An imperfect tool: contact tracing could provide valuable reductions in COVID-19 transmission if good adherence can be achieved and maintained.

Emerging evidence suggests that contact tracing has had limited success in 2 the UK in reducing the R number across the COVID-19 pandemic. We 3 investigate potential pitfalls and areas for improvement by extending an ex- 4 isting branching process contact tracing model, adding diagnostic testing 5 and reﬁning parameter estimates. Our results demonstrate that reporting 6 and adherence are the most important predictors of programme impact but 7 tracing coverage and speed plus diagnostic sensitivity also play an important 8 role. We conclude that well-implemented contact tracing could bring small 9 but potentially important beneﬁts to controlling and preventing outbreaks, 10 providing up to a 15% reduction in R , and reaﬃrm that contact tracing is 11 not currently appropriate as the sole control measure. 12

rapidly exceeded and stricter physical distancing measures required [19]. 50 Extending Hellewell et al.'s [3] UK-focused contact tracing study with 51 new insights could inform contact tracing strategy. Their key conclusion was 52 that highly effective contact tracing would be sufficient to control an initial 53 outbreak of COVID-19 in the UK, however subsequent evidence supports 54 much higher pre-and asymptomatic transmission rates than had initially 55 been considered. In particular, the original analysis only considered scenar-56 ios with 0-10% of cases as asymptomatic and 0-30% of transmission from 57 symptomatic individuals occurring pre-symptoms, compared to more recent 58 estimates of [31][32][33][34][35][36][37][38][39][40][41][42].5% [20,21,22] and 44% [23] respectively. The focus on 59 speed in the UK contact tracing programme also requires a detailed assess-60 ment of the associated trade-offs through mechanistic modelling of the testing 61 process. Up-to-date modelling studies are therefore needed to investigate the 62 feasibility of contact tracing and the conditions under which it is effective. 63 We use improved incubation period and serial interval estimates [24, 23, 64 25], consider imperfect self-reporting, adherence and tracing rates and simu-65 late the use of diagnostic tests. We explore trade-offs between testing speed 66 and sensitivity, and investigate the limitations of contact tracing and in which 67 scenarios it is likely to be most effective. 68

69
The underlying assumptions around the contact tracing logistics that have 70 been modelled and presented here are described in further detail in the Meth-71 ods; Figure 5 provides a detailed schematic of the contact tracing process. 72 Efficacy of contact tracing 73 We used data from a UK-based survey to consider three adherence scenarios 74 [12]: a scenario with low self-reporting and poor adherence to isolation (rep-75 resentative of the proportion of individuals who reported being fully adherent 76 to advice from Test and Trace); a scenario with good reporting and adher-77 ence (representing those intending to be adherent to advice); and a scenario 78 with good reporting and boosted adherence (where additional incentives to 79 adhere to isolation are introduced). CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.  CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 19, 2021. presented, aside from those with poor reporting and adherence (left column).
In scenarios with average or boosted adherence for a fixed number of total 119 cases so far, speeding up tracing reduces the probability of a large outbreak 120 and increases the relative benefit of higher contact tracing coverage.

122
Our results provide insights into why contact tracing implementation has not 123 been as effective in reducing transmission as initially hoped in the UK. We 124 conclude that the largest likely factor is adherence to the various stages of 125 contact tracing and isolation, which is believed to be relatively poor [12]. 126 However, if reasonable reporting and adherence can be achieved, then con-  At this stage of the pandemic, there are more diagnostic tools available 142 than in the initial months. In particular, use of rapid lateral flow device 143 (LFD) tests is growing due to increased speed and reduced costs compared 144 to PCR alternatives. However, our results suggest that test sensitivity is still 145 important, with contact tracing using a 95% sensitive test that takes two 146 additional days performing better than an instantaneous 65% sensitive test.

147
This effect is seen in the presence of an assumption that negatively-testing 148 individuals are not immediately released from isolation guidance but if this 149 was not the case we would expect to see an even more marked reduction 150 in contact tracing efficacy when considering the faster lower-sensitivity test.  CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 19, 2021. Our model assumes constant test sensitivity across an individual's infected period, whereas a previous study shows that testing too early or late 155 after exposure can dramatically increase false negative rates [26]. While 156 assuming a fixed incubation period of 5 days, we have ignored temporal vari-157 ation. Additionally, high between-person variance has been observed in the 158 natural history of infection [23]. It is therefore unclear what drives these 159 temporal changes in sensitivity or whether this temporal profile makes sense 160 on an individual basis. These simplifying assumptions mean we may be 161 over-estimating operational test sensitivity in some cases, leading to more 162 optimistic results around the impact of contact tracing. This reinforces the 163 conclusion that contact tracing is not currently appropriate as the sole control 164 measure. 165 We also assume the Negative Binomial dispersion, k = 0·23 [27], of sec-166 ondary cases, does not vary with R S due to different social distancing mea-167 sures. This relationship is poorly characterised, but it is believed that social 168 distancing may increase k, leading to decreased heterogeneity in number of 169 contacts across the majority of the population due to an overall reduction 170 in mean contacts, paradoxically making outbreak control harder, although 171 this effect is expected to be cancelled out by the reduction in the mean [28].

