I use a gravity model to determine the relationship between migration and development aid. The gravity model looks at the interaction between migration flow and its drivers, in this case, the types of developmental aid. The gravity model is the best choice for this analysis, because the OECD developmental finance dataset reports bilateral net aid, and gravity models can incorporate paired country time and fixed effects into a regression. The time effects will be particularly helpful in this instance because amounts of bilateral aid vary over time as countries’ situations change. I estimate the following model:
ln Migrant Stocki,j,t = β1 ln (GDP per Capitai,t) + β2 ln (GDP per Capitaj,t) + β3Conflicti,t + β4Freei,t + β5ln(Distance)i,j + β6Landlockedi + β7Common Languagei,j + β8i Former Colony of ji,j + β9ln(1 + Bilateral Aid)j-->i,t + β10ln(1 + Aid Type)j-->i,t + δi + δj + δt + εi,j,t
Continuous variables are expressed using natural log to allow for interpretation of the results as percent changes from proportional movements in the explanatory variables (World Bank 2018). The dependent variable is the migrant stock that originated in country i and is resident in country j at time t. I use migrant stock instead of migrant flow, because I can control the unique characteristics of specific time periods, destination and origin. Migrant flow measures the number of individuals entering/leaving a country during a specific period, whereas migrant stock captures the number of migrants at a given point in time. The research question is whether aid is effective in reducing the drivers of migration, but the decision to migrate depends on time, origin and available destination countries. Using migrant flow would wash out these variables in this analysis. This migrant stock data has only 2010 and 2015 estimates and the OECD’s aid data covers 2007 to 2015. To extract the most data and acknowledge that some time is needed to see aid’s effects, I split the aid data into two periods: 2007 to 2010 and 2011 to 2015 and use corresponding migrant stock from 2010 and 2015 from the UN as the dependent variable. The time periods are unfortunately unevenly split with period 1 having 3 years while period 2 has 4 years. Thus, the time periods are not meant for comparison, but rather additional information to look at the aid-migration link in two time periods.
To control for each years’ unique characteristics within each time period such as the presence of more conflict or uneven crop yields, I added a fixed effect δt to control for this. I also use fixed effects with origin δi and destination δj. I add to my regression equation to capture the network effects derived from bilateral aid from one country to another. We add ln(1 + Bilateral Aid)j-->i,t to our regression equation to capture the network effects derived from bilateral aid. The terms ln (GDP per Capitai,t) and ln (GDP per Capitaj,t) are added to reflect the abundant literature which says economic potential is a key driver of migration. GDP per capita is a proxy for potential wages migrants may earn at the destination country and reflects the economic condition of countries. I add a constant of 1 to GDP per Capita, Bilateral Aid and Aid type3 to work around the small values of these variables which would turn negative when the log is taken. These aid amounts are expressed in 2016 constant USD millions, so a constant of 1 will affect the results negligibly. I also separate my analysis into the categorized aid groups and selected subcategories within the larger category to further examine the differences between aid types. In addition to the aid type as a regressor, I add traditional variables of a gravity model such as distance between countries, the status of political and civil freedoms in a country and dummy variables to capture the following: conflict, landlocked, commonality in language and colonial history to determine the pull factors of migration. These variables have been shown to be strong determinants of migration. This is further evidenced by the regression tables below that show a strong statistical significance of these variables in predicting migration flows.
I use a quasi-poisson model with a log link or Poisson Pseudo Maximum Likelihood (PPML) model. I use PPML as opposed to an ordinary least square model (OLS), based on evidence that OLS overestimates many determinants of migration such as geographic distance between countries (Silva and Tenreyro 2006). OLS exhibits this behavior because migration data often has unequal variability across the range of variables of the predictors. Migration patterns are dependent on many factors and it is often the case that there is zero migration between a specific country pair.
3 Note that Aid is multiplied by a million to convert it to millions for an apples-to-apples comparison with the rest of aid and population statistics