Updates from the Modeling Team

The modeling team would like to thank the Cornell community for the sincerity and energy put into providing comments on our work. The webpage provides links to our reports  and updates.

Modeling Update for Gateway Testing  (August 5)

We provide this updated modeling analysis of gateway testing prompted by three recent events: COVID-19 prevalence has risen in parts of the U.S. since we wrote our report  in June; a potential lack of test access for some Cornell students in their home location; and the recently instituted requirement that people coming to NY State from some high-prevalence are as must self-quarantine upon arrival regardless of test results.

Addendum (July 17)

This provides an in-depth analysis of some of the questions posed in response to the June 15 Report. In particular, the addendum studies (a) alternative methodologies for estimating contacts / day and transmission rate, (b) the effectiveness of increased test frequency in mitigating the effect of higher-than-modeled contacts / day, (c) the effect of non-compliance with testing, and (d) the effect of offering testing to virtual instruction students.

Original Report (June 15)

Comments on the June 15 report  (and responses) are given below. Comments  are indexed for easy referral.

Issue (Last Updated July 3)


Contacts-Per-Day F28, F19, F16, F14, F10, F8, F7, F6, F1
Higher-than-anticipated transmission rate  F32, , F19, F16, F14, F10, F9, F8, F7, F6, F1
Effectiveness of Testing F9
Testing in the virtual instruction setting + testing compliance in the residential setting F29, F27, F26, F25#1, F17, F11
Off-campus students are tested in the residential scenario F31, F27A
Number of Students Returning in the Virtual Instruction Scenario
Modeling fatalities F29, F13, F12
Racial and Ethnic Disparities F25#2, F24, F23
Pressure from university leadership F18
Effect of raising transmission rate equally in virtual and residential instruction settings F21
Capacity in local hospitals F22
Framing of uncertainty F30, F9, F5, F2
Crediting of experts in disease modeling F9, F1
Impact on Tompkins County F20
Impact of students arriving early because we would start Sep 2 F15
Number of cases missed in gateway testing F32
Note On Data
Data is central to the scientific approach that is being taken. Two surveys taken early in the summer had a big impact on the university’s approach to F20.The Student survey gets at residential vs online instruction and the likelihood of coming to campus/Ithaca for the fall semester. The Faculty survey that gets at in-person vs online instruction and risk assessment of being on campus.


Print Friendly, PDF & Email

46 thoughts on “Updates from the Modeling Team

  1. While I appreciate all of the hard work that went into this report, like many others, I am deeply concerned about the assumptions of the model. By not modeling 50 or 75% compliance and comparing it to 100% compliance in the residential scenario, this report essentially denies us the opportunity to determine what level of compliance would be necessary to limit risk. Maybe 80% compliance is good enough, maybe it’s not – but we can’t know from what’s here. It is also absurd to assume 0 fatalities, based on everything we know about COVID-19 and our local Cornell community. How many members of the Cornell community, exactly, are the president and provost willing to let die to ensure a residential semester? How might different plans affect that number of deaths? Again, we’re denied the option to even explore the impacts – any discussion of “risk” that pretends fatality is not on the table is unrealistic at best, deliberately misleading at worst. Making decisions based on this model, given these limitations, is deeply concerning.

  2. “The average student has the potential to share a classroom with about 529 different students by the time they have completed one round of courses on their schedule (5,832,358*2 bi-directed arcs[“edges”]/22,051 students = 529). This does not mean that the average student will share a classroom with 529 other classmates, given attendance is rarely 100 percent. On the other hand, most classes meet more than one time per week, giving each student multiple opportunities to be in the same room…” (228-29).

    Kim A. Weeden and Benjamin Cornwell, “The Small-World Network of College Classes: Implications for Epidemic Spread on a University Campus,” Sociological Science Vol. 7 (May 2020): 222-41.

    Data like this seems to challenge the Frazier model, which assumes no increase in social contacts or R0 in a reopened campus. However, Weeden and Cornwell were not calculating contacts in a socially distanced classroom. Would be useful to get Weeden, Cornwell, and Frazier, among others, in conversation about social networks and epidemic spread in a reopened campus this fall. It seems no one on the university planning committees was tasked with examining this set of problems. Startling oversight.

  3. I feel like the university is engaging in double-speak regarding social messaging. First, they ask us to trust students to engage in safe and responsible social distancing throughout the semester if we reopen. They tell us they’re launching a campaign of social messaging to ensure students behave safely.

