I ran this for a couple of months in my own time, before I trusted my own conclusion. Strip out the noise and the actual reply rate accounts for most of the variance I saw. It stood out to a degree that mattered after I controlled for the obvious variables. Check it against your own case genuinely, before committing any real time.