Selection bias and permutation tests in fMRI Restricted; Files Only

Abstract

The exclusion of high-motion participants in functional Magnetic Resonance Imaging (fMRI) studies is a common practice to reduce motion-related artifacts. However, this exclusion can introduce biases by altering the distribution of clinically relevant variables, leading to a non-representative study sample. This paper aims to introduce a framework that employs the Average Inverse Probability Weighted Estimator (AIPWE) method to address these biases by treating excluded scans as missing data. Using simulated datasets, we tested the AIPWE method on single-region and multi-region scenarios with varying block correlations to evaluate its effectiveness in addressing selection bias. Our results demonstrate that the AIPWE method effectively mitigates the impact of the selection bias in these simulations, providing more accurate estimates of functional connectivity. We applied the AIPWE method to real-world data from 396 children aged 8-13 (144 with autism spectrum disorder and 252 typically developing) from the Autism Brain Imaging Data Exchange (ABIDE) datasets. Our findings reveal that autistic children are more likely to be excluded compared to typically developing children, suggesting that the generalizability of previous studies may be limited due to the selection of older children with less severe clinical profiles. To address data loss and resulting biases, we adapted the AIPWE method in conjunction with an ensemble of machine learning algorithms. The proposed approach identified more edges with differing functional connectivity between autistic and typically developing children compared to the standard approach, highlighting the potential of our framework to improve the study of heterogeneous populations where motion is prevalent. Overall, this study underscores the importance of addressing selection bias in fMRI studies and demonstrates the utility of the AIPWE method in enhancing the reliability and validity of functional connectivity analyses.

Table of Contents

Introduction 1

2 Statistical Methods 2

2.1 Parameter of Interest and Target Parameter . . . . . . . . . . . . . . . . . . . . . . . 2

2.2 AIPWE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2.3 Permutation Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

3 Simulations 5

3.1 Single region . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

3.2 Multiple regions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

3.2.1 Strong block-wise correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

3.2.2 Correlation from a seed-based analysis . . . . . . . . . . . . . . . . . . . . . . 9

4 Data and Methods 13

4.1 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

4.1.1 Study population . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

4.1.2 Phenotypic Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

4.2 rs-fMRI acquisition and preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . 13

4.2.1 Anatomical Data Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . 14

4.2.2 Functional and Anatomical Data Preprocessing . . . . . . . . . . . . . . . . . 15

4.3 Seed correlations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

4.3.1 Motion quality control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

4.3.2 Parcel correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

4.4 Data normalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

4.4.1 Covariates description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

4.4.2 Site harmonization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

4.4.3 Balancing diagnosis-independent variables . . . . . . . . . . . . . . . . . . . . 19

4.5 Impact of motion QC on the sample size and composition . . . . . . . . . . . . . . . 19

4.5.1 Impact of motion QC on group sample size . . . . . . . . . . . . . . . . . . . 19

4.5.2 rs-fMRI exclusion probability as a function of phenotypes . . . . . . . . . . . 19

4.5.3 Impact of motion QC on distributions of phenotypes among children with usable data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

4.5.4 Functional connectivity as a function of phenotypes . . . . . . . . . . . . . . 20

4.6 Application of AIPWE in abide data . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

4.6.1 Procedure flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

4.6.2 Procedure details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

4.7 Data and code availability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

5 Results 22

5.1 Impact of motion QC on the study sample and sample bias . . . . . . . . . . . . . . 22

5.1.1 The impact of motion QC on sample size can be dramatic and differs by diagnosis group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

5.1.2 The relationship between rs-fMRI exclusion probability and phenotype and age . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

5.1.3 Phenotype representations do not differ between included and excluded children 24

5.1.4 Phenotypes are also related to functional connectivity . . . . . . . . . . . . . 25

5.2 Application: Deconfounded Group Differences in the KKI and NYU Dataset . . . . 25

6 Discussion 29 6.1 Differences between simulation setting 1 and real data . . . . . . . . . . . . . . . . . 29

6.2 Motion quality control bias . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

6.3 Assumptions and Potential Violations . . . . . . . . . . . . . . . . . . . . . . . . . . 30

6.4 Overview and outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

7 Acknowledgement 32

About this Master's Thesis

Rights statement

• Permission granted by the author to include this thesis or dissertation in this repository. All rights reserved by the author. Please contact the author for information regarding the reproduction and use of this thesis or dissertation.

School	Rollins School of Public Health
Department	Biostatistics
Subfield / Discipline	Biostatistics - MPH & MSPH
Degree	M.S.P.H.
Submission	Master's Thesis
Language	English
Research Field	Biology, Biostatistics Biology, Neuroscience
Palavra-chave	AIPWE Move control fMRI selection bias
Committee Chair / Thesis Advisor	Benjamin Risk, Emory University
Committee Members	Raphiel Murden, Emory University Ying Guo, Emory University

Última modificação

Primary PDF

Thumbnail	Title	Date Uploaded	Actions
	File download under embargo until 18 May 2025	2023-04-09 20:47:11 -0400	File download under embargo until 18 May 2025

Supplemental Files

Thumbnail	Title	Date Uploaded	Actions
	File download under embargo until 18 May 2025	2023-04-09 20:47:17 -0400	File download under embargo until 18 May 2025