    Yet in the Frazier model, the assumption is that students living off-campus in a no-reopen scenario will not and cannot be expected to behave safely. The testing and quarantine assumed to be 100% effective if we reopen campus is not even an option; we don’t even try. Kotlikoff justified this abdication of responsibility in the town hall yesterday by saying it would harder to “enforce” testing among students not living in dorms in the no-reopen scenario. In other words, he pinned the blame on the students for not complying with a program he’s actually declined to offer them, at least in the hypothetical models under discussion.

    The notion that we can and should lock students out of their dorms to enforce testing if we reopen is a little strange. So if they miss a test, they sleep in the snow? It’s also breathtakingly naive as “enforcement” technique imagined to be highly effective. It’s quite easy to get into a locked building subject to heavy foot traffic — like a dorm — by waiting 30 seconds until somebody else opens the door. Or by calling a friend to let you in. Students will not find it difficult to get around this enforcement technique if and when they neglect to get tested on time or, as seems likely in N cases, when they choose to avoid testing, perhaps because they don’t want to be quarantined. The model does not account for any of this, just assuming “enforcement” on-campus through dorm lockouts is 100% effective. In reality, the most efficacious enforcement mechanism will likely be the same for residental as for non-residential students: shutting down their netID.

    Even absent a social messaging campaign, surely some students living off campus in a no-reopen scenario will want access to tests. Some — perhaps many — would also have an interest in having their roommates tested and provided a safe space in which to quarantine if needed. Why don’t we try surveying students about their interest? At the very least, we should not take it for granted that off-campus students in a no-reopen scenario would never participate and therefore plan to give them no access to tests or quarantine spaces, as in the model.

    A culture of conformity to public health surveillance is not impossible to achieve, although it would surely be less that 100% effective — just as on-campus compliance will be less than 100%.

    1. I’m wondering about the students living off campus in the re-open scenario. Do they simply disappear? Or does their unsafe behavior no longer matter?

  4. The report assumes 100% compliance with testing and quarantine in the reopen scenario, and zero effort to test or quarantine non-residential students (compliance not even an option) in the no-reopen scenario. The former is impossible and the latter unnecessary in the real world, so it would be fruitful to compare models a little less like apples and oranges. What conclusions do we draw, for example, if we have 80% compliance in the reopen scenario and but make a non-zero effort to test non-residential students in the no-reopen scenario? The model as it stands favors reopening, but seems predetermined to do so.

  5. First, let me say I appreciate that the university is doing this sort of work ahead of making decisions about reopening. Following the news of other universities that say they will reopen and figure out the details later has been disheartening. Engaging campus in this way, using the expertise on campus, etc. – I really appreciate this.

    1. The report considers the following scenarios: fall-reopen, no-reopen-students, no-reopen-faculty/staff. The overview says that:

    “Importantly, not reopening actually results in more infections and hospitalizations than reopening. This arises because the students returning to campus in the reopen scenario undergo test-on-return and asymptomatic surveillance, which controls epidemic spread within this population. In contrast, in the no-reopen scenario, returning students are not subject to the University’s
    asymptomatic surveillance testing protocols, allowing infections to grow rapidly in this group.” (p. 12)

    The residential instruction/no-reopen scenarios seem like two extreme ends of a spectrum of strategies the university could adopt. Why not consider a no-reopen scenario with asymptomatic surveillance for students who return to Ithaca? For example, what does the number of cases look like if the university screens students before their return to campus? Or what does the number of cases look like if the university screens students who have returned to Ithaca every 7 days for the first month of return? I understand this probably depends on student compliance but it seems at least worth exploring. Put another way: it seems ludicrous for the university to say “not my problem” if students choose to come back to Ithaca even if residential instruction doesn’t resume.

    2. The other question I have about the report is whether it is possible to consider how infections are likely to affect student populations? There are racial and ethnic disparities in the impacts of COVID-19 [e.g. https://www.cdc.gov/coronavirus/2019-ncov/need-extra-precautions/racial-ethnic-minorities.html, https://covidtracking.com/race%5D. How will this translate to students? For example, the report could begin by considering the racial and ethnic makeup of students who live on/off campus and assessing whether this changes the likelihood that particular student populations require quarantine.

    I think any reopening plan *MUST* consider this. If the models are ‘blind’ to this issue then I urge them to be explicit about that assumption and include it in the summary.

