How the estimates are made

The earlier parts report gradients, rankings and rates, and each depends on choices that a chart does not show. These include how a slope is estimated, what goes into a deprivation index, how large the area is that a number describes, and which death counts are withheld. This part takes those choices one at a time and asks how much each of them moves the results. The methods page lists the rules that every chapter shares.

Chapter 10.1, Gradients, slopes and variance shows how the slope of life expectancy on the deprivation index is estimated, finds that the slope is steeper within counties than across all tracts, and fits a separate regression for each metro, in which life expectancy is lower in more deprived tracts in every one. It splits the variation in tract life expectancy among states, counties and tracts, and most of the variation lies between tracts in the same county. It also shows that a regression run separately for each state returns coefficients of both signs for the same measure, that a correlation changes with the size of the area, and that a suppressed cell is not a zero.

Chapter 10.2, The anatomy of a deprivation index opens up the Area Deprivation Index. It rebuilds the index from census data, shows that home value and family income carry most of the weight in the published formula, and compares the index with rankings that standardize the same measures first. The index reads expensive places as less deprived, and the chapter tests how closely each ranking tracks upward mobility and life expectancy. The standardized rankings track upward mobility more closely than the published index does.

Chapter 10.3, A deprivation index at two zooms builds one index for tracts and for counties and compares what the two say about the same residents. A county's percentile follows the average of its tracts' percentiles, yet the percentile of a resident's tract can differ widely from the percentile of the resident's county, and a county percentile hides a wide spread of neighborhoods. Correlations also change with the size of the area.

Chapter 10.4, Families of places sorts tracts into groups by their deprivation profile instead of ranking them, and checks the groups against outcomes that the sort never used. The groups are regions of a continuum and not separate clusters, and the partition changes with the weights, the scale of the measures and the sample of counties. The chapter finds families that the index scores differently but that show no detectable difference in life expectancy.

Chapter 10.5, Suppressed counties in death data explains the rule that every chapter using WONDER death counts follows. CDC WONDER withholds small counts, the withheld county-years cover many places and few residents, and the places that remain are a selected set. A rate pooled over only the published years differs from the true rate in a direction that depends on the size of the county, so the book uses counties published in every year, or a modeled source, instead. The chapter also covers a second rule, which decides which counties WONDER names at all in its birth products.