U.S. 2016 Unadjusted Exit Poll Discrepancies Fit Political, not Statistical, Patterns

[Takket]

10. he's showing a probability dome

the peak of the dome is the most likely outcome, which is the actual vote matching the exit polling data. it is also likely that the exit polling is off by .1% or .2%. Right? Math tells us (by analyzing margins of error in the exit polling data) what a reasonable deviation from the exit polling data would be. If you look at figure 3, the author has made this + or - 2.5%. Meaning that if exit polls said Clinton 50.5%, it is reasonable to assume her actual total will be between 48% and 53%. He calls this the 95% confidence interval. This means that if the election were held 100 times, that 95% of the time Hillary's total would be somewhere in that interval (48% to 53%).

Her actual total of 47.6% falls outside this interval, meaning it is highly unlikely that the exit polling data was that far off.

That being said, there are 50 states, so the concept of a state being outside the confidence interval is not only believable, it is probable. The problem though is that several states are falling outside this confidence interval, meaning that "the odds of this happening are 5%" event is having many times, which is highly improbable. imagine you have 5 white balls and 95 red balls in a bin, and one at a time, you keep drawing white balls but no red ones. pretty unlikely right? The data is basically showing that white balls keep being drawn.