Abstract
Epidemiologists have never had such high-quality real-time pandemic data. Modeling CoVID-19 pandemic data became a predictive tool in-stead of an afterwards analysis. How early CoVID-19 model predictions impacted US Government policies and practices is first reviewed here as an important part of the pandemic history. It spurred independent modeling efforts, such as this, to help develop a better understanding of CoVID-19 spread, and to provide a substitute for the IHME (Institute for Health Metrics & Evaluation, U. Washington) 4-month predictions for the expected pandemic evolution, which they had to revise every couple of weeks. Our alternative model, which was developed over the course of several earlier medrxiv.org preprints, is shown here to provide a good description for the entire USA CoVID-19 pandemic to date, covering: (1) the original CoVID-19 wave [3/21/20-6/07/20], (2) the Summer 2020 Resurgence [6/07/20-9/25/20], (3) the large Winter 2020 Resurgence [9/25/20-3/19/21], (4) a small Spring 2021 “Fourth Wave”, [3/19/21-6/07/21], and (5) the present-day Summer 2021 “Fifth Wave” [6/07/21-present], which the USA is now in the midst of. Our analysis of the initial “Fifth Wave” data shows that this wave presently has the capacity to infect virtually all susceptible non-vaccinated persons who practice NO Mask-Wearing and minimal Social Distancing.
1 Introduction
The initial CoVID-19 pandemic response by the US Government and public agencies needs to be remembered for future pandemics. In spite of the availability of high-quality real-time pandemic data, the early history of the pandemic in the US shows that very little coordinated effort occurred to use this data in a timely manner, or to test the robustness of early CoVID-19 pandemic assumptions.
This report includes material from a recent presentation for the Virtual MathFest 2021, hosted by the Mathematical Association of America (MAA) on 3-7 August 2021. Figure 1, created by Andy Balk for the UK online news outlet The Independent 1, shows the US early CoVID-19 data combined with the US President’s early pandemic responses, including the President calling the pandemic a “hoax” that “wasn’t out fault”, before declaring a National Emergency in mid-March 2020.
Figure 2 captures the impact of early CoVID-19 pandemic modeling by the University of Washington IHME (Institute for Health Metrics and Evaluation), which released a 4-month projection for CoVID-19 evolution in the USA2 on 3/25/2020. It was labelled “America’s most influential coronavirus model” 3, and it initially projected a total of about 81,000 CoVID-19 fatalities. Two weeks later, on 4/06/2020, the IHME revised its projections to only about 60,000 USA deaths by August 20203. By 4/19/2020, the White House had adopted this IHME projection as a guide for their national policies4. Unfortunately, less than two weeks later, on 5/01/2020, the USA number of CoVID-19 deaths surpassed this value5, three months early. A day later, the internet blog vox.com offered the question6, “The IHME coronavirus model keeps being wrong. Why are we still listening to it?”
We found a persistent flaw in the IHME model. It is illustrated by the IHME graphs in Figure 3, which shows their expected number of daily new cases until CoVID-19 pandemic end, for both Northern Pennsylvania7 (IHME, 3/26/2020) and the nation8 (IHME, 4/22/2020). The persistent flaw is that the IHME assumed that symmetric functions for the expected number of daily new cases would always be applicable. They further assumed that these functions would be symmetric Gaussians, which drove all their predictions, as disclosed in their MedRxiv preprint of 3/25/20209.
Concerns about this IHME model and concerns with the Federal Government’s early CoVID-19 pandemic response resulted in many mathematically-minded individuals, including us, started their own independent CoVID-19 data analyses, while Sewing Circles all across America were making handmade masks for First Responders and patients.
The next section summarizes all of our CoVID-19 models, starting with our original March-April 202010 Social Distancing model. This model was then expanded11-13 to include the likely effects of Mask-Wearing. A comparison of our model to the USA CoVID-19 pandemic data up through August 2021 shows that our expanded model is applicable to all USA CoVID-19 data, and provides a fairly accurate and quantifiable picture of the four USA CoVID-19 waves that have already occurred. Initial study of present-day “Fifth Wave” data also shows that it has the capacity to become the most deadly wave of all.
2 Initial Model Background and Development
This CoVID-19 study starts with the standard SEIR (Susceptible, Exposed, Infected, Recovered/Removed) model and R0-index, as shown in Figure 4. This R0 modifies exponentials so that R0 > 1 is pandemic growth; R0 < 1 is pandemic decay; while R0 = 1 gives a persistent disease baseline in the population.
