Reference

Reading guide

The rules of evidence, the recurring patterns, the parts of a figure and the units of analysis.

Four kinds of claims

Each chapter says what kind of evidence stands behind a claim. The book uses four kinds, and each asks more of the data than the one before it. Uncertainty belongs to all four, so every claim of every kind carries its interval, its sample size or its caveat, and every chapter ends with what its data cannot show.

Kind of claim	The question it asks	What a chapter can say
Descriptive comparison	Where is an outcome high or low, and how does it change from the least to the most deprived places?	That an outcome differs across places or moves with deprivation, and by how much. Maps and gradients are descriptive comparisons.
Adjusted comparison	Does the pattern remain among places or people alike in age, race, urban or rural setting, state or county?	That a gap persists or shrinks once those differences are held fixed. Holding measured differences fixed leaves unmeasured ones in place.
Proposed explanation	Which pathways fit the pattern?	That a pathway is consistent with the data, and which other pathways are consistent with them too. Consistency with the data does not test a pathway.
Study with a stated identification strategy	Did a randomized program, a policy change or another design separate cause from selection?	A causal claim, from a cited study, inside a box labeled "Causal evidence" that names the design.

A causal box names its design in plain words. A difference-in-differences compares places that changed a policy with similar places that did not, before and after the change. A regression discontinuity compares people just on either side of a rule's cutoff. A randomized experiment assigns a program by lottery. A box labeled "Study evidence" reports a study that follows the same people over time; following people rules out some explanations, and on its own it does not separate cause from selection, so those boxes use descriptive verbs. Outside a box a chapter uses descriptive verbs such as "is higher where" and "moves with". Neighborhood data show where disadvantage is concentrated, and a cause needs a stronger design than a map.

Seven patterns that recur

Seven patterns run through the chapters. Each is a question the data can answer either way, and a chapter reports a pattern where the data show it and says so where they do not. A pattern that has a reading path links to it.

Disadvantage as a dose. An outcome changes step by step across the whole range of neighborhood deprivation, with no level below which the neighborhood stops mattering. Chapters plot the outcome against deprivation percentile and ask whether the line is straight, bent or flat.
A local law. A slope in the national average can come from a mix of unlike places. Chapters estimate the slope separately within each metro area, state or county and ask whether it points the same way in all of them.
Hyper-local differences. A county average describes no neighborhood in particular. Chapters split the variation in an outcome into what lies between states, between counties and inside counties, to show how much an average hides.
The exception as a diagnostic. An outcome that does not follow the gradient, such as a flat line, a reversed slope or a gap that no map explains, shows what the gradient measures and what it leaves out. Chapters report these exceptions and do not draw a line through them.
Different terms in the same place. Two groups raised in the same tract can reach different outcomes. Chapters compare groups inside the same tract or county, which holds the neighborhood fixed, and report what remains. Chapter 9.4, Places that beat the model asks the matching question about places that do better than their circumstances predict.
Concentrated burdens and honest nulls. Chapters measure how much of a burden sits in how few places and whether prices follow the map, and they report a result of no relationship, or of counts withheld from many places, as a result. See Chapter 6.2, Gun homicide and the causes of early death, Chapter 4.1, Home value, income and redlining and Chapter 10.5, Suppressed counties in death data.
One institution, different programs. The clinic, Medicare, Social Security, the school and the earned income tax credit each run under one national rule. Chapters ask whether the same institution delivers the same thing at every level of deprivation.

How to read a figure

A figure has a title that states its finding, a subtitle that says what is plotted (the measure, its units, the population and the period) and a caption that begins with what to look for. The source line under the chart names the dataset, the years and how the estimates are built: weighted, population-weighted, unweighted, full counts, crude rates or modeled estimates. The methods page defines each.

A shaded band or a whisker means one of two things, and the subtitle says which. Around an estimate, such as a mean, a rate or a slope, it is a 95% interval, a range built by a method that captures the true value in nearly all repeated samples, so a wider interval means a less certain estimate. Around a set of places, such as the tracts in one tenth of deprivation, a band shows how far those places differ from one another, usually as the interquartile range, and it does not narrow when more places are added. A figure built from full counts, such as every death in every county, has no sampling error and draws no interval, and its note says so. Full counts remove only that one source of error: a rate still depends on how deaths are coded and how the population is estimated, and a modeled estimate still carries the model's error.

Gray means no data. A county or tract with no value is drawn gray with a hatch, and its tooltip gives the reason, such as a count withheld for privacy or a place the source does not cover. The book never draws a missing value as zero.

Lines and bars are labeled directly where they can be, and color is never the only cue: the labels, the hover text and the "Show data" table under each chart carry the same information. Every number in the text comes from a data file. Hover over it or focus it to see its unit, population, period, source and, where it has them, its interval and sample size.

What grain means

Grain is what one row of a chapter's data describes, and it limits what the chapter can claim. Part pages and the contents list label each chapter with its grain.

Grain	One row is	Claims are about
Individuals	a person in a survey	the people surveyed and the groups they belong to
Neighborhoods	a census tract, or the area of a ZIP code	the neighborhoods themselves
Counties	a county	counties
States	a state	states
Institutions	a school district or a college	the institutions themselves

A census tract is a small area drawn by the Census Bureau, usually close to what residents call a neighborhood. A county is larger and holds many tracts that can differ from one another. A person is the smallest grain, and only chapters built on surveys of people work at it. A ZIP code is a mail route with no official boundary. Where a source reports by ZIP code, the book uses the Census Bureau's ZIP code tabulation area and says so.

A pattern across tracts can differ from the pattern among the people who live in them. Where the same question is asked at two grains, the chapters point to each other: Chapter 5.1, Health behaviors and the supply of care asks it of neighborhoods and Chapter 5.2, Health behaviors, person by person asks it of people. The methods page explains the limits of comparing places. The data sources page lists each dataset, and the glossary defines each term of art.