Basically, I made some mistakes with the code behind my election forecast that adjusted for pollsters’ performance. Thanks to Nigel for the email pointing out my mistake.
So, today there are effectively two updates to the forecast. One is a fundimental change the model I’m using to predict the outcome of the election, and the other is a new data point, yesterday’s Colmar Brunton poll.
The overall result is Greens lose a huge amount of support, and National, Labour and NZ First all gain. To the point where NZ First would eclipse the Greens as the third largest party in Parliament.
On my new numbers the only new Green MPs would be Chloe Swarbrick and Golriz Ghahraman, while Denise Roche and David Clendon would lose their seats.
Marama Fox would lose her seat – the Māori Party would be reduced to a single MP.
Labour would be almost 17 points behind National, but that would be a decent improvement on their crushing defeat under David Cunliffe.
Their new list MPs would be Priyanca Radhakrishnan, Jan Tanetti, Willow-Jean Prime, Kiri Allan, Willie Jackson, Jo Luxton and Liz Craig. Their caucus would be 48.6% female – less than the required 50%.
Apologies for the change in my model, if anyone is interested I can rerun previous predictions with my new performance numbers.
I’ve updated my election forecast with the release of the only NZ opinion poll so far in May. In short, there has been very little movement. Since my forecast was last updated, National and the Greens have both lost a seat while Labour and NZ First have both gained one. That doesn’t really change the fact that either Bill English or Andrew Little would need to do some form of deal with Winston Peters to form the next Government. Fun times.
Also, this is the first time I’ve updated my model since the Labour Party released their party list for 2017 – the first party to do so. Based on retaining all their current electorate seats and with a total of 35 MPs, Willie Jackson would be an MP and Jo Luxton and Liz Craig would narrowly miss out. They would have 16 female MPs for a 45% gender balance, less than the party constitution requires, and a whopping 11 Māori MPs – at 31% representation Māori represenation would be twice as strong as in the NZ population. Here’s what it would look like:
After a tantrum from a candidate, Labour have finally released their 2017 Party List.
I’ve updated my Labour caucus calculator, see the link at the top of this page to check out who would be in and out. Note that the default result setting is what Labour got at the last election, you might want to adjust that to be either my current forecast (34 MPs) or whatever poll you choose to believe.
For what it’s worth, I think it’s a very good list. Will have more thoughts in the days to come.
I’ve updated my polling forecast with the latest data from the only NZ public poll released in April so far (seriously, we’re less than 150 days out from a general election and there has only been one poll this month).
We’re finally starting to see some movement in the forecast, although it is still fairly minor. While NZ First are moving up in their party vote, it’s not yet enough to net them another List MP.
The Greens however have just nudged high enough to grab another List MP, at the expense of National.
This won’t have a significant impact. To form Government on the numbers predicted before this poll was released, Bill English would have had to rely on support from NZ First, that does not change).
At the last British general election the Conservatives spent £1.2m on Facebook advertising, while the Labour Party only spent £16k [source]. It still astounds me just how little focus Labour put into paid digital advertising, but then you look at the results.
It looks like, despite all their other problems, Labour may have learnt from their mistake.
The Guardian reports that they have built a custom tool called Promote which integrates with their voter database and Facebook Ads.
While it’s short on details, it sounds like they’re trying to make it as easy for a local campaign to target Facebook Ads as it would be for them to cut a list for a phone bank or door knock.
Of course, this doesn’t necessarily mean that the party will actually redirect more of their central resources into digital advertising, but this investment in infrastructure is a very good sign.
Statistics is a useful tool for understanding the patterns in the world around us. But our intuition often lets us down when it comes to interpreting those patterns. In this series we look at some of the common mistakes we make and how to avoid them when thinking about statistics, probability and risk.
1. Assuming small differences are meaningful
Many of the daily fluctuations in the stock market represent chance rather than anything meaningful. Differences in polls when one party is ahead by a point or two are often just statistical noise.
You can avoid drawing faulty conclusions about the causes of such fluctuations by demanding to see the “margin of error” relating to the numbers.
If the difference is smaller than the margin of error, there is likely no meaningful difference, and the variation is probably just down to random fluctuations.
2. Equating statistical significance with real-world significance
We often hear generalisations about how two groups differ in some way, such as that women are more nurturing while men are physically stronger.
These differences often draw on stereotypes and folk wisdom but often ignore the similarities in people between the two groups, and the variation in people within the groups.
If you pick two men at random, there is likely to be quite a lot of difference in their physical strength. And if you pick one man and one woman, they may end up being very similar in terms of nurturing, or the man may be more nurturing than the woman.
You can avoid this error by asking for the “effect size” of the differences between groups. This is a measure of how much the average of one group differs from the average of another.
If the effect size is small, then the two groups are very similar. Even if the effect size is large, the two groups will still likely have a great deal of variation within them, so not all members of one group will be different from all members of another group.
3. Neglecting to look at extremes
The flipside of effect size is relevant when the thing that you’re focusing on follows a “normal distribution” (sometimes called a “bell curve”). This is where most people are near the average score and only a tiny group is well above or well below average.
When that happens, a small change in performance for the group produces a difference that means nothing for the average person (see point 2) but that changes the character of the extremes more radically.
