Last week, I wrote this article about trying to catch witnesses doing bad things. In it, I analyzed all the blocks produced by the top-producing witnesses over a long range of block production. During that time, almost all the witnesses had nearly identical statistics, except for two: 
It turns out that neither of those two witnesses had been witnessing full-time during the whole study period. The study period began with
The basis of my analysis assumed that each witness should see the same exact number and split of transactions as every other witness. Thus, if we look at which transactions each witness actually includes in blocks, we shouldn't be able to tell blocks produced by one witness apart from those produced by another witness.
However, the actual data ended up looking like this:
See how
Why probably had a higher proportion of market orders than the others
It turns out that at the start of the study period (at block #17,000,000),
So is probably ok?
Right. Here's the same plot over a shorter time window that started around block #18,000,000: (covering about 4,000,000 blocks)
So what's the right way to approach this in the future?
One improved way to do the statistics I was trying to do in my earlier post would be to look at each witness individually, and compare him/her to other witnesses in the same witness rounds. Part of the way Steem works is that block production is organized into rounds, or sequences of 21 consecutive blocks in which each witness produces exactly 1 block. Statistically, each witness should only be compared to other witnesses that shared rounds. That would be one way to unbias the results, because it would ensure that you're only comparing witnesses to others who should look the same.