Machine-Learning Geographic Sampling Frames Underrepresent People on the Move in Fragile Settings

Political science research often studies difficult-to-locate populations: internally displaced persons (IDPs), returnees, migrants, and others who move during war, disaster, or crisis. We compare machine-learning (ML)-generated building footprints against high-resolution satellite imagery for constructing sampling frames. In displacement- affected communities in Iraq, 15.5% of ML points were invalid or duplicated, and manual additions constituted 19% of the final hybrid frame. Manual additions reached 10.2 percentage points more IDPs in rural communities and 10.1 more returnees in urban areas. We attribute differences to two mechanisms: temporary and informal shelters are harder for ML algorithms to detect, while post-bombardment reconstruction can create structures absent from the imagery underlying ML frames. ML-only frames produce systematic coverage error, and targeted satellite validation can reduce it.