Mapping Melbourne’s journeys to work

Mon 24 June, 2019

The unwritten rules of mapping data include avoiding too much data and clutter, and not using too much colour. This blog often violates those rules, and when it comes to visualising journeys to work, I think we can learn a lot about cities with somewhat cluttered colourful animated maps.

This post maps journeys to work in Melbourne, using data from the 2016 census. I will look at which types of home-work pairs have different public, private and active transport mode shares and volumes.

Although this post will focus on Melbourne, I will include a brief comparison to Sydney at the end.

Where are public transport journeys to work in Melbourne?

First I need to explain the maps you are about to see.

So that I can show mode shares, I’ve grouped journeys between SA2s (which are roughly the size of a suburb). Lines are drawn from the population centroid of the home SA2 (thin end) to the employment centroid of the work SA2 (thicker end). Centroids are calculated as the weighted average location residents/jobs in each SA2 (using mesh block / destination zone data). This generally works okay for urban areas, but be aware that actual trips will be distributed across SA2s, and some SA2s on the urban fringe are quite large.

The thickness of each line at the work end is roughly proportional to the number of journeys by the mode of interest between the home-work pair (refer legends), but it’s difficult to use a scale that is meaningful for smaller volumes. Unfortunately there’s only so much you can do on a 2-D chart.

I’ve not drawn lines where there are fewer than 50 journeys in total (all modes), or where there were no journeys of the mode that is the subject of the map. This threshold of 50 isn’t perfect either as SA2s are not consistently sized within and between cities, so larger SA2s are more likely to generate lines on the map.

To try to help deal with the clutter, I’ve made the lines somewhat transparent, and also animated the map to highlight trips with different mode share intervals. For frames showing all lines, the lines with highest mode share are drawn on top.

So here is an animated map showing public transport journeys to work in Melbourne, by different mode share ranges and overall:

Technical note: I have included journeys to work that are internal to an SA2. Usually these appear as simple circles, but sometimes they appear as small teardrops where the population and employment centroids are sufficiently far apart.

You can see that the highest PT shares and largest PT volumes are for journeys to the central city, and generally from SA2s connected to Melbourne CBD by train (including many outer suburbs).

As the animation moves to highlight lower PT mode share ranges, the lines become a little less radial, a little shorter on average, and the lowest (non-zero) PT mode shares are mostly for suburban trips.

A notable exception is journeys to Port Melbourne Industrial SA2 (also known as Fishermans Bend), which is located at the junction of two major motorways and is remote from rapid public transport (it does however have a couple of high frequency bus lines from the CBD).

The lowest PT mode shares are seen for trips around the outer suburbs. The maps above unfortunately aren’t very good at differentiating small volumes. The following animated map shows public transport journeys with a filter progressively applied to remove lines with small numbers of public transport journeys (refer blue text in title):

You can see that most of the outer suburban lines quickly disappear as they have very small volumes. Inter-suburban lines with more than 50 public transport journeys go to centres including Dandenong, Clayton, Box Hill, and Heidelberg.

Here’s another animation that builds up the map starting with low public transport mode share lines, and then progressively adds lines with higher public transport mode shares:

As an aside, here is a chart showing journeys to work by straight line distance (between SA2 centroids), public transport mode share, work distance from the CBD and home-work volume:

The black dots represent journeys to the inner 5km of the city, and you can see public transport has a high mode share of longer trips. Public transport mode share falls away for shorter journeys to the inner city as people are more likely to use active transport. A dot on the top left of the curve is 8,874 journeys from Docklands to Melbourne – which benefits from the free tram zone and the distances can be 1-2 km. Most of the longer journeys with low public transport mode share are to workplaces remote from the CBD (coloured dots).

Another way to deal with the clutter of overlapping lines around the CBD is to progressively remove lines to workplaces in and around the CBD. Here is another animated map that does exactly so that you can better see journeys in the nearby inner and middle suburbs.

As you strip away the CBD and inner suburbs, you lose most lines with high public transport mode shares and volumes. However some high public transport mode share lines remain, including the following outbound journeys:

  • Melbourne (CBD) to Melbourne Airport: 72% of 64 journeys
  • Melbourne (CBD) to Box Hill: 66% of 76 journeys
  • Melbourne (CBD) to Clayton: 57% of 82 journeys
  • South Yarra – East to Clayton: 57% of 173 journeys

Just keep in mind that these are all very small volumes compared to total journeys in Melbourne.

You might have noticed on the western edge of the map some yellow and orange lines from the Wyndham area (south-west Melbourne) that go off the map towards the south west. These journeys go to Geelong.

Here’s a map showing journeys around Geelong and between Geelong and Greater Melbourne (journeys entirely within Greater Melbourne excluded):

You can see very high public transport mode shares for journeys from the Geelong and Bellarine region to the Melbourne CBD and Docklands (and fairly large volumes), but no lines to Southbank, East Melbourne, Parkville or Carlton – all more remote from Southern Cross Station, the city terminus for regional trains.

(The other purple lines to the CBD are from Ballarat, Bacchus Marsh, Daylesford, Woodend, Kyneton, Castlemaine, Kilmore-Broadford and Warragul, with at least 60 journeys each.)

You can also see those orange and yellow lines from the Wyndham area to central Geelong, being mode shares of 20-40%. The Geelong train line provides frequent services between Tarneit, Wyndham Vale, and Geelong, and has proved reasonably popular with commuters to Geelong (frequency was significantly upgraded in June 2015 with the opening of the Regional Rail Link, just 14 months before the census of August 2016).

However, public transport mode shares for travel within Greater Geelong are very small – even for SA2 that are connected by trains. This might reflect Geelong train station being on the edge of its CBD, relatively cheap parking in central Geelong, limited stopping frequency at some stations (many at 40 minute base pattern), and/or limited walk-up population catchments at several of Geelong’s suburban train stations.

Does public transport have significant mode share for cross-suburban journeys to work?

