Abstract

The use of Geographical Information Systems (GIS) has gained importance and relevance in statistical fora. The application of grids over time, provides a tool for spatio-temporal analysis which offers a visually effective way to communicate statistical output. Geocoding of the 2021 Population and Housing Census was a regulated requirement, however, for Malta, geocoding of census data was not available prior to 2021. Spatial-temporal analysis between censuses therefore was limited to traditional outputs. The National Statistics Office (NSO) Malta initiated an exercise to geocode the 2005 and 2011 Population and Housing Census data files using the 2021 geocoded Census, as well as a spatial point layer produced inhouse, and innovative artificial intelligence (AI) tools.

For the 2021 Census, x,y coordinates were extracted either by enumerators during the canvassing stage or from a spatial point-data layer (ReBuDS) developed inhouse. The geocoding of the 2005 and 2011 Census was produced through a two-step procedure. The first process involved the assigning of the 2021 census coordinates in cases where an individual found in both or one of the 2011 and 2005 censuses, was found in the 2021 census. The 2021 geocode was assigned if the individual linked through the combination of ID card and street code. In the absence of direct matching, the ReBuDS layer was used. The census records were matched to the spatial point-data layer using street codes. Geocodes which were not assigned through the previous linking procedure were systematically distributed. It is acknowledged that the urban sprawl observed in 2021, the result of increased construction activity and population growth, was not present in the 2011 or 2005 census rounds. To ensure an accurate representation of the realities present during the different reference years, AI – in particular deep learning models – was used to generate building zones for reference year 2011 through the analysis of orthophotos. The refined ReBuDS layer for year 2011 was then used to further filter the point-data layer for the 2005 Census. Coordinates which fell outside the generated building zones were omitted prior to final assignment. Selected outputs will be presented for seven main areas of interest - population, immigration, households, health, employment, education, dwellings - at the 1km2 grid. Spatio-temporal trends will be highlighted across the three census rounds, identifying patterns and clusters.

This project marks an important milestone in spatial statistics for the NSO by extending the scope of spatio-temporal analysis of the Census beyond the 2021 round. The method developed has strong potential for back-geocoding of a wide range of data sources, thereby substantially expanding the capacity for spatial time-series analysis. The ReBuDs layer is already being used to geocode several street-based databases held by the NSO, further developing the potential for spatial analysis across multiple domains.