What Kills State Governments: Age Or Canberra?

Note (2020): An updated sequel to this article has been published, see Age And Canberra Are Still Killing State Governments

---------------------------------------------------------------------------------------------------------------------

Advance Summary:

1. This article examines the influence of two factors on the fates of state governments at elections: the age of the state government and whether the same party is in office federally.

2. Results since 1969 show a moderately strong influence of which party is in office federally, but only a very weak impact of age of government.  A state government that is of the same party as the federal government is generally strongly disadvantaged.

3. In the 1970s and 1980s two very old state governments were repeatedly re-elected.

4. Since 1989, both the age of a state government and whether or not the same party is in office federally have been strongly connected to state results.  

5. It is rare, perhaps increasingly rare, for a state government that is the opposite party to the federal government at the time to lose.  When this happens there are very strong reasons for it, with leadership instability a common factor.

6. The strength of the impact of federal politics on state politics can predict otherwise surprise results like the 2014 South Australian election win by Labor.

7.  Furthermore, if the federal government is of the same party as the state government then there is a strong relationship between the popularity of the federal government and the fate of the state government.

8. Even without any knowledge of events or polling in this parliamentary term, models based on the history of these two issues still imply that a change of government in Victoria is likely.


Warning: this article is long and pretty wonky in places; about Wonk Factor 3/5.

=====================================================================

In my recent article Victorian Liberals: Going, Going ... I claimed that a first-term Australian state government seems objectively likely to lose, and would seem that way even if we knew nothing about its polling history.

I felt a bit nervous making that claim, but now I go much further.  An element of the argument was that the Baillieu-turned-Napthine government had endured a chaotic term, and that both the loss of a Premier and going to an election in minority (as well as all the other Geoff Shaw associated nonsense) increased its chance of defeat.  I am now going to make the more radical claim that even if we did not know that this government had experienced such turbulence, and did not know anything about its polling history, we should still expect it - all else being equal - to probably lose, based on just two factors: its very slim seat margin from the last election, and the fact that it is the same party as the government in power in Canberra.

Much of the discussion about the fates of recent state governments at elections in Australia has focused on the age of the governments in question.  The 16-year-old four-term governments in NSW and Tasmania were both thrashed, the 14-year five-term government in Queensland was obliterated, the 12-year-old Victorian government lost somewhat unexpectedly, and the 12-year-old South Australian government lost the 2PP vote heavily and its majority as well but somehow clung on to office.

We haven't seen a state government survive beyond 16 years since 1986, and it's been tempting to think that the reason for that is that it's just too hard for any government in the modern world to last that long.  The idea is that incumbent governments discover that "friends come and go, but enemies accumulate".  The longer they have been in power, the less they can blame their predecessors for their sins, and the less voters remember about what the other lot ever did wrong.  But it is not so long ago that state governments could last decades: the Liberals ruled for 27 years in Victoria following the DLP split, while the Nationals (aided by a distorted electoral system) governed Queensland in some form or another for 32 years until their empire came to a very messy end in 1989.

Is it that modern times have made it impossible for state governments to last 20+ years anymore?  I'm going to say otherwise.  It's just very unlikely, because every election presents a chance of losing office, and some elections present a greater chance than others.

Methods

My interest in what I'll call "federal government drag" on state results was greatly increased by a seminal blog post by Antony Green that graphed the way in which federal and state governments seem to exist in opposing cycles.  With rare exceptions, not long after a party wins a federal election it starts shedding seats in state elections and continues to do so until it sooner or later loses office federally as well.

But could it be that these state governments are actually losing office because they tend to be a certain age by the time their own party finally wins office federally, meaning that it is actually their age and not the Canberra factor that causes them to be kicked out?

I thought I'd take a look at this, using the same starting point used by Antony (the election of the Bethune government in Tasmania in 1969, at which point the Coalition ruled in every state and federally).  I classed each of 82 state elections as to whether the same party/side held office federally, the age of the state government (rounded to a year), the result of the state election, and also the change in the party's representation in the parliament.  The latter was represented as the difference in the percentage of seats held by the government after the previous election compared to after the one being examined.  So, for instance, the worst result on that basis was Queensland 2012 (the difference between Labor's "before" and "after" seat tallies was 49.4% of the Parliament) while the best was Queensland in 1974 (the National government gained 26.8%).

