On the night of the 2023 NSW election, I along with most other observers had expected that Labor would achieve majority government.
After midnight I turned off the ABC’s predictive tools and assessed every seat solely on the votes counted. The tally I came to had Labor on a certain 45 seats, ahead in another four seats, and based on past trends had a reasonable chance of winning three seats where the Coalition was ahead on votes.
The next morning I rang the Sydney news room suggesting that news reports refer to Labor winning but step back from stating Labor had achieved a majority. On the seats remaining in doubt, you would have expected Labor would win at least two of the doubtful seats to achieve majority government. That was the position I reported on ABC news on the Sunday night.
This did not happen. Labor won only the 45 seats I had marked down as definite Labor wins on my Sunday morning check. In the four seats where Labor was ahead, and every other seat that was close, Labor went backwards with each day’s counting.
On Monday 27 March, the uncounted pre-poll centres were added to the count. A trend against Labor emerged and on the Monday evening I tweeted that Labor would probably miss out on a majority and win 45 or 46 seats. In seat after seat the addition of pre-poll votes on Monday had revealed a consistent decline in Labor’s position.
That trend went even further on the Saturday after polling day when the largest batches of postal votes were added in key seats. For the first time the Liberal Party pulled ahead in Ryde and moved even further ahead in Terrigal. Both seats had looked like Labor gains on election night.
As an election analyst, such post election night shifts set you up for criticism. You are accused of not taking into account that pre-polls and postals would favour the Coalition.
Actually, the model I use builds in a correction for postal and pre-poll voting trends. The model factors in the postal and pre-poll vote trends from the last election.
What happened in 2023 is that the 2023 pre-poll and postal vote results turned out to be very different to the election day results. The trend to the Coalition in post-election counting was much larger than in 2019.
There were also many more pre-poll and postal votes in 2023. The early vote broke more strongly for the Coalition than previously, and the impact of these votes was amplified in the final result by their greater weight of numbers.
Vote Categories in 2023
The chart below graphs the percentage of total votes cast using different voting channels. The trend in the chart is the same as for elections in other states and at Commonwealth elections. There has been a dramatic shift in the past 15 years from voting on polling day to voting before polling day. While many keep referring to the increase in postal voting as important, the real shift has been to voting pre-poll in person.
As recently as the 2007 election, 80% of NSW votes were still recorded as ordinary votes cast within district on polling day. The figure in 2023 was down to only 50.6%.
The big surge has been to in-person pre-poll voting, a six-fold increase from 5.5% in 2007 to 32.6% in 2023. The same figure in 2019 had been 22.1%, an 11.5 percentage point difference or near 50% increase in the proportion at the 2023 election.
This surge was despite the period for pre-poll voting being decreased from two weeks previously to only one week in 2023. To take more votes in a shorter period, most urban districts saw an increase in the number of early voting centres. But only one pre-poll centre was counted for each district on election night. This reduced the overall number of votes counted on election night. Without reliable historic data for some of the counted pre-poll centres, another level of complexity was added to election night predictions.
There has also been a change in early voting categories. In 2019, 5.1% of votes had been taken using the NSW Electoral Commission's iVote system, reducing the category of postal votes to just 3%. In 2023 iVote was not available for technical reasons and postal votes expanded to represent 8.5% of all votes. Some postal votes were counted for each district on election night, but would they reflect the smaller postal vote category of 2019, or an amalgam of postal and iVotes from 2019? Post-election analysis suggests combining Postals and iVotes from 2019 provides the better comparison.
Another unknown was how an increase in pre-poll voting would be reflected in Absent votes. At Federal elections, Absent votes include only votes cast at a polling places outside of a voter's home district on election day. In NSW, Absent votes include Pre-poll votes cast at early voting centres outside of the voter's home division. Would a Coalition bias in pre-poll voting flow through to a higher percentage for the Coalition in the Absent vote category?
