Poll wars
Commentators on this blog often make claims about the accuracy of various polling companies. I decided to take a look at the opinion polls within the final weeks of the last three Federal elections to see how the polling companies stacked up. I compared the predicted Coalition and Labor primary vote shares and the predicted two-party preferred (TPP) vote share with the final outcome. I also considered the predicted preference flows to the Coalition.
Where a prediction was within two percentage points of the final outcome, I scored it as a pass (shaded green in the table below). Where it was within three points I scored it as a possible pass (shaded yellow in the table below). Where it was off by three percentage points or more I scored it is a fail (shaded red in the table below).
Taking an average for the three elections, at a two per cent margin of error, ACNielsen was correct 74 per cent of the time. Newspoll was right 48 per cent of the time. And Morgan was correct 22 per cent of the time. If we take a three per cent margin, the scores improve to 96, 78 and 50 per cent respectively.
Is this conclusive? No. It is not conclusive. It was a small sample. And it was possible that three weeks out from an election actual voting behaviour may have been different to that which occurred on the election day. In addition, polling companies can change their methodology at any moment, and the methodologies used previously may not be used for the next election. Furthermore, there is always the risk that I have made a transcription or adding error. Still, it was interesting.
Update: as I had feared, there was a transcription error in my initial set of calculations. I had swapped the election primary votes for Labor and the Coalition in 1998. It has now been fixed. If there are any other errors, please let me know.