Development and validation of the CanHS: The Canada housing stock datasets for buildings and postal codes

The prevalence of different types of housing stock is rapidly changing in North America. Associations between housing stock type and social, economic, and environmental outcomes have yet to be fully explored, in part because of a lack of available data. More precisely, research into housing and social, economic, and environmental outcomes is important for public planning but is constrained by a lack of publicly available national datasets. This study introduces the Canadian Housing Stock (CanHS) datasets. The 2024 National Address Register (NAR), a dataset of all valid georeferenced civic addresses in Canada, was used as the base dataset (N= 11,690,445 buildings, N=17,293,867 units). For each NAR address, we predicted the number of storeys and classification as low-/mid-/high-rise using a combination of model- and rule-based methods. Model-based methods included linear regression, decision trees, random forests, and XGBoost. Rule-based methods used an ad-hoc series of logic rules based on apartment numbers to predict the number of storeys. Each method was trained and validated using open datasets (four cities). For buildings <50 units, random forests performed the best in predicting the number of storeys for each building: 63% perfect matches, 96% within 1 storey, and 98% classified correctly as low-/mid-/high-rise). For buildings ≥50, the rule-based algorithm performed the best: 61% perfect matches, 87% within 1 storey, and 92% classified correctly as low-/mid-/high-rise). We constructed two datasets: CanHS-PC and CanHS-BL. The CanHS-PC classifies all residential postal codes in Canada as single-unit, low-/mid-/high-rise based on the most prevalent housing stock type in the postal code. The CanHS-BL includes the building location (address), predicted number of storeys, and predicted classification (low-/mid-/high-rise) of all residential addresses in Canada. Overall correlations between the total number of occupied dwellings (2021 Census) and units (CanHS-BL) were very strong (low-rise r=0.904; mid-rise/high-rise r=0.995). The datasets can be used at the postal code or building location level to conduct rigorous, national research related to housing stock and social, economic, and environmental outcomes. For instance, spatial data on housing stock type can now, for the first time, be linked to individual health data (e.g., the Canadian Community Health Survey), social data (e.g., Canadian Social Survey), or environmental data (e.g., environmental compliance reports). The developmental method documented here may also serve as a blueprint to be adapted in other socio-geographic contexts for the purpose of deriving large-scale housing repositories where such datasets are lacking.