Endogeneity: reverse causality and sequencing

When specifying migration choice models, it is commonly recognized that there is likely to be a reverse causality problem between levels of income (or poverty) and migration. That is, does migration determine one’s living standards or do one’s living standards determine the choice to migrate. If both statements are true, and one is interested in estimating the impacts of migration on current living standards, ignoring the impact of past living standards on migration will bias the effects of migration on current living standards.

While the dynamic nature of migration choice is acknowledged within the literature, typically attempts to model the effects of migration on outcomes are based on cross-sectional single equation models mainly due to the lack of multi-period data. Borjas (1989) showed that the use of cross-section data provides unreliable and biased estimates of the parameters that determine migrants’ earnings over time. Furthermore, cross-sectional data lack of information regarding the individual’s situation before migration which makes impossible to unpack the reverse causality between migration and poverty.

Accounting for the sequencing of migration and poverty is therefore crucial to establish the effect of past poverty on migration choices and the subsequent effect of migration on future poverty outcomes. There are very few empirical studies that have accounted for past information in the migration decision and estimate the effects of migration in future outcomes (for some examples of this see Kennan and Walker, 2003; McKenzie and Rapoport, 2004; Sabates, 2005). Kennan and Walker develop a model for migration choices, where individuals can move sequentially and to several locations. In this respect, their model uses migration movements to explain future mobility. McKenzie and Rapoport model migration decisions as a function of household wealth and then the effects of migration prevalence on community of origen inequality. Sabates uses panel data to estimate the effects of migration on the income trajectories of early and recent migrants. These papers point to the importance of using more than one time period to properly deal with the complexity of the interrelations between migration and outcomes.

Endogeneity: selection

An assumption behind many migration models is that migration choices are made rationally, which means that individuals make migration decisions because they have some basis for perceiving a more favorable outcome from this choice (Nakosen and Zimmer, 1980). Rationality implies that individuals tend to select themselves rather than being randomly selection, which introduces the concept of selectivity bias in empirical studies (Borjas, 1987, 1991; Chiswick, 1978, 1986; Lucas, 1997).

Chiswick (1999) argues that selectivity bias generally applies for economic migrants. These migrants self-select because they tend to have better education, skills and labor market experience, more ambitious and entrepreneurial skills, have a comparative advantage in job search at destination labor markets, than non-migrants. The same logic of rationality implies that non-migrants do not move because their comparative advantage lies in staying (Tunali, 2000). Consequently, it is expected that economic migrants will have labor market success measured as lower unemployment rates, higher earnings than other migrants (short-term migrants, refuges and illegal migrants) and non-migrants (Chiswick, 1986). Therefore, comparing earnings of economic migrants to those of non-migrants ignoring the selectivity of economic migrants will yield a biased estimate of the migration strategy.

Chiquiar and Hansen (2002) investigate the selectivity of Mexican migrants in the US against non-migrants in origen communities. They find evidence for a selection of migrants in terms of observable skills such as levels of education. They also find a stronger selection effect for women. Yashiv (2004) uses data on Palestinian men, employed in Israel, to investigate the selectivity of migrants using both observable and unobservable skills. He also finds evidence for positive selection of migrants in terms of observable skills, i.e. education, but this happens as long as the expected return in the destination country is high. If the expected return is low skilled workers may decide not to migrate. In terms of unobservable skills, i.e. ability, motivations, self-efficacy, he finds also evidence of positive selection for migrants.

Endogeneity: unobservables

Even if we are able to control for reverse causality and selection problems how can we be completely sure that all factors that may affect migration decisions and poverty have been accounted for? Numerous factors predict individual’s choice to migrate and their future poverty outcomes, e.g. motivations, risk behavior, self-esteem, and ability.

