Assessing the robustness of brain connectivity findings in neonates - A multiverse analysis
Poster No:
295
Submission Type:
Abstract Submission
Authors:
Leonardo Zaggia1, Daniel Kristanto1, Andrea Hildebrandt1
Institutions:
1University Of Oldenburg, Oldenburg, Niedersachsen
First Author:
Co-Author(s):
Introduction:
Replication is a cornerstone of scientific research, but achieving robust results that can be replicated remains a challenge in neuroimaging. The problem stems from the high degree of flexibility in data processing and analysis choices, known as the "researcher's degrees of freedom", which can affect scientific results. Multiverse analysis addresses this problem by systematically exploring the impact of alternative analytical choices on the results of hypothesis tests. By running analyses through a range of defensible pipeline options, multiverse analysis tests robustness, a core aspect of replicability. This study examines the robustness of findings on how prematurity affects neonatal brain connectivity using multiverse analysis. The focus is on the audio-visual integration (AVI) brain network, a key system for perceptual and ultimately socio-cognitive development. Using fMRI data from the Developing Human Connectome Project (dHCP), we tested the robustness of the association between gestational age at birth and the information processing efficiency in the AVI brain network across a wide range of pipelines. The pipelines differed in the possible combinations of alternative preprocessing steps, modelling choices and sampling decisions.
Methods:
fMRI data from 301 neonates from the dHCP dataset were analysed across 11.232 different pipelines. These pipelines differed in their preprocessing and analysis choices: global signal regression (yes or no), functional connectivity metrics (correlation, partial correlation or covariance), density thresholds (13 levels), handling of negative correlations in the connectome creation(retain, zero-out or transform to absolute), edge resolution (binarization or normalization), efficiency graph metrics (local or global efficiency), statistical model flexibility (linear regression or step-wise-spline regression with varying polynomial degree: k=1, k=2, or k=3), and sample balancing(n=301, n=150, and n=95) of the full vs. preterm neonates ratio, testing the additional effect of the sample size and distribution of cases across gestational age. For each pipeline, we calculate the strength of the association between the degree of prematurity (quantified as gestational age) and the AVI network efficiency in infants. We use R² as an index of association strength and model fit.
Results:
Across pipelines with identical sample size and model flexibility, the association between gestational age and AVI brain network efficiency was consistently weak, with R² values ranging from .01 to .015 in most pipelines, consistent with the findings of Quinonnes et al. (2024). In terms of model flexibility, regression splines with linear or second-degree polynomials slightly improved fit over simple linear regression. However, adding higher degree polynomials had no additional benefit, suggesting that the association between gestational age and AVI network efficiency is not highly non-linear. Balancing the ratio of full-term to preterm infants slightly increased the R² (see Figure 1), highlighting the importance of a uniform gestational age distribution for detecting small associations. Nevertheless, the overall association is weak. Variation in the direction of the association was also found, further reinforcing the lack of a true association.
Conclusions:
In this work, we show how multiverse analysis can be used across a wide range of analysis choices as a valuable approach to testing the robustness of findings in neuroimaging. Specifically, we tested the robustness of a very weak association between gestational age and AVI network efficiency described by Quinones et al. (2024). This study highlights the importance of methodological transparency and robustness checks in neuroimaging research. The weak association between gestational age and AVI network efficiency appeared largely independent of analytical choices. Theoretical as well as methodological implications will be discussed.
Disorders of the Nervous System:
Neurodevelopmental/ Early Life (eg. ADHD, autism) 1
Lifespan Development:
Normal Brain Development: Fetus to Adolescence
Modeling and Analysis Methods:
fMRI Connectivity and Network Modeling 2
Task-Independent and Resting-State Analysis
Perception, Attention and Motor Behavior:
Perception: Multisensory and Crossmodal
Keywords:
Design and Analysis
FUNCTIONAL MRI
PEDIATRIC
Statistical Methods
Other - Multiverse analysis
1|2Indicates the priority used for review