On margins of error
Last night's YouGov poll had topline figures of CON 35%, LAB 41%, LDEM 9%, UKIP 7%. The Labour lead of 6 points is unusually low, but as ever, this in itself doesn't actually mean anything. A series of low Labour leads is meaningful, it would show a narrowing of the gap. Just one or two can be explained through normal sample variation.
I say this repeatedly - yet anytime there is a poll that is an outlier from the average I see Twitter filled with otherwise quite sensible people speculating on what might have caused the movement. The answer is almost always "normal sample error".
I think part of the problem is people forget the degree of normal volatility we should expect from polls. The quoted margin of error for polls is normally plus or minus 3 percent. This is actually quite tenuous - the 3 point margin of error refers to a genuine random sample, when actual polls are never purely random (most involve some elements of stratification or quota sampling, and even an attempt at a random sample wouldn't be random because of non-response), there are design and weighting factors, actual voting intentions are based on smaller samples once won't votes are excluded and so on. However, taking plus or minus 3 points as the margin of error is a good enough estimate for our purposes.
If we look at YouGov's fourteen polls so far this month, the average Conservative score is 33. All fourteen polls have been within 2 points of this. Eleven have been within 1 point of this.
The average Labour score has been 43. All fourteen polls have been within 3 points of this, and thirteen of them have been within 2 points of this. Eight have been within 1 point.
The average Lib Dem score has been 9 points. All fourteen polls have been within 2 points of this, thirteen have been within one point of it.
In other words, while there have been polls showing twelve point leads and polls showing six point leads, all of the polls have actually been within the normal margin of error of CON 33, LAB 43, LDEM 9 and the distribution around that has been very much what you would expect from normal sample variation.