Early Voting trends in 2019
The first two lines in the table below shows the first preference vote by party on polling day in 2019, and the same percentages for non-polling day votes. Two lines below is a Final Result %. A breakdown of party % votes for each category of non-polling day votes is also included, and the % of all votes in each category is shown.
Two further lines need explaining.
- Difference - Polling Day % minus All Other Votes %. A '+'indicates that Polling Day % was higher, a '-' that it was lower.
- Shift - Calculated as Final Result minus Polling Day %. A '+' indicates how much the party's % rose with post-election day vote counting, a '-' that the % fell.
What the table shows is -
- The Coalition polled 0.2% higher in non-polling day votes, resulting in the Coalition % vote rising 0.1% by the final count.
- Labor polled 0.7% lower in non-polling day votes producing a 0.3% shift from election night.
In two-party preferred terms, in 2019 the Coalition polled 51.6% with polling place votes and 52.6% for all other votes for an election total 52.0%. So the Coalition Two-Party Difference figure was +1.0% and the two-party shift post-election was +0.4%. These comparisons are not yet available for the state, but I have compared 2PP% in looking at individual seats in this post. The shift in individual seats is more important than the state-wide shift.
|All Other Votes||38.8||41.7||32.9||9.3||16.2|
Early Voting Trends in 2023
The table below shows the same analysis for the 2023 election. It is based on the first count of votes released by the NSWEC.
|All Other Votes||49.4||36.8||36.4||8.9||17.9|
The table shows clearly that the Coalition did much better in late counting than in 2019.
- The Coalition polled 2.7 percentage points higher in non-polling day votes compared to only 0.2 percentage points in 2019. The increase in percentage between polling day votes and the final result was 1.3 percentage points as against 0.1 in 2019.
- Support for both Labor and the Greens was lower in the post-polling day votes and support for 'Others' was lower.
While two-party preferred figures are not yet available, I suspect the two-party Difference will be about two percentage points in 2023 and the two-party Shift around 1.5 percent.
In other words, the Coalition did significantly better on non-polling day votes than in 2019, and that better performance wasn't built into election night predictions.
Let me turn to the two-party preferred figures in individual seats that were in play at the end of election night.
On election night the ABC computer had Labor on 50.6% on a 57.9% count, and projected to increase to 51.4% when all votes were counted. In a seat with an 8.9% margin, the matched swing at that point was 10.3%. The Ryde EVC had reported, while Eastwood EVC along with the two new EVCs at Epping and Macquarie Park were to be counted on Monday.
Polling day votes as a percentage of all votes in 2023 were 51.4% (2019 66.9%) and Pre-polls 27.4% (14.2%). 2023 Postals were 14.1% compared to only 4.2% in 2019 plus another 6.4% iVotes.
In 2019, the Liberal two-party % for polling day votes was 59.1%, the final result 59.0%, so the non-polling day votes had produced a tiny -0.1 percentage point shift in the Liberal vote. There are some slight discrepancies here produced by the redistribution but that should not cause major problems with measurement.
In 2023 the final figures were the Liberal Party was on 48.1% with polling day votes, 51.9% on pre-polls and 54.8% Postals, the final result Liberal 50.1%. Where the Liberal final result shifted -0.1% in 2019, it shifted +2.0% in 2023. Using the 2019 figures was why the ABC computer projected Labor to finish on 51.4% based on the matched swing.
Comparing like swings, the swing against the Liberal Party was 11.0% on polling day, but only 7.0% with the larger 2013 category of Pre-polls and 5.9% comparing Postals to a combined Postal and iVote figures from 2019.
In short the computer projection had accounted for postals and pre-polls having a different percentage from polling day votes, but not for the swing being different. That caused the computer to under-project the final Liberal result. The Liberal Party won Ryde by only 54 votes and that after a re-count.