I treated the 1995 Queensland election as inconclusive - although the Goss government briefly continued in office, its win in a seat it relied on was annulled, resulting in a by-election that determined the fate of the government.   The only other serious complication was the 1983 Queensland election.  The Nationals and Liberals won the 1980 Queensland election as a Coalition.  Before the 1983 election the coalition dissolved.  The Nationals ran as the government seeking a majority in their own right and almost achieved it, later forming government after two Liberals defected.  I have treated the Nationals solely as the government at this election.

Because some people might think of even the latter halves of the 1955-82 Victorian and 1957-1989 Queensland governments as relics of the ALP splits of the mid-1950s, I also looked at a subset of results from 1990 onwards only.

The points I looked at were:

* The win-loss rates of state governments based on whether they were the same party as the party in power federally or not.

* The win-loss rates of state governments in various age classes.

* Any correlation between either age class or party in power federally and the proportion of seats won or lost by the state government.

Federal drag correlates strongly with state outcomes

Here's the contingency table since 1969:

Overall in this time state governments have had a 65% win rate, excluding the one inconclusive election from the sample.  But for state governments that had the same party in power federally, the win rate is only 54%. For those of opposite party to the party in power federally the win rate is over 82%.  The difference is easily statistically significant.

It's worth looking at the six state governments that lost while the opposite party was in power federally.  They were:

* The Des Corcoran Labor government lost in South Australia in 1979.  The previous premier, Don Dunstan, had resigned abruptly for health reasons.  Corcoran called an unnecessary snap election seeking a mandate in his own right only two years after the last poll. 

* The Harry Holgate Labor government lost in Tasmania in 1982. This government had rolled its own Premier Doug Lowe, who quit the party with a colleague, thus reducing it to minority status and causing it to lose the confidence of the Parliament before the election.

* The Robin Gray Liberal government lost in Tasmania in 1989.  The government fell one seat short of a majority under the Hare-Clark multi-member system despite polling a primary vote of 46.9%.  Under a single-seat system the Gray government may have just won the 2PP and would probably have been returned with an outright majority even if it didn't.

* The Russell Cooper National government lost in Queensland in 1989.  This was the dreg end of the Joh Bjelke-Petersen reign and the party had been through not one but two acrimonious changes of Premier and also the Fitzgerald Inquiry findings of extensive corruption in government and policing.  The defeated government had been in office in one form or another for 32 years.

* The John Fahey Liberal government lost in New South Wales in 1995.  This government had lost its majority at the previous election and its initial Premier, Nick Greiner, had been forced to resign by the crossbench following unfavorable ICAC findings.

* The Lara Giddings Labor government lost in Tasmania in 2014.  This government had been in office for fifteen and a half years, most of the last four of these in an unpopular coalition with the Tasmanian Greens, the formation of which was widely viewed as a breach of commitments at the previous election. This government had also seen a turnover of Premier during its term.

Five of these six cases involved a turnover of Premier, and most of these governments were severely controversial.  It's possiby significant that fully half of the cases came in the state with a multi-member electoral system.

From 1990 onwards the pattern has only become more dramatic.  In this time, 21 state governments have faced the polls while their own party was in power and 13 of these have lost and seven won, for a win rate of 35% (again I'm excluding Queensland 1995).  18 state governments have faced the polls while their own party was not in power federally and of these just the Fahey and Giddings governments have lost, for a win rate of 89%.

Of those governments since 1990 that have won while their party was in power federally, four were first-term governments that had won by large margins at the previous election, and in two of the other three cases a very high proportion of seats needed to change hands for the government to change.  The most modest mountain that wasn't climbed by an Opposition in such cases was in South Australia in 2010.  A swing against the government of 10.6% of the seats could have given the Liberal Party victory, but the swing achieved was only two seats or 4.3%.

Incumbency: a relevant factor, but as strong?

The above is a similar contingency table for the post-1969 and post-1989 tables.  I've done one version where I split the age classes into five-year brackets and another where I just split it into governments in their first five years (rounded) and the rest.  Note that for the post-1969 sample, while the strike rate for governments aged 6 to 15 years is worse than for governments 5 years old or younger, the very long-lived governments in Victoria and Queensland post-split give the "Over 15" bracket a rather good survival rate.  But we don't get to see the younger days of those governments in the sample, since they were 14 and 12 years old at its startline.

