Accounting for place: confounding via geography obscures polygenic evidence on mental health and environmental exposures in the UK Biobank

This study demonstrates that explicitly accounting for local geographic context using multilevel models is crucial in UK Biobank analyses, as it reveals that traditional single-level models adjusted only for genetic principal components can produce biased associations between polygenic scores for mental health and environmental exposures like greenspace due to unaddressed geographic confounding.

The Gist

The Big Picture: The "Neighborhood Effect" on Mental Health

Imagine you are trying to figure out if a specific type of seed (a person's genetic makeup for mental health) grows better in a garden (a green, leafy neighborhood) or a concrete parking lot (an urban, built-up area).

For a long time, scientists have looked at this by planting seeds in different gardens and counting how many flowers grow. They noticed a pattern: people with certain genetic risks for depression or schizophrenia often seem to live in less green areas, while those with better genetic wellbeing seem to live in greener places.

But this paper argues that scientists have been looking at the wrong thing. They were blaming the seed for the soil, when really, the soil (the neighborhood) was influencing where the seeds ended up in the first place.

The Problem: "Confounding via Geography"

The authors call this problem "Confounding via Geography." Here is a simple way to think about it:

Imagine you are a detective trying to solve a mystery. You notice that people who live in London tend to have a different genetic profile than people who live in the Cotswolds (a rural area).

• The Mistake: If you just compare Londoners and Cotswolders without thinking about why they are different, you might think, "Oh, living in London causes this genetic difference."
• The Reality: It's not the city that changed their genes. It's that people with certain backgrounds, incomes, and life stories (which are clustered in cities vs. countryside) moved there. The city didn't cause the genetics; the history and economics of the place caused both the type of people who live there and the type of environment they live in.

The paper says that previous studies often forgot to account for this "geography trap." They treated every person as an independent data point, like individual dice rolls, when really, people in the same neighborhood are more similar to each other than to people in the next town over.

The Experiment: The "Magic Lens" (Multilevel Models)

To fix this, the researchers used a special statistical tool called a Mundlak model. Think of this as a magic lens that lets them look at the data in two different ways at the same time:

1. The "Within-Neighborhood" View: This looks at people inside the same neighborhood. It asks: "If two people live on the same street, and one has a high genetic risk for depression, do they live in a less green house than their neighbor?"
2. The "Between-Neighborhood" View: This looks at the differences between neighborhoods. It asks: "Do neighborhoods with a high average of depression genes generally have less green space?"

By separating these two views, the researchers could see if the "green space" effect was real for the individual, or if it was just an illusion caused by the fact that certain types of neighborhoods attract certain types of people.

What They Found: The Plot Twist

When they looked at the data without the "magic lens" (the old way), they found what everyone expected:

• People with genes for depression seemed to live in less green areas.
• People with genes for schizophrenia seemed to live in less green areas.

But when they used the "magic lens" (the new way), the story changed completely:

1. The "Noise" Disappeared: When they looked strictly at people within the same neighborhood, the link between bad genes and bad neighborhoods mostly vanished.
2. The Sign Flipped: For schizophrenia, the results actually flipped! When looking strictly within neighborhoods, people with a higher genetic risk for schizophrenia actually seemed to live in greener places than their neighbors.
3. The "Between" Effect is the Culprit: The reason the old studies saw a link was because of the Between-Neighborhood effect. Wealthier, greener areas tend to have different populations than poorer, urban areas. The old studies were accidentally blaming the individual's genes for the neighborhood's poverty or wealth.

The "Residual" Map: Where the Model Broke

The researchers also looked at the "leftover" errors in their model (residuals). They found that the model struggled the most in extreme places:

• The Peak District and National Parks: These very rural, green areas had huge "errors" in the model.
• Central London: These very urban areas also had huge errors.

This suggests that you cannot simply compare a person in a National Park to a person in London and assume they are "exchangeable" (comparable). The differences between these places are so massive (in terms of history, economy, and culture) that a simple genetic comparison doesn't work unless you account for the massive gap between the two worlds.

The Main Takeaway

The paper concludes that geography is a powerful confounder.

If you are a geneticist studying mental health, you can't just look at a person's DNA and their address and assume the address is a result of their DNA. You have to realize that where people live is often determined by factors that also shape their DNA distribution (like migration patterns, economic history, and local policies).

The Solution:
The authors suggest that scientists should routinely use this "magic lens" (multilevel modeling) to separate the "within-neighborhood" effects from the "between-neighborhood" effects. If the results change drastically when you do this, it means the original finding was likely just a trick of geography, not a true biological cause.

In short: Don't blame the seed for the soil. Sometimes, the soil just happens to be where certain seeds ended up, and we need better tools to tell the difference.

