1.  Contacts per day 

A number of comments expressed concern about our assumption of 8.3 contacts per day.

First, it is absolutely true that there is uncertainty about the “right” choice for the number of contacts per day. Figure 11 in the report shows how our results vary as this parameter changes. It clearly plays an important role, so the concerns expressed about its value are valid.

Second, our initial modeling views the transmission probability assumed (2.6%) as appropriate for a close contact. While definitions of a close contact vary, the CDC definition is “6 feet or less for 15 minutes or more” (https://www.cdc.gov/coronavirus/2019-ncov/php/public-health-recommendations.html).

While the literature is uncertain about the importance of transmission through interactions with smaller duration or larger distance, the CDC considers their probability of transmission low enough to treat individuals with these kinds of interactions in the same way as those in the general population.

Given that the C-TRO report calls for spacing out lecture halls so that students will be more than 6 feet apart from each other (bottom of p.9 in the executive summary), sharing a lecture hall does not constitute close contact according to the CDC definition. In response to F28, we have discussed this with Kim Weeden and Ben Cornwell. Our addendum will discuss the extent to which being in the same lecture hall further apart than 6 feet can lead to transmission if people are well-separated, masks are worn, and individuals do not cough, sneeze, sing, yell, spit or otherwise propel material into the air at high velocity. The most critical aspect of this is the extent to which SARS-CoV-2 is transmitted via aerosols, a point on which the literature is highly uncertain. Regardless, it likely is a good idea to strongly discourage students from yelling in class, and to be vigilant about coughing.

As another datapoint in specific response to questions about Weeden & Cornwell 2020, Samitha Samaranayake (Civil and Environmental Engineering) and PhD student Matthew Zalesak are undertaking a modeling study in collaboration with Weeden, Cornwell, Nathaniel Hupert (Cornell Weill) and modeling team-member David Shmoys where spread due to sharing classrooms is studied under different course scheduling policies. One aspect being modeled is the fact that transmission diminishes as the distance between individuals grows. Our understanding is that including this important biological detail results in a significantly diminished risk from within-classroom spread.

Third, the nominal value of 8.3 contacts per day is meant to reflect an average over all members of the campus community, including faculty, staff and graduate and undergraduate students. We expect some members of the campus community to have a larger number of contacts per day, while others will have smaller values.

Fourth, we especially appreciate the set of references provided by comment F16. We will discuss these and other related references in our addendum.

These references use data from a period before COVID-19 when measures like reduced classroom density, restricted access to dorms and limits on social gathering size were not in place. Moreover, not all forms of contact discussed in the literature necessarily constitute a risk of transmission as high as what we model.

In our reading of Wallinga et al., 2006, the number of contacts per day is lower than the comment’s stated value of “low 20s”. The paper studies a 1986 survey in the Netherlands that asked respondents for “the number of different persons they conversed with during a typical week, excluding household members.” In our reading of Table 1 in that paper, individuals aged 13-19 converse with an average of 8.2 other individuals per day, those aged 20-39 converse with 7.4, those aged 40-59 converse with 6.2, and those over the age of 60 converse with 4.1. Keeping in mind that our 8.3 contacts / day is meant to represent an average across the entire population, including faculty, staff, and students, the Wallinga et al. paper would seem to be roughly consistent after adding in household members. Their estimate would seem to be higher than ours after adjusting for a reduction in contacts due to social distancing.

Also, while interesting, Del Valle et al. 2007 is based on a simulation (using census data, business directory data, and national household transportation survey data) that lacks real-world information about room sizes and occupancies. The authors write, “Our simulation results (Robbins et al., 2006) show that the average degree (average number of contacts per person) is sensitive to variations in size of the sub-location.” We therefore assign less weight to this paper’s estimate of contacts / day than other papers.

Mossong et al. 2008 finds values of 18 for 15-19 year olds and in the range of 12-14 for each of the age groups between 20 and 60. Some of these contacts are less than 5 minutes in duration.

As stated in the comment, Leung et al. 2017 finds a number of contacts per day of 8.1 in Hong Kong, 13.4 in Europe and 18 in a previous survey in Hong Kong.

In Weeden & Cornwell 2020, as discussed above, the type of interaction studied does not necessarily correspond to a substantial risk of transmission.

