6 Caveats, problems, things to look out for
Replication: This guide is intended to be replicable with various types of geo-spatial data sources and country contexts - we would welcome all such replication exercises and hope that our material can be used to produce small area estimates of poverty in countries other than Bangladesh! That said, the exact raster manipulation as well as processing steps needed will depend on your specific geo-spatial data. Always check the key elements of your raster data to inform the processing you will need to do - our procedure is not to be simply copy pasted (though it provides many useful functions you can work with). CRS, spatial resolution, extent, number of cells, and dimensions are always good things to check.
Time period of geo-spatial data: whenever possible, try to obtain geo-spatial data for the same time period as the household survey data that contains the poverty variable (or any other outcome you are interested in modelling). If that is not possible - say that you can only access geo-spatial data that is older than your poverty data - you should consider whether this information is outdated, or could still be considered relevant (for instance the elevation of mountain ranges is unlikely to have changed much over time, but demographic data will change quite quickly).
Choice of poverty predictors: This list of predictors of poverty is only indicative. You should consider, when replicating the poverty modelling exercise in your country context, what sources of geo-spatial data are available and relevant to poverty prediction.
Level of administrative disaggregation: Although in this example we consider administrative level 3 in Bangladesh, small area estimation of poverty using geo-spatial data can be replicated at any other administrative level (higher or lower) - and this will vary depending on the country. Note however that the level of aggregation will inform your modelling choices (see Section B for more details).
Resampling algorithm: The choice of the algorithm used for resampling matters - and depends on the information represented by your geo-spatial data, as well as whether you are aggregating (moving to a lower spatial resolution), or disaggregating (moving to a higher spatial resolution). For example, when you aggregate demographic data - you should use the “sum” algorithm.