Ben Phillips of the Australian National University has been hawking research showing the demographic indicators that associated most clearly with the federal election swing, with the clearest patterns relating to Christianity, which correlated with a swing against Labor, and education and income, which went the other way. Evidently the Australia Institute has done something similar, with findings reported by Ross Gittins in the Sydney Morning Herald.
In considering research of this kind, one must acknowledge the perils of the ecological fallacy, whereby inferences about the behaviour of individuals are inappropriately drawn from aggregate-level data. My favourite illustration of this point relates to American politics, wherein the Republicans’ strongest states are those of the dirt-poor deep south, whereas wealthier voters favour the more conservative party in the United States as surely as they do here. As such, it should be recognised that Christian areas swinging to the Coalition need not signify that Christian voters did.
Nonetheless, the relationship between swings and the demographic features of the areas in which they did or didn’t happen is interesting in and of itself, and really all we have to go on until the Australian National University eventually publishes its Australian Election Study survey, particularly in the absence of intensive and high-quality exit polling that is conducted in the United States.
My own number crunching along these lines has involved collecting demographic measures of the areas in which each polling booth is located, and using multiple regression analysis to determine how well they predicted the primary vote swing to or against Labor. The results were as interesting for what didn’t prove predictive as for what did. In particular, an electorate’s age profile appeared to have little impact on its swing – or at least, none that couldn’t be better explained by other variables that might themselves correlate with age. This theme was picked up on in the article linked to above by Ross Gittins, which argues against the widely held notion that franking credits was the main culprit behind Labor’s poor show.
After a bit of trial and error, and whittling it down to variables that didn’t appear to be separately measuring identical effects, the most instructive variables proved to be income, home ownership, education and industry of employment, with a few ethnicity measures registering as worth-including-but-only-just.
|Estimate||Std. Error||t value||Pr(>|t|)|
|Secular/No Religion||0.105852||0.007944||13.325||< 2e-16||***|
The numbers in the “Estimate” column show the coefficients, i.e. how much each increment of that variable associated with the Labor swing. Three stars at the end means the effect is highly significant, two stars somewhat significant, one star of some significance, and with no stars we can’t say with any confidence if the relationship was positive or negative.
So, to pick one of the more striking results, for every 1% of population identifying as secular or “no religion”, Labor’s vote tended to be around 0.1% higher, independent of all other factors. Or to raise the stakes a little, Labor typically did 1% better in swing terms in places where 40% of the population identified as secular as compared with those where 30% did so. Note that the “median income” refers to weekly family income, and is measured in thousands of dollars – so an area with $2000 median family income typically did 0.3% for Labor than one with half that.
The biggest surprise for me is that “primary industry” – percentage of the workforce in mining, agriculture, forestry and fishing – had no significant explanatory power in and of itself. This doesn’t sit well with the drubbing Labor copped in central Queensland and the seat of Hunter, for which I can’t offer a ready explanation, except perhaps that I should have broken out mining and measured it independently of the others.
However, a significant negative effect is recorded for the other blue-collar industries of construction and manufacturing, together with the generally low-wage retail sector. This, remember, is independent of the effect of income, such that Labor would have suffered a combined whammy of the various effects in low-income areas with large workforces in the aforementioned industries.
On the other side of the coin, the “professional/scientific/technical” industry designation recorded a strong positive association with the Labor swing, and this too needs to be understood as part of a double whammy with the income effect. This was evident in the large-but-useless swings Labor picked up in blue-ribbon metropolitan seats. The positive effect recorded for education/health is interesting, perhaps suggesting a public-versus-private sector effect.
A fair bit has been said of Labor’s bad show with the Chinese community, but it was actually found that the “East/South-East Asian” population had a slight positive correlation with the Labor swing. However, the recorded effect is very likely drowned out by the strong positive result for “secular/no religion” variable, which records the effect of the swing against Labor in the various ethnic enclaves of Sydney and, to a less extent, Melbourne.
Finally, the “dummy” variables simply record how much of the swing could be explained by the state in which a booth was located, again independent of all other factors. Note that no measure for New South Wales is included, as it serves as the benchmark against which the other states and territories are being measured. The strong positive result in South Australian reflects that this is a primary vote measure, and both major parties rose in South Australia off the demise of the Nick Xenophon Team.
The r-squared value for the model is around 0.25, which is to say that all of this explains only about one-quarter of the variation in the Labor swing. In a future episode, I might take a closer look at what the model fails to predict by looking at individual electorates that bucked the various demographic trends just noted.
Note also: the new post below on the count for the Senate, in which only Queensland appears still in doubt, and the ongoing one dealing with close races in seat for the House, albeit that yesterday’s counting provided essentially nothing new to report.