Grouped by age into just two groupings (up to 5 years old and over 5 years old) the difference in the second table for the post-1969 table is only just statistically significant.  The difference for the post-1989 table is much more pronounced, but not quite as pronounced as the difference by party in federal power. 

There is a recent difference in the age of governments that faced elections with the same rather than a different party in Canberra, but crucially not a very big one.  In the post-89 sample, the average age of governments that went to elections with the same party in power federally is about a year and a half greater.  For the whole post-69 sample, the average ages are practically the same.

Now here's the relationship between the age of a state government and the percentage of seats in the parliament that the government loses or gains:

Governments more often lose seats at elections than gain them, but on the right we can see the dot-points for the endless Liberal government in Victoria making a small gain at age 21, and the similar National government in Queensland doing so at ages 26 and 29 before finally being removed.  I've put a dotted line of best fit through it but the relationship is very weak with less than 4% of variation explained, and the expected difference between the youngest and oldest governments about 10% of the size of parliament.  By contrast, the simple on-off variable of who is in power in Canberra explains 17% of the variation in how much a state government's share of parliament changes.  On average if the party in power federally was the same, a government lost seats worth 9.2% of the parliament (+/-10.9%) while if the party in power federally was different, a government gained 1.0% (+/-11.8%).  Different-party governments were as likely to improve their percentage share of Parliament as they were to go backwards, while same-party governments improved their share in six cases, held steady in two, and went backwards in 41.

Here's the same for the post-1989 sample:

In the post-1989 sub-sample a much stronger relationship between the age of the government and its seat swing is suggested, with 36% of variation explained and the oldest governments underperforming by over 20% of the parliament compared to the younger ones.  But the relationship between the party in power in Canberra and the seat outcome is also quite strong, explaining 32% of variation in seat share outcomes, with governments that differ from the federal party in power getting a better seat swing result by an average of 15.9% of the number of seats in a Parliament.

Incumbency and federal politics both matter

So we've seen a somewhat different pattern between the full post-1969 pattern and the post-1989 sub-sample.  In both cases the party in power federally is relevant, but it has been more relevant in the last 25 years.  The length of a party's incumbency has also been relevant, but again this pattern is much stronger once we get past the two governments that went on for decades.

The latter aspect gets rather circular.  It's hard to say for sure that a government that was 25 years old would definitely lose office today, since we haven't actually seen a government that old make it to an election in the past 25 years.  It's easy to say that the longevity of the Victorian (1955-1982) and Queensland (1957-89) governments reflects a deeper impact of the mid-50s ALP split, but the DLP was a spent force in both states from the mid-70s and yet those governments kept winning.  It might be that the relationship between term in office and chance of victory is actually not linear and that very old governments are more likely to win elections than fairly old ones.  Not enough data on this, so we don't know.

Given that there is no corellation across the whole sample between the age of  a state government at election time and whether or not the same party is in power federally, and only a possible but hardly major relationship in the last 25 years, it seems that incumbency and federal government drag are both factors.  Since they don't correlate with each other to any great degree, they're probably both part of the puzzle of predicting state election results, and it should be expected that predictive models can make use of both aspects.

On that basis, here are two multiple regressions for seat swing:

The whole sample: 

Expected size of parliament change = -.0594+.1036*(fed govt different)-.00369*(age of state govt) (+/- .109)

"fed govt different" is 1 if the party in power federally is different, 0 otherwise.

So for a 7-year-old government that is different from the party in power federally, the predicted result is that it gains -.0594+.1036-.02583=0.01837 of the parliament, or 1.8% of the seats.

The post-1989 sample:

Expected size of parliament change =.0039+.133*(fed govt different)-.019*(age of state govt) 
(+/- .093)

In this case for a 7-year-old government that differs from the party in power federally, the predicted result is that it gains 0.0039 of the parliament, or 0.39% of the seats - effectively, no seat change.

Here's what the two models look like compared to the spread of actual results:

Whole sample:

This one looks a bit odd, because it has such a small government-age term, and so most of the projections fall in two clumps.  It expects every government that is of the same party as a federal government to lose seats worth over 5% of the parliament, but it only expects different-party governments to go backwards if they are at least 12 years old.  The model explains 21% of variation.

