Estimate of new positive cases in the early days of the epidemic

The need to provide fast estimates of new positive Covid-19 cases has led us to develop a model which, although not optimal, has proven to be very effective. We are constantly improving it.

The approach we implemented tries to approximate the two steps which lead to the detection of a new “case”: a screening step and a diagnostic test. In this epidemiological setting, the screening step has the following characteristics. Alleged positive individuals either autoselect themselves by reporting their symptoms to health authorities or are pinpointed through connections with previously confirmed positive cases. Then the diagnosis is obtained via a swab test.

In the first step, the number of daily swab tests is estimated via an Integer-valued AutoRegressive (INAR) model, i.e. an integer-valued time series model. In this first step, we predict the change in the number of daily swab tests. From February to March the process, and therefore also the model, is stationary. As days go by and the amount of data at our disposal increases, we are also taking policy changes regarding swab testing into consideration.  In the second step, we instead predict the number of new daily positive cases conditional on the number of performed swab tests. More precisely, a log-linear regression model with a Negative Binomial type error is used. The rationale behind this specification is the fact that count data is largely overdispersed. We then use these estimates, together with the predicted number of swab tests, to provide a 3-day estimate of the number of new positive confirmed cases of Covid-19. Furthermore, we provide confidence intervals obtained via resampling techniques. Initial resampling techniques were very simple (jackknife). With the help of Riccardo Ievole and Lucio Palazzo, we are trying to obtain more precise uncertainty estimates.



(Translation by: Gabriele Fabozzi)