Although I admit it depends, in part, on assumptions that might not hold true*, it's looking like the party-of-four stats are based off a sample size of approximately 2500 parties, meaning there's an expected value around 2 for the "uniform distribution of non-repeat compositions" assumption. If we take into account that most four-person groups won't have both a sorcerer and a wizard and similar such things, we can reduce even further the number of expected "popular" parties, potentially down even as far as only 81 or so "popular" options and the rest literally unseen because the sample size isn't big enough.

*Specifically, I'm assuming that we can guess that the .65% to .53% gap is a gap of 3 parties, and the .53% to .49% is a gap of 1 party. This is a pretty reasonable assumption, as it would be quite strange for there to be large and exclusively even(/multiple of 3 etc.) gaps in these values--we expect relatively small gaps. Also, I assumed that these values given were exact decimal percentages, which is very unlikely, they were almost certainly rounded. But by this reasoning from the difference-of-3 between the first two clusters and a difference-of-1 between the second and third cluster, I got the same value both ways, approximately 2500. I think we can say with confidence that that's about the size of the data set for four-person parties, and it's very unlikely that any of the other data sets are particularly *larger* than that.