### Nielsen 52-48 moves poll average to 48.1% ALP 2PP, Centrebet follows

## Friday August 9, 2013

Nielsen’s 52-48 (n=1400) poll (png from GhostWhoVotes) bumped my Guardian Australia poll average a little, Labor moving from 48.2% TPP to 48.1% TPP, a new post-Rudd-return low.

Centrebet’s prices have moved this morning, Labor now out to ~~4.90~~ 5.50, the Coalition in to ~~1.18~~ 1.15, for an implied probability of a Labor win of less than 20% (~~19.4%~~ 17.3%). Galaxy must be due to drop some numbers soon, perhaps in the Sunday News tabloids. We’ll also have Morgan’s mega-multi-mode melange and Essential’s two-week rolling average released early next week. The leader’s debate might also bring out some other polling.

I’m also publishing the GA poll average here on my own blog: see the growing list of Australian election stats and links appearing in the header of my blog, above.

The model-based average gets updated as we enter new polls into the database, along with a graph showing the trajectory of the model’s estimate of ALP 2PP voting intentions over the last 90 days:

@SimonJackman Do I interpret your blog as meaning you believe all pollsters excl Reachtel have house bias to ALP, by an avg of approx 1.2%?

Wow. $5.75 at Centrebet. That can’t last. Tho Betfair is up around there too.

Those house effect estimates are one of the more interesting parts of the model. Few (!) pollsters under-estimated ALP 2PP in 2010. I could re-start the model, say with the last X weeks or months of polling data, letting bygones be bygones with respect to the house effects (forgetting the performance of pollsters in the field around the time of the 2010). ReachTEL and other newcomers begin life anew in this way, free of the “original sin” of 2010 that Newspoll/Galaxy/Nielsen/Morgan/Essential inherit. Note also that Morgan F2F and JSW bump up the average.

My model actually splits the difference. 2010 provides a fixed reference point, sure, but the model still learns about house effects by looking at the way a given pollster appears to deviate from other pollsters over the whole data stream (in my setup, 2010-present).

Newspoll’s estimated house effect is something that I spend a lot of time looking at. They almost hit the 2010 2PP exactly. So I’m always curious when my model says they are carrying a point or more of pro-ALP bias. I’m spending a lot of time trying to understand that better.

The other constraint in the model is the assumption that the house effect is constant over time. I’d almost bet that that is not true, say, for Essential, who appear to have flipped from a pro-Labor bias in 2008-2010 to be much closer to the average or even slightly pro-Coalition. Right now my model says they’ve got the same house effect bias as Newspoll, but again, its something I’m studying closely.

Simon if i were you i would deleted tha ttweet about the betting markets

because the ymade your comment in 2010 look silly

In the negotiations with the independents after the 2010 election

Katter hinted labor , will be return to governemnt

after katter’s comment

the betting markets still firm the coalition into $ 1.10

labor went out to $ 6

Does your model initially estimate house effects based on (a) a weighted average of a particular firm’s polls leading up to the 2010 election; or on (b) their election day polls? (I know that you also adjust for relative movements post 2010, and that’s logical, but I’m interested in the ‘starting point’.) If there is a trend through the campaign to the LNP (as there was in 2010) then (a) may overestimate house effects. That might be appropriate if all elections show a consistent trend towards the LNP through campaigns in the polls, but they don’t necessarily, and the trend in this election might (or might not) be different.

Of course the problem with using a single poll for each firm as the benchmark for house effects would be basing it on too small an N, and sampling error will affect all polls even on election day, so maybe you could estimate the combined house effect of all (election day) polls and use that as a benchmark.

Alternatively, treat house effects as a solely relative phenomenon, so that they by definition average out at zero. Party because you then have something you can accurately describe (a ‘weighted average of the polls, adjusting for their relative biases’), and it doesn’t build into it an assumption about how voters will behave on election day (and I find it hard to think of a theoretical reason as to why polls would inherently understate the LNP vote).

No clear answer really. But I’m a bit wary of estimates that contains a built-in assumption of a ~1% house effect.