We also note that the variation in contact rate across age groups in these papers is less dramatic than some may fear. For example, in both Wallinga et al. 2006 and Mossong et al. 2008 the contact rate in university-aged individuals is roughly 30% larger than that in the general population. This would seem to support our approach of taking a population-level R0 of 2.5, inflating it to account for the fact that our population has more young people than most, and then decreasing it by the same amount to account for the fact that we are enforcing social distancing measures.

In summary, we do not see the literature in F16 as providing clear evidence that 8.3 is an underestimate under social distancing conditions. It does, however, support an important point on which we all agree: the right number may be larger, and we should make sure we are ready for this possibility.

2. Responding to higher-than-anticipated transmission rate 

We have said previously that we can respond to higher-than modeled transmission rate through increased testing and targeted distancing / cleaning / compliance interventions.

We found particularly compelling the concerns about contact rates in residence halls. This concern about residence halls matches with the fact that high attack rates are found among individuals who live together in the same household, with low attack rates through other kinds of interaction. We will focus in particular on a scenario where contacts among those living together on a dorm floor are high. We will study the extent to which testing targeted to those individuals controls the spread of an epidemic in the addendum.

We similarly worry about contact rates in high-density off-campus housing, especially in fraternities and sororities.

One piece of intuition to keep in mind is that there is a period of roughly 2 days after exposure when new cases are not infectious. If we test frequently enough then infectious cases found will not have been infectious for long. Detecting a positive within 2 days of its becoming infectious followed by contact tracing and testing targeted to that individual’s dorm floor would allow us to isolate any secondary cases before they become infectious and create new tertiary cases. This could additionally be augmented by temporary extra cleaning and social distancing measures targeted to the dorm floor while additional testing is performed. This would dramatically reduce the spread through residence halls. Our understanding is that this strategy of frequent testing within dorms is also being pursued at Boston University. Restricting access so that only those that live on a dorm floor can gain access would complement this approach.

The paragraph above has been clarified to respond to F32. In this clarification we hope it is more clear that this paragraph does account for the fact that infections are modeled as undetectable in PCR following exposure for a random period of mean length 2 days. The above example does not quantify the effect of false negatives — it is intended only to give intuition and not to present a formal analysis. The addendum will provide this quantification.

3. Effectiveness of testing 

F9 writes, “We can respond initially by testing more often, as Peter wrote, but that only goes so far: the curves in Fig. 15 (infection rate vs. testing rate, above the nominal 20%/day) are a whole lot shallower than the ones in Fig. 11 in the relevant range (increasing R0 above 2.5 is equivalent to increasing transmission per contact above the nominal 2.6%).

These curves in Fig 15 are shallow because, with nominal parameters, testing does a good job of controlling infections before they spread, and most of the infections are due to our assumed rate of infections from the outside community.  (This is actually quite significant, in our nominal setting, perhaps pessimistically.) Thus, with the nominal parameters, once you get above a certain test frequency, there is nothing more for the testing to do. If we increase the transmission rate however, then we expect testing more frequently to provide significant value. We are studying this effect in the forthcoming addendum.

4.Testing in the virtual instruction setting + testing compliance in the residential setting 

We have analyzed both non-compliance in the residential setting and offering testing in the virtual instruction setting. This was done before the June 15 report was published but was not ready in time to be included in that published report. Some of the results in this analysis are available as a slide in the June 24 faculty senate meeting. We will include a full writeup of this analysis in the addendum.

One way to understand the impact of non-compliance in the residential setting is to look at Figure 15 in the full report, which shows results as a function of the percentage of the population tested each day. If non-compliance is distributed uniformly across the population, then failure to comply is equivalent to 100% compliance with less frequent testing. Using this reasoning, we can see that 71% compliance with 5-day testing in the residential setting is comparable to 100% compliance with 7-day testing with results pictured in Table 14 (1800 infections, 22 hospitalizations).

The analysis to be included in the addendum shows that if we achieve high compliance with once-per-5-day testing in the virtual instruction setting, then we do see fewer infections than the residential setting, with the breakeven point around 50% compliance. (A number of details will be discussed in the addendum.)

Although this falls outside of the realm of modeling, we comment briefly here on some practical reasons that testing with high compliance is difficult in the virtual instruction setting. We defer other legally-focused questions about our ability to mandate testing to University Counsel.

In the residential scenario, Cornell’s ability to ensure compliance with testing comes first from the ability to restrict access to physical property for those that miss testing, and also through an RA’s ability to talk directly to students who live in dorms. In the virtual instruction scenario, for students that aren’t using Cornell’s physical property, this first ability goes away.