To search for cross-suburban journeys with relatively high public transport mode shares, here is a map that only shows lines with public transport mode shares above 20% between homes and workplaces both more than 5 km from the CBD (yes, those are arbitrary thresholds):

Of these journeys, the highest mode shares are for journeys from the inner northern suburbs to St Kilda and Hawthorn. There’s also a 49% mode share from Footscray to Maribyrnong (connected by frequent trams and buses).

The tear drop to the north of the city is 114 people who used PT from Coburg to Brunswick (connected by two tram routes and one train route).

Most of the other links on this map are fairly well aligned with train, tram, or SmartBus routes, suggesting high quality services are required to attract significant mode shares.

But these trips are a tiny fraction of journeys to work around Melbourne. In fact 3.0% of journeys to work in Melbourne were by public transport to workplaces more than 5 km from the CBD. The same statistic for Sydney is more than double this, at 7.3%.

What about private transport journeys?

Firstly, here’s a map showing private transport mode shares and volumes, building up the map starting with low private mode share lines.

The links with lowest private transport mode shares are very radial as you might expect (pretty much the inverse of the public transport maps). Progressively less radial lines get added to the map before there is a big bang when the final private transport mode share band of 95-100% gets added, with large volumes of outer suburban trips.

For completeness, here’s an animation that highlights each mode share range individually.

There are some other interesting stories in this data. The following map shows private transport mode share of journeys to work, excluding workplaces up to 10 km from the CBD to remove some clutter.

If you look carefully you’ll see that there is a much lower density of trips that cross the Yarra River (which runs just south of Heidelberg and Eltham). There are limited bridge crossings, and this is probably inhibiting people considering such journeys.

The construction of the North East Link motorway will add considerable cross-Yarra road capacity, and I suspect it may induce more private transport journeys to work across the Yarra River (although tolls will be a disincentive).

What about active transport journeys?

Next is a map for active transport journeys, but this time I’ve progressively added a filter for the number of active transport journeys, as most of the lines on the full chart are for very small volumes.

As soon as the filter reaches a minimum of 50 active journeys most of the lines between SA2s in the middle and outer suburbs disappear. Note that journeys between SA2s are not necessarily long, they might just be a short trip over the boundary.

Then at minimum 200 journeys you can only see central city journeys plus intra-SA2 journeys in relatively dense centres such as Hawthorn, Heidelberg, Box Hill, Clayton, Frankston, Mornington, Footscray, and St Kilda. The large volume in the south of the map that hangs around is Hastings – Somers, where 882 used active transport (probably mostly walking to work on the HMAS Cerebus navy base).

Active transport journeys are mostly much shorter than private and public transport journeys – as you might expect as most people will only walk or ride a bicycle so far. But there are people who said they made very long active transport journeys to work – the map shows some journeys from Point Nepean, Torquay, Ballarat, Daylesford, and Castlemaine to Melbourne. That’s some keen cyclists, incredible runners, people who changed jobs in the week of the census (the census asks for work location the prior week, and modes used on census day), and/or people who didn’t fill in their census forms accurately. The volumes of these trips are very small (mostly less than 5).

That map is very congested around the central city, so here is a map zoomed into the inner suburbs and this time animated by building up the map starting with high active transport mode share lines.

The highest active transport mode shares are for travel within Southbank and from Carlton to Parkville, followed by journeys to places like the CBD, Docklands, South Yarra, South Melbourne, Carlton, Fitzroy, Parkville, and Carlton.

Then you see a lot of trips added from the inner northern suburbs, which are connected to the central city by dare-I-say “above average” cycling infrastructure across some relatively flat terrain. In particular, a thick red line on the map is for 471 active transport journeys from Brunswick to Melbourne (CBD) with a mode share of 17%. A second thick red line is Richmond to Melbourne (CBD) being 589 journeys with 16% active mode share.

Another way of summarising mode shares by work and home distance from the CBD

I’ve experimented with another visualisation approach to overcome the clutter issues. The next charts have home distances from the CBD on the Y axis, work distances from the CBD on the X axis, bubble size representing number of journeys, and colour showing mode shares. I’m drawing smaller journey volumes on top, and I’ve used some transparency to help a little with the clutter.

Firstly here is public transport (animated to show each mode share range individually):

The chart is roughly a V-shape with many trips on the left edge and along a diagonal (mostly representing intra-SA2 journeys), then with several vertical stripes being major suburban employment destinations (including Dandenong at 31 km, Clayton at 19 km, and Frankston at 40 km). Trips above the diagonal are roughly inbound, while trips below the diagonal are roughly outbound.

Some observations:

  • The diagonal line (mostly local journeys) has very low public transport mode shares (sometimes zero).
  • Higher PT mode shares are only seen on the far left and bottom left hand corner of the chart. Some outliers include Richmond to Box Hill (32%), Clayton to Malvern East (32%), and South Yarra – East to Clayton (57%).
  • PT mode shares of 80+% are only seen for journeys to the CBD from home SA2s at least 11 km out (with one exception of Melbourne CBD to St Kilda with 80% PT share).
  • Home-work pairs with zero public transport journeys are scattered around the middle and outer suburbs, most being longer distance journeys (home and work at different distances from the CBD).

Here’s the same chart for private transport:

The lowest private transport shares are seen for journeys to the CBD. The diagonal has many mode shares in the 80-90% range.

And here is active transport:

The highest active transport mode shares are seen in the central city area, followed by the diagonal mostly representing local journeys (with generally higher shares closer to the CBD). Some notable outliers include local trips within Clayton (1,298 active trips / 46% active mode share), Box Hill (914 / 40%), Hastings – Somers (1,762 / 27%), Flinders (240 / 24%), Glen Waverley – West (308 / 21%), and Mentone (226 / 23%).

How does Sydney compare to Melbourne?

Here is a chart with private transport mode share maps for both Melbourne and Sydney, animated in tandem to progressively add higher mode share journeys.

