The LOESS model has come back into line with the Bayesian model.
Tuesday, April 28, 2015
Steady as she goes
There is not much to report since the last fortnight. The Bayesian model has been a flat line now for eight weeks, deep in catastrophic loss territory for the Coalition.
The LOESS model has come back into line with the Bayesian model.
The LOESS model has come back into line with the Bayesian model.
Sunday, April 12, 2015
Aggregated polling update
Another set of noisy polls - moving in different directions - giving different messages on the state of the nation's voting intentions:
Using the Bayesian model to peer through the polling noise and house effects suggests there has been little movement for six weeks in the government's polling fortunes.
If we adjust the recent polls for house effects, an interesting pattern emerges. Very few of the recent polls are close to the median model prediction. A cluster of adjusted polls sees the government on 48.5 to 49 per cent. Another cluster of adjusted polls sees the government on the very uncompetitive 44.5 to 45.5 per cent. I am not sure that this is meaningful, beyond being cautious and acknowledging that the polling seems quite volatile at the moment.
I also use a less-sophisticated localised regression (LOESS) as a cross check. The end-points in the LOESS regression are very volatile, and the technique is overly sensitive to endpoints and outliers. At the moment, the LOESS regression is telling a different story to the Bayesian model. However, my experience is that the LOESS model usually comes into line with the Bayesian model as more data points are added. I would be cautious with a recovery narrative based solely on the LOESS regression. Noting the Bayesian bifurcation above, a recovery narrative for the government over recent weeks cannot be ruled out; but nor can it be proclaimed as absolute fact. All we can do is patiently wait for more data points.
- Newspoll over the weekend is unchanged on three weeks ago with the government on 49 per cent.
- Ispsos over the weekend has the government down three points on the end of February at 46 per cent.
- ReachTEL at the end of March, had the government up one point on the beginning of February at 46 per cent.
- Morgan at the start of April had the government up three points on the middle of March at 47 percent.
Using the Bayesian model to peer through the polling noise and house effects suggests there has been little movement for six weeks in the government's polling fortunes.
If we adjust the recent polls for house effects, an interesting pattern emerges. Very few of the recent polls are close to the median model prediction. A cluster of adjusted polls sees the government on 48.5 to 49 per cent. Another cluster of adjusted polls sees the government on the very uncompetitive 44.5 to 45.5 per cent. I am not sure that this is meaningful, beyond being cautious and acknowledging that the polling seems quite volatile at the moment.
I also use a less-sophisticated localised regression (LOESS) as a cross check. The end-points in the LOESS regression are very volatile, and the technique is overly sensitive to endpoints and outliers. At the moment, the LOESS regression is telling a different story to the Bayesian model. However, my experience is that the LOESS model usually comes into line with the Bayesian model as more data points are added. I would be cautious with a recovery narrative based solely on the LOESS regression. Noting the Bayesian bifurcation above, a recovery narrative for the government over recent weeks cannot be ruled out; but nor can it be proclaimed as absolute fact. All we can do is patiently wait for more data points.
Tuesday, March 24, 2015
Poll update
The polls are noisy and contradictory this fortnight. At 49-51 in Labor's favour, today's Newspoll is the best result for the Coalition since September 2014. It represents a four point move towards the Coalition over the previous fortnight. However, tempering the celebrations in the Coalition's ranks was yesterday's Morgan poll at 44-56 in Labor's favour. This was a 2.5 percentage point move towards Labor over the previous Morgan poll.
Let's look at the charts. We will start with the polling house-by-house line chart. You will note we have not seen a national Galaxy or ReachTEL poll for some time, as they focus on the NSW election.
Plugging the latest numbers into the Bayesian model, suggests the Coalition's fortunes are improving slowly, but there is a way to go before the Coalition would be in there with a chance at the next election. Nonetheless, the 2.2 percentage point improvement since the spill motion just six weeks ago is not to be sneezed at.
