How Much Impact Does The Economy Have on Australian Federal Elections?

Veteran watchers of politics may be familiar with the old saying, “It’s the economy, stupid”. It’s a saying coined by the campaign of a challenger who unseated a seemingly-formidable incumbent amidst an unexpected recession, encapsulating the simple theory that voters tend to reward governments who have presided over a strong economy while tossing out those who have presided over weak economies.

While it sounds simple on the surface, the economy/incumbent performance relationship is not as clear-cut as it may first seem. There is no one measure of the strength of the economy, some measures may contradict each other (e.g. the high inflation/low unemployment of the 1960s) and some may not correlate with government performance in the way we expect. For example, here is how the unemployment rate correlates with government electoral performance as measured by the 2-party-preferred – note how weak the correlation is:

Furthermore, there are many ways by which voters may judge economic performance during a government’s term. Broadly, they can be broken down into two categories:

1. Retrospective; or in other words, “how has the economy done throughout the term of this government?”. Retrospective theories assume that voters reward governments for having enjoyed a strong economy through their term of office, and punish them for weak economic performance. Measures such as real GDP growth over a government’s term or unemployment rate tend to be retrospective.
2. Prospective; or in other words, “what do we expect the economy to look like in the government’s next term?”. Prospective theories assume that voters extrapolate from the current economic situation, and are willing to forgive governments for a weak economy during their term if they expect the economy to improve or is already improving. Measures such as change in unemployment tend to be prospective.

In general, voters tend to act more prospective than retrospective, judging governments by economic trends rather than absolute economic performance. For example, in comparison to raw unemployment rate, the change in unemployment leading up to the election correlates much more strongly with electoral performance:

This of course doesn’t mean that voters fail to judge governments on past economic performance. Apart from unemployment, another factor which correlates decently with electoral performance is average gross-domestic product (GDP) growth over the government’s term:

I’ve compiled a list of economic variables which are plausible indicators of economic health and estimated how well they correlate with government electoral performance by calculating the coefficient of determination, or R². This is a value which estimates how much of the variation in something – e.g. the variation in government 2pp – is explained by something else (e.g. unemployment), 0 being “it explains none of it” and 1 being “it explains it perfectly”. Note that I used the word “explain” and not “predict”; as I’ve written about before, just because something explains existing data very well doesn’t mean it can predict new, unknown data equally well (I discuss predictions a little later in the piece).

Factors which correlate with government 2pp

1. Change in unemployment (R² = 0.22) – how much the seasonally-adjusted unemployment rate increased or decreased in the final quarter leading up to the election. In general, if unemployment goes up, government 2pp goes down.
   
   
   
   Underutilisation has also correlated at a similar level to unemployment (R² = 0.23). However, as far as I can tell, data on underutilisation has only been available since 1978, missing out on all of the “landslide” federal elections (1966, 1975, 1977). This means that any model or analysis which relies on underutilisation is more likely to fall apart if run in an unusual environment (i.e. is less robust).
2. Change in labour force participation rate (R² = 0.19) – how much the labour force participation rate (i.e. (total employed + total unemployed) / total civilian adult population) increased or decreased in the final quarter leading up to the election. Historically, increases in labour force participation have correlated with a weaker government 2pp.
   
   
   
   However, there are good reasons to think this may not hold. It is plausible that, in the past, decreases in labour force participation were primarily driven by retirement while increases were primarily driven by members of financially-insecure families entering the job market to support their families – in which case changes in labour force participation would serve as a proxy for how financially secure people feel.
   
   In contrast, people exiting the labour force today may be unemployed workers who have become discouraged from finding work due to poor job prospects; hence I’m skeptical that this relationship will continue to hold for further elections. This is actually reflected to some extent in the data – if I limit the analysis to just elections conducted after 1980, the negative trend reverses into a weakly-positive (though not statistically significant) one:
   
   

   

3. GDP growth over a government’s term (R² = 0.17) – the average growth in inflation-adjusted GDP per year, over a government’s term. In general, stronger GDP growth correlates with better government performance at the ballot box.
   
