Modeling the effects of contact-tracing apps on the spread of the coronavirus disease: Mechanisms, conditions, and efficiency

This study simulates the spread of the coronavirus disease (COVID-19) using a detailed agent-based model and the census data of Japan to provide a comprehensive analysis of the effects of contact-tracing apps. The model assumes two types of response to the app notification: the notified individuals quarantine themselves (type-Q response) or they get tested (type-T response). The results reveal some crucial characteristics of the apps. First, type-Q response is successful in achieving containment; however, type-T response has a limited curve-flattening effect. Second, type-Q response performs better than type-T response because it involves quarantine of those who are infected but have not become infectious yet, and the current testing technology cannot detect the virus in these individuals. Third, if the download rate of the apps is extremely high, type-Q response can achieve virus containment with a small number of quarantined people and thereby high efficiency. Finally, given a fixed download rate, increasing the number of tests per day enhances the effectiveness of the apps, although the degree of improved effectiveness is not proportional to the change in the number of tests.

interventions also change the quantitative results regarding the difference between type-T and type-Q responses. Therefore, section 6 was added to discuss these issues as follows. Figure 7 compares peak cases when type-T and type-Q responses are applied with vaccines, variants, and social distancing measures (Table 6), when the baseline results, where there are no interventions or new variants, are normalized to 1. The daily tests are assumed to be conducted on randomly selected symptomatic people with a probability of 50%, and the download rate is set to 90%. When social distancing is applied, 75% or 51% of the contacts which are randomly selected in the community and service-industry layers are negated, while their contacts in the home, workplace, school, and nursing home are maintained: Roughly, negating 75% and 51% of contacts correspond to the scenarios in which each person decreases their time outside the home by 50% or 30%, respectively, apart from commuting. Vaccines are assumed to be given to 0.7% of the population every day, with those aged over 65 years prioritized (Prime Minister's Office of Japan, 2021). Those who are vaccinated with the first dose receive their second dose 21 days later. The susceptibility of those who have received the first and second doses decreases by 80% and 95%, respectively, compared to those who have not been vaccinated (Hunter andBrainard, 2021 andPolack et al., 2020). New variants are also considered: they are 1.7 times as transmissible as the preexisting variants (Davies et al., 2021).
Results show that vaccines and social distancing substantially reduce peak cases when either response is applied. When vaccines are distributed or the number of people's contacts in the community and service-industry layers decreases by 51%, peak cases decrease by 50% to 70%; if such social distancing becomes stricter to decrease the contacts by 75%, peak cases decrease by more than 70% for each type of response. In each of these scenarios where only preexisting variants are assumed, peak cases decrease more with type-T response than with type-Q response, meaning that the relative superiority of the type-Q response decreases when compared to the baseline scenario, where no interventions are taken. This is because, even if the notified app users under the type-T response have contact with others, the chance of the virus spreading is rare because of low susceptibility thanks to vaccines or the small amount of contact from social distancing. When new variants are taken into account, transmissibility increases for both types of response. Here, the increase in peak cases with the type-Q response is 1.5 times larger than that with the type-T response, which means that the effectiveness of the type-Q response compared to that of the type-T response, again, decreases. This is because the new variant is so transmissible that even the type-Q response, which enables extreme curve-flattening in the baseline scenario, cannot sustain such effectiveness any longer. As a result, the spread of new variants makes the type-Q response less effective. In summary, when vaccines, variants, or social distancing are considered, there is less difference in peak cases under the two responses than in the baseline scenario, meaning the relative superiority of the type-Q response to the type-T response is weaker.
• Is mask-wearing and social-distancing accounted for in the model? These are very common behaviors, and including these things would likely affect your results if they are not already accounted for. The contact-tracing app can only alert an individual if they were in proximity to an infected individual, not whether or not they were 5 feet apart vs. 7 feet, nor whether or not either was wearing a mask. A population that adheres to social-distancing and mask-wearing policies would most certainly reduce the proportion of type-T individuals who become infected and spread the virus before they are tested and receive results. Modeling this could significantly reduce the difference in the effects of type-T behaviors vs. type-Q behaviors. Adding this as a variable for comparison in your model might further increase its value in terms of a tool to inform policy decisions around the pandemic.
Response: As pointed out, social distancing has two opposite effects on the results. On one hand, it suppresses virus spread because it reduces people's contact; on the other hand, social distancing makes it difficult for the apps to detect actual contact, which hinders the effectiveness of the apps. As answered in the previous bullet point, the newly added simulations in section 6 incorporate only the former effect, considering that infections rarely occur between people more than 6 feet apart, the maximum distance many apps can detect.
As for the analysis taking into account social distancing, please refer to the previous answer.
Although mask-wearing is one of the most common interventions, it is not included in this model because of the difficulty in quantifying its effects on transmissibility and susceptibility. The effectiveness of masks heavily depends on their type: cotton, surgical, or N95. There are no data available to detect the proportion of the users of each type of mask. This is mentioned in section 6.
• The author may also want to include a way to vary presumed infectiousness of the viral variant, since this is an inevitable as long as the pandemic continues. Including this may also significantly change the differences you see between type-T and type-Q populations, and again might make your model more relevant.
Response: Vaccines have been added to the analysis in section 6. As for the set-ups and results, please refer to the answer in the second bullet point.