    1. You raise an important point about the dangers of adopting a “race-blind” health model that will disparately impact minority communities at Cornell. There are medical but also legal risks here.

      The second link (covidtracking.com/race) seems to be broken.

    2. You’re right to flag the racial and ethnic disparities. We should be talking about this.

      We should also be talking about the fact that many of the reasons the CDC gives for higher risk for COVID-19 among minorities will apply to all college students in a residential semester. These risks (sourced from the CDC website linked above) include:

      “more likely to live in densely populated” spaces
      “may lack safe and reliable transportation”
      “more likely to rely on public transportation, which may make it challenging to practice social distancing”
      “where people live, work, eat, study, and recreate within congregate environments … it [is] difficult to slow the spread of COVID-19”

      Sounds like college students during any semester in Ithaca.

  6. The report says that the local hospitals can only hospitalize 16 patients only, meaning the vast majority of students 1000+ will need to fight COVID-19 off on their own. It saddens me on thinking that this will be the situation for infected students in the fall. Also, does Cornell have exclusive access to these 16 beds? It does not seem likely. There is also Ithaca College and the local community.

  7. Much of the discussion (here and in Zooms) about the case loads under residential vs. virtual instruction focus on the likelihood of a higher R0 ensuing if the campus re-opens for instruction. I would like to note that a higher R0 than the “typical community” value assumed in the modeling report, even if it is set to *exactly* the same value in both residential and virtual scenarios, could substantially change the balance between residential and virtual instruction.

    Here’s why. Under nominal parameters, in the virtual instruction scenario there are 9000 students living off-campus and unsupervised, and 7200 of them get the virus. Whatever you do to the value of R0 in the model, you can’t infect more than all of them. In the residential scenario with supervised students, there’s a lot more room for higher R0 to increase the cumulative case load.

    This observation is relevant to a question asked today in the Q&A box at the Town Hall about reopening, and to the answer that was given by Mike Kotlikoff. My answer (above) differs from what Mike said about the effect on projections if R0 were increased equally under both scenarios.

  8. The model does a good job examining the effects of Covid within the Cornell community, highlighting that there will be fewer cases with residential instruction than with online. I am concerned, however, about the impact of infections within the Cornell community on Ithaca and Tompkins County residents. Under one scenario, for instance, there is a prediction of 1254 infections among the Cornell community. What is the impact of those infections on Ithaca and Tompkins County residents? If we are concerned about the public health impact on the larger community, it would be good to know and that. Moreover, this is important information for the city, town, and county to have in making their input on Cornell re-opening.

  9. Here’s a datapoint. After closing campus and moving courses online, my average contacts per person per day fell to 1. Before that, my average was probably 50.

    The assumption that reopening campus won’t rapidly increase contacts and therefore secondary infections is absurd.

  10. The Frazier report specifies that an earlier version (as of May 27-28) of the model assumed a higher number of contacts per person per day, a key parameter. But after intervention by university leadership, for reasons undisclosed, the contacts per person per day was lowered to 8.3. What’s the justification for using this lower figure? What was the parameter prior to that intervention and what was the justification for that number?

    Page 50 of the report reveals that “the May 31 parameters predict significantly fewer infections than the May 27 and 28 parameters” as well as fewer hospitalizations. This is worrisome if key parameters were modified to less accurately represent reality. Can we see the earlier report?

  11. From page 2: “Also note (4) our nominal scenario for residential instruction assumes full compliance with testing, quarantine and isolation.” Does this imply that full compliance is not assumed in the case of virtual instruction?

  12. Thank you for what is obviously a lot of work for this report. However, as others, I have to express dismay and incredulity at the key in-built assumption of the 8.3 contacts per day for each non-quarantined/isolated person (p.19) and very especially for a population largely comprising young adults. This assumed contact level may “cause our implied R0 (calculated in Section 2.5) to match the nominal value of 2.5 recommended by the CDC” (p.19, last 2 lines), but it is below the level identified in a March 27-April 9 Gallup survey which found in the US that, even under active social distancing during a period when this was being widely pursued and social contacts were dramatically down on ‘normal’, adults in general (9.9 mean contacts per day) or working adults (13.9 mean contacts per day) were both >8.3 contacts per day. These two are both categories which would much better describe college students versus ‘adults not working’, the one category under 8.3 in the survey (https://news.gallup.com/opinion/gallup/308444/americans-social-contacts-during-covid-pandemic.aspx).

