Nerdy Sunday: Predicting campaign effects

I was pondering over a glass of red the other day whether long term political beliefs have any residual effect from one election to the next? Particularly among swing voters – including those voters that tell pollsters that they will vote for Party A even though they normally and historically vote for Party B

If a political party had a certain amount of political support before an election campaign, but during the campaign itself that level of support changed (up or down) as a result of what was going on in the campaign, as a result of the shenanigans of the campaign or simply as a result of cold feet – are those people that changed their mind more likely to go back and trust their original instincts at the next election?

Effectively I was wondering whether there was a residual political lean among swing voters that derived from their political thinking in the lead up to the previous election.

What got me pondering this was the number of times I’ve asked non political-tragics what they think about in the ballot booth before they cast their vote (yes, yes, I’m quite the hit at parties!) . A regular answer runs along the lines “I thought about who I voted for last time”. Now, that’s an interesting response because it links their previous voting behaviour with how they will cast their vote in the present.

A similarly regular response is “I really didn’t want to vote for Candidate A but…” – followed by an explanation of why they, in fact, did vote for Candidate A and an explanation that usually involves the theme of “I didn’t trust the other bloke”.

If they were going to vote for Party A but didn’t, do a sizeable proportion of those people actually end up voting for Party A at the next election? Does their longer term political lean and their link between current and previous voting behaviour end up manifesting at the following election?

It seems so.

If we look back at every Federal election result since 1990 (any further and the polling starts to get a bit erratic), there is enough data around to have a go at figuring this out.

Firstly we need to get a measure of the political support for each party at the start of, or just before each election campaign. That’s pretty simple – we can simply take the average of the three previous polls leading up to the “Election Day minus 1 month” point in time. So if an election were held on July 1st, we’d take the average of the three most recent polls leading up to June 1st.

Since we’re dealing with the two party preferred vote – we really only need to focus on one party here so we’ll use the ALP since they’re in government.

Also worth considering is if the previous election result itself plays a part in voter psychology. Do people vote for a party because they voted that way last time? We can have a look at that too.

Finally, what if both effects are in play? So we’ll also take the average of the election result and the pre-campaign polls for each election. We can then look to see whether over any election campaign, public opinion moved toward:

(a) The poll results at the start of the previous election campaign
(b) The previous election result and,
(c) The average of those two (which we’ll call Trend Average)

This is what we get: (at the bottom of the graphic under “results”, “os” means overshot)

In every one of the last 6 federal elections, public opinion moved over the course of the campaign toward what it was at the start of the previous election campaign.

Similarly, public opinion moved over each campaign toward the previous election result, as well as towards the previous Trend Average, although it overshot the previous election result on two occasions.

The reason why it’s worth looking at both the link between any election result and the previous election result as well as the link between any election result and the polling results at the start of the previous campaign is that even though they’ve all been consistent with reality so far, one of these patterns will not realistically hold for the next election unless the ALP goes into the campaign with a TPP of 57 or higher (which, mind you, is an entirely reasonable proposition considering their polling history of late) and they end up getting less than that result on election day.

If Labor goes into the next campaign with a TPP of, say, 55% – if it moves toward the previous election result (52.7%) it becomes inconsistent with public opinion shifting towards its longer term political lean of 56.6%. If public opinion moves toward that longer term political lean and the vote increases, it becomes inconsistent with it moving towards its previous election result.

One of these patterns will be shown, all things being equal, not to be true.

Hence the Trend Average comes in. What if both effects are real at the same time?
We don’t know the size of each effect, but if we make the assumption that they are equally true (a brave assumption to be sure), then public opinion should shift over the next campaign towards its trend average (the average of the last election result and the polling at the start of the last election campaign). That value is 54.65%

We can also build this information into our Pollytrend series (that little graphic at the top right of the sidebar) to make a projection of what the actual voting trend would look like if an election were called today AND the electorate behaved like they have over the past 6 elections.

