Consider a randomized experiment to compare (one or more) treatment(s) with a control. Since groups are defined by random assignment, we should expect equal variances for a null-experiment (that is, all groups given the control.)
So unequal variance would in some sense be caused by the treatments, maybe by some interaction between treatments and (unobserved or unobservable) unit characteristics. I remember seeing a paper on this, probably references on this site --- but cannot find it now, despite much searching.
Any ideas or references?