Tuesday, March 24, 2009

More on the Ipsos Poll

Sacha Peter at BC 2009 has more details on the poll. Ipsos Reid has their stuff here.

In looking at the tables, they spoke with 800 voters and 708 were decided voters. There is a significant problem here. Out of 800 people they polled in BC, I would expect about 60-65% of them to be decided voters. They got 88.5% saying they are decided voters. Only about 500 people should have come up as being prepared to vote and the decided voters should have been lower than that.

What this means is that about 200 of the respondents to the poll lied to the pollster as it is clear they will not be voting. That is 1/4 of the respondents to the poll. There is a huge systemic error here and makes the data very hard to work with. Most people will assume people lied answered in the same manner as the those that did not. I think this is an error.

I believe that my early analysis looking at the polling results looking at the very likely voters is much better representation of who will actually vote. Using this as the basis for the vote, things look worse for the NDP, among the people likely to vote, the gap is 14 to 15 points between the NDP and the Liberals.

  • Liberals - 50%
  • NDP - 35%
  • Greens - 11.5%
  • Others - 3.5%

The data is there to support this as being a realistic reflection of the public from this poll, but when looking at these numbers the margin of error at 95% confidence level is roughly +-4.7%. I come up with 430 very likely supporters in this poll, which I find realistic given voter turn out. Another 70 to 80 respondents are likely to vote but I assume they are the undecided at the moment.

The 708 decided voters makes for a margin of error of +- 3.68 at a 95% confidence level, not the 3.5% as reported, the 3.5% applies to the whole 800. I know this is a small change, but the change does flatten the curve enough to make the overall results 'fuzzier'. The margins of error at the 95% confidence level for regional breakouts:

  • Lower Mainland +- 4.65
  • Interior +- 7.63 - a VERY flat curve
  • Vancouver Island +- 9.90 - a meaningless number for regional break out
These flat curves are the reason I pay no attention to the regional numbers. The Lower Mainland has some interest, but given the fact that overall 200 more people answered they were decided voters of some sort than one should expect, the numbers are suspect.

In the end, the polling is not telling us huge amounts and will mostly be used by people to add a patina of legitimacy to their outrageous predictions for the election. Yes, I mean myself in this as well.

1 comment:

Anonymous said...

Your "very likely" voters showing a 14% - 15% Liberal spread is akin to the results (14% and 16%) of the last two Mustel polls.