Simon, does your model initially estimate house effects based on (a) a weighted average of a particular firm’s polls leading up to the 2010 election; or on (b) their election day polls? (I know that you also adjust for relative movements post 2010, and that’s logical, but I’m interested in the ‘starting point’.) If there is a trend through the campaign to the LNP (as there was in 2010) then (a) may overestimate house effects. That might be appropriate if all elections show a consistent trend towards the LNP through campaigns in the polls, but they don’t necessarily, and the trend in this election might (or might not) be different.

Of course the problem with using a single poll for each firm as the benchmark for house effects would be basing it on too small an N, and sampling error will affect all polls even on election day, so maybe you could estimate the combined house effect of all (election day) polls and use that as a benchmark.

Alternatively, treat house effects as a solely relative phenomenon, so that they by definition average out at zero. Party because you then have something you can accurately describe (a ‘weighted average of the polls, adjusting for their relative biases’), and it doesn’t build into it an assumption about how voters will behave on election day (and I find it hard to think of a theoretical reason as to why polls would inherently understate the LNP vote).

No clear answer really. But I’m a bit wary of estimates that contains a built-in assumption of a ~1% house effect.

I don’t use a single poll as benchmark for each firm. In fact, I’m surprised that my model puts as little weight on the “last pre-2010-election” poll as it does.

There is not “assumption” about house effects built-in here. The house effects are estimated. The last election is a valuable data point, but thats all.

There is a trend term in the model too.

I’m going to need a long blog post (probably over at the Guardian) to explain all this. See my article “Pooling the Polls” in the meantime.

I should also add that I’ve been house effects subject to a “sum-to-zero” constraint — i.e., the polling industry is collectively unbiased, with individual survey houses a little too one way or the other, with the model to estimate that variation around the zero mean house effect.

That was the assumption I used in my 2012 work for HuffPost, and it was wrong! The US polling industry underestimated Obama’s national 2-party vote share by over a percentage point.

I’ll run some models showing how much bias that bought in 2007 and 2010 in Australia. All the same, I’ll also jig up a “sum to zero” model for the 2013 data. I doubt “sum-to-zero” would bring Labor back to level pegging the moment, probably closer to 48.5 given that latest polling. We’ll see.

Thanks Simon. (Sorry about the double post, a technical error by me.)

Yes I’d be interested to see a “sum to zero” model estimated as well as the one you use.

Like you, I doubt it would bring Labor back to level pegging, but presumably somewhere slightly above 48.5.

Btw any thoughts you might have on why Reachtel and JWS are at opposite extremes in house effects, when they on the surface appear to have broadly similar methodologies, would be appreciated. :-)

I ran the sum to zero model. It misses the 2010 election result by well over a percentage point (2pp), puts Labor up by over a percentage point relative to the model I’m running now. Same drop since July 6, just from, say 51+ to 49 and a bit, versus, just shy of 50 to under 48 now.

Re ReachTel and JWS. JWS has just one national poll in my database, the monster 28K IVR survey they run just before the 2010 election. And ReachTEL are relatively newer entrants afaik. JWS put out numbers right before we got to see the actual 2010 target, whereas, well, we’ll see about ReachTEL — their 53-47 result the other day was on the lower side of the pack.

Interesting, thanks.

Do you know what happens if you apply the zero-sum model to earlier elections? (Assuming you have the data and time!)

That restriction doesn’t fit the data. Certainly not in 2010. I model 2007 Labor 1st prefs in my Bayes book (Example 9.3) and the 2004 case is in the 2005 AusJPS article I linked to in an earlier comment. In both those cases I’ve estimated the model with the constraint being that the recovered track has to run through the known (ex post) election result. The recovered house effects don’t sum to zero.

Morgan is/was a big part of the problem, seem to over-estimate of Labor vote share by more than anyone else.

My column this week for the Guardian Australia investigates what happens if you ignore 2010 performance and let the world start anew on Jan 1 2013 with a “sum-to-zero” constraint. Result: Labor at about 49% TPP +/ 1pp, although we’re waiting for this week’s Essential numbers.