You can see that Sydney has a lot more trips at lower private transport mode shares, and that workplaces outside the city centre start to show up earlier in the animation in Sydney – being the dense transit-orientated suburban employment clusters that are largely unique to Sydney (see: Suburban employment clusters and the journey to work in Australian cities).

If time permits, I may do similar analysis for Sydney and other cities in future posts.


How radial are journeys to work in Australian cities?

Fri 14 June, 2019

In almost every city, hordes of people commute towards the city centre in the morning and back away from the city in the evening. This largely radial travel causes plenty of congestion on road and public transport networks.

But only a fraction of commuters in each city actually work in the CBD. So just how radial are journeys to work? How does it vary between cities? And how does it vary by mode of transport?

This post takes a detailed look at journey to work data from the ABS 2016 Census for Melbourne, Sydney, and to a less extent Brisbane, Perth, Adelaide and Canberra. I’ve included some visualisations for Melbourne and Sydney that I hope you will find interesting.

How to measure radialness?

I’m measuring radialness by the difference in degrees between the bearing of the journey to work, and a direct line from the home to the CBD of the city. I’m calling this the “off-radial angle”.

So an off-radial angle of 0° means the journey to work headed directly towards the CBD. However that doesn’t mean the workplace was the CBD, it might be have been short of the CBD or even on the opposite side of the CBD.

Similarly, an off-radial angle of 180° means the journey to work headed directly away from the CBD. And a value of 90° means that the trip was “orbital” relative to the CBD (a Melbourne example would be a journey from Box Hill that headed either north or south). And then there are all the angles in between.

To deal with data on literally millions of journeys to work, I’ve grouped journeys by home and work SA2 (SA2s are roughly the size of a suburb), and my bearing calculations are based on the residential centroid of the home SA2 and the employment centroid of the work SA2.

So it is certainly not precise analysis, but I’ve also grouped off-radial angles into 10 degree intervals, and I’m mostly looking for general trends and patterns.

So how radial are trips in Melbourne and Sydney?

Here’s a chart showing the proportion of 2016 journeys to work at different off-radial angle intervals:

Technical note: As per all my posts, I’ve designated a main mode for journeys to work: any journey involving public transport is classed as “Public”, any journey not involving motorised transport is classed as “Active”, and any other journey is classed as “Private”.

In both cities over 30% of journeys to work were what you might call “very radial” – within 10 degrees of perfectly radial. It was slightly higher in Melbourne.

You can also see that public transport trips are even more radial, particularly in Melbourne. In fact, around two-thirds of public transport journeys to work in 2016 had a destination within 2 km of the CBD.

Melbourne’s “mass transit” system (mostly trains and trams) is very radial, so you might be wondering why public transport accounts for less than half of those very radial journeys (41% in fact).

Here are Melbourne’s “very radial” journeys broken down by workplace distance from the Melbourne CBD:

very-radial-trips-by-mode-distance-from-cbd

Public transport dominates very radial journeys to workplaces within 2 km of the centre of the CBD, but is a minority mode for workplaces at all other distances. Many of these highly radial journeys might not line up with a transit line towards the city, and/or there could well be free parking at those suburban workplaces that make driving all too easy. I will explore this more shortly.

Sydney however had higher public transport mode shares for less radial journeys to work. I think this can be explained by Sydney’s large and dense suburban employment clusters that achieve relatively high public transport mode shares (see: Suburban employment clusters and the journey to work in Australian cities), the less radial nature of Sydney’s train network, and generally higher levels of public transport service provision, particularly in inner and middle suburbs.

Visualising radialness on maps

To visualise journeys to work it is necessary to simplify things a little so maps don’t get completely cluttered. On the following maps I show journey to work volumes between SA2s where there are at least 50 journeys for which the mode is known. The lines between home and work SA2s get thicker at the work end, and the thickness is proportional to the volume (although it’s hard to get a scale that works for all scenarios).

First up is an animated map that shows journeys to work coloured by private transport mode share, with each frame showing a different interval of off-radial angle (plus one very cluttered view with all trips):

(click/tap to enlarge maps)

I’ve had to use a lot of transparency so you have a chance at making out overlapping lines, but unfortunately that makes individual lines a little harder to see, particularly for the larger off-radial angles.

You can see a large number of near-radial journeys, and then a smattering of journeys at other off-radial angles, with some large volumes across the middle suburbs at particular angles.

The frame showing very radial trips was rather cluttered, so here is an map showing only those trips, animated to strip out workplaces in the CBD and surrounds so you can see the other journeys:

Private transport mode shares of very radial trips are only very low for trips to the central city. When the central city jobs are stripped out, you see mostly high private transport mode shares. Some relative exceptions to this include journeys to places like Box Hill, Hawthorn, and Footscray. More on that in a future post.

Here are the same maps for Sydney:

Across both of these maps you can find Sydney’s suburban employment clusters which have relatively low private transport mode shares. I explore this, and many other interesting ways to visualise journeys to work on maps in another post.

What about other Australian cities?

To compare several cities on one chart, I need some summary statistics. I’ve settled on two measures that are relatively easy to calculate – namely the average off-radial angle, and the percent of journeys that are very radial (up to 10°).

The ACT (Canberra) actually has the most radial journeys to work of these six cities, despite it being something of a polycentric city. Adelaide has the next most radial journeys to work, but there’s not a lot of difference in the largest four cities, despite Sydney being much more a polycentric city than the others. Note the two metrics do not correlate strongly – summary statistics are always problematic!

Here are those radialness measures again, but broken down by main mode:

Sydney now looks the least radial of the cities on most measures and modes, particularly by public transport.

The Australian Capital Territory (Canberra) has highly radial private and active journeys to work, but much less-radial public transport journeys than most other cities. This probably reflects Canberra’s relatively low cost parking (easy to drive to the inner city), but also that the public transport bus network is orientated around the suburban town centres that contain decent quantities of jobs.

Adelaide has the most radial journeys to work when it comes to active and public transport.

What about other types of travel?

