Much Ado About Missingness: A Demonstration of Full Information Maximum Likelihood Estimation to Address Missingness in Functional Magnetic Resonance Imaging Data.

Eric Stice,Cara C. Tomaso,Timothy D. Nelson,Rebecca L. Brock,Sonja Yokum,Cary R. Savage

doi:10.3389/fnins.2021.746424

Abstract

The current paper leveraged a large multi-study functional magnetic resonance imaging (fMRI) dataset (N = 363) and a generated missingness paradigm to demonstrate different approaches for handling missing fMRI data under a variety of conditions. The performance of full information maximum likelihood (FIML) estimation, both with and without auxiliary variables, and listwise deletion were compared under different conditions of generated missing data volumes (i.e., 20, 35, and 50%). FIML generally performed better than listwise deletion in replicating results from the full dataset, but differences were small in the absence of auxiliary variables that correlated strongly with fMRI task data. However, when an auxiliary variable created to correlate r = 0.5 with fMRI task data was included, the performance of the FIML model improved, suggesting the potential value of FIML-based approaches for missing fMRI data when a strong auxiliary variable is available. In addition to primary methodological insights, the current study also makes an important contribution to the literature on neural vulnerability factors for obesity. Specifically, results from the full data model show that greater activation in regions implicated in reward processing (caudate and putamen) in response to tastes of milkshake significantly predicted weight gain over the following year. Implications of both methodological and substantive findings are discussed.

Highlights

Functional magnetic resonance imaging paradigms can offer powerful insights into neural processes with particular relevance to health and well-being (Insel et al, 2013)
We focused on the identification of survey variables we anticipated would be significantly correlated with blood oxygen level dependent (BOLD) signals
The variables that correlated at a meaningful level (i.e., r > 0.20) with BOLD signal – age and DEBQRestrained Eating – were retained for use as auxiliary variables in full information maximum likelihood (FIML) analyses

Summary

Introduction

Functional magnetic resonance imaging (fMRI) paradigms can offer powerful insights into neural processes with particular relevance to health and well-being (Insel et al, 2013). Suboptimal approaches to handling missing data have been the norm for most published fMRI studies, with simplistic strategies such as listwise or pairwise deletion representing the most common approach1 These approaches introduce multiple potential problems, including reduction of sample size (and power) and potential bias in parameter estimates (Enders and Bandalos, 2001). In the context of neuroimaging studies in particular, suboptimal handling of missingness may undermine the integrity of theoretical frameworks in the field of behavioral neuroscience and complicate the application of fMRI biomarkers to guide interventions (Mulugeta et al, 2017). It is increasingly clear that effectively dealing with missing data in fMRI studies is a critical step toward addressing concerns about reproducibility and rigor; examples of modern missing data approaches with robust fMRI datasets are rare

Objectives

Methods

Results

Conclusion