Abstract
Estimating the unreported cases of Covid-19 in a region/country is a complicated problem. We propose a new mathematical model that, combined with a deterministic model of the total growth of cases, describes the time evolution of the unreported cases for each reported Covid-19 case. The new model considers the growth of unreported cases in plateau periods and the decrease towards the end of an epidemic wave. We combined the new model with a Gompertz-growth model, a generalized logistic model, and a susceptible-infectious-removed (SIR) model; and fitted them via Bayesian methods to data from Cuba and Spain. The combined-model fits yielded better Bayesian-Information-Criterion values than the Gompertz, logistic, and SIR models alone. This suggests the new model can achieve improved descriptions of the evolution of a Covid-19 epidemic wave. The new model is also able to provide reliable predictions of the epidemic evolution in a short period of time. We include in the paper the steps that researchers should take to use the new model for predictions with other data.
Introduction
One of the primary challenges in managing the COVID-19 pandemic was the discrepancy between official case counts and the actual number of infections. Factors such as limited testing, asymptomatic spread, and reporting delays created a "dark figure" of unreported cases. Standard epidemiological models, like the SIR model, often fail to account for "plateau" periods—days with zero reported cases despite ongoing transmission. This paper introduces a mathematical framework that estimates the ratio of unreported to reported cases, allowing for more accurate modeling of epidemic waves.
Methodology
The researchers developed a discrete-time model for the ratio of unreported cases to reported cases, denoted as $U(n)$. This model incorporates parameters for growth during reporting plateaus ($a$), detection effectiveness ($b$), and the natural decline of unreported cases at the end of a wave ($c$). This $U(n)$ component is designed to be integrated with existing growth models, such as Gompertz, Generalized Richards (GRM), or SIR models.
The combined models were fitted using Bayesian inference, specifically the Delayed-Rejection-Adaptive-Metropolis (DRAM) algorithm, to data from Cuba and Spain. Model performance was evaluated using the Bayesian Information Criterion (BIC), which penalizes complexity to ensure the extra parameters in the new model provide meaningful improvements.
Results
The study analyzed five epidemic waves across Cuba and Spain, ranging from small-scale local data in Santiago de Cuba to large-scale national data in Spain. In all instances, the models integrated with the new unreported-case component (+U) significantly outperformed their standalone counterparts. For example, in Santiago de Cuba's first wave, the Gompertz+U model reduced the BIC from 175 to 101.
Key findings include the model's ability to maintain realistic trajectories during reporting plateaus and its strong short-term predictive power. The parameter $b$ (detection effectiveness) emerged as a potential metric for the quality of local contact tracing and testing efforts, with higher values indicating a more responsive testing regime.
Discussion
While the model shows great promise, it relies on the availability of "active case" data and the fundamental accuracy of the underlying deterministic growth model. Bayesian estimation can also be computationally complex, requiring careful tuning of initial conditions. Future work could involve integrating this framework into more complex compartmental models like SEIRD to further enhance its predictive accuracy and utility for public health planning.
Summary updated