Jonathan Hsiung

The following article does not reflect the views of my employer, nor any political affiliation. These are simply my own views - at this point in time - on this particular subject.

"Coronavirus model [IHME] projects 134,000 deaths in US, nearly double its last estimate ." My line of work has taken me through various types of forecasting analyses: sales, costs, adoption rates, and on the rare occasion human behavior. As a disclaimer, I have no expertise in epidemiology nor am I experienced with the specific code behind the University of Washington IHME model. But I am experienced in reconciling between two halves commonly at war: justifying that modeling accuracy is sometimes elusive based on underlying data used, versus understanding "the approach" of how a model is constructed which is key to understanding why it may underperform. Ultimately, the combination of a thoughtful approach combined with good data lends itself to a successful model.

See that red cloud? There's quite some variability, even with two months of data and information from other countries.

(Source: IHME. covid19.healthdata.org)

R0 measures transmissibility Not all viruses are created equal. One main concept before hopping in is R0 ("R-naught", with a subscript zero that LinkedIn doesn't permit), which describes the transmissibility of the particular virus. Airborne measles: 12-18 (high R0). Bodily fluids-borne Ebola or HIV: 2-4 (low R0). Coronavirus is around the 5.7 range . The Coronavirus R0 score was revised upward as scientists have discovered more about how easily transmitted it can be, which factors in methods of transmission and livelihood on surfaces, among other factors. A quick way to ruin a model is to get this number wrong.

1. Historical Curve-fitting/Extrapolation models To understand why the IHME model was off, we look at how it predicts hospitalizations and deaths. Originally, the model was based on the infection curve adopted from China, Italy, and Spain . In the absence of additional data, you take the same curve and extrapolate it over the US population - like taking your shrunken sweater out of the dryer and gifting it to your little sister. Turns out though she also has two arms and a head, general differences in body types matter too. The geography of Wuhan may have shown a few differences from Nantucket, just as TV sometimes leads us to believe that Pauly D from Jersey Shore is representative of the residents in Milan. The IHME model recently evolved to a "multi-stage hybrid model" where they brought in the concept of SEIR.

2. Process-based SEIR models No it's not the rating for your air conditioner; SEIR refers to the four states: susceptible (S), exposed (E), infected (I), and resistant (R). This simplifies the infection process and rolls blocks of populations into these tiers, modeling how transition occurs over time. As you move populations between each of the phases, you account for assumptions regarding social interactions, quarantine policies, and health systems factor into the model. SEIR modeling appears very formulaic and graphs are very simple.

3. Simulation Agent-based models This is when we start getting into more advanced predictive-modeling, simulating actual interactions between people. This involves tracing where people go, how they behave, and occasionally some details like knowing how many times you touched fruit and put it back into the display bin at Whole Foods. The advantage of agent-based models is that it can synthesize many variables and provide strong scenario testing. The problem is when you have so many features you've realized you now no longer know what was useful anymore. As with more advanced models, it requires more effort to explain than simple models.

Conclusion These three traditional models may be the primary methods in epidemiological modeling, but with today's technological advances in statistical modeling and computational abilities, we are living in a world rife with possibilities of applications. At the very least, there should be some ensemble models mashing the best predictions of each model. In addition, it would be a shame to ignore new capabilities such as building a live feedback loop to instantly adjust models such as with the lifting of SIP restrictions, perform some Principal Component Analysis (PCA) to identify and analyze the highest determinant features, or tie near-real time GPS movements directly into Agent-based models.

Since models are used for public policy, I would encourage that the code be shared on Github. These models should not be black boxes; some rocks may thrown its way but if the model is logical and robust it should withstand scrutiny. Some articles debate whether the use of particularly conservative models by the government to set policy may be inadvertently dangerous ("oops, you'll have to stay in longer"), and so are the alternative effects on public perception by overestimating a pandemic ("sorry, society is broke now but fewer people died"). Good analytics managers must be conscious of these downstream so-called "business" effects, though should be careful as to not bias model results. The government policy debate should be reserved for another discussion.

Jonathan Hsiung

Irvine, California, United States

2K followers 500+ connections

View mutual connections with Jonathan

Welcome back

Email or phone

Password

Forgot password?

or

New to LinkedIn? Join now

or

New to LinkedIn? Join now

Experience

KPMG US

Education

University of Southern California - Marshall School of Business

Licenses & Certifications

Certificate in Finance

UCLA Extension

Certificate in Project Management

OSP International

View Jonathan’s full profile

Other similar profiles

Kshitij Agrawal

Jonathan Y Lee

Ryan Hogan

June Kang

Kristina Nguyen

Stephanie C.

Kenneth D. Stewart

Brian Amante

Ria Vidhate

Victoria Nguyen

Abdul-razaq Momodu

Daniel Kepler

Helga Farjadi , Ph.D

Hailing Xu, CETF®

Katie (Thomas) Dahler

mimi jacob

Igor Voynov

Mary Nguyen

Shura Durakova, CPA

Atul Shahi

Explore collaborative articles

Others named Jonathan Hsiung

Jonathan Hsiung

Jonathan Hsiung

Jonathan Hsiung

Jonathan Yeap Han Hsiung

Jonathan Hsiung

Irvine, California, United States 2K followers 500+ connections

See your mutual connections View mutual connections with Jonathan Sign in Welcome back Email or phone Password Show Forgot password? Sign in or New to LinkedIn? Join now or New to LinkedIn? Join now

Experience

KPMG US

Education

University of Southern California - Marshall School of Business

Licenses & Certifications

Certificate in Finance

UCLA Extension

Certificate in Project Management

OSP International

View Jonathan’s full profile

Other similar profiles

Kshitij Agrawal

Jonathan Y Lee

Ryan Hogan

June Kang

Kristina Nguyen

Stephanie C.

Kenneth D. Stewart

Brian Amante

Ria Vidhate

Victoria Nguyen

Abdul-razaq Momodu

Daniel Kepler

Helga Farjadi , Ph.D

Hailing Xu, CETF®

Katie (Thomas) Dahler

mimi jacob

Igor Voynov

Mary Nguyen

Shura Durakova, CPA

Atul Shahi

Explore collaborative articles

Others named Jonathan Hsiung

Jonathan Hsiung

Jonathan Hsiung

Jonathan Hsiung

Jonathan Yeap Han Hsiung

Irvine, California, United States

2K followers 500+ connections

View mutual connections with Jonathan

Welcome back

Email or phone

Password

Forgot password?

or

New to LinkedIn? Join now

or

New to LinkedIn? Join now