Powerful Behavioral Data Science

PANDEXIT is a detailed agent-based simulation, augmented with a geographical information system, that models the whole population of Qatar one-to-one. Thus, for every inhabitant of the country a corresponding agent exists in the simulation, and for every physical place in Qatar (whether a house, an apartment, a work office, a restaurant, a metro station, etc.) a corresponding digital place with geographical coordinates is included in the simulation. Transport means like buses, metros, private cars, and ride sharing services, are also represented inside the virtual world and the simulated agents use them to follow their everyday routine which is based on their socioeconomic profile.

In order to obtain the following curves, we have set up a sequence of stages for a simulation starting on February 15th, 2020, with the intention of shadowing inside the virtual world of PANDEXIT the most important real world policies adopted by the authorities of Qatar during the early lockdown and later relaxation phases. Each stage includes a set of parameters related to the number of foreign arrivals and the subsequent mandatory quarantine of arriving passengers, regional policies determining which public and private buildings remain open and at which capacity, economic policies related to the specific sectors allowed to remain open and the applicable workforce restrictions, early detection policies like random testing and required usage of thermal cameras, fever assessment, and contact tracing applications at the entrance of public places, religious policies like closing and partial openings of worship sites, as well as parameters related to the behavior of the agents like friends and family visits, mask policies compliance, and respect for social distancing norms.

Over every iteration of the simulated scenario, five random cases are bootstrapped inside the country at the starting date and let loose to continue with their daily routines, unaffected. Following the turn of events, schools are closed on March 9th and air traffic is severely restricted from March 18th onward, while a full lockdown is established on the beginning of April. Vigilance and policy enforcement is increased until the peak is reached near the beginning of June. Then, following the official guidelines for the relaxation of the lockdown, four successive stages starting on June 15th, July 1st, August 1st, and October 1th are applied, that comprise gradual opening of religious places, shops, and restaurants, as well as the consistent increase in returns to on-premises work schedules. On these notable dates the parameters of the simulation are modified to reflect the policy changes, which in turn determine the allowed actions of the simulated agents and their whereabouts.

The granularity of the simulation is such that intervals of ten minutes of virtual time are used mostly, except for modeling the transport network which requires a granularity of one minute. For every time slice, agents are moved between locations if needed and their interactions are computed to determine whether infected agents, symptomatic or asymptomatic, spread the infection. Without a need for fine tuning, the curves resulting from the data gathered by the simulation match their real world counterparts to surprising accuracy, even though the objective of the tool is not to improve on the state of the art curve fitting models but to enable decision makers to evaluate expected outcomes of complex scenarios that a statistical model cannot encompass.

Three main curves can be seen in the daily changes plot at the right. Even though there are five plots in total, the dead count and the infected foreign arrival count are negligible in comparison with the other three, and thus cannot be seen clearly. In yellow the total number of infected agents is shown while black displays the number of detected cases. The latter is always lower than the former because some agents are allowed to be asymptomatic, never develop a fever high enough to be detected by thermal cameras, or do not present strong symptoms that force them to go to the hospital instead of proceeding with their normal life, uninterrupted. Public reports are necessarily based on detected cases; this is why the real world curve should be compared to the black line showing detected cases inside the simulation. If adequately configured, PANDEXIT can also offer a hint on the number of undetected cases, which still fuel the spread of the pandemic.