not repeated grade 777 228 362 187

Panel B. Integration outcomes

Has relatives as best friends 772 228 356 188 Has non-Roma as best friends 773 227 359 187

Panel C. Aspiration outcomes

All “job wanted” outcome variables 748 213 357 178 Notes: Total sample size is 791 children (474 families) with children aged 7 to 14, as shown in Table 3.7. Each row represents a different outcome variable, presented in column 1. The sample size for each outcome variable varies depending on the number of missing values for that variable. Family typesOnly 1 child(column 3) andMore than 1 child(columns 4 and 5) refer to the number of children aged 7 to 14 in the fam- ily. Columns 4 (No variation) and 5 (Variation) refer to variation in the key dimension variable described in Table 3.7: The variableResettled before age 6, which takes value 1 (variation in key dimension) if the family has children resettled before age 6 and children resettled when they were 6 years old or older, and takes value 0 otherwise (no variation in key dimension). The “job wanted” outcome variables inPanel Care:The child wants to work in a traditional Roma job,The child wants to be a skilled worker –Apprenticeship,The child wants to be a skilled worker –University,The child wants to work in the formal economy, andThe child wants to work in the informal economy.

(3.1) varies slightly for each outcome variables, once we discount missing values; as Table 3.8 shows, it varies from 187 children (for the education outcome) to 178 children (for the set of aspiration outcomes).