Abstract
Reports of crises of reproducibility have abounded in the scientific and popular press, and are often attributed to questionable research practices, lack of rigor in protocols, or fraud. On the other hand, it is a known fact that-just like observations in a single biological experiment-outcomes of biological replicates will vary; nevertheless, that variability is rarely assessed formally. Here I argue that some instances of failure to replicate experiments are in fact failures to properly describe the structure of variance. I formalize a hierarchy of distributions that represent the system-level and experiment-level effects, and correctly account for the between-and within-experiment variances, respectively. I also show that this formulation is straightforward to implement and generalize through Bayesian hierarchical models, although it doesn't preclude the use of Frequentist models. One of the main results of this approach is that a set of repetitions of an experiment, instead of being described by irreconcilable string of significant/nonsignificant results, are described and consolidated as a system-level distribution. As a corollary, stronger statements about a system can only be made by analyzing a number of replicates, so I argue that scientists should refrain from making them based on individual experiments.