if you look at figure 2 and don't see something worth massive investigation, you don't need to be here. You're either not understanding the charts, or you're on someone else's team.
Ouch. I look at figure 2 and scratch my head because I'm not sure what to expect from graphing the cumulative difference between a candidates vote total and the average of his delegate race totals, plotted against the cumulative average of total votes cast at a precinct with precincts sorted by the total votes cast.
But I'm not one to give up THAT easily. So I first tried to make sure I could replicate the result, to see if I at least understood the calculation being performed even if I didn't have a prior expectation for what that calculation should yield. I initially had it wrong, too, because I hadn't noticed that the y-axis is also cumulative. My result doesn't have a sharp "knee" though. Is that just the effect of having straight lines superimposed over it? I calculated the difference at each precinct, then graphed the cumulative sum of that. Anyway this is close:
I didn't have any prior expectation for what this graph would look like so it's hard to be surprised. But I can try to dig further and understand what's going on.
I also did a scatter plot, not cumulative on either axis:
The vertical blue line separates the precincts that come before the 300,000 "crime happens here" point from the ones that come after. Now I'm not clear on what flipping theory says would be done to the votes at each precinct, but it's a small number of points (with big vote counts) that are doing a lot of the work here.
I started wondering about those large precincts that were doing the work here. Was there perhaps some geographical clustering, like there was in Va Beach City? I wondered how the votes on the right-hand side of a cumulative graph were distributed around the state, compared to the votes on the left-hand side. A calculation that I posted earlier showed that the top few counties account for a significantly larger proportion of the 100 largest precincts than they do overall. Not surprising really. But I wanted to look closer.
I grabbed the 2011 census data for Alabama counties. They range in size from 9k to 659k. I sorted the precincts by vote count, and with a moving window of 100 precincts I plotted the percentage within that window that came from the 20 (out of 68) most populous counties:
Can I get a "just wow"? No? Maybe if I drew some straight lines to emphasize a slope change at around the 300,000 point where it flattens out?
What does this tell us? Over on the left-hand side of the many cumulative graphs we've seen, the precincts are predominantly from the less-populous regions of Alabama. That gradually changes, and right around the "crime" point of 300,000 the geographic composition has changed to being around 80% from the most-populous regions of Alabama, and continues that way to the end of the graph. After you cross that 300k point nearly half the counties in Alabama are not represented at all, and just 5 counties account for 58% of the remaining precincts even though those 5 account for only 24% of all precincts. (I also did the same graph with county population density. Very similar graph.)
People keep saying that the demographic argument is dead. With this graph in mind, tell me why that is. Why would you expect there NOT to be demographic differences from a sample weighted heavily toward the least-populous areas of the state (the left hand part of the graph) and another sample weighted heavily from the most-populous areas of the state? It's true that even a large county can have some tiny precincts, right next to very large precincts. But where those tiny precincts from populous counties are showing up on the left side of the graph, they're far outnumbered by tiny precincts from sparsely populated areas.
And considering that the curve of the Mitt minus dels graph slopes more sharply upward (when suitably smoothed on both the x and y axes) over the same region where the third graph flattens out because from there on out we're drawing >80% of the precincts from the most populous counties, can anyone at least entertain the possibility that there could be a demographic difference at work here? That if your samples are 80% from the most populous counties you might just be looking at some significant demographic differences compared to samples drawn mostly from less populous counties?
