I'm not sure that these experiments tell the whole story. I took one of these tests once, and I couldn't reliably detect distortion in a 15-second sample of a song unless it was -30dB or greater. However, I do notice that over time, when I listen to equipment that has higher distortion, I am more likely to reach for the volume knob, and eventually shut if off, even when that distortion is below my -30dB threshold. This is independent of whether or not there are tonal imbalances and whether or not they have been corrected.
It would be interesting if a study could be conducted in which the participants would be divided into different THD% groups, were given access to a diverse library of music, all pre-distorted to a fixed THD, and told they could listen to whatever they want, adjust the volume as needed and quit whenever they got tired of listening. The researchers would then compile data for each group for average listening volume overall, average listening volume at start, after 10 minutes, 20 minutes, etc., total listening time, number of songs listened to from start to finish, and the number of songs that were started but ended before completion. Ideally, the size of the music library and number of participants would have to be quite large. Even more ideally, there could be groups assigned not only to different levels of THD, but also emphasis on even-order, odd-order, and higher-order harmonics at each level, as well as IMD.