Rob Salmond is an expatriate Kiwi who knows a fair bit about politics and a lot about statistics. His Pundit article is certainly worth reading.
At the start of an election year, many folk like to try and project the future. Some do it with gut instinct like “Key has the nation’s trust”, others by simple historical precedent like “New Zealanders favour multi-term governments.” These are good fun to read but aren’t usually very rigorous. Others use well-regarded bases in science, like “National is miles ahead in the polls, therefore they’re likely to win,” or “the economy is in the crapper, so the incumbents will lose seats”. These are an improvement, but sometimes get you woolly, ambiguous predictions.
A third group in New Zealand has started using polling trends, as opposed to polling levels, to project an election result.
“The gap is closing, gradually but surely. And it could be very close indeed by late 2011…” said Marty G at The Standard in December, an argument he has continued to make since. I think this style of psephology is seductively sciency but sadly shallow, and should be resisted rather robustly. Here’s why:
1. You can get almost any result you want.
Look at Figure 1. It shows the trend of right-leaning support (National + ACT) as compared to left-leaning support (Labour + Greens) from Pundit’s Poll of Polls from March 2008 to December 2010, along with four projections of the 2011 election result. All are simple linear regressions, based on all the data subsequent to some starting point. All project the result of a November 26, 2011 election. And they project wildly differing results. Line 3 projects a squeaker, with the right winning by only two points, and the election result up to the Maori party (this was basically the version used at The Standard), while line 1 projects a right-leaning landslide. Indeed, if you took the 2008 result as a fixed gauge of support as opposed to an estimate, and then ran your regression on all subsequent data, you would project an even bigger thrashing.
2. Trends don’t last a year.
Every projection, no matter the start point, is based on the premise that the broad trend we are following now will more or less continue through until the election. Without that kind of assumption, you cannot project. But, as Figure 2 shows, trends in New Zealand politics do not last a year. In fact, they last only four months on average.