Avoid this error by reflecting on whether you’re dealing with extremes or not. When you’re dealing with average people, small group differences often don’t matter. When you care a lot about the extremes, small group differences can matter heaps.
4. Trusting coincidence
Did you know there’s a correlation between the number of people who drowned each year in the United States by falling into a swimming pool and number of films Nicholas Cage appeared in?
If you look hard enough you can find interesting patterns and correlations that are merely due to coincidence.
Just because two things happen to change at the same time, or in similar patterns, does not mean they are related.
Avoid this error by asking how reliable the observed association is. Is it a one-off, or has it happened multiple times? Can future associations be predicted? If you have seen it only once, then it is likely to be due to random chance.
5. Getting causation backwards
When two things are correlated – say, unemployment and mental health issues – it might be tempting to see an “obvious” causal path – say that mental health problems lead to unemployment.
But sometimes the causal path goes in the other direction, such as unemployment causing mental health issues.
You can avoid this error by remembering to think about reverse causality when you see an association. Could the influence go in the other direction? Or could it go both ways, creating a feedback loop?
6. Forgetting to consider outside causes
People often fail to evaluate possible “third factors”, or outside causes, that may create an association between two things because both are actually outcomes of the third factor.
For example, there might be an association between eating at restaurants and better cardiovascular health. That might lead you to believe there is a causal connection between the two.
However, it might turn out that those who can afford to eat at restaurants regularly are in a high socioeconomic bracket, and can also afford better health care, and it’s the health care that affords better cardiovascular health.
You can avoid this error by remembering to think about third factors when you see a correlation. If you’re following up on one thing as a possible cause, ask yourself what, in turn, causes that thing? Could that third factor cause both observed outcomes?
7. Deceptive graphs
A lot of mischief occurs in the scaling and labelling of the vertical axis on graphs. The labels should show the full meaningful range of whatever you’re looking at.
But sometimes the graph maker chooses a narrower range to make a small difference or association look more impactful. On a scale from 0 to 100, two columns might look the same height. But if you graph the same data only showing from 52.5 to 56.5, they might look drastically different.
You can avoid this error by taking care to note graph’s labels along the axes. Be especially sceptical of unlabelled graphs.
I know it’s a cliché, but in the 2017 NZ general election, social media is going to be more important than ever. Political parties are able to communicate with fewer voters than ever before through traditional means, and the vast majority of the country actively uses social networks.
Any serious candidate or party needs to have a digital communications strategy.
For a little over a year now I’ve been collecting public information on a variety of politicians and their parties via Facebook’s Open Graph API. I’ve got scripts that are constantly running, recording what politicians are posting online, and how many people are liking them.
The data is automatically updated every day. It shows how many people like each page, how many people have liked it in the last week, a metric that Facebook uses called Talking About, which basically shows how many people are engaging with the page, and how many posts they’ve done on Facebook.
Please let me know if I’ve missed any public Facebook pages off this, and I’ll add them asap. Note that it takes a week for the new likes metric to be updated, so that’s why it shows zero for some.
The updated predicted outcome is exactly the same, in that each party would still get the same number of seats as they did before this poll was released. However, some movement has been observed.
Labour is now up 0.5 per centage points, so they must be very close to getting a 35th seat. Also, both the Greens and Māori Party are also observing an increasing trend in support with both of them increasing 0.4 percentage points and also being close to pickup an additonal seat.
I think this result shows the value in a model like this – significant new movement will still be observed, but will only end up showing in the results if it eventually forms part of a trend. In essence, we’re doing what we can to nullify “rogue polls”.
There are few things that annoy me more than how blogs and Twitter light up after the release of a single political poll.
Pundits will make the huge inferences from statistically insignificant changes, or attribute meaning to an event that occurred after polling finished.
Today I’m proud to release something I’ve been working on for a while, a forecast model for the 2017 New Zealand General Election. It is a mathematical model for analysing polling and determining what Parliament would look like if an election were held today.
It takes all available public polling, adjusts for historical data (for instance, known bias’ that individual pollsters have), produces a weighted average based on recency and sample size.
It then produces an estimate of each party’s seat count in the Parliament.
No doubt people will have a ton of questions, hopefully the following will answer them. If you have any further questions, or ideas or suggestions, please either leave them on here, or email me directly on email@example.com
Lastly, my thanks to the many people on both sides of the Tasman (you know who you are) who have helped me with the coding, maths, and design. Hopefully you’ll find it useful!
I have to admit I’ve been a bit surprised with the response to my Labour MP calculator – it seems to be a hit, even though at the moment it’s only working with a modified version of the current caucus rankings.
I’ve now plugged the National Party’s caucus rankings and selected candidates into the back end, and have the first version of my National MP calculator working…
There a bug with the gender calculations for the Nats, so I’ve removed it for the time being. They don’t have a constitutional gender requirement so it will create far fewer headaches for them, but I’ll include it as soon as I can get it working as a point of comparison.
Also, the Nats haven’t finished many of their selections, and have been very sporadic about issuing releases when candidates are selected. If you know of any National candidates that have been endorsed but are missing from my list, please let me know.