Let N(t) be the total number of pandemic cases, with dN / dt being the expected number of daily new cases. The simplest SEIR model is:
The {KR R0} combination sets how fast an infected person spreads CoVID-19 to others; No is the number of infected people at time t = 0; and tdbl is the pandemic doubling time. As shown in Figure 4, the early USA CoVID-19 data5 followed this exponential growth, with a very short doubling time of tdbl ≈ 2.02 d. The form of the Eq. [2.1a] differential equation makes this SEIR model a local model, so it does not automatically account for changes in the collective behavior among a large population of uninfected people.
These collective behavior changes can arise from government mandates, or from new individual choices among uninfected people, such as Social Distancing and Mask-Wearing. These factors add a new non-local dimension to pandemic evolution, requiring extension of the Eqs. [2.1a]-[2.1c] SEIR exponential growth models.
On March 19, 2020, Governor Gavin Newsom of California ordered a CoVID-19 “stay-at-home” lockdown of virtually all of California’s ∼ 40 million residents14. Similar statewide CoVID-19 lockdowns were next ordered by the Governors of Illinois, New York, Indiana, Michigan, Ohio, Washington, West Virginia and Wisconsin15. Within days, this nationwide change slowed the CoVID-19 pandemic growth, as shown5 in Figure 5. This change corresponds to an increasing tR or tdbl, making these parameters time-dependent. We next showed10 that the simplest tdbl(t) model, using a linear function for tdbl(t):
naturally predicts pandemic end, since:
The αS parameter measures the amount of Social Distancing. Instead of being a symmetric function, the dN / dt from Eq. [2.2] for the expected number of daily new cases is very asymmetric, with a predicted ∼1/t2 long-term tail:
This asymmetry in dN / dt closely matches the bing.com CoVID-19 USA data up through 4/19/202017, as shown in Figure 6.
Our original preprint was sent to multiple organizations on 4/29/2020, including the IHME. Within a week, on 5/4/2020, the IHME substantially revised their entire reported modeling effort16, and offered only a range of possibilities, in lieu of a specific prediction. As reported by Alan Boyle of geekwire.com17, the IHME researchers acknowledged on 5/4/2020 that their previous CoVID-19 modeling wasn’t “sophisticated enough”.
Figure 7 shows the results of applying Eq. [2.2] to the ongoing bing.com CoVID-19 USA data up through 6/07/2020, seven weeks later than in Figure 6. The {No[3/21/2020] tR, αS} values changed only by 8% –10% between the 4/19/20 and 6/07/20 analyses. The NFINAL value for the total number of USA CoVID-19 cases predicted at pandemic end, changed ∼18%, from NFINAL ≈ 5.46 million to NFINAL ≈ 4.50 million. It represents a significantly smaller model prediction change over a much longer time, compared to the 3/25/2020 IHME model.
Figure 8 shows the world-wide early CoVID-19 data, as assembled by The Royal Society of London for their early CoVID-19 pandemic 8/24/2020 review18. Various preset tdbl values are also shown as a visual guide. Using a logarithmic ordinate, virtually all these data show a downward curvature away from the straight-line function of pure exponential growth.
In Figure 9, the Eq. [2.2] model was applied to the early pandemic growth for various countries and the World, showing it provided a fairly good approximation for most cases examined10. However, as Figure 10 shows, the early CoVID-19 data from Italy was somewhat different, in that it had a post-peak dN / dt tail that was almost a pure exponential decay down to fairly low values, which made it difficult for the Eq. [2.2] model to handle.
As NPR reported19, the People’s Republic of China, as part of their early CoVID-19 pandemic assistance to Italy, recommended three large-scale changes, which the Italian government quickly adopted: (1) significant mask-wearing, (2) more aggressive business shutdowns, and (3) more social distancing. It is doubtless that these aggressive changes contributed to a different early CoVID-19 pandemic evolution in Italy.
A second parameter was then added to the Eq. [2.2] model12, to explicitly enable dN / dt long-term tails to exhibit a nearly exponential decay with time: but it needs an added restriction that this pandemic wave ends whenever the calculated dN / dt < 0 first occurs using Eq. [2.5].
Using Eq. [2.5] gives the Italy datafit shown in Figure 10. Since the main difference between the USA and Italy CoVID-19 responses at that time was significant Mask-Wearing, the δo size is likely to primarily be a Mask-Wearing metric.
