Is this why you have to be careful about extrapolation?
One of the main innovations in modern polling is the mechanism to ensure a politically balanced sample by asking how people voted last time and then attaching a weighting.
Sometimes this can produce very strange results and there have been none stranger than what happened for former Lib Dem voters in the ICM Glasgow East poll at the weekend. Here, for some reason, the pollster only found six people who said they voted Lib Dem at the 2005 general election. This was far below what actually happened and ICM made its standard adjustment.
So those six 2005 LDs were weighted up to 31 – a staggering increase but in line with the procedure. Then in the course of the calculation on certainty to vote that got reduced to 24 or an unweighted figure of just four.
The result can be seen in the detail above. Those four split three to the Lib Dems to one to the SNP and as a result of the multiplier the latter resulted in six going onto the SNP total – a sizeable chunk in a survey with a sample size of just over 500.
Although this example is extreme it does not undermine my faith in the core principle – that past vote weighting produces more accurate polls because the sample is not based on demographic factors alone.