Thursday, October 15, 2015

if you're in calgary or edmonton, or red deer, you can get a good guess of where conservative support is by taking their 2011 number and multiplying it by 0.6. i've been over the details elsewhere. this is the short answer.

so, if your candidate got 65% in 2011, you can guess where they're at now by calculating 65*.6 = 39.

there are a lot of seats in play. 75*.6 = 45.

i can't tell you who's leading in any specific riding. but if you're in edmonton or calgary, and you want to vote against the conservatives, you're being given a chance to have your vote count that you may never see again, if you don't wipe them out. and, i think you *can* wipe them out this time. even kenney. even harper.

don't squander it.

www.huffingtonpost.ca/2015/10/15/calgary-confederation-riding_n_8299638.html

Shana Proulx
you can also check the riding by riding predictions at threehundredeight.com for an idea of who is best positioned to beat Harper in your riding.

jessica amber murray
no!

that model is not going to work well this election, as it bases support on the previous election and distributes it uniformly. that's the long version.

one example is edmonton centre. the liberals performed higher there in 2011, so the models suggest the liberals are ahead. but riding polls suggest it's a two-way race between the conservatives and ndp.

if you follow the models, it will produce vote splitting.

consult riding polls if you can find them. do not rely on models - they are going to be very flawed, this time around, because we're seeing massive, unpredictable swings almost everywhere.

i've sent emails to leadnow to get more riding polls in calgary and edmonton. i hope they do it. it could be the difference between albertans taking advantage of the election of a generation, or them banging their heads against the wall for the next fifty years.

there are riding polls here.

but be careful of dates; don't take anything done before oct 1 seriously, and take the newest ones the most seriously.

https://en.wikipedia.org/wiki/Opinion_polling_in_the_Canadian_federal_election,_2015_by_constituency#Alberta

Steve Furr
And therein lies the problem with strategic voting -- ppl can't even agree on the proper strategy. Heck, they don't even understand the data they are relying on so casually.

The riding model from three hundred and eight does NOT base support on the previous election and distribute it uniformly --- whatever that's supposed to mean.

The model uses a proportional swing to adjust the previous election's results by the difference in each party's regional polling results from their proportion of the vote from the last election. They then further adjust by factors for incumbency, star candidates, etc.

Hard to say that local riding polls are better. In order for strategic voting to work, you have to have a threshold (lead) for the CPC that can be overcome by strategic voters shifting allegiances. In many / most cases, this will be a three-way race.

There's no way a local riding poll with a small sample (<500) is going to give you a reasonable enough margin of error to correctly guess (in a distributed thinking fashion) who to shift from or to. This is compounded by the relative infrequency in which local riding polls are conducted, meaning they won't often track momentum effectively.

If I had to choose, I'd probably go by 308. I hope it's not a sucker's game.

If you are fortunate enough to live in a riding where there's a clear two-way race between the CPC and Libs or NDP (say, when 308, tooclosetocall and riding polls all agree), then by all means make sure you vote for the second place candidate. If not, good luck strategic voting.

.. and if you live in the Lower Mainland and you're looking at a poll that's more than a week old? Forget it. The momentum shift in regional polls, exceeds anything local riding results were likely to tell you. Ditto Ontario. At least in AB, you're likely to see something more stable (Lib in Calgary; NDP in Edmonton).

jessica amber murray
"The model uses a proportional swing to adjust the previous election's results by the difference in each party's regional polling results from their proportion of the vote from the last election. They then further adjust by factors for incumbency, star candidates, etc."

it uses a uniform proportional swing as a base, adds it to the last election's results and then makes things up regarding incumbency and "star candidates".

i'm going to ignore the making things up part, because it's just making things up, and focus on the parts that i can argue are poor modelling.

first, i will again point you to edmonton centre. the reason the models have the liberals ahead there is simply because they polled higher than the ndp in 2011. is that a reasonable assumption? it can be sometimes. but a model like this would be unable to predict a situation where the liberals are up province wide, and you have a swing from the liberals to the ndp in a specific riding. i want to be clear: i'm not saying it would have a hard time predicting this. it would be unable to predict this. ever.

he might argue it's unlikely. but in a situation like alberta, it really isn't. the conservatives are dominant. historically, the liberal and ndp votes have swung around all over the place. it is never a good idea to try and guess existing support for the ndp or liberals based on previous support in calgary or edmonton. there is too much volatility. the model requires stability, which does not exist there.

