Sunday, October 09, 2011

Ontario election polls postscript

The Ontario election produced a minority, or bare majority if the legislature selects an opposition member to be speaker, who is then obligated to vote with the government in the event of tied vote.

Overall, the polls were fairly close to the actual result. The day after analysis in the Globe, in TC's view, was somewhat muddled on this point.

The table below illustrates how the pollsters performed.  It is ranked by the sum of the errors with respect to the PCs, Liberals and NDP (the Top Three column). The All column adds in the error for Greens and others:





Liberal
PC
N.D.P.
GP
Other
Total Error

Election Result
Oct. 6, 2011
37.62
35.40
22.70
2.90
1.30
All
Top 3
Oct. 5, 2011
Forum Research
Oct. 4-5
-0.62
-2.40
0.30
0.10
-0.30
3.72
3.32
Oct. 4, 2011
Abacus Data
Oct. 3-4
-0.62
-1.40
1.30
1.10
-0.30
4.72
3.32
Oct. 5, 2011
EKOS
Oct. 3-5
0.98
-1.40
-1.60
2.40
-0.30
6.68
3.98
Oct. 4, 2011
Nanos Research
Oct 1-3
0.08
-2.20
3.10
0.50
-1.30
7.18
5.38
Oct. 5, 2011
Angus Reid*
Oct. 4-5
-0.62
-2.40
3.30
0.10
-0.30
6.72
6.32
Oct. 4, 2011
Angus Reid*
Oct. 3-4
-4.62
0.60
3.30
2.10
-1.30
11.92
8.52
Oct. 4, 2011
Ipsos Reid
Sept 30-Oct 3
3.38
-4.40
2.30
0.10
-0.30
10.48
10.08

Ipsos, which is at the bottom of this list, had the most accurate poll in the federal election.  Given the inherent error in polling one should be very cautious in concluding which polling firm is 'better' than others.

However, there is an asterisk beside Angus Reid as they delivered two 'final' polls in less than a day leading to this account in the Globe after the election:
Ipsos Reid said the Liberals would receive 41 per cent among decided voters, the Tories 31 per cent and the NDP 25 per cent.
That same day, the polling firm Angus Reid announced in a press release, titled “Tories Edging Liberals But Ontario Race Could Turn in Final Hours.” that an online survey it conducted from Monday to Tuesday morning showed the Tories were at 36 per cent and the Liberals at 33 per cent.

“The Progressive Conservative Party has recovered some grounds,” the firm said of its survey conducted for the Toronto Star.

The Star published the results, saying it was the final major poll of the campaign.

But later that day, Angus Reid began to conduct another online poll, which yielded very different results.

“We wanted to track the election right to the last minute so we want back in the field Tuesday afternoon,” Angus Reid managing director Jaideep Mukerji said in an interview.

“There was volatility among Ontarians . . . We knew it was going to a very tight race.”

The later poll ,which wasn’t commissioned by the Star but conducted on Angus Reid’s own account, Mr. Mukerji said, predicted the Liberals would get 37 per cent, the Conservatives 33 per cent and the NDP 26 per cent.

“We’re happy we caught that shift,” Mr. Mukerji said. “There were some Liberals at the last minute who got cold feet about voting conservatives.”

The pollster who worked for The Globe and Mail and CTV, Nanos Research, however released similar numbers two days before Angus Reid’s final poll.

Federal election polls in Ontario
The polls on the provincial election were quite good overall. This was was not the case with the Ontario sub-samples on polls taken during the federal election. The average error in Ontario understated Conservative support by 6.1 percentage points while overstating all the others, the NDP by the 3.2 points the Liberals by 1.7 and the Greens by 1.1.  TC is still puzzled by the error in polling during the federal election although one can find some interesting discussion of this on Pundit's Guide.  Ekos pollster Frank Greaves is the only one who seriously addressed his own mistakes after the election.

TC observes that the average Conservative support from the beginning of the federal election to the approximate beginning of the NDP surge in Ontario was 41.8% (not too far from 44.4% they actually won) and thereafter it averaged 38.1%. However, I am not sure what that tells us other than perhaps the instability in preferences reflected in the shift had some impact on the measurement difficulties actually experienced.