On election night the ABC computer finished with Labor on 51.3% on a 57.2% first preference count (54.0% 2CP counted) and projected to finish on 51.1%. In a seat with a 12.3% margin, the matched swing at that point was 13.4%. The Terrigal EVC had reported with an 11.5% swing with three more EVCs to be counted on Monday. No Postal count had been received.
Polling day votes as a percentage of all votes in 2023 were 55.2% (2019 62.2%) and Pre-polls 26.1% (23.2%). 2023 Postals were 9.8% compared to only 3.5% in 2019 plus another 4.0% iVotes.
In 2019, the Liberal two-party % for polling day votes was 62.2%, the final result 62.3%, so the non-polling day votes had produced a tiny +0.1 percentage point shift in the Liberal vote. Terrigal was untouched by the redistribution.
In 2023 the final figures were the Liberal Party was on 48.4% with polling day votes, 53.8% on pre-polls and 62.4% Postals, the final result Liberal 51.2%. Where the Liberal final result shifted +0.1% in 2019, it shifted +2.8% in 2023.
Comparing like swings, the swing against the Liberal Party was 13.8% on polling day, but only 11.0% with Pre-polls and only 2.3% comparing Postals to combined Postal and iVote figures from 2019. The Absent vote swing was also low at only 4.0%.
In short, Labor romped it in on election day, did not make up as much ground on Pre-poll votes, but completely failed to generate a swing with postal votes. The ABC's prediction had factored in Labor going backwards on Pre-polls and Postals, but had not factored in that Labor would be flogged on Postal votes.
Kiama was a complex contest. The computer was comparing Gareth Ward's support as an Independent versus Labor, against his 2019 support as a Liberal MP versus Labor. The computer model did factor in the potential for greater variability in the swing.
On election night the ABC computer had Labor on 51.9% and projected to finish on 51.5%. Looking at the figures with more time and the benefit of hindsight, the count in Kiama was not far enough progressed to be confident Labor would win. On the night we gave Kiama to Labor, but having re-considered the election night setting in our system, a different setting would have left Kiama in doubt. After reviewing the 2023 result, that different setting will become the default for future elections.
The count at the end of election night had a 41.3% primary vote completed but only a 32.3% preference count. Crucially on election night we did not receive a preference count for the 2,200 postal votes counted. We also received only the small Kiama Election Manager's Office pre-poll with 350 votes, not the three big pre-poll centres in Albion Park, Kiama and Nowra.
Polling day votes as a percentage of all votes in 2023 were 45.1% (2019 51.9%) and Pre-polls 41.0% (33.3%). 2023 Postals were 7.8% compared to only 2.9% in 2019 plus another 4.0% iVotes. Kiama was untouched by the redistribution.
In 2019 the final figures were Gareth Ward as the Liberal candidate on 62.0% with polling day votes, 63.1% on pre-polls and 70.9% Postals, the final result Liberal 62.0% so no shift in the Liberal final result compared to polling day votes. In 2023 the shift for Ward as an Independent compared to polling day was +2.2%.
Comparing like swings, and again comparing Ward as an Independent versus Ward as a Liberal, the swing against Ward was 13.4% on polling day, but only 10.4% with Pre-polls and only 5.9% comparing Postals to a combined Postal and iVote figures from 2019.
There are several lessons here. First, as already mentioned, a slightly different internal setting in the predictive model would have left Kiama in doubt on election night. Second, the model always performs more erratically when you are not doing a like versus like comparison. You must be cautious using a Labor versus Liberal comparison with a new Labor versus Independent pairing. Two-party preferred variance from polling place to polling place reflects variance on the primary vote for the two-parties. In Kiama, where Ward polled well as did the Greens, there is another variance distribution describing how those candidate's vote varied across polling places, which complicates analysis of variance on a traditional two-party basis.
Finally, the missing Postal 2CP count and the fact the three big EVC's weren't in was not given enough attention in the hurly-burly of election night. The lesson is in a seat like Kiama, someone with good knowledge of on the ground figures will perform better than a standard computer model.
