How the Canadian pollsters performed

May 12th saw voting in the British Columbia elections. This Canadian provincial election saw Gordon Campbell’s governing Liberals re-elected and also a few barbs thrown about internet polling. As in the UK, internet polling has both its fans and its detractors. So does the Canadian experience help shed any light on the UK polling debates?

Perhaps the most striking fact about Canadian pollsters is that they haven’t yet had a “1970”. In Britain, pollsters got the 1970 general election badly wrongly with much of the blame being placed on having fieldwork that ended too far ahead of polling day. As a result, a last minute swing to the Conservatives was missed and in subsequent elections the final round of polling has included fieldwork up until just before polling day.

In the BC elections, however, fieldwork finished some way before polling day. Given the general trend during the campaign for the NDP to be closing the gap on the Liberals, and the strong performance of the NDP leader Carole James in a TV debate on May 3rd, it is quite plausible that there was a swing to the NDP in the time between polling stopping and voting starting.

The final actual results were: Liberal 46%, NDP 42%, Greens 8% and in the referendum on using STV in future elections the no vote was 61%.

Of the three final polls, Mustel had the earliest fieldwork for their phone poll: 29 April – 6 May. Their scores were Liberal 47% (+1), NDP 38% (-4), Green 12 (+4), for an average error of 3%. On the STV they gave the no vote 52% (after re-allocation of don’t knows), an error of 9%.

IPSOS-Reid also did phone polling. They had slightly later fieldwork, but still finishing well before polling day: 4-7 May. Their scores were Liberal 47% (+1), NDP 39% (-3), Green 10 (+2), for an average error of 2%. For the referendum they gave a no vote of 61% (after re-allocation of don’t knows), which was dead on.

Finally, we have the internet polling of Angus Reid. Their polling took place around the same time as IPSOS-Reid, but was done more quickly: 5-6 May. To add to the difference in methodology, Angus Reid also gave people the list of candidate names for their own constituencies. Their scores were Liberal 44% (-2), NDP 42% (0), Green 10 (+2), for an average error of 1.3%. For the referendum they gave a no vote of 55%, an error of 6%.

Conclusions? The least accurate poll – Mustel – also had the fieldwork the furthest from polling day, having started in April. In particular their under-recording of NDP support, which was just higher than the standard margin of error, may have been due to a late rally in NDP support, particularly as much of their polling took place before the TV debate won by the NDP’s leader.

Angus Reid were more accurate than IPSOS-Reid on the election result, but before fans of internet polling chalk this up as evidence in their favour, it is worth remembering that Angus Reid was significantly worse on the STV referendum result, and the use of individual candidate names may explain the difference in the results rather than the merits of internet versus phone.

In other words, perhaps the best conclusion to draw is about the importance of polling right up until the election, rather than anything about the merits of phone versus internet.