now, we need to be clear on the next point: there is no serious debate over whether modelling is preferable to direct riding polling. direct riding polling is always preferable. there are no exceptions to this. experiments are always preferable to theories. and, when experiments contradict theories, we conclude that the theories are wrong. the polling that puts the ndp ahead in the riding proves that the model is not working. it's not a debate. it's hard evidence v. a guess. this is clear, settled and beyond question.

the second problem is the uniformity of the proportional swing. i know that may be confusing to you. uniform proportional? you thought it was proportional, not uniform. it's not proportional in the sense of how it distributes, it's proportional in the sense of how it measures. it's uniform in how it distributes. so, i'll explain by example.

the conservatives are down 13 points in alberta, roughly, according to a few models. they were at 66 last election. they're at 53 this election. so, if you were naive, you could say they're down 13 points in every riding. but, that would mean they'd be down proportionally more in some ridings, and it's probably not true. when he says he's using proportionality what he means to say is that he wants to calculate what proportion of 66 that 13 is. it's about 20%. so that's the proportional part. i'm in support of this aspect of the model.

but, then he goes and says that they're down 20% in EVERY riding. that's the uniform part, and what i meant when i said it's distributed uniformly. is that a reasonable assumption?

in alberta, i claim it is not. the reason is that rural alberta never changes. if you look over the last hundred years, they always win by huge margins. 70. 80. it doesn't change. there is essentially no variability in rural alberta.

however, there is a reasonable amount of variability in calgary and edmonton. it doesn't always lead to the conservatives losing a seat. but they do from time to time.

so, given that there is no variability in rural alberta, and some variability in urban alberta, i would argue that the proportional swing should not be distributed uniformly, but proportionally to where the variability is - calgary and edmonton.

there's not any algebra required, because calgary and edmonton are, together, half the population of alberta. it's easy to calculate that a 40% swing in calgary and edmonton, combined with a 0% swing in the rural areas, adds up to a 20% swing province wide.

this is how i calculated that you can guess that they're at about 60% of where they were. and, this is verified through consulting multiple recently conducted riding polls in edmonton mills, edmonton centre and elsewhere. we need more in calgary, especially.

but, because there is historically so much volatility between the liberals and ndp, it is not a good idea to assume that the beneficiary will be the same as the second-place finisher in 2011.

we need riding polls to determine that.

in summary, if you want to use a model, you need to make assumptions. and, the outcome of the model is only as good as the assumptions are. in an election with small swings that are predictable, the models being used would be roughly accurate. however, this is an election with very large swings that are very unpredictable. and, by definition, the model will be unable to get much of it right.

the people pushing these models should not be surprised by this, and if they come out next week and offer shock at the poor performance of their models, it will only serve to demonstrate that they do not truly understand them. they should be coming out and explaining that the assumptions required for their models to work are not present, and that they expect their models to perform poorly. that admission, before hand, would save their credibility.

fwiw, i took the bare statistics requirements when i was completing my bachelor of mathematics, and focused mostly on computer mathematics when i was studying at the graduate level. my focus was on what is called pure mathematics - abstract algebra, and what calculus turns into once you get past a certain level.

another way to think of this is that if both parties are starting off under 25%, or even under 20%, it is not reasonable to assume that the one with the higher vote share has a significantly greater chance of taking advantage of a decrease in support for the incumbent. that is, because they both have a long way to go from a point of relative obscurity, it is never going to be clear which one is actually moving until the riding can be measured directly. a more complex model would need to play off these probabilities, and then argue there's a 34% chance the liberals are ahead, a 31% chance the ndp is ahead and a 35% chance the vote is split - or whatever the numbers say. but, then you can see that it's not so useful, is it?

so, i will reiterate: we can conclude the conservatives are likely down substantially in the urban ridings, and ballpark it to about 60%. we cannot use modelling techniques to reasonably suggest who is benefiting. this must be measured directly.

as it is, we have riding polls that suggest that the model is performing poorly - which should be expected in this election, if you really understand how it works.

strategic voters need to consult riding polls, please.

so, as an aside, what's the best way to do this, then? well, i think that the key is how the data is organized. we organize by provinces, which ignores the fact that provincial boundaries really aren't cultural boundaries, and it simply doesn't distribute well.

rather, i might suggest creating new categories.

1) western bc
2) vancouver / lower mainland
3) eastern bc / rural alberta / rural saskatchewan / rural manitoba
4) calgary-edmonton-regina-saskatoon-winnipeg [or maybe even cut that up a little]
5) northern ontario
6) southern & eastern rural ontario
7) southern & eastern urban ontario
8) western quebec [including gatineau & montreal]
9) eastern/southern quebec
10) maritimes

same number of categories, actually. but, it would increase accuracy dramatically.