   
   
   Real GDP per capita growth correlates at a very similar level, however, while I have data on GDP going back to 1959, data for GDP per capita is only available going back to 1973 and hence for the sake of completeness I’ve opted to use GDP instead. In any case growth in both GDP per capita and GDP have been highly correlated (R² = 0.97) so the differences between the two are likely minimal.
   
   I will note that GDP growth in the final quarter leading up to the election and in the election year also correlates fairly strongly with government performance (R² = 0.15 and R² = 0.1 respectively).
4. Change in household disposable income (R² = 0.13) – the inflation-adjusted change in estimated income not consumed by rent or groceries over the final quarter prior to the election. Generally, strong growth in household disposable income leads to better government performance while weak growth or reduction in household disposable income has resulted in a weaker government 2pp:
   
   
5. Inflation rate (R² = 0.1) – the quarterly inflation rate as measured by the Consumer Price Index in the quarter leading up to the election. In general, a high inflation rate has historically been correlated with weaker government performance; however the relationship isn’t particularly strong:
   
   
   
   Instead of a linear relationship, we can get a somewhat better fit by using a quadratic curve – which assumes that a small amount of positive inflation is good for the economy but no inflation or massive inflation is bad (which aligns with most economic theory):
   
   
   However, it’s fairly apparent that we have fairly little data on the extremes of the inflation rate (note the big outlier to the right – that’s the 1975 election aka the Dismissal election) and voting intention during times of very high inflation.^1x There’s also the fact that inflation over the year leading up to the election correlates much more weakly with government 2pp, R² = 0.0423. Furthermore, I suspect the relationship between inflation and economic health, as felt by voters, is more complex than it is for the other factors I’ve listed here. For example, it’s entirely plausible that voters may be perfectly fine with a higher inflation rate as long as wage and asset growth outpace price inflation, or may be dissatisfied with the targeted 2 – 3% annual inflation rate if it is accompanied by stagnant wages.
   
   Another particular issue with inflation is the massive difference between inflation prior to economic reforms taken in the 1990s (average 7% per year) and inflation in the modern era (average 2% per year). This may make it harder to make direct comparisons; in 1977, 9% annual inflation was only slightly above average for the era and allowed the Coalition to be handily re-elected (54.6% 2pp) whereas I suspect an inflation rate of 9% would be considered an absolute disaster if replicated at the 2021/2022 federal election.

The economic data I use in this piece is available here; it’s been sourced from the Australian Bureau of Statistics (ABS – GDP, labour statistics, inflation) and the Reserve Bank of Australia (RBA – household disposable income).

Modelling the impact of the economy on government 2pp

Out of the above factors, I’ve picked final-quarter change in unemployment and average annualized GDP growth over government’s term for an economic fundamentals model to explain variation in governments’ popular vote (combining a prospective and a retrospective variable).^2x I’ve mentioned above how inflation and labour-force participation changes can have a complex and hard to define relationship with “economic health” as felt by voters and hence may not be a robust indicator of voters’ opinions on the economic situation; hence I decided to leave those out.

This leaves final-quarter change in unemployment, average annualized real GDP growth over term, and final-quarter change in real household disposable income.

The reason I went with change in unemployment and GDP growth is because when you control for change in unemployment, the correlation between GDP growth and government 2pp remains positive. This makes sense – if unemployment drops, governments should do better, and if the economy has grown over a government’s term, they should also do better.

On the other hand, when you control for change in unemployment, the correlation between change in disposable income and government 2pp becomes negative – which makes no sense. As I noted in my piece on model overfitting, it’s important to think about whether a model makes sense given your underlying theory; a regression which says government 2pp decreases both when unemployment goes up as well as when disposable income increases doesn’t make sense in that regard and is most likely the outcome of statistical artifacts or random noise in the data. Together, they explain about a third of the variation in governments’ re-election performance:

Interestingly, this is a fairly similar result to what FiveThirtyEight once found in their analysis of economic fundamentals in USA elections (with regards to how well they explain election results). Such a model would have correctly called the winner of the 2-party-preferred vote correctly in about two-thirds of elections (15 of 23). I’ve marked out the elections with the three biggest differences between the expected 2pp and the actual 2pp; a cursory look into the history of each serves to demonstrate how non-economic factors can influence an election:

