This is an idea about what might happen if different types of information were attended to. Consider this thought experiment:
There is an item A, and another item B. A and B both contain sub-features A1...An, B1...Bn. When we attend to A or B, we are in fact binding A1...An, and/or B1...Bn, to represent A, B.
Now, we have limited bandwidth. Which means, we can only process a limited amount of information at any one time. Consider for the moment that we can only process 4 bits of information. So, if we attend to A, we only process A1-A4, and if we process B, we process B1-B4. We could, by way of divided attention, process A1,A2,B1,B4. Assuming that there is minimal cost in having to dissociate between two different groupings of features (which is rarely the case, but lets just assume that this is possible for argument's sake). This also means, we do not process the other information about the other features that are present.
Now, consider another type of processing, or rather, another level. If in fact we process something called A-B. That is, we bring the binding function up from the item level of A and B, to a higher representation that binds both A-B. What would this result in terms of the amount of information we can process at a time?
This is now only 1 bit of information. We would have more capacity left over (3 bits) from our initial 4 bits. Furthermore, within the 1 bit, we might be able to reinstate the original A1-A4, and B1-B4 via past experience. However, we will suffer from interference in this case, since we did not explicitly process A1-A4 or B1-B4, but rather A-B. Thus, there should be a cost of attending to this higher level at the expense of the lower levels. Likewise, there is a cost of attending to the lower levels at the expense of the higher levels. This is also known in the literature as chunking.
Thus, in summary is that the level of binding should be inversely related to the bandwidth, or the amount of information we can process at any one time.