Nice overview. Thanks Keith.
There seems to be a bad link in for GFYS in:
Signalling that Management, Commerce and Arts should GFYS: Other higher-cost subjects may see an increase in funding if necessary.
of am I just being punished for not using Chrome? I presume this is Go Fund Your Self? ;-)
I feel proud that this has finally made it into law, but that is tinged with a bit of sadness that our friends across the ditch have practically no chance of following suit when Captain Catholic and the Conservatives get back into power. *sigh*
Maybe this will help with visualizing the 92% thing as a distribution. I just noticed on the right hand side of the methodology page there is a graph called Electoral Vote Distribution. It shows the probability that Obama receives a certain number of Electoral College votes. Sorry I don’t know how to easily imbed that graphic here, you have to go and look for yourself.
There are three obvious peaks, the highest at about 20% is 330-ish (it’s such a small graph it’s hard to read). That’s your distribution. I presume it’s so lumpy because of the effect of FPP in each state (except 2). Yet not too lumpy because different states get different numbers of votes.
The to left of 270 are all the outcomes in red which elect Romney, to the right are all the blue outcomes which elect Obama. Add up all the probabilities on the red side and you would get about 8%, and on the blue side gives you about 92%. Minus something for “exact tie” and because we’ve been rounding along the way.
I am of course embarrassed to say I read all the words on the left and didn’t look at all the interesting graphics on the right. I used to be such a proponent of graphical data analysis, yet I got stuck into the words. Too many tea leaves for me.
And then I remembered this about what the 72% means back from when the number de jure was 72% rather than 92%. There are some interesting links there to follow up.
Note I don’t have trouble with anybody saying “72% chance of winning” and “too close to call” and believing in both (Red Queen anybody?) for a different reason (at least I think it’s different) from the ones given. I’d just say the “too close to call” is in fact a personal probability statement which reflects how much somebody might bet on the outcome (if at all) and at what odds. For me that’s a different domain from the 72%. The “too close to call” comes from how risky you as an individual are willing to be – and that gets us back to “how certain do you want to be that the real percentage is 72%?” Every “hard number” turns into a pool of probability underneath if you look closely enough. Fortunately in this case the cat is still alive when the box is opened. Even if the fox isn’t even aware that it is dead.
Some context is that in "gold star" science you usually want to be 95 or 99 percent certain. Nate went with 92%, but he is more of a poker player I suspect.
The number of combinations of parties isn’t too huge. The number that could form a government is even less. But how to work out what the chances are of one of the combinations being more likely than another is where I lose the plot. A matrix of friendliness between the parties?
Perhaps a matrix like that could be polled off the population, using the assumption that people that vote for a party are represented by the representatives, so some kind of averaging of their individual matrices of preference could represent what the representatives are likely to feel. How to assign confidence to these figures, though? Would need a lot more past data than I can presume exists.
Actually, I was thinking more of having to create a neural simulation of the brain of Winston Peters, given that I seem to remember he has prior form in saying something like "Vote for me. I would never go into coalition with Party A" and later going into coalition with Party A. The idea of that puts me right off trying.
Which, on my understanding, is a rewording of the very definition of “X% likely”. Is there a better word for what this kind of confidence used in the 538 modeling is? Or is a phrase like “92% likely” simply something that doesn’t convey any clear information because the definition of %likely is not set?
I don't know exactly what the 92% likely to refers to. Hopefully it means there is a 92% chance that the statement "Obama gets 270 or more Electoral College votes" is true. What's missing of course is reporting of confidence intervals and such. Remember that when I talked of enumeration of scenarios before, that was when I hand't read the methodology section and thought more of it was done by simulation modeling based on what had been said here.
Maybe the whole basis of what the 92% means has been formally stated somewhere and I haven't seen it. But in most situations you wouldn't lay out the careful definition of all your probability statements in something for general public consumption. I'm sorry to be the bearer of bad tidings, but Journalists can't handle something as complex as percentages and more generally, numbers with denominators in many cases.
But while we are linking to StatsChat here is an interview re: predicting things accurately here in NZ
Brad Luen has put together an open script for aggregating NZ’s polling data, weighted to actual election results and DimPosted. It’s an excellent effort. Unfortunately we lack the fine polling that the US has, which would be equivalent to polling at the electorate level. Does anyone know if this has been done by parties in NZ?