172
Furthermore it is also possible that less heterogeneity in contacts may make 173 tracing of individual contacts more feasible, allowing for a higher coverage.

174
The vaccine roll-out is currently in progress in the UK, with over 50% 175 now believed to have COVID-19 antibodies through either vaccination or 176 prior infection [29]. This will have the effect of reducing R and, eventually,  Overall, we conclude that well-implemented contact tracing could bring CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.  Table 1.
CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.  CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 19, 2021. ; https://doi.org/10.1101/2020.06.09.20124008 doi: medRxiv preprint stantially lower that the intentional adherence reported by individuals who 261 had not yet developed symptoms or been traced, which was taken as a 'good' 262 compliance scenario: 70% individuals said they would isolate following symp-263 toms; 40-50% would self-report following onset; and 65% would isolate for 264 the full duration if contacted by Test and Trace. We also considered a sce-265 nario with a boosted adherence to tracing of 90%. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

Simulation process
Results presented are the combined output of 5,000 simulations for each pa-298 rameter combination or scenario, and each simulation is run for a maximum 299 of 300 days. These results are used to derive the probability of a large out-300 break given a range of conditions. A large outbreak is defined at 2,000 cases: 301 this threshold was chosen from experimental runs with a maximum of 5,000 302 cases and noting which of the simulated epidemics went extinct; 99% of ex-303 tinctions occurred before reaching 2,000 cases. The model was written in R 304 with pair code review and unit tests [39]. The code is available from a public 305 GitHub repository (github.com/timcdlucas/ringbp). . This project has received funding from the European Union's Horizon 2020 research and innovation programme -project EpiPose (101003688: PK). Royal Society (RP/EA/180004: PK). Wellcome Trust (210758/Z/18/Z: JH, SA). Views, opinions, assumptions or any other information set out in this article should not be attributed to BMGF or any person connected with them. TC is funded by a Sir Henry Wellcome Fellowship from the Wellcome Trust (215919/Z/19/Z). TMP's PhD is supported by the Engineering & Physical Sciences Research Council, Medical Research Council and University of Warwick (EP/L015374/1) and thanks Big Data Institute for hosting him. All funders had no role in the study design, collection, analysis, interpretation of data, writing of the report, or decision to submit the manuscript for publication.

9
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Code Availability
The code used in this study is available in Zenodo with the DOI: 10.5281/zenodo.4752369.

Data Availability
The data that support the findings of this study are available in Zenodo with the DOI: 10.5281/zenodo.4752369.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 19, 2021. ; https://doi.org/10.1101/2020.06.09.20124008 doi: medRxiv preprint . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 19, 2021

15
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 19, 2021.

17
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

18
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 19, 2021. ; https://doi.org/10.1101/2020.06.09.20124008 doi: medRxiv preprint Figure 4: Outbreak thresholds. Probability of a large outbreak (>2000 cases) by total number of cases so far (observed and unobserved). Sensitivity = 95%, self-reporting proportion = 50%, time to test from isolation = 1 day. Error windows: 95% confidence intervals from output variation of 5,000 simulations.

19
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 19, 2021. ; https://doi.org/10.1101/2020.06.09.20124008 doi: medRxiv preprint Figure 5: Contact tracing schematic. Overview of the contact tracing process implemented in our model. Person A isolates and self-reports to the contact tracing programme with some delay after symptom onset, by which time they have infected Persons B, C and D. When Person A self-reports they isolated and are tested, a positive result initiates contact tracing. Person B was infected by A prior to their symptom onset and is detected by tracing after some delay. After isolating they are tested, with a false negative result. This leads to B either a) stopping isolation immediately or b) finishing a minimum 7 day isolation period. Both may allow new onward transmission. Person C was infected by A but not traced as a contact. Person C does not develop symptoms but is infectious, leading to missed transmission. Person D is traced and tested. The test for D returns positive, meaning that D remains isolated, halting this chain of transmission.

20
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 19, 2021. ; https://doi.org/10.1101/2020.06.09.20124008 doi: medRxiv preprint