To update our effort from 2008, let’s see what we might know when the first polls close.
At 7pm, the polls will be closed in the following states: KY, GA, IN, NH, SC, VT, VA.
Let’s list these in order of projected Trump/Clinton vote share: KY, IN, SC, GA, NH, VA, VT.
I’ll use Kremp’s updating program to compute Trump and Clinton’s probabilities of winning, under his model, for several different scenarios.
First, with no information except the pre-election polls:
> update_prob() Pr(Clinton wins the electoral college) = 90% [nsim = 100000; se = 0.1%]
Clinton has a 90% chance of winning
Now let’s consider the best possible scenario for Trump at 7pm, in which he wins Kentucky, Indiana, South Carolina, Georgia, New Hampshire, or Virginia (but not Vermont, cos let’s get serious):
> update_prob(trump_states = c("KY","IN","SC","GA","NH","VA"), clinton_states = c("VT")) Pr(Clinton wins the electoral college) = 2% [nsim = 1340; se = 0.4%]
Next-best option for Trump, he wins all the states except Virginia and Vermont:
> update_prob(trump_states = c("KY","IN","SC","GA","NH"), clinton_states = c("VA","VT")) Pr(Clinton wins the electoral college) = 28% [nsim = 3856; se = 0.7%]
Most likely scenario, Trump wins Kentucky, Indiana, South Carolina, and Georgia, but loses New Hampshire, Virginia, and Vermont:
> update_prob(trump_states = c("KY","IN","SC","GA"), clinton_states = c("NH","VA","VT")) Pr(Clinton wins the electoral college) = 93% [nsim = 88609; se = 0.1%]
Or Trump just wins Kentucky, Indiana, and South Carolina:
> update_prob(trump_states = c("KY","IN","SC"), clinton_states = c("GA","NH","VA","VT")) Pr(Clinton wins the electoral college) = 100% [nsim = 5240; se = 0%]
P.S. Kremp writes:
If you remember, I have a polling error term in my forecast, so all polls can be off in any given state by the same amount. And these polling errors are correlated across states. I picked a 0.7 correlation—which may be a bit high. It made the model more conservative about Clinton’s chances, but today, it’s going to make it jump to conclusions when the first results come in.
Interesting. I’m not sure if Kremp’s model really is overreacting: I’d guess that errors across states will have a very high correlation. I guess we’ll see once all the data come in.
P.P.S. I forgot Kentucky in my first version of this post. Kentucky was never going to be close so including it does not change the numbers at all. But for completeness I updated the code.
This seems to make the election depend on New Hampshire to an implausible degree.
Beamish:
It’s not that the election depends on New Hampshire, it’s that the outcome in New Hampshire is informative of the national swing.
Well, it seems to make the outcome in New Hampshire informative to an implausible degree. But maybe not. You’re the expert.
Beamish:
To be fair, we haven’t tested this model so we could have some bugs…
Actually, that makes sense – most of the uncertainty is about polling errors. The only way Clinton loses is if polls nationwide are biased against Trump, and he does systematically better than expected. If he’s not doing systematically better across the board than the polling showed, he’s going to lose.
To make the obvious update, a year later, the errors were regionally correlated, not nationally correlated.
I’m glad you summarized this information.