I noted that all cohorts showed an unexpected 5% leap in divorce from 1999 to 2004. Prof. Stevenson elaborated in a gracious response to my query:
The problem is that for every group there is a fraction (about 10%) who haven't had a chance to make it to the last anniversary assessed by the census. The 2004 survey was conducted from July to September 2004, and hence it is impossible for around one-in-ten of those surveyed to have reached the last assessed anniversary. For instance, a couple who married in October 1994 were counted as part of the 1990-94 marriage cohort, but even if they stayed together forever, they could not have reached their tenth anniversary by the survey date. Thus the percent in that cohort reaching the 10th anniversary was understated. This problem affects all of the numbers along the diagonal--these numbers reflect only the percent of couples in a group that have celebrated that anniversary. Those not "making it" can be divorced, dead, or simply still waiting for their anniversary to roll around.
Stevenson and Wolfers will be releasing a more thorough critique soon. I plan to report on it when it appears.