Cornell can also restrict access to our electronic property, e.g., Cornell email, canvas, for those that aren’t tested. In the virtual instruction setting, achieving this second kind of control over electronic property requires knowing where a student is. This is because we can only expect those students in Ithaca to be tested.  Students elsewhere will not have access to testing, at least not on the cadence that we would require. We do not currently have a mechanism for determining where students are. While we could simply ask students to self-report their location, a student who has their email & canvas access revoked due to not being tested recently would plausibly be tempted to simply say they have left town as a way to regain access.

We emphasize that a residential student that has their netid & canvas access revoked because they were not tested would not have an easy way to get access back, other than to be tested.

5. Off-campus students are tested in the residential scenario 

F31 writes, “But in the reopen scenario the off campus students are also not included in the surveillance testing.”

Off-campus students are included in surveillance testing during the residential scenario in which Cornell reopens.

As described in the section above, one reason that it is easier to mandate testing for off-campus students in the reopen scenario is because Cornell will know students’ locations. This allows us to enforce testing simply by turning off email and canvas access for any enrolled student that does not comply. In the virtual instruction scenario, Cornell will not be able to reliably know whether a student has not been tested because they are living elsewhere or because they are living in Ithaca but misrepresenting their location.

In addition, as described in Appendix 4 of the C-TRO report, Cornell’s legal authority to mandate testing (and behavior changes) for off-campus students is broader in the residential scenario than for “students living off campus and taking classes online only”, even though the off-campus students would not reside on campus.

6. Number of students returning in the virtual instruction scenario

The June 15 modeling report studies what happens when 9,000 students come back to Ithaca, unmonitored, in a virtual instruction setting. Although there are no community comments here on this topic, questions in other fora have asked about the sensitivity of our results to this number.

To address this, we studied what happens if fewer than 9,000 students return. We presented this in a Cornell faculty senate meeting on June 24. Slide 6 of this slide deck shows the number of infections and hospitalizations as we vary (1) the number of unmonitored students returning and (2) contacts / day in the virtual instruction setting relative to the residential setting.

If virtual instruction is unable to reduce contacts / day despite reduced density (e.g., because of the reduced legal authority to mandate behavior called out in Appendix 4 of the C-TRO report), then it remains worse than residential instruction down to under 1,000 students.

If virtual instruction is able to reduce contacts / day by roughly 40%, then the breakeven point improves to 2,000 students. In other words, if more than 2,000 students return, then infections and hospitalizations are larger in the virtual instruction scenario.

Thus, if we assume that we would achieve a reduction in contacts / day somewhere between 0% and 40% relative to the residential setting, an appropriate threshold to examine is 2,000 unmonitored students returning. Although we think it is entirely plausible that fewer than 9,000 students would return, especially if Cornell, the City of Ithaca, and Tompkins County all strongly discouraged students from living in Ithaca, student survey results and conversations with local landlords suggest it would be substantially harder to achieve fewer than 2,000 unmonitored students returning.

If virtual instruction is able to reduce contacts / day by substantially more than 40% (i.e., achieve fewer than 3.32 close contacts / day among the unmonitored students), this corresponds to an assumption that the R0 among the unmonitored student population is below 1, or is close enough to 1 that contact tracing alone is sufficient to control epidemic growth, while the R0 in the monitored residential setting is 2.5. Under these assumptions, virtual instruction becomes safer than residential instruction even if more than 2,000 students return.

7. Modeling fatalities 

Unfortunately, fatalities may happen in any of the scenarios considered (residential instruction, virtual instruction) and in no way did we mean to imply otherwise.

We did not model them in the June 15 version of the report because they depend in a critical way on factors that are challenging to model, especially access to ventilators and the fraction of the infections in each age group that strike those with underlying health conditions. In particular, access to health care in Ithaca and the health of the population are both better than among the general population, especially for the time period on which most fatality data is based.

For these reasons, we felt that hospitalizations could be modeled more reliably than fatalities and would be a better guide as to the magnitude of the health consequences. In doing this, we meant for the reader to understand that any hospitalization can result in death. We believe that this was clear to members of the C-TRO committee when discussing the results, and we talked many times about measures to reduce mortality and morbidity (for example, pausing programs in which students volunteer in nursing homes).