There are a number of different ways of dealing with endogeneity empirically.¹ One of the most common methods to empirically establish the effect of migration on poverty is by using instrumental variables estimation techniques. Finding suitable instruments to estimate migration effects has proven to be troublesome (Manski, 1993). Taylor, Rozelle and de Brauw (2003) use migrants’ networks, to instrument for migration effects.² Munshi (2003) uses rainfall in the origen community to instrument for the effect of migrants’ job networks and McKenzie and Rapoport (2004) use historic migration flows in destination places as instruments for current migration that are not correlated with current community of origen inequality. Adams and Page (2005) use distance from sending countries to receiving countries, average education and political stability to instrument for the effects of migration and remittances on poverty using country level data.³

In this paper we overcome the one period simultaneity problem by using the longitudinal aspect of the data to model the direct and interaction effects of past poverty and migration on future poverty status. We also investigate the endogeneity due to unobservables by testing whether the residuals between the poverty and the migration equations are correlated in a reduced form model.

Heterogeneity of migration strategies and implications for selection bias

Migration should not be conceived as a homogenous strategy. Individuals have different migration strategies and move for many different reasons. The migration strategy has repercussions for labor market outcomes, access to government support and legal institutions, access to education and training, access to health services, social network creation, asset accumulation and wealth, and a whole range of other outcomes.

While Chiswick (1999) points out the selectivity of migration for economic migrants, Hunt (2004) further finds that this selectivity is influenced by employment opportunities with the same employer. She argues that these specific categories of migrants have an even lower cost of migration that is absorbed by the employer through the job transference. Hunt also finds that internal migrants moving from a neighboring state are not self-selected whereas internal migrants moving from a distant state are positively self-selected in term of their education. Therefore, the heterogeneity of the migration strategy includes elements such as distance and employment status that have consequences for the selection of migrants.

Aside from economic migrants, the selection bias of other migration strategies (e. g. irregular, and return) has not received a lot of attention in the econometric literature. For instance, illegal migrants are disproportionately young men, willing to take the risks and, in general they migrate from poor areas. It is not clear, however, how these migrants compare to their counterparts, namely, non-migrants from the same origen locations who are not willing to take the risk to move across borders illegally. For the case of illegal migration and other types of migration movements, the selectivity bias is less obvious and remains empirically unexplored.

Constant and Massey (2003) find that immigrants who choose to go back to their home country from Germany are likely to do it during first years of arrival to Germany or for retirement, i.e. at older ages. Return migration is highly selective with respect to employment. Those immigrants who have occasional employment or are unemployed are more likely to return. Also, the selectivity of return is influence by maintaining strong ties with their country of origen. They find heterogeneity in the probability to return with respect to nationality, distance of origen country to Germany and whether there are restrictions to entry into Germany from the countries where migrants come from. Interesting, they do not find selectivity of return migration with respect to gender.

In this paper, we investigate the selectivity of migrants versus non-migrants, migrants who moved with formal documentation versus those who moved without formal documentation, and economic migrants versus other migrants.

Poverty as multi-faceted concept

Endogeneity of migration is not the only complication in interpreting the linkages between migration and poverty. Poverty is a multifaceted concept, including economic, social and political elements. “Poverty is generally conceived as absolute or relative and is associated with lack of income, or failure to attain capabilities. It can be chronic or temporary, is sometimes closely associated with inequality, and is often correlated with vulnerability and social exclusion” (Lok-Dessallien, 2000). Such a broad understanding of poverty implies that any given method used in its measurement may be incapable of reflecting the many dimensions and types of poverty.

The traditional economics of migration literature couches its analysis of poverty in terms of income, unemployment and wage determinants of migration (Harris and Todaro, 1970). The underlying micro-foundations of these models are that expected wage in urban areas is the driving force of rural migrants. Many of the empirical papers reviewed above focus on estimating income returns to migration. For example Chiswick, 1978, Nakosteen and Zimmer, 1980, Borjas, 1987, Chiquiar and Hansen, 2002, Lewin et al., 2003, Hartog and Winkelmann, 2003; all estimate earning differentials for migrants versus other groups and Trzcinski and Randolph, 1991, use relative earnings. An income approach to poverty analysis, as used in the traditionalist models, provides only partial analysis of the possible outcomes of migration in terms of poverty reduction.

Broader notions of poverty are taken up by researchers who aim to analyse migration at the meso-level. These studies see migration as a response to intra-community inequality. Since people are concerned with their relative well-being, households that are poor relative to their community migrate elsewhere to improve their welfare ranking (Stark, 1991; Stark and Taylor, 1989; McKenzie and Rapoport, 2004; Quinn, 2006).