Post-1989 sample:

This is a much neater-looking model than the full sample one.  It predicts a more natural-looking spread of outcomes, which is not surprising given that in this case having the same party in power federally is only worth seven years of incumbency, compared to 28 years in the full-sample model.  It's also a much more effective model, explaining 57% of variation based on two variables for a sample of 39 items, and doing vastly better than either variable alone. 

How These Models See Recent Elections

The recent South Australian state election was a surprise result on the surface.  A bit like the 1993 federal election, it was a classic model-buster for predictions (and an especially telling blow against the "betting odds' paradigm).  A very long-lived government had a slim seat margin, had had a change of Premier and was polling very poorly.  It was at double-figure odds to be returned on election day, with the Opposition at $1.02.  Even with my familiarity with federal government drag, I thought the government would probably lose (but that it did have realistic survival chances).

The models above see that in a very different light.  The full-sample model's projection for South Australia 2014 is a seat change of zero, ie that the government would be returned.  The post-1989 model's projection is a seat change of -0.09 of the parliament, ie that the government would lose about four seats (which would have still left it with a hung parliament, albeit one in which the Liberals would most likely have taken power).  If we take a loss of three seats as the cut-off point for Labor to retain office then these models imply Labor's chance of retaining office was about 75% for the full-sample model and 43% for the post-1969 one.  (Admittedly, the SA 2014 data are already included in both models, but as the result is so close to what both predicted, that makes very little difference.) 

The Tasmanian result surprised no-one except a few political insiders and far-leftists, who move too much in limited inner-urban circles and failed to accept the mood of the voters outside those circles.  The government's losses were worth 12% of the parliament, although a case can be made for treating Labor and the Greens as a coalition government (despite the staged dissolution of that coalition) and counting the losses as 20%.  The Hare-Clark system tends to mute the damage of a given swing in seat terms, but it also tends to amplify the size of a swing to begin with, because voters shift their voting behaviour to try to avoid hung parliaments.

Anyway, if we take a loss of two seats (8% of parliament) as being the level likely to lose government (given that the Greens were expected to lose one) then the post-1989 model gets it more or less right, predicting a loss of four seats and giving the ALP government a generous 18% chance of having been returned.  However the weaker full-sample model doesn't weight sixteen years in office as heavily as there being an opposite government in Canberra, so it wrongly expects that the Giddings Government would continue.  A possible weakness of both models is that they don't incorporate how long the federal government in question has been there for (in this case just six months.)

These sorts of models can't be a substitute for looking at the polls to see what voters are thinking, and it was polls and a view about the way Tasmanian voters react to coalitions involving the Greens that enabled me to be very confident about the Tasmanian election result in advance.  But as seen in the modelling posted here in the lead-up to the federal election, the percentage of variation in election results that is explained by polling takes until the last few weeks to rise above about a third.   So in cases where a government is not clearly dead in polling, and it is a long time til election day, they may provide useful default predictions.

Outliers:

Looking at recent results, the part-sample model flags some of them as exceptionally good or poor performances.  The five strongest performances by state governments after factoring in age of government and party in federal office are: Queensland 2001 and Victoria 2002 (Beattie and Bracks make massive second-term seat gains capitalising on opposition disunity), WA 1996 (Court government improves standing with Howard government still in honeymoon phase), Qld 2009 (initially successful switch to Anna Bligh saves seats despite Labor being in power federally) and SA 2010 (the model says Rann should have lost, but he won). 

The five worst performances were Queensland 2012 (the model predicts a landslide but only one half as big as actually happened), SA 1997 (massive swing against first-term Liberal government that rolled its own Premier), Tas 1996 (first-term Liberal government loses its majority even with Labor still just in power federally), WA 2001 (two-term Court government walloped) and NSW 1991 (Greiner government loses majority in one term despite unpopular Hawke/Keating government). 

Coming Elections:

The next Victorian (by Nov 2014), NSW (scheduled March 2015), and Queensland (by June 2015) state elections are all extremely likely to be held with the Coalition in power federally, while the next Western Australian (March 2017), South Australian and Tasmanian (both March 2018) may or may not be.

The two models give the Victorian government chances of 28.2% (full sample) and 25.7% (part sample) of winning at least 50% of the seats, and each project that it will lose about six seats.  They make similar seat loss predictions for the Queensland and New South Wales parliaments and consider the return of each government more or less certain, but those two projections are worthless because defending governments with such huge majorities as these are outside the model's data set.  The only real takeaway is that those governments are highly likely to return because they are young and have very large margins.

