2025 · Nature Communications
Global gridded population datasets systematically underrepresent rural population
Verdict
SCIENTIFIC
The study follows the scientific method and does not operate inside a theoretical framework. The verdict is scientific.
1·Study methodology · no theoretical framework
SCIENTIFIC
Follows the scientific method. No theoretical framework attached.
2·Paradigm · Basic Epidemiological Comparison
NOT APPLICABLE
No theoretical framework. The study verdict decides.
How the verdict is decided
This study does not operate inside a theoretical framework. Using hospital records, billing codes, or diagnostic names as labels on visible events is not the same as depending on that framework. The verdict is the study methodology score alone.
1·Study methodology
Did this study test its claim with methods that are independent, falsifiable, and non-circular?
SCIENTIFIC
This study claims that five global population maps underestimate rural populations by 53% to 84% when checked against the number of people actually relocated from 307 areas flooded by large dams. All three tests pass.
Independently Verifiable
Pass
On-the-ground surveys counted the people physically moved out of 307 valleys before dams flooded them, and anyone can compare those headcounts to the numbers printed on a map without accepting any modeling theory.
Falsifiable
Pass
If the map headcounts inside the 307 reservoir boundaries had matched or exceeded the number of relocated residents, the claim of rural underestimation would have been disproven, and nothing in the design blocked or explained away positive numbers where they occurred.
Non-Circular
Pass
The resettlement headcounts came from local dam-planning surveys conducted before flooding, completely separate from the census models and satellite rules used to build the five global population maps.
Why
When a valley is flooded for a large dam, every person living inside the reservoir footprint has to move out. Dam authorities record those relocated residents from ground surveys taken before the water rises. The authors took resettlement counts from 307 dams across 35 countries and matched each reservoir to its outline on satellite water maps. Because satellite snapshots often catch reservoirs below full water level and miss 18.8% of their reported surface area on average, the authors boosted the map headcounts inside each outline by 23% to give the maps credit for the missing shoreline. They then pulled the population numbers from five global maps for the years before dam completion and compared them to the actual relocation counts. Every map missed more than half the residents.
2·Paradigm · Basic Epidemiological Comparison
The three tests do not apply. There is no theoretical framework.
NOT APPLICABLE
Why
This study does not rely on a theoretical framework to make its comparison. The five global population datasets being tested are built from computer models that use night-time lights, road networks, building footprints, and census totals to guess where people live. The authors do not assume those allocation rules are true. Instead, they treat the five maps as predictions on trial and test them against an outside physical record: the number of people who actually had to be moved out of 307 rural valleys before dam reservoirs filled with water. Subtracting a ground headcount from a map's published number for the same piece of land is straightforward arithmetic on visible events, so there is no underlying paradigm to score.
From the study
The resettlement numbers typically stem from comprehensive on-the-ground surveys as part of impact assessments carried out during the planning and authorisation procedure of dam construction projects .
An inquiry we sent to ICOLD Central Office confirmed that the numbers represent only physical resettlements from areas that were later inundated and occupied by reservoirs and do not include secondary displacements of people residing outside the reservoir areas, e.g., due to livelihood loss induced by the project (personal communication, 28.8.2023).
Figure expand the analysis to the other reference years and reveals a systematic and significant negative bias in all five datasets, ranging from −53.4% (WorldPop), −65.0% (GWP), −66.9% (GRUMP), and −68.4% (LandScan) to −83.8% (GHS-POP).