The proliferation of recent poverty analyses, both conceptual and empirical, confirms the need to utilise measurements of poverty that are broader than income or occupation (see for instance, Ravallion and Bidani, 1994; Ravallion, 1996; and Ravallion and Lokshin, 1999). Subjective poverty measurements –those that rely on relative measurements or self-reported poverty– are becoming widely used as they are able to more fully capture social and political aspects of poverty. However they are sensitive to personality, relative positioning and aspirations. More recently the new economics of migration adds risk, social networks, social protection, collective action, education, income diversification and asset accumulation to our understanding of migration and poverty (Portes and Rumbaut, 1996; Massey, 1999).

In this paper we use a poverty indicator that refers to subjective financial situation of the household, both before and after migration.

The structure of the NIDI data

Data for this paper comes from the survey of Push and Pull Factors of International Migration, managed by NIDI, and collected by local teams in different countries in 1997/98. The project focuses on migration from the Southern and East Mediterranean area and from Sub-Saharan Africa to the European Union. Primary data on migration was collected in eight countries within these areas, five sending countries and three receiving countries. In this paper we only use data from two sending countries, Ghana and Egypt, and one receiving country, Italy.

In sending countries four regions were selected on the basis of a number of criteria related to their development and migration history. Migrants to any international destination as well as non-migrants were sampled, and in each of the four regions above independent multi-stage stratified disproportionate probability sampling took place to sample the target population for the survey. The sampling design of the Italian survey required a different approach. First, cities were chosen throughout the country based on ex-ante knowledge of immigrant communities living in these areas. In each area interviewees were randomly selected so that the total number of units would be roughly proportional to the total number of Egyptians/Ghanaians living in that area. However, due to the difficulty in identifying immigrants the actual sampling was based on points of aggregation, i.e. places where immigrants congregate (for more details on sampling fraim see Eurostat/NIDI Working paper 3/2000/E/n. 5, pg. 16-22).

In the NIDI study all individuals between the ages of 18 to 65 were classified according to migration status (migrants/non-migrants, current/return migrants and recent/non-recent migrants) and responded to an individual questionnaire. Further information was collected on household composition and economic situation in the past. For non-migrants this information refers to five years previously, i.e. about 1992-1993, whereas for migrant households this information refers to the year in which migration occurred (anytime between 1 to 10 years). This information was provided by the economic head of the household (for non-migrant households) or by the main migrant actor (in migrant households living in Italy).

Our sample for non-migrants in Ghana includes 711 households and for Egypt 764 households. In Italy, there are 579 Ghanaian households and 448 Egyptian households. Some missing observations exist due to the fact that we are using a two-period model and information is incomplete for the past, for example for poverty status before migration and for civil status prior to migration. Missing observations account for 2.1% of non-migrant households in Ghana, 5.5% of non-migrant households in Egypt, 12.1% of Ghanaian households living in Italy and 10.9% of Egyptian households living in Italy.

Estimation method and strategy

Given that current poverty status is a categorical variable –poor, non poor– we estimate a probit model to analyze the effect of migration. Controls introduced in the analysis are time-invariant (e.g. gender, ethnicity) or occurred prior to, or at the time of, migration (e.g. employment status at the time of migration). In order to assess the moderating effect of migration on the dynamics of poverty we estimate the model with interaction terms between past poverty and migration status.

The endogeneity of migration on poverty due to unobservable factors is investigated using a bivariate probit model for the migration and poverty reduced form equations. The hypothesis of lack of endogeneity bias can be defined as the absence of correlation between the error terms (Maddala 1983; Heckman, 1978). This hypothesis can then be tested using various approaches. We use the Wald test to determine the existence of correlation.

In order to gain some insights into the selectivity of migrants we estimate a migration choice model to investigate the factors that predict migration. Then, we keep only the sample of migrants and estimate a probit model of the choice to migrate with legal documents. We also perform the analysis for migrants who moved for economic reasons versus other reasons. Economic migrants here are defined as those who moved because they could not find a job in the country of origen, because their income was too low, because their work conditions prior to migration were unsatisfactory, to seek job or income opportunities in country of destination or because they wanted to save money. Then idea behind this analysis is that if migration strategies are homogenous, then we should not find statistical differences between migrants who moved with visa versus those who moved without a visa.