And an aside on Kiama. On Monday 27 March, Labor dropped behind in Terrigal and Miranda with the addition of pre-poll votes, Holsworthy and Goulburn did not budge, the Labor's lead in Ryde narrowed as it did with EVC counting in Kiama. That evening I spoke to Gareth Ward who gave me the unreported Nowra EVC count where he polled 61.0% out of 6,500 votes. Once that centre was reported the next day, and after I prodded the NSWEC to have the 2CP count for the Postal votes added, it was clear Ward was going to win Kiama. So it was local count information ahead of the computer that was behind my Monday night tweet predicting Labor would finish with only 45 or 46 seats.
Kiama, Ryde and Terrigal were the difference between the ABC computer predicting Labor reaching a majority and the actual result of Labor winning only 45 seats. Miranda was a fourth seat where Labor led but it was kept in doubt. Three other seats listed below had Labor trailing but, based on 2019 trends, still in with a chance of winning.
I'll deal with these seats in less detail.
On election night the ABC computer had Labor ahead in Miranda with 50.3% but projected to finish on 49.7%. The seat was correctly left in doubt. No postal vote counts had been received, no 2CP count had been received for Sylvania EVC, and there were four more EVCs to report. Additional counting of the missing votes on the Monday made clear the Liberal Party would hold Miranda.
Miranda had a margin of 14.4%. The polling day swing was 14.2%, pre-poll swing 10.6, and Postal swing only 7.1% comparing to Postals and iVote in 2019.
The difference in the Liberal 2PP polling day % vote and final result was -0.7% in 2019 but +1.3% in 2023.
On election night the ABC computer had Labor behind in Goulburn on 49.7% but projected to reach 50.7%. The seat was classed as in doubt. Preference counts were missing from five polling places, a thousand postal votes had been counted showing a 3.7% swing to Labor, the same as the overall swing. The new Crookwell EVC had reported but there was no history available as a comparison. The big Yass and Goulburn EVCs were to be counted on Monday.
In 2019 the Liberal position had slipped 0.1% between polling day votes and the final count. In 2023 the difference was +2.3%.
Goulburn had a margin of 3.1%. The polling day swing was 4.8% to Labor. The Pre-poll swing was 0.7% to the Liberal party and Postals were 4.7% to the Liberal Party compared to Postals and iVotes in 2019. Goulburn was rightly left in doubt on election night.
On election night the ABC computer had Labor on 49.2% but projected to finish on 50.6%. The margin was 6.0%, the election night matched swing 6.7%, and the first batch of 2,000 postal votes showed a 10% swing to Labor. Holsworthy had gained areas around Liberal voting Menai in the redistribution. The new Menai EVC with 3,000 votes had reported a primary count but not yet a preference count. Holsworthy was rightly left in doubt.
Early vote counting on the Monday firmed up the Liberal position and postal vote counting the following Saturday confirmed the Liberal victory.
On different boundaries in 2019, the difference between the Liberal poling day result and the final result was -1.3%. In 2023 it was +0.1%, another example of a strong Liberal improvement in counting after election day.
Measuring election day swing by vote type is not reliable in Holsworthy due to changes in electoral boundaries.
On election night the ABC computer had Labor on 49.6% but projected to finish on 50.6%. The margin was 6.8% and the election night matched swing 7.4%, the same swing as on the first batch of postals. 3,500 pre-polls had been counted at the external Hurstville EVC showing a 15% swing but this was probably unreliable given it was external. The Labor gain at Hurstville EVC was more than reversed by the Mortdale EVC counted on Monday.
On different boundaries in 2019, the difference between the Liberal poling day result and the final result was -0.6%. In 2023 it was +0.3%, not as dramatic a shift to the Liberals as in other seats, but enough to turn a close seat on election night into a Liberal retain. The eventual Liberal margin of victory was 0.5%.