Saturday 10 March 2012

YouGov's figures don't add up (again)

YouGov is widely regarded as a pretty accurate weathervane for voting intention but how accurate are their results?

They've got a pretty good record of predicting election outcomes (although not as good as Survation) but not necessarily the numbers involved.  And deciphering YouGov's research is an art in itself.


Take their latest voting intentions research as an example: by totting up all the voting intentions they've accounted for 103% of voters.  The "others" voting intention is 12% in the table but the intention by "other" party totals 13%.  That's the sort of maths you normally find in Labour's rotten boroughs where the postal votes have Pakistan post marks on them.


You would ordinarily put the odd few percent down to rounding errors but expressing the LibLabCon/Other headline as a percentage of 100% makes quite a big difference (relatively speaking) to the Labour and Tory voting intention - CON 37%, LAB 42%, LD 8%, Other 12% unrounded and corrected becomes CON 35.9%, LAB 40.7%, LD 7.7%, Other 11.6%.

The difference is even greater as a proportion of the total "other" votes: SNP/PC 4%, UKIP 4%, Green 3%, BNP 1%, Other 1% becomes SNP/PC 3.7%, UKIP 3.7%, Green 2.8%, BNP 0.9%, Other 0.9%.

What really makes the results interesting is if you include the "Don't Know" and "Wouldn't Vote" numbers to get an idea of support for the parties.  The LibLabCon/Other headline changes from CON 37%, LAB 42%, LD 8%, Other 12% to CON 29.1%, LAB 33.1%, LD 6.3%, Other 9.4%.  The "other" vote changes from SNP/PC 4%, UKIP 4%, Green 3%, BNP 1%, Other 1% to SNP/PC 3.1%, UKIP 3.1%, Green 2.4%, BNP 0.8%, Other 0.8%.

By eliminating the rounding errors, the outcome remains the same but the gap does narrow slightly.  As an expression of actual support for the parties, YouGov's results are wildly inaccurate.  The LibLabCon parties are supported by only 70% of the electorate as opposed to 85% of those who are decided and intend to vote.

There's also the question of YouGov's weighting.  Some of the differences in unweighted (what people actually said) and weighted (what YouGov have adjusted) are high which means some of the numbers have a large amount of guesswork or manipulation associated with them.  Most notable of these is the identification with a political party where 89 people who identified with an "other" political party was corrected to just 22.

Until YouGov's maths starts adding up, it's impossible to take their research too seriously.