Cross National Comparisons of Internal Migration Flows By Marital Status and Gender

Nayoung Heo, Asian Demographic Research Institute, Shanghai University
Guy Abel, Asian Demographic Research Institute, Shanghai University

Past studies of internal migration patterns suggest distinct heterogeneity in the levels of migration by marital status and gender. Understanding these differentials is essential for explaining past patterns and projecting future population sizes and compositions. Exploiting migration variables from Integrated Public Use Microdata Series data for 65 countries over the period 1960-2011 we first describe migration intensity by marital status and gender. Subsequently, we fit a series of weighted multilevel gravity-type spatial interaction model. We utilize a range of variables to study differences between migration of population subgroups (i.e. marital status and gender) from country specific contextual factors as well as regional push and pull factors. We find distinct patterns in the migration levels through different marital status, where, for example, divorced and widowed migrants are associated with longer distance moves. Within marital status variations by gender are also apparent. For example, separated or widowed females migrate at higher levels to more populated and urban areas compared to their male counterparts.

1        Data

The internal migration, socioeconomic, demographic, andgeographic data used in the current study were obtained from the IntegratedPublic Use Microdata Series International (IPUMSI) database  (Minnesota Population Center, 2015). Each set of census microdata contains a random sample (0.4%-10%) of unidentified private households and associated persons based on a full census conducted by the national statistical agency in each country. The countries and years used in this study, shown in Figure 1, are based on censuses collected between 1960 and 2011. In total, the current study utilized the data from 295,742,164 individual records in 198 censuses of 65 countries

As shown in Figure 1, there is a consistent pattern in thefemale share of internal migration flows across marital status groups in bothspace and time. Migration flows of widows and separated/divorced/spouse absentcategories are predominantly female across all of the countries. There are somenotable exceptions in this general pattern in China, Pakistan, and India.  Migrationflows of those in single and married statuses tend to be more balanced, withequal amounts of male and female migrants. Single male migration flowsoutnumber their female counterparts in many developing countries. In some ofthese cases the male dominance is only present in earlier censuses, fading awayin more recent years (e.g. Kenya, Cameroon, Indonesia and Ecuador).

Figure 1: Countries and census years used in the study,along with the proportion of female internal migrants by marital statuscalculated from the IPUMSI data.


2        Measurement and Model Specification

In order to further investigate these patterns we usedmultilevel regression models fitted to eight sets of internal bilateralmigration flow data, one per combination of the four marital statuses and twogender categories. 2.1       Response Variable

The count of migration flows between the pair of origin and destination were taken from thequestion which asks individuals where they lived previously. 2.2       Model Specification

A sequence of spatial interaction models were fit to each ofthe data set using a cross classified multi-level regression model. Wespecified our model to account for additional variation generated from a widerange of origin (i) and destination (j) regions, the set ofcountries (c) in our study, duration intervals (d) to definemigration and the years (t) of the censuses used. This was operationalizedin our model using random effect parameters which cross classifies a givenobserved flow by each of these characteristics. Prior to fitting the spatialinteraction model we also control for country level economic development andincome inequality. We then use further fixed effects parameters to study therole of geographic, demographic and socioeconomic variables derived fromregionals summaries using IPUMSI. In order to cope with differentadministrative geographies in each country we weighted observations by thenumber of possible migration corridors. This allowed migration flows withincountries with fewer regions to weigh more than flows within countries withmany regions. We had to delete the model equations due to website restrictions:( 3        Results

The fixed effect parameter estimates from the full model onthe first of the multiple imputation data sets is plotted in Figure 2. Theuncertainty in the parameter effects is illustrated using the error bars whichrepresent 1.96 times the standard error.

Figure 2: Fixed effect parameter estimates and theirstandard errors from the full model on the first of the multiple imputationdata sets. Note, error bars represent 1.96 times the standard error

  4        Conclusion and Future Work

We have explored differences in the levels internalmigration by marital status and gender. From IPUMSI we constructed a nearexhaustive database of migration moves within countries and potentially relatedexplanatory variables. In our initial exploratory analysis we foundconsiderable variation in the female share of internal migration flows across martialstatuses in both space and time. We identified some general patterns in thedata; female shares of migration totals tended to be higher in divorced/separatedand widowed categories.

In order to further investigate these patterns and toexplore variation at the regional level we used a multilevel regression modelsfitted to eight sets of internal bilateral migration flow data, one percombination of the four marital statuses and two gender categories. We areundergoing work to provide full robustness checks for our results. In thispaper, regression models were run on only the positive migration flows, wherezero counts were dropped, following a similar approach as (Beine, Docquier, & Özden, 2011; Kim & Cohen, 2010). Initial results for models that include the zero flows provide similar results albeit with larger parameter estimates. We assumed our residual errors follow a Gaussian normal distribution. We are also currently experimenting with multilevel Poisson regression models with extensions for over-dispersion.


Presented in Session 1221: Internal Migration and Urbanization