In retrospect it would have been better to be more explicit about this in our modeling report. To get a sense for the relationship between hospitalizations and fatalities, refer to the CDC planning scenarios (https://www.cdc.gov/coronavirus/2019-ncov/hcp/planning-scenarios.html). Take the ratio between the Symptomatic Case Fatality Ratio and the Symptomatic Case Hospitalization Ratio to see that the fatality rate among those hospitalized is modeled by the CDC as being between 3% and 18% in their “best estimate” scenario. Thus, 16 hospitalizations in the nominal residential instruction setting would correspond to an expected number of fatalities between 0.5 and 2.9 fatalities. But again, please keep in mind that we have a significant amount of uncertainty about this number.

One of the comments refers to a sentence on page 32 that errantly suggested that fatalities would be 0. This referred only to our code, not to our beliefs. Our code does not model fatalities because we did not have time to add this feature and not adding them did not affect other results, but again should not be considered to imply that fatalities would not happen in reality.

It is a mischaracterization to claim that fatalities are not top of mind for the modeling team. Like all of those reading this comment, we are extremely concerned for the health and well-being of all members of the Cornell community. We ourselves are a part of this community, as are many of our dearest friends, neighbors, colleagues and family.

8. Racial and Ethnic Disparities 

It is unfortunately true that there are significant racial and ethnic disparities in the impact of COVID-19. We will investigate what data is available for modeling and then discuss either in the addendum or a follow-up comment. We also think it is important to assess disparities in quarantine and health impacts in ongoing monitoring that Cornell would do as part of a reactivation.

Racial and ethnic disparities in health outcomes and quarantine among Cornell students are likely to be different than in the overall U.S. population. Indeed, some of the most important drivers of these disparities in the overall population, such as lack of access to healthcare and an over-representation of racial and ethnic minorities among essential workers, are likely to drive outcomes less significantly in our student population.

9. Suggestion of pressure from university leadership 

One comment expresses a concern that the contacts / day parameter was changed because of pressure from university leadership. This concern is unfounded.

This comment is referring to section 7 of the June 15 report, which writes, “Based on feedback from President Brown and his team on the May 27 version, we reexamined the literature on asymptomatic rates, which led us to make a number of modifications to other parameters that significantly altered the results.”

President Bob Brown is the President of Boston University.  Here, “his team” refers to David Hamer, who is Professor of Global Health and Medicine at BU involved in BU’s plan for COVID-19 response. Thus, the interaction which led to the change to our parameters did not come from Cornell University leadership — it came from BU.  Indeed, Cornell University leadership was unaware that we were making this change and only became aware when we provided them with an updated report.

Moreover, the suggestion that Bob and David made was to increase the fraction of cases that are asymptomatic. We had been using a number based on early data from Italy and China which was subject to underreporting bias. They did not ask us to change the number of contacts per day. After reading the literature about asymptomatic rates, especially data from the CDC, we realized that there was an opportunity to improve other aspects of our parameters. We changed the number of contacts per day based on this reading, as explained in Section 7.

10. Effect of raising transmission rate equally in virtual and residential instruction settings 

F21 discusses what happens if we raise the transmission rate in both the virtual instruction and residential instruction setting.

First, we should clarify that the people infected in the virtual instruction scenario are not just the 9000 unmonitored students living in Collegetown but also the graduate students / faculty / staff living on campus who are being tested regularly (with full compliance). They become infected because they are subject to outside infections from Tompkins County, just like in the residential scenario. We do not model the infections from the unmonitored Collegetown students, but they would present a substantial risk.

Despite this clarification, what F21 says is true. If we hold the transmission rate equal in the two scenarios and raise it, then what we’ll see is that the fraction of the population infected grows in both settings. The fraction infected will be smaller  in the residential setting because of the testing in place, but if we bring the transmission rate high enough then the fraction infected will approach 100% in both settings. Then, because the overall population size is larger in the residential setting, nearly 100% of the population represents more people.

With this said, we currently believe that the transmission rates that we are most likely to face in reality are ones where interventions available to us in the residential setting will prevent having a high fraction of the population infected. The addendum will investigate further this critical question.

With regard to Provost Kotlikoff’s response regarding this in the town hall, we were not able to find that specific response and so we are unable to comment.

11.  Capacity in local hospitals 

The comment is mistaken about what the report states. The report estimates that there would be 16 people hospitalized during the fall semester. It does not comment on the local hospital’s capacity to hospitalize. This was top of mind for President Pollack and has been considered elsewhere in the C-TRO process, but the important task of understanding healthcare capacity is not part of the scope of the modeling report.