In a future post, I’ll look at the radialness of general travel around Melbourne using household travel survey data (VISTA), and answer some questions I’ve been pondering for a while. Is general travel around cities significantly less radial than journeys to work? Is weekend travel less radial than weekday travel?

Follow the blog on twitter or become an email subscriber (see top-right of this page) to get alerted when that comes out.


Visualising the components of population change in Australia

Sat 27 April, 2019

Australian cities are growing in population as a result of international migration, internal migration, and births outnumbering deaths. But which of these factors are most at play in different parts of the country?

Thanks to ABS publishing data on the components of population growth with their Regional Population Growth product, we now have estimates of births, deaths, internal/international arrivals, and internal/international departures right down to SA2 geography for 2016-17 and 2017-18.

This post aims to summarise the main explanation for population change in different parts of the country.

This post isn’t much about transport, but I hope you also find the data interesting. That said, it’s possible that immigrants from transit-orientated countries might be more inclined to use public transport in Australia, and that might impact transport demand patterns. We know that recent immigrants are more likely to travel to work by public transport than longer term residents, but that probably also has a lot to do with where they are settling.

How is population changing in bigger and smaller cities?

First up, I’ve divided Australia into Capital Cities (Greater Capital City Statistical Areas), Large regional cities (Significant Urban Areas with population 100,000+, 2016 boundaries), small regional cities (Significant Urban Areas, with population 10,000 to 100,000, 2016 boundaries ) and “elsewhere”.

Here’s a chart showing the total of the six components of population change in each of those four place types. I’ve animated the chart (and most upcoming charts) to show changes in the years to June 2017 and June 2018, with a longer pause on 2018.

There were significant internal movements in all parts of Australia (shown in green) – even more so in 2018. These include people moving between any SA2s, whether they adjacent within a city or across the country.

International arrivals and departures were much larger in capital cities and there were more arrivals than departures in all four place types. International arrivals declined between 2017 and 2018, while international departures increased slightly between 2017 and 2018.

Births also outnumbered deaths in all place categories in both years.

Here’s a look at the larger capital cities individually:

The chart shows Sydney, Perth, and Adelaide had more internal departures than arrivals. These cities only grew in total population because of natural increase and net international immigration. Melbourne and Brisbane had a net increase from internal movements in both 2017 and 2018, while Canberra has been a lot more even.

International arrivals outnumbered international departures and births significantly outnumbered deaths in all cities. Melbourne and Canberra were the only cities to see a significant increase in international arrivals between 2017 and 2018.

Here is the same chart but for medium sized cities:

Again, there were much larger volumes of internal migration in 2017-18 compared to 2016-17.

The Gold Coast is the only medium-sized city to have significant volumes of international movements. The fast population growth of the Gold and Sunshine Coasts is mostly coming from internal arrivals.

What is the dominant explanation for population change in different parts of Australia?

As mentioned the ABS data goes down to SA2 statistical geography which allows particularly fine grain analysis, with six measures available for each SA2. However it is difficult to show those six components spatially. They can be consolidated into three categories: net natural increase, net internal arrivals/departures, and net international arrivals/departures, but that is still three different metrics for all SA2s.

One way to look at this is simply the component with the largest contribution to population growth (or decline). Here is a map showing that for each SA2 in Melbourne:

You can see that international arrivals dominated population growth in most inner and middle suburbs, while internal arrivals dominated population growth in most outer suburbs. There are also some SA2s where births dominated (often low growth outer suburbs).

This representation is quite simplified, and doesn’t show what else might be happening. For example, here is a summary of the population changes in Sunshine for the year to June 2018:

Population change in Sunshine, year to June 2018

Overseas arrivals dominated population growth (net +313), but the otherwise hidden story here is that they were largely offset by net internal departures of 279.

So to add more detail to the analysis, I’ve created a slightly more detailed classification system that looks at the largest component and often a secondary component, as per the following table.

Explanation summaryLargest componentOther components
Growth – mostly births replacing localsNatural increaseNet internal departures more than 50 and net internal departures more than net overseas departures
Growth – mostly birthsNatural increaseNet internal and overseas departures of no more than 50
Growth – mostly immigrants replacing localsNet overseas arrivalsNet internal departures of at least 50 and/or natural decrease of at least 50.
Growth – mostly immigrationNet overseas arrivalsNet internal departures less than 50 (or net arrivals).
Growth – mostly internal arrivals replacing deathsNet internal arrivalsNet natural decrease of 50 or more, and bigger than net overseas departures
Growth – mostly internal arrivalsNet internal arrivalsNet internal arrivals greater than net overseas arrivals and natural increase

Decline – mostly internal departures

Net internal departures

Natural increase and net overseas arrivals both less than 50
Decline – mostly internal departures partly offset by births Net internal departuresNatural increase of at least 50, and natural increase larger than net overseas arrivals.
Decline – mostly internal departures partly offset by immigrantsNet internal departures Net overseas arrivals of at least 50, and net overseas arrivals larger than natural increase.
Decline – mostly deathsNatural decrease

There are no SA2s where net international departures was the major explanation for population change.

Here’s what these summary explanations look like in Melbourne (again, animated to show years to June 2017 and June 2018):

Technical notes: On these maps I’ve omitted SA2s where there was population change of less than 50 people, or where no components of population change were more than 1% of the population. Not all classifications are present on all maps.

You can now see that in most middle suburbs there has been a net exodus of locals, more than offset by net international arrivals (light purple). Also, many of the outer suburbs with low growth actually involve births offsetting internal departures (light blue).

Turning near-continuous data into discrete classifications is still slightly problematic. For example the summary explanations don’t tell you by how much one component was larger than the others. For example if there were 561 net international arrivals and 560 net internal arrivals, it would be classified as “Growth – mostly immigration”. Also, SA2s are not consistently sized across Australia (see: How is density changing in Australian cities?), so my threshold of 50 is not perfect. At the end of the post I provide a link to Tableau where you can inspect the data more closely for any part of Australia.

