Sunday Requestathon

Far too long ago, a new (ahem) “regular” feature was launched to supplement Nerdy Sunday – Requestathon where you folks can make requests about how polling and electoral data plays out with other things. Today we’ll start to get some answers for those requests.

Dave55 asked:

Any stats on where people think each party falls on the (admittedly largely irrelevant) political spectrum. With Humbull trotting out the old communist red book stuff today, it got me wondering just how many people actually think Labor is socialist.

The Australian Election Study runs post-election surveys on these types of things, and while I can’t find the distribution of where people see the various political parties sitting on the ideological spectrum (which is a pity, I’d like to see estimates on how many people still reckon the ALP are a bunch of unreconstructed socialists), they do publish two pieces of data; Where the average person sees themselves on the Left/Right spectrum (where zero is far left, 10 is far right) and the average score of each political party on that same spectrum from the same survey.

An anonymous person asked:

I’d be keen for some kind of analysis of the average sentence and word length used by Rudd through time, and to see whether this corresponds with his approval/satisfaction rating, PPM of even primary/TPP vote for Labor. The hypothesis would be that his approval/PPM/Labor vote is higher when he speaks clearly and directly than when we get a barrage of bureaucratese.

There’s quite a bit of data monkeying involved to get an answer for that, and while I can’t answer it comprehensively over that period,  I did run every speech the PM has made this year through a Flesch-Kincaid Grade Level test (which attempts to estimate the number of years of education an average person would require to understand what is being said), and took the average result for each month.

If we compare that measure to the monthly all pollster average of the ALP two party preferred vote we get:

There doesn’t seem to be any relationship between ALP political support and Rudd’s language complexity – at least with his speeches. However, it’s probably his interviews that most people listen to when it comes to paying attention to the PM and there’s so many of them since just January that it’s a PhD waiting to happen all by itself!

Another person asked if there was any way to calculate our Pollytrack series in a way that didn’t punish the phone polls because of their smaller sample size (even though we know that phone polls are generally accurate every election) compared to the large 1800+ online panel that Essential Report uses and the big face to face samples that Morgan face to face often uses.

Essential Report and Morgan Face to Face polls do lean toward the ALP compared to the phone polls. We don’t know if that lean is correct or not, just that it exists. The problem we have in Australia compared to, say, the recent US election is the amount of polling we get is relatively small and prevents us from doing robust analysis of which pollster leans toward which party and by how much.

But what we can do is change the weighting regime of our polls. In Pollytrack we weight purely by sample size, meaning the polls that have larger samples are expected to theoretically be more accurate in terms of minimising their sampling error, and hence having a smaller margin or error.

A poll of 2000 is always theoretically more accurate than a poll of 500, all else being equal.

Our problem is that we don’t know if all else is equal. We do know that Federal election results nearly always end up within the margin of error of our phone pollster’s final polls. What we also know is that we cant say that for Morgan Face to Face polls because Morgan use phone polls for their final pre-election day poll.

So instead of weighting purely by sample size and using a simple pooled sample, we can weight by sample size and methodology. If we give our two regular phone pollsters (Nielsen and Newspoll) 25% weighting each, and then weight the non-phone pollsters by sample size to make up the other 50% (and inlcude the irregular Morgan Phone polls in that second group), our new weighted Pollytrack looks very similar to the old one, but isn’t as strong for the ALP vote at the higher ends of the spectrum.

If we compare our original Pollytrack which we’ll call “Pollytrack 1” against our newly weighted Pollytrack which we’ll call “Pollytrack 2”, it looks like this for Primary and TPP votes.

If we take just the ALP Primary Vote and also run our low sensitivity LOESS regression through it, we can see how the three measures track:

So the big question is – should we add this new weighted Pollytrack to our standard Pollytrack measurements?

Is there anything else you can think of doing that would improve the tracking?