Impact of analytic decisions on test–retest reliability of individual and group estimates in functional magnetic resonance imaging: A multiverse analysis using the monetary incentive delay task

Abstract

Abstract Empirical studies reporting low test–retest reliability of individual blood oxygen-level dependent (BOLD) signal estimates in functional magnetic resonance imaging (fMRI) data have resurrected interest among cognitive neuroscientists in methods that may improve reliability in fMRI. Over the last decade, several individual studies have reported that modeling decisions, such as smoothing, motion correction, and contrast selection, may improve estimates of test–retest reliability of BOLD signal estimates. However, it remains an empirical question whether certain analytic decisions consistently improve individual- and group-level reliability estimates in an fMRI task across multiple large, independent samples. This study used three independent samples (Ns: 60, 81, 119) that collected the same task (Monetary Incentive Delay task) across two runs and two sessions to evaluate the effects of analytic decisions on the individual (intraclass correlation coefficient [ICC(3,1)]) and group (Jaccard/Spearman rho) reliability estimates of BOLD activity of task fMRI data. The analytic decisions in this study vary across four categories: smoothing kernel (five options), motion correction (four options), task parameterizing (three options), and task contrasts (four options), totaling 240 different pipeline permutations. Across all 240 pipelines, the median ICC estimates are consistently low, with a maximum median ICC estimate of .43 – .55 across the 3 samples. The analytic decisions with the greatest impact on the median ICC and group similarity estimates are the Implicit Baseline contrast, Cue Model parameterization, and a larger smoothing kernel. Using an Implicit Baseline in a contrast condition meaningfully increased group similarity and ICC estimates as compared with using the Neutral cue. This effect was largest for the Cue Model parameterization; however, improvements in reliability came at the cost of interpretability. This study illustrates that estimates of reliability in the MID task are consistently low and variable at small samples, and a higher test–retest reliability may not always improve interpretability of the estimated BOLD signal.