Follow-on analyses for the successive waves of USA CoVID-19 infections are shown in Figures 11-15. In addition to covering: (a) the initial Spring 2020 pandemic [Figure 7], this Eq. [2.5] few-parameter model reasonably fits: (b) the Summer 2020 Resurgence [Figure 11], (c) the large Winter 2020 Resurgence [Figures 12 and 13], (d) the smaller Spring 2021 “Fourth Wave” [small dN / dt peak in Figures 12 and 13], and (e) the present-day Summer 2021 “Fifth Wave” [Figures 14-15].
Since each new CoVID-19 wave sits atop the tails from all the prior CoVID-19 waves, the CoVID-19 overall progression in Figure 15 shows that this multi-wave analysis has retained its overall validity, from the initial 3/21/2020 start date, up through the latest data. Model parameters that were derived from each prior wave did not need any revisions, once the next new CoVID-19 wave became established as exceeding the prior baseline.
The various {tR, αS, δo}values associated with each of these CoVID-19 stages are summarized in Figure 16. Those parameter values provide additional insight as to what factors were likely driving the CoVID-19 infection rate for each USA pandemic wave.
3 What the Parameter Values Indicate
The first weeks of the CoVID-19 pandemic (3/7/2020-3/21/2020) had an N(t) doubling time of ∼ 2.02 d, as shown in Figure 4. As given in the Figure 16 table, the entire pandemic first-wave, from 3/21/2020 through 6/07/2020, showed a best-fit value of tR = 2.8804, corresponding to an intrinsic doubling time of: which is virtually identical to the original Figure 4 early-pandemic value. Since this tR parameter is closest to the Eqs. [2.1a]-[2.1c] basic SEIR model, it shows that the intrinsic USA-wide CoVID-19 spreadability remained constant. Thus, virtually the entire USA pandemic mitigation for this time interval was due to the Social Distancing measures that were implemented. The calculated value of αS = 0.066l8 / day gives a final CoVID-19 pandemic end of NFINAL ≈ 4.4995 million USA cases total.
Unfortunately, loosening controls led to a CoVID-19 Summer 2020 Resurgence, which likely started around 6/07/2020, 78 days after the start of USA-wide Social Distancing on 3/21/2020. Other factors such as different CoVID-19 variants circulating could have also contributed to the overall size of this resurgence. The Eq. [3.1] tR-value and intrinsic doubling time for the Summer 2020 Resurgence was 40% – 50% larger than the CoVID-19 initial wave, indicating that SEIR parameters for this resurgence were actually less aggressive. The calculated Social Distancing value of αS = 0.058 / day was similar to the initial CoVID-19 pandemic wave, indicating a similar amount of Social Distancing mitigation.
People in the USA started to engage in Mask-Wearing, so that the post-peak decline for this Summer 2020 resurgence was faster than for the initial Spring 2020 CoVID-19 wave, as shown by the Figure 11 inset. Our analysis gave δo = 0.0108 / day for Mask-Wearing. Although this value is significantly smaller in size than αS, its unit-for-unit impact is greater. Thus, Mask-Wearing is more powerful as a CoVID-19 pandemic stopping agent than Social Distancing. Had all of those controls remained in place, and had the weather not changed, these calculations would have predicted a CoVID-19 pandemic end with NFINAL ≈ 9.643 million USA cases total.
But the seasons changed, the USA weather changed, and the dreaded, almost unavoidable, Winter 2020 Resurgence then occurred. Our best estimate for its starting date was 9/25/2020, 110 days after the 6/07/2020 start of the Summer 2020 Resurgence. Its Eq. [3.1] tR-value and intrinsic doubling time was almost 2X larger than the Summer 2020 Resurgence, indicating that SEIR parameters continued to become even less aggressive. However, both the Social Distancing parameter αS, and the Mask-Wearing index δo were much less, which made the Winter 2020 Resurgence very long and very deadly.
It overlaps with a small Spring 2021 “Fourth Wave”, which started around 3/19/2021, 175 days after the Winter 2020 Resurgence started on 9/25/2020. This small “Fourth Wave” likely ended 0 days later, around 6/07/2021, which ironically covered virtually the same calendar interval as the original CoVID-19 pandemic start. Combining the Winter 2020 Resurgence and this small “Fourth Wave”, with the prior CoVID-19 waves now predicted a final CoVID-19 pandemic end at NFINAL ≈ 34.8l7 million USA cases total.