The inner city area of Melbourne was a little congested with data marks on the above map, so here is a map zoomed into inner and middle Melbourne:

You can see significant population growth in the Melbourne CBD and surrounding SA2s, particularly in 2017. The main explanation for inner city growth is international immigration, although internal arrivals came out on top in Southbank in 2018. Curiously, net internal arrivals were larger the international migration in Brunswick East in both years. And natural increase was dominant in Newport in the inner-west.

Zooming out to include the bigger regional centres of Victoria (note: many regional SA2s don’t show up because of very little population change):

In most regional Victorian cities, internal arrivals account for most of the population growth, although the net growth in “Shepparton – North” of +222 in 2017 and +152 in 2018 was mostly made up of international arrivals.
The only other SA2s to show international arrivals as the main explanation were in inner Geelong.

(I haven’t shown all of Victoria because few SA2s outside the above map had significant population change).

Heading up to Sydney, the picture is fairly similar to Melbourne:

Like Melbourne, internal arrivals accounted for most of the population growth in outer growth areas.

International immigration dominated the inner and middle suburbs in 2017, but in 2018 immigration eased off, and births became the main explanation for population growth in more SA2s.

The middle SA2s of Homebush Bay – Silverwater and Botany are noticeable exceptions to the pattern, dominated by internal arrivals.

Zooming out to New South Wales:

Central Newcastle, central Wollongong, Armidale and Griffith saw mostly international immigration led population growth. Most larger regional towns saw growth from internal arrivals, but further inland there was population decline – mostly from internal departures.

Next up, Brisbane:

Population growth in Brisbane’s inner suburbs is much more of a mix of internal and overseas arrivals. There are also more SA2s where births dominate population growth. There were also some SA2s with slight population decline for various reasons.

Zooming out to South East Queensland:

International arrivals dominated areas on the Gold Coast closer to the coastline, but much less so on the Sunshine Coast and in Toowoomba.

Looking at other parts of Queensland:

There was population decline in several areas, including Mackay and Mount Isa. Rockhampton and Cairns saw population growth mostly through internal arrivals. Townsville was dominated by internal arrivals in 2017, and births in 2018.

Airlie – Whitsundays stands out as having population growth mostly from international arrivals in both years.

Next up, Perth:

Like other cities, population growth in the outer suburbs was dominated by internal arrivals. There were a lot more SA2s showing population decline, and this was largely due to internal departures, partly offset by natural increases or net overseas arrivals.

Zooming out to Western Australia:

Population growth on the south-west coast was mostly dominated by internal arrivals, while in many other centres around the state there was population decline, mostly due to internal departures, however in many areas this was offset partly by births.

Next up, Adelaide:

Firstly, keep in mind that there has been relatively slow population growth in Adelaide (the scale is adjusted). The inner and middle suburbs mostly show population growth from international arrivals (often offsetting net internal departures), and the outer growth areas were again mostly about internal arrivals.

Zooming out to South Australia:

In 2017 there was considerable population decline in Whyalla and Port Augusta. Murray Bridge is another rare regional centre where population growth was largely driven by almost 400 overseas arrivals each year.

Next is Tasmania:

Note the circle size scale is even smaller. Overseas arrivals dominated population growth in central Hobart and Newman – Mayfield in Launceston (possibly related to university campuses), while internal migration dominated most other areas.

Here is Canberra:

International immigrants dominated population growth around Civic and the inner north. Internal arrivals dominated Kingston and Griffith and most outer growth areas. The outer suburbs saw a mixture of births and internal arrivals as the dominant explanation.

And finally, Darwin, which actually saw net population decline in the year to June 2018:

Palmerston South saw the largest population growth – mostly from internal arrivals. International arrivals were significant around Darwin city in 2017, but were much less significant in 2018. Most of the northern suburbs saw population decline in the year to June 2018.

Didn’t see your area, or want to explore further? You can view this data interactively in Tableau (you might want to filter by state as that will change the scale of circle sizes).

Where were international arrivals most significant?

I’ve calculated the ratio of international arrivals to population for each SA2. The SA2s where international arrivals in the year to June 2018 make up a significant portion of the 2018 population are all near universities and/or CBDs. Namely:

  • Melbourne CBD and neighbouring Carlton at 20% (Melbourne Uni, RMIT, and others)
  • Brisbane CBD at 18% and neighbouring Spring Hill at 20% (QUT and others)
  • Clayton in Melbourne at 18% (Monash Uni)
  • Sydney – Haymarket – The Rocks at 15% and neighbouring Pyrmont – Ultimo at 17% (near to UTS, Sydney Uni, and various others)
  • Acton (ACT) at 17% (ANU)
  • Kingsford (in Sydney) at 16% (UNSW)
  • St Lucia (Brisbane) at 15% (UQ)

I hope you’ve found this interesting. In a future post I might look at internal migration origin-destination flows, including how people are moving within and between cities.


How is density changing in Australian cities? (2nd edition)

Sun 21 April, 2019

[updated April 2020 with 2019 population data]

While Australian cities are growing outwards, densities are also increasing in established areas, and newer outer growth areas are some times at higher than traditional suburban densities.

So what’s the net effect – are Australian cities getting more or less dense? How has this changed over time? Has density bottomed out? And how many people have been living at different densities?

This post maps and measures population density over time in Australian cities.

I’ve taken the calculations back as far as I can with available data (1981), used the highest resolution population data available. I’ll discuss some of the challenges of measuring density using different statistical geographies along the way, but I don’t expect everyone will want to read through to the end of this post!

[This is a fully rewritten and updated version of a post first published November 2013]

Measuring density

Under traditional measures of density, you’d simply divide the population of a city by the area of the metropolitan region.

At the time of writing Wikipedia said Greater Sydney’s density was just 4.23 people per hectare (based on its Greater Capital City Statistical Area). To help visualise that, a soccer pitch is about 0.7 hectares. So Wikipedia is saying the average density of Sydney is roughly about 3 people per soccer field. You don’t need to have visited Sydney to know that is complete nonsense (don’t get me wrong, I love Wikipedia, but it really need to use a better measure for city density!).

