The Issues With How Minnesota’s Model Works, Part 4

By April 22, 2020 Commentary

In this post we will discuss how the model handles adjustments caused by “non-pharmaceutical interventions” or mitigation of spread measures.  The supposed justification for these measures is that they will prevent health care resources from being inadequate for the number of patients who need treatment at any point in time.  The model as run in Minnesota purports to measure the number of people who should be in the Hospitalized bucket or the ICU bucket on that day and compares that to capacity.  The model concludes there is a shortfall.

There are two ways in which that modeling could be in error.  One is whether the capacity is measured correctly.  You can’t have a lot of confidence in this when from the first model run to the second, ICU capacity was moved up tenfold.  We have a serious problem in the country in terms of financial harm to hospitals from lack of utilization; they are filing for bankruptcy and laying people off as they become revenue-starved.  So it is a little hard to buy the story of inadequate resources, especially at the hospital level.  And I have made the point repeatedly about the flexibility of all hospital units and even the ability to turn non-hospital facilities into hospitals quickly.  This isn’t a speculative point, it occurred in New York City.  And for those who are worried about ventilators, here is more evidence on the recommendation to stop using them so often for coronavirus patients.  (STAT Article)  So, particularly when you consider the errors in the modeling regarding how many patients will actually need a hospital or an ICU-type bed, I don’t take the concern about overwhelming capacity as a realistic problem.

Be that as it may, the modelers looked at the effects of several mitigation of spread strategies.  Keep in mind that these strategies do not prevent infections or deaths, they only delay them.  The only circumstance in which deaths are prevented is if there was an overrun of needed health resources that caused a patient to not get care necessary to prevent death.

The mitigation of spread measures work through the contact model, and reduction in contacts will reduce the people flowing into the Exposed, and eventually the Infected, buckets on any given day.  We don’t have the formulas for the contact model, so we can’t see exactly how they accounted for the changes induced by mitigation measures, but the paper says that “social distancing”, which is not defined, reduced contacts by 50% and shelter in place reduced them by 80%.  In the  slides, four mitigation strategies are set out and six scenarios based on duration of a set of mitigation measures. The first set of mitigation strategies is a low form of social distancing, which reduces contacts 20%; the second, physical distancing which reduces contacts by 50%; the third, requiring vulnerable people to stay at home, which reduces contacts by 50%; and fourth, a universal stay-at-home order, which reduces contacts by 80%.

For the model runs in regard to mitigation of spread measures, Scenario 1 is no mitigation.  Scenario 2 is stay home order plus physical distancing, with some pieces lasting different lengths of time.  Scenario 3 is a long-term stay-at-home order for the most vulnerable and the other pieces still in place, generally lasting longer than in Scenario 2.   Scenario 4 is extending the stay-at-home order while keeping other pieces.  Scenario 3.1 is extending physical distancing by 6 weeks.  Scenario 3.2 is long-term slowed contacts for all.

The slides describe outcomes of model runs from some of the scenarios.  This has been gone over ad nauseam, but from the first time they ran the model, the estimates of mortality dropped by over half.  What is described as Scenario 3 shows a central estimate of 22,000 deaths.  What is described as Scenario 4 has 22,000 deaths.  No f***g difference, but in one scenario most people aren’t subject to the stay-at-home order and in the other the most vulnerable are protected.  I don’t know what more I can say, when the cost of one in terms of jobs lost is so much obviously worse than the other.

Since the other thing the Governor did was order businesses to close, it is very surprising that no modeling was done on the effect of that measure, which has cost hundreds of thousands of jobs in the state.  I don’t know how it is good decision-making, or the Governor’s favorite, being “data-driven”, when you don’t even consider in the modeling what effect closing certain businesses has.

Another interesting observation comes from looking at the curves for ICU demand in the slide deck.  These are basically the epidemic curves, since ICU demand is just derived from the flow into Infected and out of Infected. These modelers are telling us that under the most stringent mitigation scenario, the epidemic is over in 30 weeks, meaning natural immunity exists.  If that weren’t the case, there would be a much longer tail, since they have consistently said that, other than the overrun of health resources issue, we are only delaying, not preventing infections.  And since the mildest mitigation measure curve has the epidemic over in 20 weeks, an unmitigated scenario must have the epidemic ending in even less time.  So now I am scratching my head about why the Governor is calling for 18 months of mitigation measures.

It would be nice if the modeling was a little more granular about the effects of more isolated and specific mitigation measures, so that each could be evaluated separately.  In other places people have compiled lists of all possible tactics.  At a minimum it would be helpful for the public to understand the incremental gain, and cost, of each specific measure.  So future modeling should flesh that out further.

There are a number of caveats and limitations in both the technical paper and the slides, which should be read carefully.  One of the most important is the note that because many of the deaths are clustered in nursing home settings, the whole model is likely thrown off.  The most obvious solution is to take the population of Minnesota residing in nursing homes and similar settings out of the general model and run them as a separate population, with different parameters and different mitigation approaches.  Given the residential setting for these individuals, their contact patterns certainly are very different.  And if that group is out of the model, we would have a more accurate projection of the course of the epidemic for the general population.  (Here is a hint, it will show very few deaths.)

The slides specifically say that the most effective measure to mitigate spread are the measures that protect the vulnerable.  I don’t know how any one at the State could say they have done a good job of doing that, when you look at the deaths from nursing homes and other group settings for seniors.

And that pretty much wraps up the discussion of the State of Minnesota’s epidemic modeling effort.  I think we are all eagerly awaiting the next model run and explanation.  We will see if the modelers have appropriately changed parameters, and moved away from Chinese data to support assumptions for parameters, to at least using US data, if not Minnesota data.  If the model is adjusted appropriately, the level of deaths will certainly be far lower than we are being shown.  But that might be embarrassing to certain political leaders who have used these scary numbers to justify very extreme and restrictive actions.  So we will see what the modelers are permitted to do.

Leave a Reply