On our current trajectory, California has more than an adequate number of hospital beds to withstand the pandemic.
This Post outlines the revisions in methodology that constitute my Model 3.0 for predicting the number of hospital beds California requires to manage the COVID-19 outbreak. (I developed a Model 2.0 earlier this week, but it immediately became obsolete when California changed its reporting of pandemic statistics on March 31st).
Background:
The key justification for the extraordinary statewide stay at home order issued by Governor Newsom on March 19th was that it was necessary to “flatten the curve” in the number of new COVID-19 cases each day to ensure the healthcare system had the necessary number of beds to properly care for COVID-19 (and all other) patients.
Given this rationale, the critical question is how many beds is California likely to need given the actual trajectory of the virus.
Using publicly reported data on new cases, hospitalizations, and fatalities, I built a model to answer this exact question. The State has predicted it will need 20,000 to 50,000 beds through May. It reiterated those predictions on March 30th.
My modeling shows we need far fewer beds. I do not believe the State’s prediction is supported by the publicly released data.
(In certain locations, it is possible that there could be a bed shortage, but this is becoming less likely (and less likely to be an ongoing constraint) as we enter our 14th day under the stay at home order.)
I published my California healthcare capacity Model 1.0 and the assumptions upon which it is based on March 26th.
The key predictions from Model 1.0:
- There are essentially three waves of the COVID-19 pandemic in the data based on significant changes in public health measures through March 19th (the issuance of the statewide stay at home order):
o Wave 1 ended on March 8th when decentralized public health measures began across the State;
o Wave 2, which was marked by decentralized public health measures, ended on March 19th with the issuance of the statewide stay at home order; and
o Wave 3 (which will end whenever the stay at home order is replaced with less disruptive public health measures). - It will take roughly 18 days from the end of Waves 1 and 2 for their impact to be fully represented in the data.
- The State would need roughly 6,300 hospital beds to manage the pandemic indefinitely under the stay at home order.
As I noted at the time, the model would be updated based on better available data.
Model 1.0 has performed reasonably well.
But as I anticipated, there have been significant changes in the quality and quantity of data reported over the past several days, including the introduction of new statistics by the State of California. As a result, I am updating my model to version 3.0.
Model 3.0
The key changes in assumptions and caveats from Model 1.0 to Model 3.0 are:
- Rate of spread
o Revised the initial (end of Wave 1) rate of spread slightly higher based on an analysis of fatality data that was not previously availabe.
o Revised slightly the rate of spread for Wave 2 based on an analysis of fatality data that was not previously available (model assumption did not change – rate trends downward on a gradually reducing curve).
o Revised the rate of spread slightly downward for Wave 3 based on the slower spread now reflected in Wave 2 and in light of newly available data from Italy. - Hospital admissions have been revised significantly to conform to the State’s new reporting. There are two categories of hospital admissions: “Suspected” and “Confirmed.” The State has offered little guidance on how these categories are defined. I am assuming the following definitions and relationships:
o Suspected: There are “suspected” COVID-19 cases in the hospitals that have not yet tested positive for the virus, but are being treated as if they are positive.- The model assumes a 3-day lag between hospitalization and a test result.
- The model assumes that roughly .4 of this population will test positive and that .6 will test negative. This is an assumption for calculating certain statistics. Positives are added to confirmed cases. Negatives are not. However, this differentiation does not alter the combined bed total calculations as set forth below.
- Confirmed: There are positive tests that are followed by hospitalization. For simplicity, the model assumes that a person who tests positive will develop symptoms requiring hospitalization on the day the test results are confirmed (3-day lag). This does not have a meaningful impact on bed capacity. In this population, the model assumes a .2 hospitalization rate (rather than .14 as in Model 1.0). This is based on California data over the past several days. This changes the number of hospital beds in use on Day 1.
- In order to build the model, I had to estimate certain starting values for the new statistics the State is reporting by working backwards in time from limited known data. These starting values may need to be revised.
- Backlog
o Structural backlog – the difference between tests conducted and reported each day remains accounted for in the model by separately calculating and reporting “suspected” COVID-19 cases in a separate category, just as the State is now doing.
o I believe that any actual positives resulting from the 39,000 test backlog that suddenly appeared in the data on March 25th would now be accounted for in either actual or suspected cases. I made no additional allocation for this backlog as there has been no additional guidance from the State on how to view these cases. - Average hospital stay duration: I increased this to 12 days based on new guidance from the CDC noting the median is 10-13 days.
- I did not calculate the ICU bed requirement, but it appears the State will need .4+ ICU beds for every hospital bed required for a confirmed case. It will likely need a lower ratio for suspected cases because at least some percentage of this population will test negative and not require such extraordinary care.
What is now being reported from Model 3.0:
[UPDATE April 6, 2020: I am creating charts to track the performance of my active models and present the results more clearly. When they are done, you will be able to find them here.]
I have predictions running past April 16th in my model, but they are not shown yet because I predict the curve will be flat or bent by April 9th – a condition that will continue so long as the stay at home order is in place.
The key predictions from Model 3.0:
1. The State would need roughly 2,800 hospital beds to manage confirmed positive hospitalizations resulting from the pandemic indefinitely under the stay at home order.
2. The State would need roughly 1,700 hospital beds to manage suspected hospitalizations resulting from the pandemic indefinitely under the stay at home order.
3. In total, the State needs roughly 4,500 hospital beds to manage the pandemic while the stay at home order is in place.
Even if my estimates and assumptions are wildly inaccurate – a view that plainly runs contrary to the initial performance of Model 3.0 – it is difficult to understand from the publicly reported data how the State believes 50,000 hospital beds will be required for COVID-19 patients in May, let alone 20,000. There is no apparent basis in the data for asserting we could see a wave of cases after April 6th.
I will track the performance of Model 3.0 as noted above.
Updated April 6.
Updated April 9: In preparing to plot performance curves, I recognized that I had inadvertently transposed confirmed and suspected hospitalization numbers. This has been corrected. There was no change to predicted bed counts or total capacity requirement.
April 3, 2020
The world wide rate of change in the death rate started a decline about March 26th. This matches your predictions. I do see a major difference in NY and Ca patterns. Puzzled why LA hasnt had an explosive breakout like NY. I also wonder too about Mexico’s pattern since I would think it would be like Italy or Spain.
Your comments are much appreciated.
April 5, 2020
These are insightful questions. I have not followed Mexico’s pattern, so I cannot comment on their trajectory. However, I believe the differences between LA and NYC, narrowly, and CA and the rest of the US, more broadly, can be partially explained by demographics (CA is relatively young and healthy and our smoking rates are among the lowest in the U.S. and far lower that Europe, for example). The WHO has recently acknowledged that smoking may be one of the strongest predictors for hospitalization rates (a possibility I raised many days ago). In addition, despite its population, LA lacks the density of NYC and is far less dependent on mass transit. I suspect epidemiologists will explore the significance of these differences when they study the trajectory of this pandemic in various locations. Thank you for your comment.