Assessing the potential of social media data to support remote sensing data in migration analysis – An explorative modelling approach for Ghana, 2015-2020

Kurzfassung

Human migration in West Africa is a complex phenomenon driven by both environmental and social factors. In Ghana, north-south migration has been linked to rainfall variability and soil degradation. In a previous study, remote sensing data and microdata from Ghana’s 2021 census were used to measure the migration effectiveness index (MEI). This index is the net to gross migration ratio and measures the ability to attract, avoid or balance migration in a way that is comparable between administrative units. The aforementioned study reached an R² of 52.5% using purely remote sensing data-driven factors. Including additional factors derived from census data increased the R² to 60%, a substantial improvement. Census data is a valuable resource, but it is often unavailable in many countries. Even when it is accessible, its limited frequency and lack of comparability across countries present significant challenges. Social media data could potentially fill this gap with its capacity to match Earth Observation data in frequency and comparability. If geolocated social media data could serve as a substitute for some census variables, this would allow migration researchers to assess physical and socioeconomic factors at flexible time and spatial scales. This, in turn, could lead to a better understanding of the impacts of climate change, natural disasters, and policies. In this follow-up study, therefore, we examined if the socioeconomic variables from the Ghana census can be substituted with geolocated social media data from Twitter (now X). From a dataset of 20,179,946 geolocated social media posts collected for Ghana over a 5-year period, we calculated the following statistics at district-level: Total number of posts, number of authors, diversity of languages, and trend in the districts’ share of all posts over time. The majority of posts (90.5%) were found to be geolocated at relatively low precision (level of cities or coarser) and were excluded to prevent distortions between rural and urban districts. Our analysis was twofold. Firstly, we assessed correlations between Twitter statistics and the socioeconomic metrics from the census. Secondly, we applied a modelling approach to predict the MEI for recent migration based on residence five years earlier. In this modelling approach, we included the same drivers that were previously identified in a stepwise multiple linear regression procedure and measured the improvement in explained variance (R²) gained by adding Twitter statistics. Results showed that twitter statistics correlated with many of the census variables. Of the tested combinations of variables, more than half (51 of 77) were significant at the 95% level. To give some examples: Particularly strong were the correlations between proportion of the economically active population and post- and author counts, and language diversity (r = 0.76, 0.75, and 0.67 respectively). Further, proportion of population employed in the agricultural sector was inversely related to post and author counts (r = -0.80 and -0.72 respectively), and the proportion of population born outside the region was inversely related to the proportion of English posts (r = -0.39). Adding these twitter statistics into the linear model leads to a marginal improvement of up to 0.5% adjusted R², compared to 8.6% that is gained by including census variables. When excluding districts with low amounts of posts (typically less populated districts), the improvement is larger, with up to 4.4% improvement in R² gained by including the logarithmic number of authors per district. Our study shows that substituting census variables with social media statistics is promising but not straightforward. The high correlations with census-derived variables suggest that social media statistics contain information related to socioeconomic situations, which can contribute to a better understanding of migration. However, this information does not seem to relate to migration effectiveness to the same degree as the census data, which is likely to remain a precious and valuable data source. While we do not find social media statistics to be a direct substitute for census-data in modelling migration effectiveness, it is a promising data source in itself, as we find it to be plausibly related correlated to socioeconomic statistics such as employment rate, proportion of in-migrants, and the importance of the primary sector. Future research should investigate ways to address the limitations of social media data and develop targeted statistics to address specific knowledge gaps in migration research.