Thursday, July 8, 2010

EKOS: Tories Led by 4.4 and 10.5 in Past Two Weeks

Update: So EKOS (which, by the way, replied very professionally and thoroughly to be inquiry: thanks and good job!) did not make a mistake: the issue is that the way their sample sizes vary from week to week differs a lot across regions. For example, in this poll, Alberta and BC had more than 3 times as many respondents in the first week as in the second, while the ratio for MB/SK was only about 1.6 (and the overall ratio for Canada was 2). Thus, the weight applied to each AB/BC respondent was lower in the first week than the second in order to calculate the national averages for each of the weeks (the opposite is true for MB/SK, and so on for all regions). But when you pool the two weeks, within each region, you would naturally apply the same weight to responses from the first week and responses from the second week. [By the way, the same applies to other demographic classifications like male/female, etc.]

So this effect can indeed cause certain two-week numbers to be below or above both of the corresponding one-week numbers, but is only significant if sample sizes from week to week are substantially disproportional across demographic categories. I had not thought through this yesterday because I assumed their samples from one week to the next would be roughly proportional, but that's clearly not the case as explained above!

__________

This week's (actually these past two weeks') EKOS poll is out. EKOS published figures for the full two weeks on its website, while the CBC gave the regionals for the second week, so I was able to back out the regionals for the first week as well. Thus, I'm treating today's release as two separate polls for projection purposes. While there was little movement at the end of June, the Liberals have lost significant ground in the first week of July. The most drastic change is in Ontario, where the Liberals led by 3% at the end of June, but supposedly trailed by a whopping 11.3% in early July! The caution is that EKOS only polled 1166 respondents in the most recent week, so a bump like this could be mostly statistical noise.

Also, there's a problem with the poll. Here's what I left as a comment on EKOS' website:

I have a concern about the most recent poll: you report 18.3% as the NDP national support for 6/22-29, 17.9% for 6/30-7/6, and yet somehow, their two week average, 17.5%, is lower than both. How is that possible?

I think that you might have flipped the Liberal and NDP Atlantic numbers for the computation of the 2nd week national averages. According to the tables posted by the CBC, the NDP got 42% on 6/30-7/6. Given that it got 19.4% there in the two-week roll-up, that implies they only polled 8.7% in the first week of polling. Even with small Atlantic sample sizes, that is extremely unlikely. Furthermore, swapping the NDP 42% with the Liberal 24.8%, and recalculating the 6/30-7/6 national averages accordingly, would make all your numbers consistent.

Such a mistake would mean that EKOS' reported national figures for the 2nd week are 1.3-1.4% too low for the Liberals and too high for the NDP.

The current aggregate predictions takes the numbers reported by EKOS as such:

CON - 136
LIB - 77
BQ - 55
NDP - 40

However, if I'm right about them swapping the NDP and Liberal numbers in Atlantic Canada for the 2nd week, the Liberals would take two seats from the Tories in the Atlantic.

The Tory national lead is estimated at 7.4%; if I'm right about EKOS' mistake, it would be 7.1%.

Note: As discussed previously, the lack of polls is forcing me to change the formula for the summer. I have lengthened the depreciation period for a poll from 27 to 34 days; I plan to leave it this way for the rest of the summer and switch back to 27 days in the fall, but of course things might change as the situation warrants. Furthermore, for this update alone, I made an ad hoc adjustment due to the fact that the most recent EKOS poll has a much smaller sample size than their previous polls.

2 comments:

The Pundits' Guide said...

Good catch, Watcher!

Election Watcher said...

Thanks! It turns out there was no mistake, and I've updated my post to explain why.