The LOESS model has come back into line with the Bayesian model.
Let's look at the charts. We will start with the polling house-by-house line chart. You will note we have not seen a national Galaxy or ReachTEL poll for some time, as they focus on the NSW election.
Plugging the latest numbers into the Bayesian model, suggests the Coalition's fortunes are improving slowly, but there is a way to go before the Coalition would be in there with a chance at the next election. Nonetheless, the 2.2 percentage point improvement since the spill motion just six weeks ago is not to be sneezed at.
The LOESS model has come back into line with the Bayesian model.
Thursday, March 12, 2015
Polling update
This week's Newspoll was unkind to the Coalition (but not as bad as the early February poll). The Morgan poll this week was the most favourable from the polling house since late November 2014.
Collectively, through the Bayesian model, the polls are suggesting a one and a half percentage point recovery for the Coalition since the Prince Philip purple patch at the beginning of February. But they are also suggesting that the Coalition's recovery momentum is stalling.The Coalition could not win an election with these numbers.
Collectively, through the Bayesian model, the polls are suggesting a one and a half percentage point recovery for the Coalition since the Prince Philip purple patch at the beginning of February. But they are also suggesting that the Coalition's recovery momentum is stalling.The Coalition could not win an election with these numbers.
Monday, March 2, 2015
February poll update
Since the spill motion on 9 February, the polls have improved for the Coalition. If an election was held now, it would still be a thumping loss, just not as thumping as early February.
I should point out that LOESS regressions can be overly influenced by outliers, especially at the end-points. We will need to see more polls to know whether the recent change in voting intention is as dramatic as that suggested by the previous chart.
I should point out that LOESS regressions can be overly influenced by outliers, especially at the end-points. We will need to see more polls to know whether the recent change in voting intention is as dramatic as that suggested by the previous chart.
Monday, February 9, 2015
Newspoll 43-57
The Australian's Newspoll from 6-8 February 2015 continues the challenging numbers for the Coalition: with a two-party preferred voting intention of 43 to 57 in Labor's favour.
Dropping these numbers into the Bayesian model yields the following charts and a headline, aggregate TPP voting intention of 44.2 for the Coalition and 55.8 for Labor.
At this point in the blog, it is my normal practice to remind people that I anchor the above Bayesian aggregation with the assumption that the net bias across all of the polling houses sums to zero.
The LOESS model yields 43.7 per cent for the Coalition and 56.3 per cent to Labor.
Both models are suggesting a sizable decline in voting intention for the Coalition since the New Year.
Dropping these numbers into the Bayesian model yields the following charts and a headline, aggregate TPP voting intention of 44.2 for the Coalition and 55.8 for Labor.
At this point in the blog, it is my normal practice to remind people that I anchor the above Bayesian aggregation with the assumption that the net bias across all of the polling houses sums to zero.
The LOESS model yields 43.7 per cent for the Coalition and 56.3 per cent to Labor.
Both models are suggesting a sizable decline in voting intention for the Coalition since the New Year.
Sunday, February 8, 2015
Galaxy 43-57
Another published poll ahead of (now) Monday's spill motion. Today's Galaxy poll is no change on last week's poll: 43 to the Coalition, 57 to Labor.
I have made a pre-processing change to the Bayesian model. Before I explain that change, let me provide some context. Rather than resurrect my code base from the 2013 Election (I have about 40 or 50 model and processing files in that directory), I decided to code for the 2016 Election from scratch. Which is what I have done.
I had been thinking about the model output, and in particular how choppy that output appeared. To address the choppiness, I added a Henderson moving average. It worked. But it was not the most comfortable solution.
Last night, I noticed that in the lead-up to the 2013 Election, I reduced the sample size for the Morgan polls down to 1000 as an adjustment for the observed over dispersion given the sample size of the Morgan multi-mode polls. I have made this same adjustment for the 2016 models. It has had the effect of reducing the choppiness in the model. But it also means the outlier poll from the middle of 2014 no longer has such influence on the model. Consequently, the middle of 2014 is no longer the Bayesian nadir for the Coalition.