We are now in the CoVID-19 pandemic “Fifth Wave”. We estimate that it started around 6/07/2021, and it is a pernicious Summer 2021 Resurgence. Our initial analyses shows that the basic SEIR model tR parameter continues to increase, which is good. But any positive tR value means that the CoVID-19 pandemic can still exponentially grow at any time. CoVID-19 pandemic shutoff needs a relatively large αS Social Distancing value, and a finite Mask-Wearing index δo to hasten the pandemic end.
Our model tells us who is getting infected in this “Fifth Wave”. As the Figure 16 Table shows, the infected people in this “Fifth Wave” are associated with virtually NO Mask-Wearing, and very little αS Social Distancing. The αS and δo values are so small for the beginning of this “Fifth Wave” that our model predicts an NFINAL value in the billions, assuming an arbitrarily large susceptible population. What such a large NFINAL means is that, unabated, this “Fifth Wave” can infect virtually every susceptible non-vaccinated person in USA who is in this “minimal Social Distancing” and “NO Mask-Wearing” sub-group.
4 Summary
Let N(t) be the total number of pandemic cases, with dN / dt for the expected number of daily new cases. Figures 1-3 here summarize elements of the early USA CoVID-19 response. One of the most influential early USA CoVID-19 models by the IHME (Institute for Health Metrics and Evaluation) was likely wrong, which led us to develop an alternative model for USA CoVID-19 spread.
For the early CoVID-19 pandemic, a basic SEIR (Susceptible, Exposed, Infected, Recovered/Removed) epidemiology model accurately predicted an N(t) exponential growth : where tR is the pandemic growth rate, and tdbl is the N(t) doubling time.
Because SEIR models are local models that track and predict the number of infected persons, they do not automatically account for changes in the collective behavior among the large population of uninfected people. These changes can be due to new government mandates, or changes in the individual choices among uninfected people, such as Social Distancing and Mask-Wearing. These factors can add a new non-local dimension to pandemic evolution, requiring an extension of the basic Eqs. [4.1a]-[4.1c] SEIR exponential growth models.
The USA CoVID-19 data showed that these large-scale interventions convert tdbl in Eq. [4.1c] into a function of time tdbl(t). To model these changes, early data supported using a nearly linear function of time for tdbl(t):
A simple interpretation of αS is that it measures the amount of Social Distancing among uninfected persons, as a pandemic mitigation measure. When tdbl(t) is a nearly linear function of time, the total number of infected persons at the pandemic end, N(t → ∞), will converge to a finite value, automatically shutting the pandemic off. Having a sub-linear tdbl(t) growth would only slow the pandemic down, but not stop it from eventually infecting all susceptible persons. Social Distancing is needed, and it has to be robust enough so that the long-time behavior of tdbl(t) is at least a linear function of time to enable pandemic shutoff.
This model provided a good fit to the early USA CoVID-19 data, holding its predictive power for nearly two months. The Eqs. [4.2a]-[4.2c] functions were found to provide acceptable CoVID-19 datafits for many other countries around the world, except for Italy.
Italy differed from many other countries that mandated Social Distancing, Non-Essential Business Closures, and Lockdowns, in that they also instituted significant Mask-Wearing. As a result, their post-peak decline in dN / dt for their first major CoVID-19 wave showed nearly an exponential decay, which is significantly faster than the Eq. [4.2c] predictions. A second parameter was added to Eq. [4.2a], giving: to cover this new case, but an additional restriction is needed that the end of the pandemic wave occurs when dN / dt < 0 is first calculated in Eq. [4.3]. The simplest interpretation for this δo second mitigation parameter is that it measures the amount of Mask-Wearing among uninfected persons.
While the Eq. [4.2a] model fits (a) the initial USA Spring 2020 pandemic [Figure 7], the Eq. [4.3] enhanced few-parameter model fits virtually all the observed follow-on USA CoVID-19 waves, including (b) the Summer 2020 Resurgence [Figure 11], (c) the large Winter 2020 Resurgence [Figures 12 and 13], (d) the smaller Spring 2021 “Fourth Wave” [small dN / dt peak in Figures 12 and 13], and (e) the present-day Summer 2021 “Fifth Wave” [Figures 14-15].
The calculated αS and δo values for this “Fifth Wave”, given in the Figure 16 Table, are so small that, unabated, virtually every susceptible non-vaccinated USA person in this “minimal Social Distancing” and “NO Mask-Wearing” subgroup can become infected. Hopefully, as this “Fifth Wave” evolves, more aggressive personal responsibility by this susceptible sub-group will raise the pandemic-ending αS and δo values.
Data Availability
All data used is in the Public Domain or was maintained by bing.com.