Does Citizenship Still Matter? Second Birth Risks of Migrants from Albania, Morocco, and Romania in Italy

Abstract

In this article we investigate the propensities to have the second child in Italy for foreign women from Albania, Morocco, and Romania. Our study contributes to the international debate on migrant fertility by testing the main competing hypotheses present in literature, using the Italian case as an illustration. Italy is an important case study because it has been a country of immigration for only a few decades and because the literature on this topic was limited in Italy by the difficulties in obtaining proper longitudinal data. An important component of our work was therefore to build a new data set, using record linkage procedures that allow us to improve the information from Survey on Birth and Resident Permit Registers and to study the individual childbearing trajectories. Our results confirm the importance of the mother’s citizenship. The impact remains strong after controlling for the main demographic and migratory characteristics. We found that older cohorts experience a disruption effect but that a native Italian partner can promote an adaptation process such as a convergence in fertility behavior toward that of native Italian women.

Résumé

Cet article examine les probabilités d’avoir un deuxième enfant en Italie pour les femmes étrangères venant d’Albanie, du Maroc et de Roumanie. Cette étude vise à contribuer au débat international sur la fécondité des migrants en testant les principales hypothèses concurrentes présentes dans la littérature et en les illustrant par le cas de l’Italie. L’Italie est une étude de cas intéressante car c’est un pays d’immigration depuis seulement quelques décennies et parce qu’en Italie les études sur ce sujet sont limitées du fait de la difficulté à obtenir des données longitudinales adéquates. L’une de principales tâches de ce travail a donc été de construire une nouvelle base de données à partir de procédures d’appariement qui nous ont permis d‘améliorer les informations issues de l’Enquête sur les naissances et des Registres de titres de séjour, et d’étudier les trajectoires individuelles de fécondité. Cette étude confirme l’importance de la citoyenneté de la mère. L’effet reste important après contrôle des principales caractéristiques démographiques et migratoires. Les résultats montrent que les cohortes les plus âgées ont subi des effets perturbateurs mais qu’un partenaire né en Italie peut favoriser un processus d’adaptation conduisant à une convergence du comportement de fécondité des femmes étrangères vers celui des femmes nées en Italie.

Appendices

Appendix 1

The aim of record linkage is to identify the same individual, object, or event in different databases by using common characteristics and the information available in single files (Fellegi and Sunter 1969). These techniques were first introduced in epidemiologic studies in the late 1950s. A group of geneticists implemented a new methodological system to link different datasets (Newcombe et al. 1959). Fellegi and Sunter (1969) formalized this new procedure. Later it was used for different applications, including the construction of longitudinal datasets, internal links within one data set to manage duplicates, and the estimation of the size of a population with capture–recapture methods. Most commonly, it was used to match and merge information on one individual from multiple data sources (Winkler 1995; Cibella et al. 2008). Record linkage techniques are particularly useful for migration studies for which there are little available data. Let us assume a data set A of dimension N _A and a second data set B of dimension N _B. The aim of the record linkage is to find all the pairs (a,b) in the space Ω = {(a,b), a ∈ A and b ∈ B} of size N = N _A  × N _B that refer to the same individual (Fellegi and Sunter 1969). The final aim is to classify the pairs into two groups. The first group is the set of matches (M) where (a = b), and the second group is the set of non-matches (U) where (a ≠ b). If the characteristics taken into account in the analysis, known as matching variables, have errors or missing values, it is not possible to simply join the two data sets. For these cases, a comparison function has to be constructed (Cibella and Tuoto 2009). For each pair (a,b) ∈ Ω, the comparison vector γ = (γ₁,…,γ _j ,…,γ _k ) is defined where the k elements are the variables under comparison. The results of the comparison are generally equal to 1 if they agree and are 0 otherwise: γ = (1,…,0,…,1). Fellegi and Sunter (1969) used γ to distinguish matches and non-matches to calculate the likelihood ratio

$$ r = \frac{{P\left( {\gamma \left| {(a,b) \in M} \right.} \right)}}{{P\left( {\gamma \left| {(a,b) \in U} \right.} \right)}} = \frac{m(\gamma )}{u(\gamma )}{\text{ with }}M \cup U = {\Upomega} \,{\text{and}} \, {M} \cap U = \emptyset . $$

All pairs are then ranked according to the ratio r.¹⁸ In order to identify the real matches, two thresholds T _m and T _u are assigned where T _m  > T _u . They represent the thresholds for link, non-link, and possible link. The pairs where r is greater than T _m are considered link and non-link where r is smaller than T _u . The group of possible links includes all cases where r lies in the interval between T _m and T _u . They have to be checked manually. Both thresholds are chosen in order to minimize the occurrence of false matches λ (matches of different individuals) and false non-matches μ (non-matches of the same individuals), and are usually identified by empirical methods comparing the two estimated distribution functions of u and m (Cibella et al. 2008). Both distributions are shown in Fig. 5. The distance between these two distributions, which are hypothesized to be binomial, discriminates between matches and non-matches. The larger the distance between the two distributions, the easier it is to distinguish between them. In general, the distribution of the matches is smaller than the distribution of non-matches.

