Tracking Covid-19 in Australia



The Covid-19 outbreak in Victoria

For each of the three time series shown, the green segment on the left shows a period of time that is substantively settled. The brown/orange segment is a "nowcast" - still heavily impacted by delays in incubation and reporting, but with at least some partial data. The final blue segment is a forecast.

The darker part of each ribbon reflects the estimated 50% credibility interval - the range within which a particular variable (infections, cases, or effective reproduction number) is believed to have a 50% chance of occurring, were the truth to be known. The wider, paler part of each ribbon is a 90% credibility interval.

The vertical pale grey bars represent confirmed cases, but with a small adjustment for test positivity (ie the proportion of tests for Covid-19 that return positive on any one day). Victoria's test positivity over the last few months is shown in this chart, which makes clear that the proportion of positive tests is still low (ie good) by international standards:

The Covid-19 outbreak in New South Wales

For each of the three time series shown, the green segment on the left shows a period of time that is substantively settled. The brown/orange segment is a "nowcast" - still heavily impacted by delays in incubation and reporting, but with at least some partial data. The final blue segment is a forecast.

The darker part of each ribbon reflects the estimated 50% credibility interval - the range within which a particular variable (infections, cases, or effective reproduction number) is believed to have a 50% chance of occurring, were the truth to be known. The wider, paler part of each ribbon is a 90% credibility interval.

The vertical pale grey bars represent confirmed cases, but with a small adjustment for test positivity (ie the proportion of tests for Covid-19 that return positive on any one day). New South Wales' test positivity over the last few months is shown in this chart, which makes clear that the proportion of positive tests is very low (ie good) by international standards:

The Covid-19 outbreak in Australia as a whole

Treat the estimates for Australia as a whole with particular caution. After all, the average reproduction number will be an average over a large, diverse area. Generally, the state-specific estimates will be more useful.

For each of the three time series shown, the green segment on the left shows a period of time that is substantively settled. The brown/orange segment is a "nowcast" - still heavily impacted by delays in incubation and reporting, but with at least some partial data. The final blue segment is a forecast.

The darker part of each ribbon reflects the estimated 50% credibility interval - the range within which a particular variable (infections, cases, or effective reproduction number) is believed to have a 50% chance of occurring, were the truth to be known. The wider, paler part of each ribbon is a 90% credibility interval.

The vertical pale grey bars represent confirmed cases, but with a small adjustment for test positivity (ie the proportion of tests for Covid-19 that return positive on any one day). Australia's test positivity over the last few months is shown in this chart, which makes clear that the proportion of positive tests is still low (ie good) by international standards:

The positivity adjustment could be improved by doing it separately for each state. More generally, the effective reproduction number for Australia as a whole should be treated as even more indicative than for individual states; it is more useful to consider the states as separate outbreaks for this purpose.

What is this?

The main content on this page is a set of estimates of time-varying effective reproduction number in various states and for Australia as a whole. In essence, this is an estimate of the average number of people in an outbreak that each infected person infects in turn. If it is more than one, the epidemic expands. If it is less than one, it is under control and soon (if it stays under one) the case incidence numbers will begin to shrink.

This is a standard and important measure widely used in epidemiology. Unfortunately, it is difficult to estimate in real time, with incomplete data and significant delays between infection and a case showing up in the official confirmed count. Hence the emphasis on this site on a credibility interval showing the range within which we think the number is, not just a point estimate. Focus on the range of possibilities rather than expecting the precise number to be right.

Source data

The case numbers for Victoria are collated by The Guardian, from media and data releases by the Victorian Premier and the Department of Health and Human Services. I use the daily increase in cumulative cases as my initial estimates of incidence. Note that these are sometimes slightly less than the daily figures announced by press release, because of adjustments to previous days' case numbers which cannot be allocated to specific dates. The estimates would be marginally improved by using a cleaner set of data that nets out these values closer to when they happened.

The case numbers for New South Wales come directly from NSW Health. There is about a 24 hour lag between the latest media announcements of "today's confirmed cases" and appearing in this data source.

The data on number of tests comes from The Guardian.

Estimation methods

This analysis is done with the EpiNow2 R package, whose primary authors are at the London School of Tropical Hygiene and Medicine. It takes into account estimated delays from infection, incubation and reporting. The approach is described in this blog post from July 2020.

The NSW data are published by likely source of infection, but currently the EpiNow2 method cannot take this into account in estimating effective reproduction number. I am sticking with the EpiNow2 approach because it has the best available adjustments for uncertain delay and right-truncation of case numbers, despite this limitation for now.

The approach to adjusting for positivity is as described in this blog post, with a power parameter of 0.1.

Code

The code for this analysis is available on GitHub.