More relevantly, if the Coalition is still in power federally, then the part-sample model predicts the boot for the Barnett government (40% chance of retaining government in coalition with the WA Nationals) although the full-sample model is more optimistic for it (68% chance of at least a coalition government).  If, however, Labor is back in the Lodge by then, then both models give the WA Liberals about a 90% chance of trucking on, with probably little seat change.

For Tasmania if the Federal Coalition remains in power then both models see the Hodgman government continuing in majority with a loss of about two seats (I'd suggest the 60% majority win chance is a bit pessimistic given the differences created by Hare-Clark).  But if Labor is in power federally then the models give Hodgman about a 90% chance of a second majority term.

Finally, South Australia: the part-sample model's projections for a 16 year-old government should be treated with caution since there are so few in the data range.  The full-sample model gives the Weatherill government a roughly even chance of yet another term if the Coalition is still in Canberra, but only about a 12% chance otherwise.

These projections are, again, no substitute for a poll taken the day before the election - but it will be fun to see how the model pans out compared to actual results. 

Why is it so?

I should note here that I haven't proven that federal drag will continue to affect state election results to the indicated degree.  I've just observed that it has done so over a couple of time scales, and predicted this will continue to be the case to something like that degree over time. That prediction can now be followed and tested by future data to see if it was right.  (The time scales are arguably cherrypicked, but the post-1989 one could have been made more so by starting it after the 1986 Queensland election, while it is important to get clear of most of the impact of the DLP split when looking for a starting point overall.) 

However, the apparent result makes sense.  There are two obvious reasons why the party in federal power should influence the outcome of state elections.  Firstly, voters who are disgruntled with the performance of a federal government may use a state election to send the federal party a message.  Secondly, many voters seem to like having different parties in power at state and federal level.  They believe that if the parties are the same, the state government will be too compliant and will not stand up for the state against the federal government's decisions.

Added: Does Federal Government popularity matter?

Matthew Bowman (comments) asks whether the popularity of the federal government matters as well as which party is in power.  So I thought I'd look at that, just for the post-1989 data set so I could wheel out my trusty Newspoll rolling average and be applying the same polling standard across all the elections included.

I split the results into two groups.  Firstly, those when the state government was of the opposite party to the federal government:

 It might be expected here that if the federal government was on the nose the voters would give a state government of the opposite party a boost.  In fact there's nothing to see here; the Bracks government made a huge seat pick-up on re-election with the Howard government ahead in the polls federally, while the Fahey government lost in NSW with Keating's ALP polling below 45% 2PP.

Now when the state government was of the same party as the federal government at the time:

Now that looks more impressive. For state governments of the same party as the federal government, the federal government's 2PP at the time of the state election explains 29% of variation in the state government's results.  If the national 2PP is worse than 50% then a state government of the same party will typically lose over 10% of its seats.  The break-even point is 59% 2PP (somehow I don't think the current federal government will be polling that in November) and below that a state government expects to lose nearly 1.8% of the state parliament for every point of federal 2PP. 

Of course, there are exceptions; the Goss government in Queensland suffered no drag from the Keating government in late 1992.  (The purple dot near the 45 might look impressive too except that it was Tasmania 1992, and the seat loss is muted by Hare-Clark; that government was actually heavily defeated). 

It's true here that Queensland 2012 is an outlier that has a large influence, but even if it is removed the relationship still explains 20% of variation and still costs a state government 1.25% of its parliament for every point of federal 2PP loss (and this time from an implied break-even point of 61.5% 2PP). 

(It's possible some of this is because unpopular state governments can drag down the polling of a same-party federal government, and testing the direction of causation on this would be a messy business.  My suspicion is that the drag is much more federal-on-state than the other way round.)

Another thing that is notable here is that for state governments of the same party that go to elections, the average federal government Newspoll 2PP at the time has been 51%.  When the state government going to an election is of a different party, the average federal 2PP has been 49%.  (Given the remaining findings of this article, this isn't surprising.) So if the Abbott government is polling worse than 51% 2PP come November, the Napthine government's chances of re-election based on historic results should be lower still than the 25.7% figure given above. 

If the Coalition remains unpopular federally, it will be a remarkable feat by historic standards should the Napthine government survive.