Public access to your model:
I understand from your statement that the specific data from the Japanese census that you used to populate your model are not available publicly. However, is your model itself publicly accessible for people to populate with their own data? While it is not possible to run simulations using the exact dataset that you used, it would be possible to validate the model itself using a different dataset if it were publicaly available (for example, the Covasim model of Kerr et al.) Response: Thank you very much for your suggestion. The Python codes used in the simulations has been uploaded in a zip file as Supporting Information.

Related to point 2 above, it would be helpful for the author to make explicit in the paper that the Covasim model by Kerr et al. is in fact open-source. The link to the model should also be included, perhaps in a footnote.
Response: In the third paragraph in section 1, I highlighted the fact that Covasim is an open-source tool. In addition, the outline of Covasim and the description of what is new in the present model is clarified in the first paragraph in section 2.

Throughout the article, the author refers to the apps "requiring" a particular behavior (testing alone, or quarantine). The apps themselves do not require any particular behavior; they are simply a source of information, and it is the individual who determines their behavior. For this reason, a more precise orientation would be to refer to a "type-T response", or a "type-Q behavior" as this reviewer has done in the summary section above.
Response: Thank you for pointing out this imprecise language. The responses are referred to as "type-T/Q response" in the revised paper, and the phrasing "requiring" has been revised throughout.

I found Section 4.1 on comparison between SIR and ABM very interesting and informative. These are things that I do not have expertise in. However in the context of the author's article, this discussion seemed out of place. SIR is not mentioned at any other point in the article, and comparison of the relative merits of SIR and ABR is not listed as an objective of the study. I believe that the use of ABR for your study is appropriate and stands on its own in the context of this work, and there is no need to describe or justify why this method was chosen instead of SIR. If you feel it belongs in the paper, I would suggest a short explanation in your conclusions section, or a note in your appendix.
Response: Thank you for your suggestion. Figure 2 in section 4.1 in the original paper was intended to explain how the assumption of heterogeneity in virus transmission, including super-spreading environments, changes the simulated speed of the outbreak. As the high-dimensional heterogeneity is the feature of this analysis, I chose to keep the figure. However, as you pointed out, the original expression, which frequently used the term "SIR" without its explanation, was inappropriate. Therefore, in the revised paper, the section was moved to appendix B, and a qualitative explanation for the analytical models was added as follows: This section discusses how agent-based models are different from analytical epidemiological models, namely, susceptible-infected-removed (SIR) models (Anderson andMay, 1979 andKermack andMcKendrick, 1927). In many analytical models, the number of newly infected people is determined by the product of the number of the infected and the number of the susceptible. The number of the severely ill and of the recovered are determined as a certain proportion of the infected. Thus, the overall mechanism of the transmission of virus in analytical models is equivalent to that in agent-based models, as presented in this paper. An obvious difference is that agent-based models take a bottom-up approach, meaning that the smallest unit in their structure is an individual, which enables analyses on inter-relationships among people. Fundamentally, this can be interpreted as high-dimensional heterogeneity. That is, in agent-based models, each individual is characterized by a variety of attributes and contacts. In other words, an agent-based model without this diversity is substantially the same as analytical epidemiological models.
Other suggested improvements: 1. Clarify imprecise or confusing terminology: • For example, "superspreader" is not well-defined in its current usage. As the reference the author cites points out (Cave, 2020), "superspreader "is sometimes used to refer to an individual who may shed much more virus than most infected people and who is therefore more infectious, and it is sometimes used to refer to an event in which conditions are such that a large number of individuals become infected (a very crowded event with a long timeframe, for example). It was sometimes unclear what the author was referring to when the term was used. I ultimately determined that the author was referring to an individual, but this needs to be made clear from the start. Additionally, there is one place in the paper [p. 3, Section 2 (Model), line 4] where the author specifically uses the term "superspreading environments" which further adds to the confusion.
Response: Thank you for pointing out this confusion. The simulations assumed super-spreading environments, which refers to a phenomenon in which a small fraction of environments are high-risk compared to the total. The description and rationale behind this assumption are explained in section 2.3 in the revised paper, with some new citations, as follows: A super-spreading environment refers to the phenomenon in which a small amount of cases accounts for a large amount of transmissions (Cave, 2020). Moreover, it has been argued that the occurrence of this phenomenon depends on the environment rather than on the infected individuals (Majra et al., 2020) and that the Pareto principle, namely that 80% of all consequences come from 20% of the causes, could apply in the case of COVID-19 super-spreading environments (Kumar et al., 2020). Therefore, it is reasonable to consider that transmission probability for 20% of contacts selected at random is 50 times as high as that for the remaining contacts in the simulations. In fact, super-spreading environments are one of the determinants of coronavirus spread. As described in Appendix B, introducing super-spreading environments into the model decreases peak cases by 40%-50%. This means that the simulations overestimate cases if super-spreading environments are not considered. Accordingly, the term "super-spreader" was rephrased to "super-spreading environments." • "Lockdown" as it is used in the paper is another confusing term. I believe what the author means by this term is isolation or quarantine of an individual. However the common understanding of the term "lockdown" (in American culture at least) In addition, at one point [p. 1, line 2-3] the author uses the term and associates it with not only quarantine but also with behaviors such as social-distancing (in the US this term is specifically used to indicate individuals from different households staying at least 6 feet apart while in public).