12. Framing of uncertainty 

F9 writes, “So I would hope to see messaging and planning that acknowledge the vast uncertainty. For example, the Executive Summary of the modeling report highlights numbers (percentiles of projected infection rates) that take no account of parameter uncertainty, which is very misleading.”

Regarding the executive summary of the modeling report, we included a discussion of parameter uncertainty in the executive summary.  See, for example, the bullet beginning “In all of our modeling results, modifying modeling parameters by only a modest amount from nominal values can result in substantially different numbers of infections and hospitalizations.”  About half of the bullets below this describe uncertainty, limitations, or assumptions in one way or another.  However, such bullets do not begin until the bottom of page 2 in the executive summary and we should have raised their prominence.

F30 points out that “some of the comments here about the ‘executive summary’ being misleading” may actually refer to the executive summary of the C-TRO report, rather than the modeling report. F9 does specifically refer to the “Executive Summary of the modeling report” and F5 offers general advice and context rather than criticism, but F30’s statement could be true for F2.

Regardless of comments’ intent, we agree that the executive summary of the main C-TRO report could have done more to articulate the full range of uncertainty in our modeling efforts. For example, in the often-cited phrase, “Paradoxically, the model predicts that not opening the campus for residential instruction could result in a greater number of infected individuals affiliated with Cornell”, the uncertainty conveyed by “could” is easy to miss next to the certainty suggested by the typographical emphasis applied to “not”. While we devoted substantial effort toward conveying caveats and nuance in our own modeling report, we could have devoted more effort to ensuring that more of this nuance was retained by translations and summarizations.

13. Crediting of experts in disease modeling 

F9 writes, “The report doesn’t credit Yrjo, Ivana, or anyone with prior human disease modeling expertise for any involvement before June.”

The report actually does credit them, by name in Section 6.  It also credits them indirectly in the executive summary, writing “This effort is supported by … a set of reviews provided by experts both within and outside Cornell on a previous version of this report.”  However, we should have credited them in a more prominent place.

14.  Impact on Tompkins County 

This is an excellent point.

Qualitatively, we think that the impact on Tompkins County will be mild in the nominal residential setting: gateway testing is quite effective at identifying cases before they infect others, and then asymptomatic screening is able to isolate the infectious cases we miss quite quickly. A substantial number of the 1254 cases in this setting are actually due to infections originating outside of Cornell.

However, it is also possible that a small number of cases that leak out of Cornell could grow substantially in other communities that lack access to regular testing. This could cause significant negative effects in those communities, and the infections created there could also come back and reinfect Cornell students, staff and faculty.

Quantitatively, we have only recently developed the capability to study this, through a new simulation framework that can simulate multiple groups of people. We plan to use this to study the question you ask. It likely will not be in the first version of the addendum, but we do plan to include it in a follow-up version.

In the nominal virtual instruction scenario, although we do not model it the negative effect on Tompkins County is likely to be quite substantial.

15. Impact of students arriving early because we would start Sep 2

This is also a good point.  Rather than modeling all of the students as arriving on Aug 27, it may be better to model them as gradually arriving. It is also important (critical, even) to model our gateway testing procedure for these students. We will investigate what our plans are here for testing before conducting any modeling work. We will provide an update either as a comment here or in a follow-up to the addendum.

Even without modeling, we think that having a significant number of people returning from afar without having gateway testing in place can be quite dangerous. The sooner we can have the mechanisms in place to include all returning off-campus students in our asymptomatic testing protocol, the better.

16. Number of cases missed in gateway testing 

F32 writes that “a back of the envelope computation shows that this two 2 undetectable period will lead to about 10 case of infected students which will not be detected by the rigorous testing during the move-in days. This number grows to 15-20 if one includes the the test has false negatives.”

Our analysis of gateway testing includes both (1) cases missed because they are too soon after exposure to be detectable in PCR; and (2) cases missed due to false negatives.

This is discussed in detail in the full modeling report in Section 2.7 and 3.1 with the spreadsheet used to do the analysis available for examination from a link provided in Section 3.1. In that spreadsheet, people who are in the undetectable state and thus missed in gateway testing are in cell E28. This is 0.07% of the population of returning students. False negatives contribute an additional 0.02% of the population that are infectious and missed in gateway testing.