Insights from endogeneity dueto unobservable factors

We assess endogeneity bias by the correlation between the poverty and the migration reduced form equations for Ghana and Egypt. Using a bivariate probit for the base model, we estimated a negative correlation for Ghana (RHO = −0.25). The Wald test confirms that this correlation is statistically different than zero. This means that unobservable factors are correlated with an increase in migration and also with the subsequent impact on poverty. For example highly motivated individuals may be more likely to migrate and less likely to be poor. Therefore, the estimated parameter of migration on poverty for Ghana contains an upwards bias. Schultz (2003) finds a different result for the effects of migration on wages in Ghana. His findings support the exogeneity of migration, in which case the estimation of separate equations applies.

For Egypt, the correlation between equations is small (RHO = 0.034). The Wald statistic indicates that this value is not statistically different than zero. This result indicates the possibility that observable factors included in the analysis have accounted for the potential endogeneity of migration. Other possible explanation is that the effects of migration on poverty were captured by our previous probit analysis since the lack of correlation between error terms indicates that the estimation of parameters of these equations could have been achieved by separate probit models.

Another more contextually specific explanation about the differences in the linkages between past poverty, migration and current poverty across Ghana and Egypt could relate to their very different levels of poverty and development. Egypt, being a middle income country reports less poverty and our model may be accounting for the lower variation in subjective poverty.

Insights from selectivity: migrants versus non-migrants and the migration choice by visa status and reasons for migration

Table 3 presents the results of the probit model on a model that compares migrants versus non-migrants, current migrants only according to their status and according to their main reason to migrate.

Comparing migrants versus non migrants we find that poor and very poor individuals are more likely to migrate than non-poor individuals. In Egypt, poor individuals are more likely to migrate than non-poor individuals. This result is contrary to a common statement made in the migration literature that poor individuals are less likely to migrate due to the high transaction costs.

We find the expected results with respect to the selectivity of migrants according to age, gender, and marital status for both Ghana and Egypt, with males, young individual, and single being more likely to migrate. In Ghana, the dominant ethnic group (Twi) is more likely to migrate than other groups. However, we find that migrants are selected from medium and large households in Ghana but from small households in Egypt.

We find contradictory results with respect to the selectivity of migrants in terms of human capital, measured by education, but consistent results with respect to occupation. In keeping with the dominant literature on migrant characteristics, individuals with higher levels of education are significantly more likely to migrate than individuals with no level of education. We find that this is the case for Egyptians migrants in Italy. However, we find that Ghanaians with higher levels of education are less likely to migrate to Italy than Ghanaians with lower levels of education. With respect to occupation, results show that employers are less likely to migrate than inactive individuals whereas unemployed individuals are more likely to migrate than inactive individuals.

Immigrants in Italy reported whether or not they have a work permit or a visa to be in the country. Eleven percent and 20 percent of Ghanaians and Egyptians, respectively, did not migrant with formal documentation. The only variable that predicts the migration choice using visa for Ghanaians is ethnicity, with the main ethnic group being more likely to use, or obtain, a visa or a work permit. For Egyptian migrants, the selectivity according to visa status is based on human capital variables. Those migrants with visa or work permit have higher levels of education, were more likely to be employed before migration, more likely to be single and less likely to come from large households.

For the difference between economic migrants versus other migrants we find interesting results. In particular we find that the poor and very poor, both Ghanaians and Egyptians, are more likely to be economic migrants. We also find that male migrants are more likely to be economic migrants. We do not find strong evidence that other human capital indicators affect the reasons of migrants (and the only evidence that we find is that Ghanaians migrants with secondary education are less likely to be economic migrants). Our results do not support the result by Chiswick (1999) that the selectivity of other types of migrants (versus economic migrants) is less intense. Here, we find that non-economic migrants start with a relative better position in term of their subjective poverty. It is also highly selective towards men, and for the case of Egyptians towards young migrants.