Fig. 5

The mixture model of the distributions m and u. Source Cibella et al. 2008

To apply the probabilistic record linkage, we used the program RELAIS (REcord Linkage At ISTAT),¹⁹ which allows us to deal with these different record linkage techniques in every phase of the procedure (Tuoto et al. 2007). We applied record linkage techniques between the mothers from Albania, Romania, and Morocco who were registered in the Survey on Live Births in 2003, and who were already linked with the information about the children they had in the period 2003–2006 (see Mussino et al. 2010), and the residence permits registers following two different approaches. The matching variables that we used were (1) the information on the name—name combined with the surname but without vocals; (2) the date of birth (day, month, year); and (3) the mother’s citizenship. The citizenship is used to reduce a priori the dataset so it is not inserted in the model to compare the pairs in the space Ω. In this study, we used as a comparison function equality, and on the residuals of the first links, the Jaro–Winkler function (considering the threshold of 0.8).²⁰ This comparison function is an extension of the one created by Jaro (1989) in that it also takes into account a partial agreement between two strings. In fact, Jaro accepts “spelling deviations.” Winkler (1995) is even more suitable when the person’s name is used because it gives more importance when there is a common prefix. To reduce the search space of the link candidate, the sorted neighborhood procedure (Baxter et al. 2003) was performed. The aim of this technique is to sort the two datasets according to a variable (the name) and to fix a window, which is large w. The size of w depends on the frequency of the sorting variable, so that the comparison is evaluated inside a window and not for the entire space. We used the decision model proposed by Fellegi Sunter, which estimates the parameters with the help of the EM algorithm mentioned earlier. At the end of the procedure, the distributions of the residuals by the main characteristics were analyzed. The distributions appear to be similar to one of the links, so they seem to be random.

The mothers were supposed to have residence permits, so we expected a 100 % linkage between A and B. For several reasons, this was not the case. First, if there are missing data in the matching variables, the individuals are not considered in the model. Second, there are administrative reasons: in order to be recorded in the Municipal Population Register, a foreigner needs a residence permit (except those from 2007 who have EU citizenship). To record a birth in the municipal register, one of the two parents has to be registered, but this may sometimes be the father of the newborn, and not the mother. A mother may still be in the register even if the permit has expired because she may not have reported the expiry to the municipal authorities. Table 4 shows all mothers who are considered in the analysis, indicating the numbers of links with the residence permit register for the period 2002–2004. The percentage of matched records is considered as an indicator of quality for the record linkage procedure (Lariccia et al. 2011). Although the proportion of links is below 100 % and varies with citizenship, it is very high, thus confirming the high-quality linkage.

Table 4 Links between cohort of mothers with first recorded child in 2003 and residence permit information

To better investigate the quality of the results, we ran some tests on the distributions of sets M and U. Fortunately, the distance between the two distributions seems to be large and the area where they overlap is small, indicating again that the results are of high quality. Focusing on Albanians, when the year 2003 is considered for the residence permit, we looked at the data on 6,257 mothers in file A (RHD) and 81,203 women in file B (resident permits for 2003). Owing to the lack of other discriminating variables, we only fixed one threshold to assign pairs to sets M and U, avoiding manual revision for the possible link. This threshold was fixed at a value corresponding to the posterior probability of a true match p post = 0.9999. In Fig. 6, the m- and u-distributions are represented in respect to their ln(r) value. The distance between the two distributions seems to be large and the area where they overlap is small.²¹ This indicates a high quality of results.

Fig. 6

Estimates of m- and u-distributions for Albanian mothers

Appendix 2

See Tables 5, 6, 7.

Table 5 Transition to a second child—person-months (exposures) for the 2003 cohort of primiparous mothers with Albanian, Moroccan, and Romanian citizenship

Table 6 Transition to a second child—person-months (exposures) and conceptions (occurrences) for the 2003 cohort of primiparous mothers with Albanian, Moroccan, and Romanian citizenship

Table 7 Hazard ratio of having a second birth in the period 2003–2006 for the 2003 cohort of primiparous mothers with Albanian, Moroccan, and Romanian citizenship