But true to form, it looks like FiveThirtyEight has actually increased Beto's chance of winning from 27.9% to 28.6%. They turned Cruz's 6-point lead into a 3.9-point lead and decreased his margin in the polling average. I understand how they did it, but I've been a little skeptical of their methodology.
Sometimes, I think algorithmic complexity can just get in the way of the obvious.
I think you're seeing a malicious intent where there isn't one. The reason that it was shifted down to a 3.9% lead is because of YouGov's previously established "House Effects". Essentially, 538 is looking at polls Yougov has done in previous elections and has found that they are, on average, 2.1% too favorable to the Republican candidates. Also, if they were really trying to make it look like O'Rourke was doing better then he actually was, why adjust Public Policy Polling, Quinnipiac University, Reform Austin, etc. away from O'Rourke?
Finally, doesn't it seem naturally counterproductive for a odds website to purposely obfuscate what's going on? From a business perspective, wouldn't that just make them lose traffic? If anything, they should be adjusting toward O'Rourke to make it look like more of a horse race then it actually is.
I don't see any malicious intent, it's just that I've found their model to be odd at times. And I've always questioned their use of house effects to adjust polls. House effects evaluate an individual poll against the aggregate of all other polls. For example, a poll that shows R+4 versus a polling average of R+1 has a house effect of R+3. An alternative is to use mean-reverted bias, which evaluates polls against actual election results.
It's actually possible for a pollster to have a house effect in favor of one party while having a mean-reverted bias in favor of the other party. FiveThirtyEight's article on pollster ratings explains it like this:
FiveThirtyEight chooses to use house effects for their model, but I've done my own experimentation and have found that using mean-reverted bias comes up with more accurate results.