Shared DNA Tracks Neighborhoods and Health Gaps

Ancestry labels draw boundaries through continuous variation. This study instead begins with stretches of DNA inherited from common ancestors and asks how the resulting groups relate to place and health. ^[1]

The finding

Relatedness beyond ancestry labels

Genetic studies usually sort people by ancestry labels or by self-reported race. The authors point out that such categories draw hard lines through continuous variation, and that automated ancestry labels depend on which reference populations were used. Self-reported race is also often used in U.S. clinical settings as a stand-in for social and environmental factors. ^[9]

A different approach starts from identity by descent (IBD). Two people share an IBD segment when they inherited the same stretch of DNA from a common ancestor. Each generation, recombination (the shuffling of DNA when egg and sperm form) breaks these stretches into shorter pieces, so two people who share more IBD DNA are, on average, more closely related. Earlier work used IBD groups to study population history and, in a Los Angeles health system, to link groups to diagnoses. That study analyzed 35,968 patients, found 376 IBD clusters, and reported 1,218 significant diagnosis associations. ^{[5, 8]}

According to the paper, what was missing is a way to ask how much of a group's health difference runs through measured environmental exposures. ^[9]

Measuring shared surroundings

Health data are often sorted by race labels that blur together ancestry and surroundings. This study shows a way to group people without reference panels and to ask, in numbers, how much a gap runs through measured environments. Its most useful result may be the unexplained remainder, which marks where better exposure data are needed. Groups defined by relatedness describe populations, not any single person.

The evidence

Building groups from inherited DNA

The data come from the Biorepository for Integrative Genomics (BIG), a Tennessee cohort that links genomes, health records, and neighborhood data. The detailed figures in this post come from the full text of the bioRxiv preprint of the paper. ^[9] The study analyzed 13,143 participants from BIG, a predominantly pediatric cohort, and the preprint reports that 87% of BIG participants overall were under 18. ^{[1, 9, 7]}

Step 1: find shared DNA. The software hap-ibd scans every pair of people for shared segments, and its authors report it handling 485,346 UK Biobank samples in 24.4 hours. ^[3] The BIG study kept a 2 centimorgan minimum for the starting seed and removed segments under 0.6 centimorgans. A centimorgan is a unit of genetic distance along a chromosome. ^[9]

Step 2: build a network. Each person is a dot. A line connects two dots, and its weight grows with the total length of DNA they share. ^[9]

Step 3: find communities. The Leiden algorithm sorts the network into groups of densely linked people. It improves on its predecessor, Louvain, which can produce poorly connected or even disconnected groups. ^[2] Applied twice, it gave four communities and 17 subcommunities. ^[1] Clustering on principal components, a standard summary of genetic variation, is a common alternative but can miss recent relatedness. The authors also tested nearby settings and a Louvain run, and the groupings stayed close. ^[9]

Step 4: link to neighborhoods. Each participant's ZIP code was matched to public data on poverty, smoking, obesity, urban residence, unemployment, and fine-particle air pollution (PM2.5). ^[9]

Step 5: split the risk. For asthma, dermatitis (inflamed skin), epilepsy, and influenza, the authors used an interventional mediation analysis. It estimates how much of a group's extra risk would change if its members' exposures were set to the reference group's values. The part flowing through the chosen measured factors is the environmental effect. The remainder is called the independent factor. Comparisons were against a reference subcommunity, with age and sex as covariates. ^[9]

The authors also computed E-values, which state how strong an unmeasured confounder (a hidden factor linked to both group and illness) would have to be to erase an observed link. ^[4]

Neighborhood exposure and asthma differences

Groups match ancestry, not perfectly. Communities lined up closely with inferred ancestry, with a Cramér's V (a 0 to 1 agreement score) of 0.81. Some people who sit between clusters on a principal components plot were still assigned to one community by IBD. ^[9]

Relatives live close. About 49% of first- and second-degree relatives shared a neighborhood, against 4.45% for relatives eight or more degrees apart. ^[9]

Health tracks pollution across groups. Across subcommunities, PM2.5 and respiratory-condition prevalence were strongly correlated (Spearman 0.85). The same measure showed almost no link to neurological conditions (0.10). This is a correlation among a small number of groups, not proof that air causes illness. ^[9]

Most of the gap stays unexplained. The mediation analyses reported 1,757 people with asthma, 742 with dermatitis, 1,502 with epilepsy, and 3,378 with influenza. Averaged over the nine subcommunities of one community (labeled C2), asthma risk was 9.9 percentage points above the reference. On average, 33% of that ran through the measured environment, and about 67% did not. For dermatitis the shares were 26% and 74%, and for influenza 28% and 72%. ^[9] For epilepsy, the mean total risk difference across the C2 subcommunities was 2.6 percentage points below the reference, with subcommunity estimates ranging from 6.7 to 0.7 points below. ^[1]

Race does not erase it. Adjusting for self-reported race shrank the asthma and dermatitis effects but not the influenza effect, and some subcommunities kept significant effects. In a simulation where exposures were equalized, the model projected about 33 fewer asthma cases per 1,000 people in high-risk subcommunities. ^[9]

ZIP-code data blur individual exposure, which pushes the leftover share upward, so it is an upper bound. The mediation assumptions cannot be tested in observational data, and hidden confounders may remain. The cohort is a single urban pediatric cohort from Tennessee, so results may not transfer elsewhere. ^[9] One fair reading, offered by this post and not by the authors, is that the leftover share is a catch-all and the design cannot say what is inside it. This post did not replicate any analysis.

Future directions

The authors propose testing IBD groups as a way to account for population structure in association studies. That use remains open: a comparison with principal-component adjustment could measure whether the community labels change association results and how well they control confounding. ^[1]

They also note that ZIP-level exposure data blur differences between individuals. Repeating the mediation analysis with person-level exposure measures would test how much of the leftover share moves into measured environmental effects. The observational assumptions would still need to accompany that estimate. ^[9]