Three weeks after the start of city's Delta outbreak, Covid-19 modellers from Te Pūnaha Matatini delivered officials a report, now publicly released, that charted possible "optimistic" and "pessimistic" scenarios.
Under their "optimistic" scenario, Auckland stayed at alert level 4 for 60 days, at which time almost all trajectories for spread were either contained or eliminated.
In another scenario, the modellers assumed Auckland moved to level 3 on September 16 - six days before the city did in real-life - and then looked at different potential level 3 trajectories.
Interestingly, their modelled "pessimistic" level 3 scenarios tracked closely to what we've now observed throughout the past few weeks.
Did this mean the Government made a mistake with the step-down, and that we might have contained the outbreak by now?
Modeller Dr Dion O'Neale explains his team's work.
First off, can you tell us about the modelling you've used here? What models did you use, how did they work, what can they tell us, and how reliable are they in a "real-world" setting?
The report uses two models – a Branching Process Model (BPM) and a Contagion Network Model (CNM).
The BPM has the advantage of being relatively quick and easy to produce results, using some limited assumptions about what the reproduction number is expected to be.
The CNM is much more detailed, as it includes explicit representation of individuals and their interactions, along with processes that influence transmission.
This means that it is able to try to directly include the effect of different aspects of different alert levels.
At level 4, for instance, around 30 per cent of workplaces are operating on-site, depending on sector, but at level 3, this increases to about 50 per cent.
A difficulty of the CNM is that all of these different aspects need data to best estimate what their effect will be.
You looked at different scenarios for level 4 and 3. What did these scenarios assume?
One of the pieces of information that the Contagion Network Model needs is an estimate of the percentage of non-work, non-household interactions that would not occur at different alert levels.
We can estimate this parameter from survey data for previous lockdowns, but there is still a degree of uncertainty.
A key difference between the optimistic and pessimistic scenarios in the CNM is that the optimistic scenario assumes that, for interactions that would be classified as "close contact", 80 per cent of non-work, non-household events don't happen at alert level 3, compared with only 70 per cent not occurring in the pessimistic scenario.
We also assumed stronger transmission reduction measures - like more people wearing masks, better ventilation and rostering workers on non-overlapping shifts - for workplaces operating on-site for the optimistic level 3 scenario.
For the BPM, the optimistic and pessimistic scenarios simply correspond to changing the estimate of the reproduction number to be slightly better or worse than the effect we saw for past periods at level 3.
Can you tell us about the trajectories that both of these scenarios took over time?
Both the BPM and the CNM are "stochastic" models - that is, both include effects of random chance in their simulations, with variations from run-to-run.
The Contagion Network Model predicted that, under an optimistic scenario, case numbers were likely to continue at low levels - typically less than 10 cases per day but rarely reaching zero cases.
In contrast, it predicted that, for the pessimistic scenario, case numbers would start to climb, with this trend becoming apparent about two weeks after the de-escalation from level 4 to level 3.
The spread on the contagion network model takes place on an "interaction network" that has areas with different amounts of connectivity.
Because of this, the model showed that, for both the optimistic and pessimistic scenarios, there are a small number of simulations where the contagion finds its way into a particularly "high-spread" part of the network, and case numbers take off.
The other side of this is that both scenarios also have some examples where the contagion finds its way into low spread parts of the network, with many dead-ends, and where lots of the infection pathways die out.
What matters is the relative number of such high-spread and low-spread simulations for the two scenarios.
We might infer from this work that, had Auckland remained at level 4, this outbreak could have been stamped out, given it started 63 days ago, and that, instead, the "pessimistic" level 3 scenario is largely playing out as modelled. Can we leap to this conclusion?
This is at least partly a case of the model getting the "right answer for the wrong reason".
Although the predicted case numbers from the pessimistic scenario are reasonably close to what we observed later in real-world case data, the model did not predict the specific scenario where the contagion spread through a particularly "hard-to-reach" community, where alert level interventions would have little effect on the transmission environment.
This is an example of how when case numbers are small, the most important details are those from the contact tracers on the ground who have individual case details.
But for the initial weeks of the outbreak, both models did well at indicating where case numbers were heading.
This was during the period when the models could be more directly related to the situation on the ground.
In the case of the CNM, we were able to input parameters that described the level 4 intervention as best we understood it.
The resulting spread in the output followed roughly the predicted patterns, for the predicted reasons, with case numbers decreasing at the rate we would expect for the Delta variant combined with level 4 and vaccination levels at the time.
The model predicted that staying at level 4 would have led to elimination or containment in the vast majority of cases.
This relied on the assumption that the level 4 intervention continued to apply to the situation on the ground where spread was taking place.
But in the later stages of the outbreak, this assumption became less accurate as spread moving into parts of the community where alert level interventions were no longer so relevant.
Are there any other caveats - or otherwise important pieces of context - that we need to consider here?
It's always important to be considering details of how an outbreak is progressing to find cases when assumptions in models may no longer apply, or when they might need updating.
Almost by definition, the epidemic will spread in those environments where it is easiest to.
This could mean situations like communities with low vaccination rates, more crowded households, or with people who are less likely to seek a test, or less able to self-isolate, if they first get symptoms.
In this sort of situation, the contagion that we observe in case data is going to look more like what we would expect for a pessimistic scenario, rather than for an "average" transmission environment.
Where does the modelling predict the outbreak going from here, compared with what we're observing presently?
Most of the analysis suggests that we currently have a doubling time of around 10 days to a fortnight, which corresponds to a reproduction number of around 1.3.
This means that we would expect the number of new cases to double around every 10 to 14 days.
With about 60 cases per day recently, and more than 90 today, it is easy to see how we could easily have hundreds of cases per day in a month's time.
Increasing vaccination coverage will help to slow that increase somewhat but the modelling tells us that interventions like alert levels will still be needed in conjunction with vaccination if we want to suppress this spread.
This is in addition to public health interventions such a contact tracing and isolation of symptomatic cases.
From here, a lot depends on the details of how closely people adhere to the intention of the current level 3 intervention and how they choose to act on the guidelines for Auckland's new three-step roadmap.
One of the key aspects of this is the number of non-work, non-household interactions that people have.
If people are meeting up with multiple other households - and even if they are making those connections one household at a time - then we can very quickly see rapid growth in case numbers.
When other countries have implemented something like our "picnics rule", they have included expectations, such as only meeting with a fixed total number of other people or households over the period of a week or two.
This means that the number of unique interactions that people have is lower, even if they are still having the same number of interactions.
Another key factor will be how much transmission reduction we can get within those interactions that do still go ahead.
This will come down to things like people being good about wearing masks properly when they are in the workplace - and not only when they are in customer-facing roles
In fact, given that most customer-facing Auckland businesses have outdoor click-and-collect, transmission is more likely through non-customer facing contexts in backroom settings where mask-wearing is not mandated.
While it's been considered, there's still no apparent appetite for a jump back to level 4 to allow vaccination rates to get to where they need to, and to keep case numbers low enough to be managed by public health controls. Is this still something that could be worthwhile?
While case numbers are increasing, a circuit breaker style lockdown gives us an opportunity to press pause, or maybe even decrease case numbers a bit, while we get vaccination coverage as high as possible - remembering that it takes a couple of weeks for the protection of vaccination to kick in.
If we wait until later, when case numbers are higher, before we reach for the handbrake of level 4 or similar, then it is likely that we will need that extra intervention for longer to bring things under control.
Aotearoa's original strategy of "go hard, go early" turns out to still be relevant, even if our goal is suppression rather than elimination.