The current winter COVID-19 surge in California, despite ongoing lock-down and mask mandates, has resulted a death rate close to 2X that of the summer surge. We present a detailed analysis to help understand both the driving causes and potential outcomes of this current surge. We have applied our COVID Decision Model, a discrete agent analysis tool, to simulate the California COVID-19 outbreak from its inception in Q1 of 2020. We model the initial outbreak, the summer surge and the winter surge in continuous fashion using a scaled agent population of 3 Million. We make the following observations:
- The level of activity (connectivity) necessary to drive the current surge in the winter months EXCEEDED the pre-shutdown levels in March 2020 (the assumed Ro baseline).
- The mortality rate reduction gains due to therapeutic advancements likely degraded during peak hospital utilization and hospital stress effects on infrastructure and staff.
- By March 2021 CA will have largely reached HIT for normal seasonal connectivity. Our models show that there is little difference between a return to normal and a hypothetical full vaccine scenario.
- Vaccine rollout is too slow to affect this current wave and should be targeted at vulnerable people if we wish to save lives.
- Vaccine programs will help prevent a future winter wave. We will investigate this in subsequent analysis.
- Aggressive stay at home orders and mask mandates have little effect on slowing the spread and are likely counter productive. Home transmission within multiple generations is largely driving this wave.
- See this publication by Jeff Harris: Understanding the Los Angeles County Coronavirus Epidemic: The Critical Role of Intrahousehold Transmission
- See this reference post by Nicholas Lewis for a high level analysis explaining why lockdowns are counter productive: COVID-19: why did a second wave occur even in regions hit hard by the first wave?
Time Dependent Variables
The key input variables that drive the rate of the infection are the connectivity levels by age group over time. Therapeutic and improvements in treatment reducing mortality is varied over time. In addition, extra protection in terms of reduced contact for vulnerable people is varied over time. The actual death rate, back fit to actual date of death is the calibration standard. The dynamic input parameters are varied to best match the actual death rates. In the case of this particular analysis, to match the severe rise in deaths seen in December, the connectivity during the winter months was increased ABOVE the initial baseline going into March 2020 AND the gains in mortality reductions due to therapeutic measures were also degraded. We examined 4 cases:
- Case 0 has a hard shut off on Dec 25th to illustrate the delay between infectious contact and actual date of death.
- Case 1 assumes the connectivity rise, peaking in mid/late December has a symmetrical fall to September levels.
- Case 2 assumes a return to the pre-shutdown baseline on March 1st 2021 (matching the initial R0) with everyone getting back to normal.
- Case 3 assumes a hard shutdown in March 1st 2021 to simulate a full vaccination scenario.
The following graphs present the daily and cumulative deaths for each case. Actual deaths are fit to the date of death using our algorithm described below. Peak death reporting will rollover in early February, but the actual peak is mid January. Overall, the cumulative deaths in California will approach be between 55,000 and 60,000. If vaccines are effectively deployed to vulnerable people, this number can be reduced somewhat. Overall, given the anticipated connectivity factors, seasonal herd immunity will be largely established, as evidenced by the small differences between case 1, 2 and 3. Note that the fit of actual deaths will be updated to reflect current counts.
The R(t) shows the dynamic nature of the infection activity.
Our model shows that the actual number of infections in the winter surge are not grossly excessive when compared to the summer surge. The level of testing is much higher and there are many duplicate and false positives, so the actual case data needs to be used cautiously. The demographic of those being infected has shifted to attack more vulnerable people, including more elderly.
Key Lesson: Ro at the Outbreak May NOT represent the Bounding Case
The key lesson is that human connectivity driving infectious spread varies seasonally as a function of weather and human behavior. We had assumed that the initial R0 conditions for a given geography / demographic were a static upper bound. However, our assessment of the winter outbreak required an increase of connectivity to a level ABOVE the levels used to match the initial R0 conditions. In simple terms, the level of potentially infectious human interaction during the cold winter months, especially during the holidays, exceeds the levels in early spring. Likewise, very hot weather also drives an increase in connectivity as people spend more time inside in air=conditioned homes, but likely not to the extreme of the colder winter months during the Thanksgiving, Christmas and New Year’s holiday season.
Key Lesson: Quality of Care Likely Degrades as Hospitals Reach Capacity
As hospital capacity is reached, especially in the ICU, the quality of care appears to degrade and the mortality rate increases. Our model indicates that the reduction in mortality for critical care patients was largely reversed when hospital capacity was stretched thin. This could be due to a combination of factors.
- Staff are burned out and stretched thin. This stress on staff appears to be the key element driving ICU and hospital capacity limits.
- Beds are unavailable. This is a secondary limit behind staff.
- People defer seeking treatment until too late. The longer a seriously ill COVID-19 patient waits to seek treatment, the less likely treatment will be effective.
- COVID death classification has broadened. We assume that the methodology for classifying COVID-19 deaths has not changed, but if this changed it would explain the increase mortality.
Date of Death Extraction
Determination of actual date of death is critical to calibration of the analysis. The graph below shows the derived date of death the reported death date data sets found at the COVID Tracking Project. Our approach is far more accurate than a rolling average of reported deaths and is detailed here: Reported versus Actual Date of Death. The most recent 2 weeks represent incomplete counts.
Initial Conditions at Outbreak: March R0
The starting point for the simulation, the baseline connectivity, drives the base reproduction factor, Ro. This is the measure of how many infections each infected person causes at the onset of the infection. A hypothetical unmitigated calibration run utilizing the initial conditions for March yields an Ro of 2.75 and a peak infection point where R(t) crosses 1 on day 80. The herd immunity threshold here is 32% and the final size of the epidemic (FSE) is 56% of the population infected.
Seasonal Winter Contact Rates and Reference Ro
We also examine the outbreak parameters for the condition of peak connectivity in December 2020 that fits the death curves. A hypothetical unmitigated calibration run utilizing the peak conditions for December yields an Ro of 3.3 and a peak infection point where R(t) crosses 1 on day 74. The herd immunity threshold here is 37% and the final size of the epidemic (FSE) is 62% of the population infected.
Input Human Characteristic Distributions
Each agent (person) in the model is assigned a unique set of characteristics, including mobility/connectivity, health, susceptibility and age. These characteristics are distributed across a reasonable range of variation, as occurs in real life. This heterogeneous variation of individuals allows results in a lower herd immunity threshold when compared to models that assume all people are the same. For a detail analysis see our post: Are We Closer to Herd Immunity than Most Experts Say?
Input Infection Outcome Probability Tables
In the simulation, human interactions are modeled by random chance, and the transmission of the disease is determined by a random number fit against a probability function. Once infected individual outcomes are chosen from an outcome tables aligned with current data as well as disease timers for symptoms and recovery or death. The probability tables are set so that 90% of fatal outcomes are for those with comorbidity conditions.
Outcome Result Summaries: