How Precise Is the Q-Matrix?
A quintile is a band, not a point: what bounds the error, what does not, and how it fared against local figures.
FAND places every Root by two ratios, people per km² and income per capita, and reads each by the population-weighted fifth it falls in (the Q-Matrix). A place’s fifth is a band, not a point. Two things bound the error; two do not.
What bounds it
- Country totals hold. Each country’s places are scaled so that their population-weighted average income equals the country’s GNI per capita (Atlas method). An income error in one place is offset elsewhere in the same country; the country’s level is not in doubt.
- A fifth is a wide band. Each middle fifth spans about a factor of two in income and two to three in density. A place’s figure has to be badly wrong to move it more than one fifth.
What does not
- Spread within a state. A place with no income figure of its own carries its state’s or its nation’s. That compresses the spread within the state and pulls its places toward the state’s fifth: errors cancel in the country’s total, not place by place. About one person in five lives in such a place; among the densest fifth of humanity, one in three.
- Density, and relatives. The country scaling works on income only. Places whose land area is still inferred from their parent’s density (426 million people in 70 countries) carry that error untouched. And scaling fixes the level, not the relatives: a district’s product divided by a population of another year stays wrong relative to its neighbors.
- Cut-points. A place near a cut-point moves on a small error, and a better figure for some places moves the cut-points for all: a block of places sharing one figure can change fifth with no change in its own data.
The test against local figures
674 places in India and China that carried their state’s or province’s income were given their own district or prefecture product. 64 percent stayed in their fifth, 35 percent moved one, and 5 places (0.7 percent) moved two or three. The largest move, Gautam Buddha Nagar (Noida), counts output where it is produced against the people who live there.
In Egypt the earlier figures were survey-based regional relatives, not a parent’s figure. Official governorate product spread them from 1.3 to 1 to 12 to 1, and moved 14 of 23 governorates one fifth: most moved, none by more than one.
Across both tests, 99 percent of places stayed within one fifth. The headline findings held: the densest, poorest cell went from 100 to 99 places, the sparsest, richest from 2,782 to 2,760, and the measurement asymmetry stayed at 28 to 1.
Two limits on the test. It swaps a borrowed or survey-based figure for an official one; it does not test imagery, which has yet to show it can rank places as well. And it covers the states that publish district product; three large Indian states whose district figures are not openly published are outside it.
What follows
Better local data should mostly re-rank places within one fifth of where they sit. Where a parent’s figure was borrowed, two places in three keep their fifth; where a different construct is replaced, as in Egypt, most move, but by one. That is why the public Where Explorer shows quintiles, not figures.
Directions for further work: give the 1,131 places that borrow their income a reading of their own, testing lights at night where people are crowded and land cover where they are sparse against the places that already publish local product; and fill the land areas still inferred from a parent’s density.