The major problem with metropolitan boundaries – in Australia we use now Greater Capital City Statistical Areas – is that they include vast amounts of rural land and national parks. In fact, in 2016, at least 53% of Greater Sydney’s land area had zero population. That statistic is 24% in Melbourne and 14% in Adelaide – so there is also no consistency between cities.

Below is a map of Greater Sydney (sourced from ABS), with the blue boundary representing Greater Sydney:

One solution to this issue is to try to draw a tighter boundary around the urban area, and in this post I’ll also use Significant Urban Areas (SUAs) that do a slightly better job (they are made up of SA2s). The red boundaries on the above map show SUAs in the Sydney region.

However SUAs they still include large parks, reserves, industrial areas, airports, and large-area partially-populated SA2s on the urban fringe. Urban centres are slightly better (they are made of SA1s) but population data for these is only available in census years, the boundaries change with each census, the drawing of boundaries hasn’t been consistent over time, they include non-residential land, and they split off most satellite urban areas that are arguably still part of cities, even if not part of the main contiguous urban area.

Enter population-weighted density (PWD) which I’ve looked at previously (see Comparing the densities of Australian, European, Canadian, and New Zealand cities). Population-weighted density takes a weighted average of the density of all parcels of land that make up a city, with each parcel weighted by its population. One way to think about it is the average density of the population, rather than the average density of the land.

So parcels of land with no population don’t count at all, and large rural parcels of land that might be inside the “metropolitan area” count very little in the weighted average because of their relatively small population.

This means population-weighted density goes a long way to overcoming having to worry about the boundaries of the “urban area” of a city. Indeed, previously I have found that removing low density parcels of land had very little impact on calculations of PWD for Australian cities (see: Comparing the residential densities of Australian cities (2011)). More on this towards the end of this post.

Calculations of population-weighted density can also answer the question about whether the “average density” of a city has been increasing or decreasing.

But… measurement geography matters

One of the pitfalls of measuring population weighted density is that it very much depends on the statistical geography you are using.

If you use larger geographic zones you’ll get a lower value as most zones will include both populated and unpopulated areas.

If you use very small statistical geography (eg mesh blocks) you’ll end up with a lot fewer zones that are partially populated – most will be well populated or completely unpopulated, and that means your populated weighted density value will be much higher, and your measure is more looking at the density of housing areas.

To illustrate this, here’s an animated map of the Australian Capital Territory’s 2016 population density at all of the census geographies from mesh block (MB) to SA3:

Only at the mesh block and SA1 geographies can you clearly see that several newer outer suburbs of Canberra have much higher residential densities. The density calculation otherwise gets washed out quickly with lower resolution statistical geography, to the point where SA3 geography is pretty much useless as so much non-urban land is included (also, there are only 7 SA3s in total). I’ll come back to this issue at the end of the post.

Even if you have a preferred statistical geography for calculations, making international comparisons is very difficult because few countries will following the same guidelines for creating statistical geography. Near enough is not good enough. Worse still, statistical geography guidelines do not always result in consistently sized areas within a country (more on that later).

We need an unbiased universal statistical geography

Thankfully Europe and Australia have adopted a square kilometre grid geography for population estimates, which makes international PWD comparisons readily possible. Indeed I did one a few years ago looking at ~2011 data (see Comparing the densities of Australian, European, Canadian, and New Zealand cities).

This ABS is now providing population estimates on a square km grid for every year from 2006.

Here is what Melbourne’s estimated population density looks like on a km square grid, animated from 2006 to 2019:

The changes over time are relatively subtle, but if you watch the animation several times you’ll see growth – including relatively high density areas emerging on the urban fringe.

It’s a bit chunky, and it’s a bit of a matter of luck as to whether dense urban pockets fall entirely within a single grid square or on a boundary, but there is no intrinsic bias.

There’s also an issue that many grid squares will contain a mix of populated and non-populated land, particularly on the urban fringe (and a similar issue on coastlines). In a large city these will be in the minority, but in smaller cities these squares could make up a larger share of the total, so I think we need to be careful about this measure in smaller cities. I’m going to arbitrarily draw the line at 200,000 residents.

How are Australian cities trending for density using square km grid data? (2006 to 2019)

So now that we have an unbiased geography, we can measure PWD for cities over time.

The following chart is based on 2016 Significant Urban Area boundaries (slightly smaller than Greater Capital City Statistical Areas but also they go across state borders as appropriate for Canberra – Queanbeyan and Gold Coast – Tweed).

Technical notes: You cannot perfectly map km squares to Significant Urban Areas. I’ve included all kilometre grid squares which have a centroid within the 2016 Significant Urban Area boundaries (with a 0.01 degree tolerance added – which is roughly 1 km). Hobart appears only from 2018 because that’s when it crossed the 200,000 population threshold.

The above trend chart was a little congested for the smaller cities, so here is a zoomed in version without Sydney and Melbourne:

You can see most cities getting denser at various speeds, although Perth, Geelong, and Newcastle have each flat-lined for a few years.

Perth’s population growth slowed at the end of the mining boom around 2013, and infill development all but dried up, so the overall PWD increased only 0.2 persons/ha between 2013 and 2018.

Canberra has seen a surge in recent years, probably due to high density greenfield developments we saw above.

How is the mix of density changing? (2006 to 2019)

Here’s a look at the changing proportion of the population living at different densities for 2006-2019 for the five largest Australian cities, using square km grid geography:

It looks very much like the Melbourne breakdown bleeds into the Sydney breakdown. This roughly implies that Melbourne’s density distribution is on trend to look like Sydney’s 2006 distribution in around 2022 (accounting the for white space). That is, Melbourne’s density distribution is around 16 years behind Sydney’s on recent trends. Similarly, Brisbane looks a bit more than 15 years behind Melbourne on higher densities.