For those that can’t be fluffed to remember or look it up, Pollytrend is a locally weighted polynomial regression that we run through the real time Pollytrack data. It is a quadratic regression with a bandwidth of 0.2, making it quite an aggressive LOESS equation that we use as our smoothing algorithm.

But the really important part is the underlying Pollytrack data that it smooths.

Pollytrack is a polling aggregation where we use the latest poll from each of the 4 major pollsters – Newspoll, Nielsen, Essential Report and Morgan. Each poll result is weighted by its sample size so if one poll has a sample of 2000 and another a sample of 1000, the first poll will have twice as much attention paid to it than the second one. After we’ve weighted the polls, we take the average two party preferred vote estimate to get out Pollytrack figure.

As each new poll comes in to replace the previous poll in the Pollytrack equation, it generates a new Pollytrack number. We then run our regression equation through those Pollytrack numbers to give us our Pollytrend.

Pollytrend is important not for its point estimates, but for its behaviour over time. It tells us whether the support for a party is increasing, decreasing or staying the same as the weeks and months move on. While we all may have our favourite pollster for point estimates and many of us might believe that some pollster or another leans too far toward the ALP or the Coalition – Pollytrend cuts through that by identifying voter movements that occur across all pollsters at the same time.

We can build this reversion to trend average and reversion to past election behaviour into our Pollytrend figure by simply treating the trend average and the previous election result as hypothetical polls.

Having played around with the data from the last 6 elections, giving these new additions a polling weight of 30% seems to do a pretty good job, producing a figure for the voter trend at the start of the campaign that was closer to the actual election result for both the ‘reversion to trend average’ projection and the ‘reversion to election result’ projection than the Pollytrend was at the time.

Effectively, both projections predicted the direction of the trend that would occur during the election campaign itself. So what do they say at the moment?

We can replicate the Pollytrend chart with the Reversion to Trend Average projection and the Reversion to Election Result projection overlaid. We can also put them in a chart that uses just the TPP vote as the y-axis.

The way to read these gets a bit tricky because we are projecting how the trend would change over the very short term of a campaign and not the actual point estimates of an election result.

If an election were called today, we would expect the Pollytrend line to move towards both the Reversion to Trend and Reversion to Election lines if the electorate behaved the same as it did in the previous 6 federal elections. We expect this because the current Pollytrend value is higher than both projections.

The Reversion to Trend projection tries to pull Pollytrend toward the 54.65% mark, while the Reversion to Election projection tries to pull Pollytrend towards the last election result of 52.7%. You’ll notice that the gap between the Pollytrend line and the Reversion to Trend line closed dramatically as the Pollytrend line approached 54.65%? That’s a good visual explanation of how the projections work. They attempt to account for the way each of the two effects we’ve been talking about make the voter trend structurally change over the election campaign.

That’s not to say that voters won’t fully decide or change their political opinion on the basis of events that occur in the campaign itself, such as various policy releases by each party or whatnot – far, far from it. Merely that any change in voter behaviour because of those events would happen from the blue or black trend projections lines rather than from the red Pollytrend.

Depending on which projection effect you believe is most important, that is the one you should pay attention to.

If you believe the Reversion to Election effect is more powerful – that is, you believe that the way people voted last election is an important determinant in how they will vote at the next election, then the black line is the estimate of where the actual trend in voting behaviour currently sits once adjusted for that effect.

Similarly, if you believe that the Reversion to Trend effect is more powerful– that is, you believe that the longer term political lean of voters is also an important determinant in how people will vote at the next election, then the blue line is the estimate of where the actual trend in voting behaviour currently sits once that behaviour is adjusted for.

Whether any of these effects hold at the next election will be interesting in a nerdy kind of way, since they’ve both held for the last 6 elections. I have no idea at all which one if any is true, but I thought it was interesting to play around with them considering the amount of times people say “I was going to vote for Party A but…”