Discordance Between Spatial and Population Correlations From Human Brain Imaging Data

TL;DR

This paper warns that correlations between brain imaging data across regions may be misleading and not reflect true biological relationships across individuals.

Contribution

The study demonstrates that across-regions correlations in brain imaging data can be biased and misinterpreted as meaningful biological relationships.

Findings

Across-regions correlations showed significant positive results, but region-specific participant correlations were not significant.

Simulations revealed that regional mean and variance differences can bias correlations and increase false positives.

The paper argues that across-regions correlations are ambiguous and not reliable proxies for participant-level relationships.

Abstract

It has become increasingly common to probe correlations between human brain imaging measures of receptor/protein binding and function using population‐level brain maps, typically drawn from independent cohorts to estimate correlations across regions. This strategy raises issues of interpretation that we highlight here with both an empirical multimodal brain imaging dataset and simulation studies. Twenty‐four healthy participants completed neuroimaging with both [11C]Cimbi‐36 positron emission tomography and magnetic resonance imaging scans to estimate receptor binding potential (BP) and cerebral blood flow (CBF), respectively, in 18 cortical/subcortical regions. Correlations between BP and CBF were estimated in four ways: (1) Pearson correlation across regions of mean regional BP and CBF from a single or separate cohorts (ρ1.1 and ρ1.2, respectively), to mimic studies using data from independent cohorts; (2) Pearson correlation between BP and CBF across participants in each region (ρ2); or (3) the correlation between BP and CBF across participants across all regions within a single linear mixed effects model (ρ3). We observed a significant positive correlation across regions (ρ^1.1 = 0.672, p = 0.0023; ρ^1.2 = 0.659, p = 0.0030). Region‐specific correlations across participants were substantively lower and not statistically significant (ρ^2: mean = 0.140, range = −0.112–0.336; all p > 0.10), nor when estimated simultaneously within a linear mixed model (ρ^3 = 0.138, p = 0.26). Our simulation study illustrated that regional differences in BP or CBF mean and variance can substantially bias across‐regions correlations and inflate the type‐1 error rate. Our observations allude to ambiguity in the meaning of across‐regions correlations and suggest interpreting them as evidence for a biological relation, which implies a relation across participants, is problematic. Without validated methods that handle confounding and other biases, we urge caution in how future studies interpret across‐regions correlations of population‐level brain maps.

Correlation analyses across regions made between population‐level brain imaging maps are common yet raise issues of interpretation.We show that such correlations do not align with across‐participants correlations, evidencing high bias and inflated type‐1 error rates.Across‐regions correlations have an ambiguous meaning and are not straight‐forward proxies for across‐participants correlations.

Correlation analyses across regions made between population‐level brain imaging maps are common yet raise issues of interpretation.

We show that such correlations do not align with across‐participants correlations, evidencing high bias and inflated type‐1 error rates.

Across‐regions correlations have an ambiguous meaning and are not straight‐forward proxies for across‐participants correlations.

We show that the across‐regions correlation of population‐level brain maps does not align with the typically desired across‐participants correlation. Across‐region correlations are variably biased with potentially large type‐1 error rates, raising issues of their interpretation and generalizability.

Figures