Enhancing Long-term Forecasting: Learning from COVID-19 Models (preprint)

Long-term forecasts are hard, but also indispensable in personal and policy planning. How could long-term predictions of complex phenomena, such as COVID-19 contagion, be enhanced? While much effort has gone into building predictive models of the pandemic, some have argued that early exponential growth combined with the stochastic nature of epidemics make the long-term prediction of contagion trajectories impossible. We leverage the diverse models contributing to CDC repository of COVID-19 death projections to identify factors associated with prediction accuracy across different projection horizons. We find that better long-term predictions correlate with (1) capturing the physics of transmission (instead of using black-box models); (2) projecting human behavioral reactions to an evolving pandemic; and (3) resetting state variables to account for randomness not captured in the model before starting projection. A very simple model, SEIRb, that incorporates these features and includes few other nuances offers predictions comparable with the most accurate models in the CDC set. Key to the long-term predictive power of multi-wave COVID-19 trajectories is endogenously capturing behavioral responses: balancing feedbacks where the perceived risk of death continuously changes transmission rate through the adoption, and relaxation, of various Non-Pharmaceutical Interventions (NPIs).

Simulation-Based What-If Analysis for Controlling the Spread of COVID-19 in Universities (preprint)

A simulation model is developed to analyze the spread of COVID-19 in universities. The model can be used to conduct what-if analysis and to estimate infection cases and the probability of death for students and faculty/staff under different policies. For proof-of-concept, the model is simulated for a hypothetical university of 25,000 students and 3,000 faculty/staff in a US college town. In this case, students arrive on campus in September and the semester is planned to last for a 90-day period. In base run, absent major policies other than reactive testing at 500 tests per day (sensitivity=.8, specificity=.998), limited quarantine of 170 beds, and R0=3, the disease quickly spreads and the probability of having at least 1, 2, and 5 deaths by the middle of the semester reaches .94, .78, and .16 for students, and ~1, ~1, and ~1 for faculty/staff. Simulation results show that there is no silver bullet to avoid an outbreak and, instead, a combination of policies should be carefully implemented. The effectiveness of proactive testing highly depends on testing capacity (to maintain high test frequency) and, inversely, on the delay between symptom onset and test results. Contact tracing and quarantine only when combined with other policies such as rapid, frequent testing and mask use enforcement are effective. To decrease death likelihood, universities should ask all staff/faculty of over 60 years old, and a large fraction of 30-60 years old to work remotely. Simulation results suggest these alternatives: 1) (almost) full remote operation from the beginning, 2) remote operation for high-risk individuals (all over 60s and most of 30-60s) in addition to frequent rapid tests, contact tracing with high capacity for quarantine, enforcing mask use, and social distancing. Results show that the system is highly vulnerable, and considering implementation challenges, many universities are likely to close and switch to remote classes to avoid catastrophic outcomes. A simulation platform for what-if analysis is offered so marginal effectiveness of different policies, and different decision making thresholds for closure can be tested for universities of varying populations. The model in Vensim is available. A web app is provided at https://forio.com/app/navidg/covid-19-v2/ and an instructional video is available at https://youtu.be/PrYarrpqa4Y for further analysis.

Weather Conditions and COVID-19 Transmission: Estimates and Projections (preprint)

Understanding how environmental factors impact COVID-19 transmission informs global containment efforts. We studied the relative risk of COVID-19 due to weather and ambient air pollution. We estimated the daily reproduction number at 3,739 global locations, controlling for the delay between infection and detection, associating those with local weather conditions and ambient air pollution. Controlling for location-specific fixed effects and local policies, we found a negative relationship between the estimated reproduction number and temperatures above 25oC, a U-shaped relationship with outdoor ultraviolet exposure, and weaker positive associations with air pressure, wind speed, precipitation, diurnal temperature, SO2, and ozone. We projected the relative risk of COVID-19 transmission due to environmental factors in 1,072 global cities. Our projections suggest warmer temperature and moderate outdoor ultraviolet exposure may offer a modest reduction in transmission; however, upcoming changes in weather alone will not be enough to fully contain the transmission of COVID-19.

Simulation-based Estimation of the Spread of COVID-19 in Iran (preprint)

Background: The 2019 Coronavirus (COVID-19) has turned into a global pandemic with unprecedented challenges for the global community. Understanding the state of the disease and planning for future trajectories relies heavily on data on the spread and mortality. Yet official data coming from various countries are highly unreliable: symptoms similar to common cold in majority of cases and limited screening resources and delayed testing procedures may contribute to under-estimation of the burden of disease. Anecdotal and more limited data are available, but few have systematically combined those with official statistics into a coherent view of the epidemic. This study is a modeling-in-real-time of the emerging outbreak for understanding the state of the disease. Our focus is on the case of the spread of disease in Iran, as one of the epicenters of the disease in the first months of 2020. Method: We develop a simple dynamic model of the epidemic to provide a more reliable picture of the state of the disease based on existing data. Building on the generic SEIR (Susceptible, Exposed, Infected, and Recovered) framework we incorporate two behavioral and logistical considerations. First we capture the endogenous changes in contact rate (average contact per person) as more death are reported. As a result the reproduction number changes endogenously in the model. Second we differentiate reported and true cases by including simple formulations for how only a fraction of cases might be diagnosed, and how that fraction changes in response to epidemic's progression. In estimating the model we use both the official data as well as the discovered infected travelers and unofficial medical community estimates and triangulate these sources to build a more complete picture. Calibration is completed by forming a likelihood function for observing the actual time series data conditional on model parameters, and conducting a Markov Chain Monte Carlo simulations. The model is used to estimate current "true" cases of infection and death. We analyze the future trajectory of the disease under six conditions related to the seasonal effects and policy measures targeting social distancing. Findings: The model closely replicates the past data but also shows the true number of cases is likely far larger. We estimate about 493,000 current infected cases (90% CI: 271K-810K) as of March 20th, 2020. Our estimate for cumulative cases of infection until that date is 916,000 (90% CI: 508K, 1.5M), and for total death is 15,485 (90% CI: 8.4K, 25.8K). These numbers are significantly (more than one order of magnitude) higher than official statistics. The trajectory of the epidemic until the end of June could take various paths depending on the impact of seasonality and policies targeting social distancing. In the most optimistic scenario for seasonal effects, depending on policy measures, 1.6 million Iranians (90% CI: 0.9M-2.6M) are likely to get infected, and death toll will reach about 58,000 cases (90% CI: 32K-97K), while in the more pessimistic scenarios, death toll may exceed 103,000 cases (90% CI: 56K-172K). Implication: Our results suggest that the number of cases and deaths may be over an order of magnitude larger than official statistics in Iran. Absent extended testing capacity other countries may face a significant under-count of existing cases and thus be caught off guard about the actual toll of the epidemic.