1. 1966: Harold Holt leading the federal Coalition to a landslide against Labor, led by Arthur Calwell. While the economy was fairly strong, Holt’s performance was unusually strong (56.9% of the 2pp, the largest in the postwar era) which most likely reflected factors such as Calwell’s public support for policies regarded at the time as being outdated e.g. White Australia and nationalization, as well as Calwell endorsing socialism amidst the Cold War.
2. 1972: Gough Whitlam leads Labor to its first electoral victory over the Coalition in the postwar era. Despite a decent economic tailwind for the Coalition government, Whitlam primarily campaigned on social issues such as healthcare and education (in Whitlam’s own words being focused on “cities, schools and hospitals”) in the “It’s Time” campaign, which may have shifted voter focus away from economic issues.
3. 1977: Whitlam’s failed attempt to return to the Lodge after his landslide defeat in 1975. While the economy was weak at best, votes may have given Malcolm Fraser credit for the fact that it was doing better than it was at the last election. On Labor’s end, Whitlam’s continued leadership of Labor may have reminded voters of the poor economic situation during his last term and led to Labor doing worse than it otherwise would have.

However, as I’ve mentioned before, a model which does decently at explaining existing data doesn’t necessarily translate into a model which does well at predicting new, unseen data. While the best way to test how something predicts data is to make it predict unseen elections and see how it does; as I currently don’t have the ability to summon up new elections on command, the next best thing is to train the model on only some data and get it to predict the rest of the data and see how it does (also known as cross-out validation). Comparing the “predicted” 2pp to the actual 2pp result for each election:

Models generally do worse when applied to out-of-sample data

The most important takeaway here is the drop-off in R²; which goes to show how a model which can explain variation in government vote doesn’t necessarily perform as well when tested on elections it hasn’t “seen” yet. That aside, such an economic regression would have predicted 14 of 23 (61%) elections correctly with an average error on the government 2pp of about 2%; in contrast the final pre-election polls have an 75% correct-call rate and an average error of 1.6% on their 2pp estimates.^3x Although that includes some time periods when polls used the less-accurate respondent-allocated method of estimating two-party-preferred; if I calculate last-election estimates for those polls (as most of our pollsters currently do), the average error goes down slightly to 1.5% and the correct-call rate goes up to 80%.

In other words, a purely economic model of voting is unlikely to outperform the polls when it comes to accurately predicting the results of a coming election. Which makes sense – not only are there non-economic factors which play a role in voters’ decisions (e.g. voters’ approval of the Prime Minister and/or Opposition Leader), any model (economic or otherwise) which attempts to predict voting intention using non-polling factors will have to assume that the factor – e.g. changes in unemployment – will affect voter behaviour in much the same way as they have in the past, an assumption which doesn’t always hold. In contrast, voting-intention polls, for all their faults, have the advantage of directly asking voters who they intend to vote for; a process which requires fewer assumptions than non-polling modelling.^4x You do have to assume the sample (or your weighted sample) is representative of the voting population at large, and if you’re doing poll-related analysis (e.g. election forecasts and whatnot) you do have to make assumptions about how well polls conform to statistical theory.

On the other hand, with non-polling models, you similarly have to assume that the elections you have in your dataset are a good sample of all elections (e.g. if half of the elections in your sample featured leadership chaos, but over a long period of time only a tenth of elections feature leadership chaos, then your election sample will be skewed by voters’ responses to leadership chaos) and make assumptions about what statistical model to apply in addition to assuming the relationship between the non-polling factor and the popular vote holds for future elections.

That being said, economic models do have an advantage over polling-only models: they provide information on why voters voted the way they do instead of simply predicting how they intend to vote. Furthermore, an economic model seems to be able to explain about a third of the variation in election results, which is more than enough to give a serious boost to governments running for re-election in times of plenty while hampering governments who face the electorate amidst poor economic conditions.