In Perth up until 2013 there was a big jump in the proportion of the population living at 35 persons / ha or higher, but then things peaked and the population living at higher densities declined, particularly as there was a net migration away from the inner and middle suburbs towards the outer suburbs.

Here’s the same for the next seven largest cities:

Of the smaller cities, densities higher than 35 persons/ha are only seen in Gold Coast, Newcastle, Wollongong and more recently in Canberra.

The large number of people living at low densities in the Sunshine Coast might reflect suburbs that contain a large number of holiday homes with no usual residents (I suspect the dwelling density would be relatively higher). This might also apply in the Gold Coast, Central Coast, Geelong (which actually includes much of the Bellarine Peninsula) and possibly other cities.

Also, the Central Coast and Sunshine Coast urban patterns are highly fragmented which means lots of part-urban grid squares, which will dilute the PWD of these “cities”.

Because I am sure many of you will be interested, here are animated maps for these cities:

Sydney

Brisbane

Perth

Adelaide

Canberra – Queanbeyan

Sunshine Coast

Gold Coast

Newcastle – Maitland and Central Coast

Wollongong

Geelong

What are the density trends further back in time using census data?

The census provides the highest resolution and therefore the closest measure of “residential” population weighted density. However, we’ve got some challenges around the statistical geography.

Prior to 2006, the smallest geography at which census population data is available is the collector district (CD), which average around 500 to 600 residents. A smaller geography – the mesh block (MB) – was introduced in 2006 and averages around 90 residents.

Unfortunately, both collector districts and mesh blocks are not consistently sized across cities or years (note: y axis on these charts does not start at zero):

Technical note: I have mapped all CDs and MBs to Greater Capital City Statistical Area (GCCSA) boundaries, based on the entire CD fitting within the GCCSA boundaries (which have not yet changed since they were created in 2011).

There is certainly some variance between cities and years, so we need to proceed with caution, particularly in comparing cities. Hobart and Adelaide have the smallest CDs and MBs on average, while Sydney generally has larger CDs and MBs. This might be a product of whether mesh blocks were made too small or large, or it might be that density is just higher and it is more difficult to draw smaller mesh blocks. The difference in median population may or may not be explained by the creation of part-residential mesh blocks.

Also, we don’t have a long time series of data at the one geography level. Rather than provide two charts which break at 2006, I’ve calculated PWD for both CD and mesh block geography for 2006, and then estimated equivalent mesh block level PWD for earlier years by scaling them up by the ratio of 2006 PWD calculations.

In Adelaide, the mesh block PWD for 2006 is 50% larger than the CD PWD, while in the Australian Capital Territory it is 110% larger, with other cities falling somewhere in between.

Would these ratios hold for previous years? We cannot be sure. Collector Districts were effectively replaced with SA1s (with an average population of 500, only slightly smaller) and we can calculate the ratio of mesh block PWD to SA1 PWD for 2011 and 2016. For most cities the ratio in 2016 is within 10% of the ratio in 2011. So hopefully the ratio of CD PWD to mesh block PWD would remain fairly similar over time.

So, with those assumptions, here’s what the time series then looks like for PWD at mesh block geography:

As per the square km grid values, Sydney and Melbourne are well clear of the pack.

Most cities had a PWD low point in 1996. That is, until around 1996 they were sprawling at low densities more than they were densifying in established areas, and then the balance changed post 1996. Exceptions are Darwin which bottomed out in 2001 and Hobart which bottomed in 2006.

The data shows rapid densification in Melbourne and Sydney between 2011 and 2016, much more so than the square km grid data time series. But we also saw a significant jump in the median size of mesh blocks in those cities between 2011 and 2016 (and if you dig deeper, the distribution of mesh block population sizes also shifted significantly), so the inflection in the curves in 2011 are at least partly a product of how new mesh block boundaries were cut in 2016, compared to 2011. Clearly statistical geography isn’t always good for time series and inter-city analysis!

How has the distribution of densities changed in cities since 1986?

The next chart shows the distribution of population density for Greater Capital City Statistical Areas based on collector districts for the 1986 to 2006 censuses:

You can more clearly see the decline in population density in most cities from 1986 to 1996, and it wasn’t just because most of the population growth was a lower densities. In Hobart, Canberra, Adelaide, Brisbane and Melbourne, the total number of people living at densities of 30 or higher actually reduced between 1986 and 1996.

Here is the equivalent chart for change in density distribution by mesh block geography for the capital cities for 2006, 2011, and 2016:

I’ve used the same colour scale, but note that the much smaller geography size means you see a lot more of the population at the higher density ranges.

The patterns are very similar to the distribution for square km grid data. You can see the how Brisbane seems to bleed into Melbourne and then into Sydney, suggesting a roughly 15 year lag in density distributions. This chart also more clearly shows the recent rapid rise of high density living in the smaller cities of Canberra and Darwin.

The next chart shows the 2016 distribution of population by mesh block density using Statistical Urban Area 2016 boundaries, including the smaller cities:

Gold Coast and Wollongong stand out as smaller cities with a significant portion of their population at relatively high densities, but a fair way off Sydney and Melbourne.

(Sorry I don’t have a mesh block times series of density distribution for the smaller cities – it would take a lot of GIS processing to map 2006 and 2011 mesh blocks to 2016 SUAs, and the trends would probably be similar to the km grid results).

Can we measure density changes further back in history and for smaller cities?

Yes, but we need to use different statistical geography. Annual population estimates are available at SA2 geography back to 1991, and at SA3 geography back to 1981.

However, there are again problems with consistency in statistical geography between cities and over time.

Previously on this blog I had assumed that guidelines for creation of statistical geography boundaries have been consistently applied by the ABS across Australia, resulting in reasonably consistent population sizes, and allowing comparisons of population-weighted density between cities using particular levels of statistical geography.

Unfortunately that wasn’t a good assumption.

Here are the median population sizes of all populated zones for the different statistical geographies in the 2016 census:

Note: I’ve used a log scale on the Y-axis.