    Most studies in fact find that young adults are at the busier end of social contacts per day numbers (e.g. Mossong et al. 2008). For example, studies from the pre-COVID-19 era suggest much higher contacts per day (versus 8.3) as typical for the college-age group, for example in the low 20s (e.g. Wallinga et al. 2006; Del Valle et al. 2007). Data does of course vary. Whereas previous studies found averages of 13.4 contacts per person (all ages) per day in Europe and 18 for Hong Kong, another study did get a figure of 8.1 (see Leung et al. 2017). Regardless, two findings stand out in the Leung et al. (2017) paper: (i) “The contact intensity was highest among school-aged children aged below 20”, and (ii) the data were strongly right-skewed (some of the sample reported large numbers of contacts)—and even a small but substantive such super-contacter/spreader group undermines modelling based on low numbers of contacts.

    The campus context and the issue of repeated contacts all exacerbate such a situation as your ref. [37] by Weeden & Cornwell 2020, based on Cornell data, makes apparent.

    Student housing (all forms), any common dining facilities (however social distanced), and campus-life (however social distanced) all aggregate contacts and probabilities (contrary lack of such assumptions in this study)–physical changes to dorm structures (double to single, pod model, mentioned p.49, but not included in the modelling could of course be important). There is further the unproven assumption that a large set of 18-21 year olds will all behave as model, disciplined agents. This seems unlikely from any perspective across the range of accommodation scenarios in question. The past period has indicated many older adults do not. Further, however, important COVID-19 and social distancing may be as a recognized aim for person and society, other fundamental concerns and issues will potentially take precedence—as recent very justified waves of protest around BLM illustrate—potentially dramatically changing vector relations on campus from the modelled scenarios.

    For all these reasons I struggle not to find this study fundamentally flawed, and potentially a dangerous basis on which to make decisions for the Cornell community.


    Del Valle, S.Y. et al. 2007. Social Networks 29: 539-554.

    Leung, K. et al. 2017. Scientific Reports. 7: 7974.

    Mossong, J. et al. 2008. PLoS Medicine 5(3): e74. https://doi.org/10.1371/journal.pmed.0050074

    Wallinga, J. et al. American Journal of Epidemiology 164: 936-944

    Weeden, K. A. & Cornwell, B. 2020. Sociological Science 7: 222-241

  13. Let’s say we have 9000 off-campus students in town for N days waiting for residential classes to start. Shouldn’t the no-reopen model be applied to predict what happens during those N days? The reason I am asking is that the C-TRO calendar starts Sept 2, 6 days later than the originally planned Aug 27 start date. Considering apartment leases and other pressures for soon-as-possible move in, it seems that we are encouraging a larger N with the delayed start.

    Thus, I would like to see what the model says about a Aug 27 start vs a Sept 2 start.

  14. The modeling assumes that an online semester would preclude asymptomatic testing of Ithaca-based students; why can’t we test all Cornell students in Ithaca regardless of whether they’re taking online courses?

  15. Please produce a table that has columns associated with duration of class (using T = 50min, 75min, 120min) and whose rows specifiy class size (N = 10, 20, 30, 40, 50). The table entries should be how many contacts that counts for. E.g., I am in a 50 minute class with 20 classmates and that adds 5 contacts to my daily contact total.

    Make reasonable assumptions about the teaching space, e.g., reasonable airflow, the room has Nx100 square feet etc.

    To the layperson, 8.3 contacts per day is shockingly low. Help us reason about safe teaching spaces and contacts!

  16. Thanks again to the modeling group for their enormous efforts. Without your work our discussion would be entirely based on intuition and guesses.

    Yesterday’s Senate meeting and the “hallway discussion” clarified some things. The report doesn’t credit Yrjo, Ivana, or anyone with prior human disease modeling expertise for any involvement before June. It’s good to hear that they played a larger and earlier role.

    However, the discussions also confirmed that the model’s projections about the safety of residential instruction, in absolute terms and relative to virtual instruction, rest on the assumption that campus will not be an atypically hospitable environment for disease spread. Peter said that the break-even point for residential vs. virtual in the model is at twice the CDC estimate for typical/average situations, and beyond that the report shows that residential instruction quickly gets a whole lot worse. I, and many others, think that re-opened colleges campuses are likely to be very atypical hot-spots for viral spread. Peter and the modeling team evidently disagree, and they’re a pretty sharp group.

