This fortnightly poll-mixing differs from
Reuters and other poll averages in several ways.
It is not done
every fortnight, but more regularly than that
calculation, it finds the most recent poll and looks at all polls over the
fortnight up to then
It weights for time passed,
so a poll conducted today counts
twice as much as one held two weeks ago. But "time-weight" only
applies to most recent fortnight, not past
It weights for sample size
I'm getting, from two
major pollsters, numbers to one
decimal place rather than to the nearest integer
I estimate two party preferreds from the
pollsters' data. Ie I don't use the pollsters' 2pps
Go to recent poll-mixes
Read on only if you
wish to get more technical
Crunching the polls
First, a word of warning. I am not a statistician, and some of my crunching
might appear, to practitioners of that trade, a little unsound. I welcome
suggestions for improvement - but in reasonably plain English please.
Reuters poll trend
You might be aware of the Reuters Poll Trend, which comes out once a month. It uses something called the “ Henderson moving average” to crunch together three regular
polls. You can read more about it
(The ABC's 'Insiders' program also runs a weekly trend graph, but I can't
find any reference to it online.)
The Reuters graph is a fine thing, but it has some shortcomings. For one
thing, it doesn't include Galaxy. Reuters have their logistical reasons for
this, but Galaxy's excellent record since materialising out of nowhere in 2004
makes them a must-include.
Nor does Reuters include any of the irregular, small polls that appear from time to
They also use pollsters' published numbers that are rounded to the nearest integer
(except Morgan, which goes to the nearest .5), do their Henderson magic and then
publish the result to 0.1 percent. This seems unsound to me.
I reckon we can do better. From now until election-day I’ll be regularly mixing up the polls, using my own herbs and spices, and publishing headline two party preferred numbers to the nearest half a percent.
The biggest coup is that two pollsters have very kindly agreed to provide to me, from now until the election, their voting intention data to one decimal place.
I won’t be publishing those numbers, but they will greatly improve the quality of my aggregates. (One of those helpful pollsters is Galaxy; the other prefers to remain anonymous but is also one of the Big Four.)
I'm not using the pollsters' two party preferred numbers, but am estimating from their primary numbers. Three of the four main pollsters arrive at 2pp
like this anyway, although using different strategies.
Test-driving my formula on the final published polls before the 2004 election gives 52.5 to 47.5 in the Coalition’s favour - very close to the actual result of 52.8 to 47.2.
I'm weighting for sample size, which allows me to include wonky little
surveys that appear from time to time.
I'm also weighting for time that has passed since the poll was taken, which
is admittedly iffy and subject to revision. That is, on any given day I look at
all polls taken in the last fortnight, and the further back the poll was taken
the less it counts.
Notionally distributing preferences by minor party is tricky, because
different pollsters give different minor party data. Everyone publishes a number
for the Greens, but not necessarily for Democrats, Family First and One Nation.
Yet whether these parties are extracted also affects the makeup of the
I allocate .75 percent of Greens to Labor. This is actually less than the
.81 which flowed in 2004, but I reckon with Bob Brown not as close to Kevin Rudd
as he was to Mark Latham, .81 is too high. The 2001 number was about .75.
For the preference flows from other minor parties and the rest, I used (roughly) the numbers from this
Those who are that way inclined can see my Foxpro/DBase formula here;
field names are self-explanatory. (But like sausages and the law, it's not something I