It seems there is a lack of understanding by many of how polls work. In particular, accusations of push-polling keep popping up. In the UK, if a polling company is a member of the British Polling Council you can be sure they never do push-polls. It is unethical and accusations of such behaviour will rightly be met with, shall we say, a vigorous response! For example, YouGov and ComRes have both been accused of such behaviour recently.
Another common criticism is that the sample size is too small or unrepresentative.
There are two methods of ensuring a sample is representative. First, “random” sampling involves a polling company using a list of randomly-generated telephone numbers (usually the last digit is randomised to ensure the sample includes listed and unlisted numbers), email addresses or names and home addresses (e.g. drawn from an electoral register).
Second, “quota” sampling involves setting quotas – for example, of age and gender – and seeking out those people in each location who, together, match those characteristics. This type of poll is frequently used by internet-based pollsters who use quota samples to select from a database of people who have already provided information about themselves.
All polling samples seek to be representative of the total electorate of Great Britain by carefully selecting samples from a variety of social groups.
Usually, Northern Ireland is omitted from opinion polls as it has a different party system.
Accuracy depends on the absolute sample size rather than the proportion of the population it represents. This means that a survey of 1,000 people for the whole country is as reliable as a survey of 1,000 people for a single constituency.
But, surely 1,000 people isn’t enough? I’m not going into the statistics (mostly because I’m too lazy to fully research them), but it is widely accepted that polls of 1,000 people will be accurate to within 3%, with one caveat: as the samples are random, there is a one-in-twenty chance a particular poll has picked an abnormal group that lies outside that 3% margin of error.