Actually, this study (http://ops.fhwa.dot.gov/wz/workshops/accessible/McCoy.htm) lends credence to the claim. Late merges have better throughput than early merges when encouraged by traffic signs and uniformly adopted.

The study does not directly address the effect of "cheating" late merges which can create waves, but did note that traffic signs which strongly encouraged early merging by everyone increased overall transit time through construction. That suggests that the default case is better.

That makes a little sense to me: a single lane can bear less traffic, so extending the amount of distance everyone has to spend in a single lane might result in higher latency overall, if not necessarily lower throughput.