Two time points poorly capture trajectories of change: A warning for longitudinal neuroscience

Emerging neuroimaging studies investigating developmental changes in the brain aim to collect sufficient data points to examine trajectories of change across key developmental periods. Yet, many published studies using these data include only two time points, while data collection is ongoing. We offer conceptual and empirical arguments that this approach will often fail to meet the stated goal. First and foremost, analyzing data when subsequent waves are planned is analogous to “peeking” at data with half the sample collected. Practical concerns of funding and pressure to publish might counterbalance these downsides if analyses of early waves are good indicators of full developmental trajectories. We demonstrate empirically that this is unlikely to be the case, because of the imprecision of two time point change models to characterize underlying trajectories of change. Even under ideal simulated conditions (i.e., high signal-to-noise ratio, known functional form), two time points are not sufficient to accurately capture individual differences in trajectories of change. We generated underlying longitudinal data and fit models with 2, 3, 4, and 5 time points across 1000 samples. Recovery was poor for the two time point model, correlating typically at r = 0.41 with the true individual parameters -- meaning that these scores share only 16.8\% of variance -- and even lower reliability (ICC(2) = 0.28). As expected, models with more time points recovered the growth parameter more accurately. Yet, parameter recovery for the three time point model was still low, correlating typically around r = 0.57 (ICC(2) = 0.51). We argue that preliminary analyses on early subsets of time points in longitudinal analyses is detrimental to the ultimate goals of studies of change over time.