Friday, March 27, 2009

YouGov sub-sample : Labour lose support but extend lead

In the Scottish sub-sample of the latest YouGov poll for the Daily Telegraph, Labour have slipped three points but have nevertheless extended their lead over the SNP from seven to twelve points. Here are the full figures -

Labour 36% (-3)
SNP 24% (-8)
Conservatives 18% (+3)
Liberal Democrats 16% (+7)
Others 6% (+1)

Note that the percentage changes are based on the figures from the last UK-wide YouGov poll, rather than the full-scale Scottish poll that was published on the same day.


  1. The sample size is far too small to be taken seriously - you need a poll of at least 1000 not 182. You get wide variations with these small samples.

  2. Yup, it would be more useful if you combined a number of these sub-samples together to create a combined decent sample size. With each new sample you could lose the oldest one to give a rolling prediction.

  3. Anon no. 1 - Actually, I used to make that point myself every single time I mentioned a sub-sample, but I thought it was getting a bit tedious! They're considerably less meaningful than a full-scale poll (margin-of-error of 7-12%), but not completely meaningless. Mike Smithson at has been known to pore over sub-samples of fifteen people, so if it's good enough for him..!

    Anon no. 2 - I'm instinctively against that idea because it would be such an artificial exercise. For instance, would I give each sub-sample the same weight, even though Populus and ComRes typically have Scottish sample sizes of just 60 or so, while YouGov have samples of almost 200? The methodologies are significantly different as well - I think it's better just to let the raw numbers speak for themselves.

  4. James, thanks for the response (anon no 2 here). I agree, giving equal weighting to the sub samples doesn't make sense. What you would need to do is take the actual samples out and combine them. So take the 60 ComRes samples and combine them with the 200 YouGov samples and so on till you get to >1000 samples. Then look at the voting intentions (can take into account party association and voting likelihood etc). As a new sample comes in, add them to the pool and lose the oldest set - while keeping the sample size reasonable and taking samples from a range of methodologies might actually be a good thing. A bit of work, but could be useful and give us a bit more information than we get by looking at them in isolation. I'll probably try to do this myself, if no one else does, in a few weeks time when I've got some spare time.