    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. Re-opening (if it happens) might go very smoothly. But I think the model implies a real potential for uncontrolled outbreak and we need to be planning how to shut down fast if we need to, with minimal spillover to the Ithaca community. 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%).

  17. In Figure 11 on page 38, where they graph “average contacts per person per day” against “percentage of population infected” and you do not even consider the possibility of over ~13 contacts per day, by which point things are already pretty bad. Please explain.

  18. What’s the empirical justification for a model based on 8.3 contacts per day?

    I gather Frazier’s team imports this figure from the CDC, but the CDC isn’t modeling what we’re trying to model. They’re looking at broad populations; we’re looking at a residential college. Seems dangerously misleading to take abstract inferences from the former generic analyses and apply them as “facts” to our specific situation. Instead we should attempt to estimate the actual contacts per person per day if we reactivate campus with in-person classes, dorms and dining halls. We should not assume students never socialize or party.

    We should model infections and our capacity to control them based on our best estimates of conditions that will actually obtain, then reopen this conversation.

  19. Hi Peter, thanks for all the work you and your team have put into this.

    Re: your comment on what residential contact rate would cause greater infection than virtual instruction, you say “this would be predicted to occur if the daily contact rate increased to roughly double what we expect.” Does that mean twice 8.3 = 16.6 contacts per person per day?

    You further state that “If the contact rate is much larger than expected in the fall in the residential scenario … the larger-than-expected contact rate is likely to have a specific cause into which we would have visibility via contact tracing.” Perhaps in-person classes would be the specific cause. Even assuming we eliminate in-person instruction for classes in the hundreds, we routinely gather dozens of students and faculty together in classrooms for hours per day. Strangely, this fact doesn’t appear to factor into your model of contagion in a residential semester.

  20. Thanks for your questions.

    Experts on infectious disease modeling were consulted, as described in section 6 of the full report. This included Yrjo Grohn and Renata Ivanek. They both provided detailed reviews of the modeling report.

    In regard to the question of daily contact rate on campus, Figure 11 on page 38 shows sensitivity of the number of infections and hospitalizations to this parameter, holding the test frequency and social distancing measures fixed. More contacts / day increases infections and hospitalizations, with an inflection point near double the nominal value. This is roughly the daily contact rate at which we would lose control of the epidemic if we took no action, although as discussed below we would see this happening and would respond.

    I understand part of your question as asking how much the residential contact rate would need to increase to make that scenario on par or worse than virtual instruction. The median number of infections overall in the virtual instruction scenario is 7200, which is ~20% of the full population of 34K. From Figure 11 under the nominal parameters, we see that this would be predicted to occur if the daily contact rate increased to roughly double what we expect, which would correspond to an R0 of 5.

    If the contact rate is much larger than expected in the fall in the residential scenario, we’ll see this in the asymptomatic screening results and also through contact tracing. We would then take action — the larger-than-expected contact rate is likely to have a specific cause into which we would have visibility via contact tracing. If this cause is addressable, we would do that. If not, we would also have the ability to fall back on increased test frequency. Figure 15 shows the effectiveness of increased test frequency.

    Although it isn’t emphasized, this visibility and the opportunity to respond that it provides is an important capability provided by asymptomatic screening. In a virtual instruction scenario where we don’t do screening, we risk not knowing until later that an epidemic is underfoot.

    For reference, here is a link to the full report:

    -Peter Frazier

    1. Thank you for posting the full report. Especially important since the executive summary is pretty misleading about the parameters used and therefore conclusions one should draw from the report.

  21. The “paradoxical” finding that residential instruction will decrease the total number of cases among students is based on scenarios (p. 44 of the full report) where it is assumed that living on campus and mixing with other students in classes, dorms, dining, etc., will not cause any increase in the daily contact rate or expand students’ contact networks. Is that assumption reasonable? Given that assumption, the finding is almost unavoidable. It’s hard to predict how much contact rates might increase among students on campus and in dorms. How much of an increase would it take to offset the benefits of testing and contact tracing, or to overwhelm the system so that the testing/tracing/isolation plans become infeasible?

    Experts on infectious disease modeling (several from the Cornell Vet School, and one external expert on human disease modeling) were only consulted after-the-fact (May 31), too late for their inputs to affect the Committee reports or recommendations. Why was it decided to do that, instead of getting some of them involved from the start?

Leave a Reply to F28 Cancel reply

Your email address will not be published.