Good old R. I quit having anything to do with political polling a few elections ago. I did work on behalf of a party, but all such details are confidential of course. And that's one of the problems: I am 95% confident (you asked a statistician after all *wink* ) that polling of individual electorates is still done in "marginal seats" by parties with enough money. Perhaps they can now afford to do all electorates. But that sort of thing is never made public.
The real fly in the ointment for predicting outcomes in our MMP environment goes deeper than just lack of electorate level polling. We can't always predict how a particular party might cobble together enough seats from other parties to form a Government. That adds another level of uncertainty which is currently in the domain of the Pundits. I wouldn't know where to begin to do a computer simulation of that. Maybe Nate Silver could; he seems much more clever than I am.
So having read the methodology for the 538 project (thanks David Hood), it is more based in doing good meta analysis rather then simulation. The point estimates are done by "traditional statistical methods", as are most of the confidence limits for each point estimate.
The use of simulation seems to be limited to the construction of confidence limits in the case where we want to know the confidence limits for a compound event (N things happening together). This use of simulation is the domain of resampling methods which is what I alluded to earlier. It is basically a way to get confidence limits for a statement like "Obama will achieve N Electoral College delegates" which are corrected for the fact that we are using point estimates for a number of independent races (within each state) but each of the state level point estimates is itself subject to confidence limits.
So we start out with a stand alone proposition like "Obama takes the Electoral College delegates for California" and then we keep adding propositions which creates longer and longer chains of compound events such as "Obama takes the Electoral College delegates for California" AND "Obama takes the Electoral College delegates for Illinois" AND... which leads to having enough delegates to win. As that NY Times decision tree had it, there are a number of scenarios to consider about who takes which state.
This is why concerns about being "wrong N times out of so many coin tosses" get tossed out. That has been corrected for in the calculation of the confidence limits for the proposition "Obama will achieve N Electoral College delegates" as a compound event. Note that there may well be two levels of this correction going on, one for counties (districts within states -- I guess that's "electorates" to us) and one at the state level. I may have simplified that discussion too much.
Would that I could express it better so that the "intelligent reader" would not have their eyes glaze over.
Thank you. Now I have no excuse not to get on it right away.
Keeping some of the finer detail of some of his adjustments private isn't playing the academic game properly, but if he wants to sell his results to somebody to earn a crust (he did say he was a liberal libertarian or something was it?).
In case anybody is still interested in the 92% all these hours later...
I'm your friendly resident statistician. That's the good news. The bad news is I haven't read Nate Silver's methodology yet. But if the description of it as a Monte Carlo simulation is right then I can help you. I used that technique for my Masters Thesis, although they made me do "real work" for my PhD because "simulation" wasn't considered real enough back then before the revolution.
Then they add up the number of times they see a given result, in this case 92% of the times they ran the model it said Obama won.
Exactly so. The 92% is an enumeration of the outcomes for the simulation in which Obama won as a percentage of all outcomes. This stands in for a probability. Again, the important thing is that the model is properly constructed and range testing on assumptions is adequate. It is even possible (but much more work) to construct confidence limits for the 92%.
But you simply can’t treat it like a dice roll or coin toss and say they’ll be wrong once every 8 or 9 elections. It just isn’t that kind of stats.
Spot on. The concerns about being wrong 1 in N elections are not well founded. Sounds like the law of averages at work to me. 2016 is not another "coin flip" following on from 2012, and there is no chain of probability. Once you get past very simple probability models it is advisable for professionals like doctors and lawyers to consult professional statisticians. We wouldn't want another Meadow incident would we?
And I wouldn't be as kind (diplomatic?) as Russell is:
I would just like to point out that when Corin Dann explained this story to One News viewers this week, he said that Nate Silver had forecast that “Obama will win by 92%”.
This data journalism thing clearly has a way to go yet.
This kind of reporting does not meet the professional standards for accuracy. Allowing "Obama has a 92% chance of obtaining enough Electoral College votes to become the President" to morph into “Obama will win by 92%” isn't good enough. Sloppy. Not uncommon. But still Sloppy.
Thank you Russell, what a great laugh.
Thankfully Jon Stewart jumped in and wrapped up the interview with Nate Silver before Nate finished telling the Republican strategists the list of things they needed to change to win the next election. Not that they would necessarily listen, but hey, let's not make it any easier.
"Statistical thinking will one day be as necessary for efficient citizenship as the ability to read and write." – H.G. Wells