While there isn’t a huge amount of variation between medians at mesh block and SA1 geographies, there are massive variations at SA2 and larger geographies.

SA2s are intended to have 3,000 to 25,000 residents (a fairly large range), with an average population of 10,000 (although often smaller in rural areas). You can see from the chart above that there are large variances between medians of the cities, with the median size in Canberra and Darwin below the bottom of the desired range.

I have asked the ABS about this issue. They say it is related to the size of gazetted localities, state government involvement, some dense functional areas with no obvious internal divisions (such as the Melbourne CBD), and the importance of capturing indigenous regions in some places (eg the Northern Territory). SA2 geography will be up for review when they update statistical geography for 2021.

While smaller SA2s mean you get higher resolution inter-censal statistics (which is nice), it also means you cannot compare raw population weighted density calculations between cities at SA2 geography.

However, all is not lost. We’ve got calculations of PWD on the unbiased square kilometre grid geography, and we can compare these with calculations on SA2 geography. It turns out they are very strongly linearly correlated (r-squared of over 0.99 for all cities except Geelong).

So it is possible to estimate square km grid PWD prior to 2006 using a simple linear regression on the calculations for 2006 to 2018.

But there is another complication – ABS changed the SA2 boundaries in 2016 (as is appropriate as cities grow and change). Data is available at the new 2016 boundaries back to 2001, but for 1991 to 2000 data is only available on the older 2011 boundaries. For most cities this only creates a small perturbation in PWD calculations around 2001 (as you’ll see on the next chart), but it’s larger for Geelong, Gold Coast – Tweed Heads and Newcastle Maitland so I’m not willing to provide pre-2001 estimates for those cities.

The bottom of this chart is quite congested so here’s an enlargement:

Even if the scaling isn’t perfect for all history, the chart still shows the shape of the curve of the values.

Consistent with the CD data, several cities appear to have bottomed out in the mid 1990s. On SA2 data, that includes Adelaide in 1995, Perth and Brisbane in 1994, Canberra in 1998 and Wollongong in 2006.

Can we go back further?

If we want to go back another ten years, we need to use SA3 geography, which also means we need to switch to Greater Capital City Statistical Areas as SA3s don’t map perfectly to Significant Urban Areas (which are constructed of SA2s). Because they are quite large, I’m only going to estimate PWD for larger cities which have reasonable numbers of SA3s that would likely have been fully populated in 1981.

I’ve applied the same linear regression approach to calculate estimated square kilometre grid population weighted density based on PWD calculated at SA3 geography (the correlations are strong with r-squared above 0.98 for all cities).

The following chart shows the best available estimates for PWD back to 1981, using SA3 data for 1991 to 2000, SA2 data for 2001 to 2005, and square km grid data from 2006 onwards:

Technical notes: SA3 boundaries have yet to change within capital cities, so there isn’t the issue we had with SA2s. The estimates based on SA2 and SA3 data don’t quite line up between 1990 and 1991 which demonstrates the limitations of this approach.

The four large cities shown appear to have been getting less dense in the 1980s (Melbourne quite dramatically). These trends could be related to changes in housing/planning policy over time but they might also be artefacts of using such a coarse statistical geography. It tends to support the theory that PWD bottomed out in the mid 1990s in Australia’s largest cities.

Could we do better than this for long term history? Well, you could probably do a reasonable job of apportioning census collector district data from 1986 to 2001 censuses onto the km grid, but that would be a lot of work! It also wouldn’t be perfectly consistent because ABS use dwelling address data to apportion SA1 population estimates into kilometre grid cells. Besides we have reasonable estimates using collector district geography back to 1986 anyway.

Melbourne’s population-weighted density over time

So many calculations of PWD are possible – but do they have similar trends?

I’ve taken a more detailed look at my home city Melbourne, using all available ABS population figures for the geographic units ranging from mesh blocks to SA3s inside “Greater Melbourne” and/or the Melbourne Significant Urban Area (based on the 2016 boundary), to produce the following chart:

Most of the datasets show an acceleration in PWD post 2011, except the SA3 calculations which are perhaps a little more washed out. The kink in the mesh block PWD is much starker than the other measures.

The Melbourne SUA includes only 62% of the land of the Greater Melbourne GCCSA, yet there isn’t much difference in the PWD calculated at SA2 geography – which is the great thing about population-weighted density.

All of the time series data suggests 1994 was the year in which Melbourne’s population weighted density bottomed out.

Appendix 1: How much do PWD calculations vary by statistical geography?

Census data allows us to calculate PWD at all levels of statistical geography to see if and how it distorts with larger statistical geography. I’ve also added km grid PWD calculations, and here are all the calculations for 2016:

Technical note: square km grid population data is estimated for 30 June 2016 while the census data is for 9 August 2016. Probably not a significant issue!

You can see cities rank differently when km grid results are compared to other statistical geography – reflecting the biases in population sizes at SA2 and larger geographies. Wollongong and Geelong also show a lot of variation in rank between geographies – probably owing to their small size.

The cities with small pockets of high density – in particular Gold Coast – drop rank with large geography as these small dense areas quickly get washed out.

I’ve taken the statistical geography all the way to Significant Urban Area – a single zone for each city which is the same as unweighted population density. These are absurdly low figures and in no way representative of urban density. They also suggest Canberra is more dense than Melbourne.

Appendix 2: Issues with over-sized SA1s

As I’ve mentioned recently, there’s an issue that the ABS did not create enough reasonably sized SA1s in some city’s urban growth areas in 2011 and 2016. Thankfully, it looks like they did however create a sensible number of mesh blocks in these areas, as the following map (created with ABS Maps) of the Altona Meadows / Point Cook east area of Melbourne shows:

In the north parts of this map you can see there are roughly 4-8 mesh blocks per SA1, but there is an oversized SA1 in the south of the map with around 50 mesh blocks. This will impact PWD calculated at SA1 geography, although these anomalies are relatively small when you are looking at a city as large as Melbourne.