With today's Galaxy poll factored into the mix, the Bayesian model is now reporting the national voting intention at 44.6 to the Coalition, and 55.4 to Labor. This is pretty much the same result as Kevin Bonham, who puts it at 44.4 to 55.6.
Because I have made a change to the model, I will provide a fairly comprehensive set of charts.
I have made a pre-processing change to the Bayesian model. Before I explain that change, let me provide some context. Rather than resurrect my code base from the 2013 Election (I have about 40 or 50 model and processing files in that directory), I decided to code for the 2016 Election from scratch. Which is what I have done.
I had been thinking about the model output, and in particular how choppy that output appeared. To address the choppiness, I added a Henderson moving average. It worked. But it was not the most comfortable solution.
Last night, I noticed that in the lead-up to the 2013 Election, I reduced the sample size for the Morgan polls down to 1000 as an adjustment for the observed over dispersion given the sample size of the Morgan multi-mode polls. I have made this same adjustment for the 2016 models. It has had the effect of reducing the choppiness in the model. But it also means the outlier poll from the middle of 2014 no longer has such influence on the model. Consequently, the middle of 2014 is no longer the Bayesian nadir for the Coalition.
With today's Galaxy poll factored into the mix, the Bayesian model is now reporting the national voting intention at 44.6 to the Coalition, and 55.4 to Labor. This is pretty much the same result as Kevin Bonham, who puts it at 44.4 to 55.6.
Because I have made a change to the model, I will provide a fairly comprehensive set of charts.
Friday, February 6, 2015
Sunday, February 1, 2015
Bayesian updates for Galaxy and IPSOS
Over the weekend we have two new polls. Both cast more light on the political impact of awarding a Knight of the Order of Australia to the Prince Consort: IPSOS (46-54; -2 on the first week of December for the Coalition) and Galaxy (43-57; also -2 for the Coalition on the first week of December). Individually, both polls suggest a movement of voting intention away from the government of two percentage points since the first week of December.
However, the Bayesian model was not convinced. In part, given the small number of polls we have from both these houses, it has responded by adjusting the house bias for these houses. But, to be fair, the January 27 ReachTEL poll only had a one point movement away from the government between 20 November and 27 January. Also the December Morgan poll and the latest Morgan poll show a one point movement in favour of the Coalition. Consequently, the Morgan and ReachTEL results would have mediated the latest IPSOS and Galaxy movements in the Bayesian model.
While the Bayesian model might not have moved as shockingly for the government as the two most recent polls, it is still moving in the wrong direction for the government. If an election was held now, it would be a thumping landslide win for Labor.
As a very rough rule-of-thumb, we can use the cube rule to estimate the number of seats Labor would win with 55 per cent of the two party preferred vote. That estimate is 97 seats.
However, the Bayesian model was not convinced. In part, given the small number of polls we have from both these houses, it has responded by adjusting the house bias for these houses. But, to be fair, the January 27 ReachTEL poll only had a one point movement away from the government between 20 November and 27 January. Also the December Morgan poll and the latest Morgan poll show a one point movement in favour of the Coalition. Consequently, the Morgan and ReachTEL results would have mediated the latest IPSOS and Galaxy movements in the Bayesian model.
While the Bayesian model might not have moved as shockingly for the government as the two most recent polls, it is still moving in the wrong direction for the government. If an election was held now, it would be a thumping landslide win for Labor.
As a very rough rule-of-thumb, we can use the cube rule to estimate the number of seats Labor would win with 55 per cent of the two party preferred vote. That estimate is 97 seats.
Updated
- 5 February 2015 - added maxima, minima and endpoint statistics to the charts (at the request of Andrew Catsaras). Also added an observation based on the cube rule.
Subscribe to:
Posts (Atom)