The default null hypothesis in typical statistical modelling software is that a parameter's value is equal to zero. However, this may not always correspond to the actual conditions that would hold if the effect of interest did not exist. In two case studies based on recent research in cognitive science and linguistics, we illustrate how data simulation can shed light on unspoken, sometimes even incorrect, assumptions about what the null hypothesis is. In particular, we consider two experiment-specific measures of changes in probability distributions where the null condition is not always equal to zero change; and an investigation of a cognitive bias for skewed distributions based on the assumption that, without such a bias, distributions would always remain uniform. All in all, simulating null conditions not only improves each researcher's understanding of their own analysis and results, but also contributes to the practice of "open theory". Formalising one's assumptions is, in itself, an important contribution to the scientific community.

Authors:

Aislinn Keogh: University of Edinburgh; Elizabeth Pankratz: University of Edinburgh