DarrenBarefoot.com

Go Darren!

[Reply]

Mandatory nitpickery: you’re referring to the 95% confidence interval being roughly 14% for that sample size.

There are other problems with exit polling. First and most obviously, they’re harder to conduct, being as media activities in and near polling places on election day are regulated. The second is to get a representative sample from across the city to cover all sides of the northwest-southeast political axis.

It’s useful to get a sense of how much specific communities or demographic characteristics support the various candidates. I’m not convinced that exit polls will tell the story unless many media outlets pooled their money to pay for a study that hit all corners of the city, and surveyed enough voters to get subsamples that were large enough to be useful. That’s especially important in an election thiat was as close as the one we had.

Even among the “ethnic Chinese” I suspect there’s a gap between recent immigrants and those who are third-generation Canadians. A study that finely grained is beyond the resounces of most newspapers; it tends to be the domain of academics or political pros.

My horse sense tells me that Sullivan did outpoll Green amongst Chinese-Canadian voters due to his knowledge of Cantonese and the fact that recent Chinese immigrants skew to the right, but whether it’s 71% or not I don’t know. Certinaly not enouigh evidence there to call Sullivan’s advantage ‘massive.’

[Reply]

Ian: I’m no mathematician–can you explain a little more about the nitpickery? I’m not being argumentative–I want to understand what you mean.

[Reply]

No prob — I’m no mathematician myself, but I did sail through stats well enough, found it interesting, and still use ‘em occasionally.

What people are usually referring to is the sampling error that one has to live with; this is ritualistically expressed as being within x% of true, x% of the time. The “margin of error” referred to in most poll data means that the probability is 95% that the results will be within however-much-percent of the population as a whole; this figure is derived from figuring out how wide of an interval one must establish to cover 95% of the sampling distribution; as it turns out, that’s 1.96 standard deviations in a normal distribution, which leads to the handy-dandy equation of 0.98/sqrt(N) where N is the sample size to figure out the magnitude of the 95% confidence interval, or what’s popularly called the margin of error. Plug 1000 into the equation, and you get 0.031, which is why a poll of 1000 voters is usually stated as having a margin of error of +/- 3.1%. Seriously — that’s all there is to it.

The other 5% of the time, the sample results will fall outside this interval and be even less representative. (In that case, the results are usually not too far outside that interval, but you’ve got to draw a line somewhere.) Thus the “19 times out of 20″ that frequently follows the +/- figure. In fact, that value is based on 50% of the population expressing whatever is being measured; the width of this interval becomes smaller as one approaches the extremes.

It’s possible to establish higher confidence levels, but 95% is by far the most widely used — it’s high enough to be useful, yet not so wide as to be overly fuzzy; a 99% confidence interval is about 30% wider than one of 95%.

What I’m more curious about is which polling stations were targeted by the survey, considering how widely the results varied across the city.

[Reply]