Technical Summary

Technical Summary: Accounting for Place: Confounding via Geography in Genetic Epidemiology

Problem Statement
This study addresses the issue of "confounding via geography" in genetic epidemiological research, specifically within the context of the UK Biobank. While traditional analyses of polygenic indices (PGIs) for mental health traits typically adjust for genetic principal components (PCs) to control for population stratification, they often fail to account for local geographical context. The authors argue that because both mental health outcomes and environmental exposures (such as greenspace) exhibit strong geographical patterning, common causes may cluster by location. Ignoring this structure can induce bias, leading to spurious associations between genetically predicted traits and environmental exposures. The paper posits that treating individuals as independent observations in geographically structured datasets risks misattributing variance explained by geographically structured confounds (e.g., local policy, demographic differences) to individual-level genetic exposures.

Methodology
The authors utilized data from the UK Biobank (N ranging from 209,391 to 293,851 depending on the outcome) to investigate the relationship between genetically predicted mental health traits and two measures of residential greenspace:

1. Percentage Greenspace: Land-use derived data within a 300m radius.
2. Normalised Difference Vegetation Index (NDVI): Satellite-derived greenness within a 500m radius.

The study constructed PGIs for three traits: depression, wellbeing, and schizophrenia, using summary statistics from Genome-Wide Association Studies (GWAS). These PGIs were generated across a spectrum of p-value thresholds (from $p < 0.5$ down to genome-wide significance $p < 5 \times 10^{-8}$) to test the robustness of signals under varying levels of SNP filtering.

To address geographical confounding, the authors compared two modeling approaches:

1. Single-level Linear Regression: Standard models adjusting for age, sex, and the first 25 genetic PCs.
2. Multilevel Mundlak Models: These models decompose the effect of the PGI into:
   • Within-area effects: The association between an individual's PGI and greenspace, holding the area mean constant.
   • Contextual (Between-area) effects: The association between the mean PGI of a neighbourhood and the greenspace outcome, holding individual characteristics constant.

This formulation allows for the explicit modeling of geographically structured variance rather than treating it solely as a nuisance. All models were adjusted for 25 genetic PCs.

Key Results

• Phenotypic Associations: Consistent with prior literature, phenotypic analyses showed that individuals with depression or schizophrenia diagnoses, or lower wellbeing, tended to reside in areas with less greenspace.
• PGI and Greenspace (Single-Level): In single-level models using less strictly filtered PGIs (higher p-value thresholds), genetically predicted depression and schizophrenia were associated with lower greenspace, while wellbeing was associated with higher greenspace.
• Attenuation and Sign Reversal: When stricter p-value thresholds were applied to the PGIs, the associations attenuated significantly. Notably, for schizophrenia, the sign of the association changed entirely in stricter models.
• Within-Area vs. Single-Level: The within-area estimates derived from the Mundlak models for less-filtered PGIs aligned more closely with the estimates obtained from strictly filtered PGIs in single-level models.
• Contextual Effects: The study found strong negative contextual effects, particularly for schizophrenia. This suggests that living in an area with a higher average schizophrenia PGI is associated with less greenspace, whereas within a specific area, individuals with a higher schizophrenia PGI may actually be closer to more greenspace.
• Geographical Residuals: Analysis of model residuals revealed that geographically naïve single-level models performed poorly in extreme environments, such as National Parks (high residuals) and central London (low residuals), indicating that these models fail to capture structural differences between these distinct geographical contexts.

Key Contributions

1. Demonstration of Bias: The paper provides empirical evidence that ignoring geographical context in genetic epidemiology can lead to biased estimates of the relationship between genetic predispositions and environmental exposures.
2. Methodological Framework: The authors advocate for the routine use of multilevel Mundlak models to decompose PGI effects into within-area and contextual components. This approach helps distinguish between individual-level causal pathways and geographically structured confounding.
3. Sensitivity Analysis Proposal: The study suggests that comparing single-level estimates with within-area estimates (analogous to a Hausman test) serves as a critical sensitivity analysis. If these estimates diverge, it signals the presence of confounding via geography that is not captured by standard genetic PC adjustments.
4. Recovery of Genetic Signal: The results indicate that explicitly modeling geographical context can help recover robust genetic signals from noisier, less-filtered PGIs, suggesting that the "true" individual-level effect may be obscured by between-area confounding in standard analyses.

Significance and Claims
The paper claims that the observed associations between mental health PGIs and greenspace in previous studies likely reflect confounding via geography rather than direct causal pathways. The authors argue that genetic epidemiologists working with spatially structured populations must routinely investigate local context. They contend that assuming conditional exchangeability between individuals in vastly different geographical settings (e.g., central London vs. the Peak District) is often incorrect.

The study concludes that while single-level models adjusted for genetic PCs are insufficient, the inclusion of local context via multilevel modeling offers a mechanism to infer about complex causal structures and mitigate bias. The authors recommend that such geographically informed sensitivity analyses become a standard practice in the field to prevent the misattribution of genetic causes to upstream environmental or social origins.