Tag Archives: lies damned lies and statistics

A brief reflection on BC election polling

Once the disappointment over the BC election isn’t so painful, there’s something that needs to be talked about regarding polling. ┬áThis is a quick, stream-of-consciousness post, and I plan on writing more later.

While Mario Canseco of Angus Reid said, this morning, that it’s not a “methodology” problem, I think it is. Forum Research, while still off, was the closest to the results, with a survey sample of about 1,100 using telephone voting. Angus Reid, Ekos, Ipsos, etc., all used web-based voting with samples of about 800, and were wildly off.

The polling firms claim that web panels can work well, because you can weight respondents based on demographics. Sure, but you’re weighting a non-random sample, which can be really problematic scientifically. And it’s web based. Everyone knows that political parties send out e-blasts to freep polls – “go here and click this.” Because everyone thinks that polls lead opinion, rather than the other way around.

I know that a lot of people point at telephone polls and flail, because “cellphones aren’t polled.” Sure, they’re not. Normally. Some are, depending on where you put your phone number and such. But more importantly, the demographic most likely to be excluded from telephone polls because of over-reliance on landlines – the 18-34 demographic – is also the demographic least likely to vote. So, perhaps telephone polling acts a filter, removing the least likely voters from the poll sample to start with?

Even more importantly, 800 respondents is a dismally low sample, especially, in my mind, for a web-based vote. If we take a voting population of about 3.1 million, for a +/- 3%, 9 times out of 10 confidence level, you want to have more than 1,800 respondents.

So, for a confident poll of BC voters, you need 1,800 respondents. Forum was closest to that sample, and they were closest in their projections. Angus Reid and the rest, while they claim to weight their samples based on demographics and other fancy calculus… their sample is less than 1/3 of a number statistically required for confidence.

This doesn’t even get into the larger campaign issues.