Effective contact tracing for COVID-19: A systematic review

Contact tracing is commonly recommended to control outbreaks of COVID-19, but its effectiveness is unclear. Following PRISMA guidelines, we searched four databases using a range of terms related to contact tracing effectiveness for COVID-19. We found 343 papers; 32 were included. All were observational or modelling studies. Observational studies (n = 14) provided consistent, very-low certainty evidence that contact tracing (alone or in combination with other interventions) was associated with better control of COVID-19 (e.g. in Hong Kong, only 1084 cases and four deaths were recorded in the first 4.5 months of the pandemic). Modelling studies (n = 18) provided consistent, high-certainty evidence that under assumptions of prompt and thorough tracing with effective quarantines, contact tracing could stop the spread of COVID-19 (e.g. by reducing the reproduction number from 2.2 to 0.57). A cautious interpretation indicates that to stop the spread of COVID-19, public health practitioners have 2–3 days from the time a new case develops symptoms to isolate the case and quarantine at least 80% of its contacts.


Introduction
Late in 2019, the first outbreak of coronavirus disease (COVID- 19) was discovered in Wuhan, China. On March 11, 2020, the World Health Organization (WHO) characterised COVID-19 as a pandemic. In April and May 2020, it published guidelines on public health control measures, including contact tracing [1].
Contact tracing involves identifying, notifying, and quarantining people who have had close contact with new cases in order to prevent further transmission within the community [1]. Prior to COVID-19, it had been used for other infectious diseases, such as HIV, tuberculosis, Ebola, SARS, and H1N1 influenza, for which it was estimated 4363 times more cost-effective than school closure ($2260 vs. $9,860,000 per death prevented) [2]. In May 2020, initial WHO guidelines for contact tracing were: "At least 80% of new cases have their close contacts traced and in quarantine within 72 hours of case confirmation." [1] However, some studies already suggested that tracing needed to be more thorough and prompt to be effective [3,4]. WHO guidelines did not cite peer-reviewed evidence.
The US and European Centres for Disease Control and Prevention (CDC) also recommended contact tracing, but offered seemingly conflicting advice in the face of widespread transmission, when thousands of daily new contacts must be traced. Indeed, in May 2020, the US CDC wrote: "When a jurisdiction does not have the capacity to investigate a majority of its new COVID-19 cases, [it] should consider suspending or scaling down contact tracing." [5] In contrast, in April 2020, the European CDC advised: "Contact tracing should still be considered in areas of more widespread transmission, wherever possible, and in conjunction with physical distancing measures." [6,7] This apparent contradiction and lack of evidence in official recommendations may have left contact tracers wondering which guidelines to follow, and how effective tracing was. Case in point: the UK spent ten billion pounds on its test and trace programme, which may not have been effective due to limited case detection and compliance in naming contacts [8]. Therefore, in this systematic review, we aimed to examine contact tracing effectiveness and to identify characteristics of effective tracing efforts.

Methods
We prepared this systematic review in accordance with PRISMA guidelines [9]. Its protocol was registered with PROSPERO (CRD42020198462).

Eligibility criteria
All studies evaluating the effectiveness of contact tracing efforts in the community (alone or in combination with other interventions) were included. Effectiveness was defined as stopping or slowing the spread of COVID-19, or reducing the burden of infection. As Nussbaumer-Streit et al. [10] have, we operationalised "burden of infection" as including symptoms, complications, disability, hospitalisation, health-related quality of life, unintended health-related harms of interventions, infected individuals, and reproduction numbers. Randomised trials, observational studies, and modelling studies were included. All methods of contact investigation were considered, including tracing by telephone or via mobile apps. Articles in English, French, Spanish, and Portuguese from all countries were included. Studies were included only if they focused on SARS-CoV-2, as this virus poses unique challenges (such as asymptomatic and presymptomatic transmission). Abstracts, letters, protocols, preprints, and other unreviewed research were excluded, as well as studies limited to hospitals, nursing homes, prisons, and other enclosed spaces where transmission dynamics may not reflect that of the community. Reviews were also excluded, but their reference lists were checked for additional studies.

Data extraction and synthesis
Two investigators screened all studies. Discrepancies were solved by mutual agreement. Characteristics of studies were recorded in a spreadsheet. Those included: first author, publication date, study design, population, characteristics of contact tracing efforts, and main findings. We found substantial differences in study design, settings, outcomes, and effect measures. For example, in some settings, contact tracing was implemented along with lockdowns and other interventions (e.g., China). In other settings, it was not (e.g., South Korea). In addition, across studies, effect measures varied widely. We concluded that meta-analysis was not feasible [11]. Therefore, to promote transparent reporting, we followed synthesis without meta-analysis (SWiM) guidelines [12]. We synthesised results using the vote-counting method and narratively. For narrative synthesis, we focused on studies with the lowest risk of bias. We reported results separately for observational and modelling studies in a GRADE evidence profile table [13].

Risk of bias
We assessed risk of bias using methods similar to a Cochrane review on a related topic (quarantines for COVID-19) [13]. Briefly, for observational studies, a Cochrane tool was used [14]. This tool assesses internal and external validity across eight criteria (Table 1). Overall risk of bias was then rated as low (when all validity criteria were met), moderate (when at least one criterion was unclear), or serious (when at least one criterion was not met). For modelling studies, we used three criteria for best practices recommended by the International Society for Pharmacoeconomics and Outcomes Research (ISPOR) and the Society for Medical Decision Making (SMDM). These were: (1) Was the model dynamic? A model is dynamic when infection risk at a given point in time is dependent on the number of infectious individuals [15]. (2) Were uncertainty analyses conducted on key model parameters and assumptions? Uncertainty analyses quantify the influence of variations in model inputs on predicted effects (3) Did results include a comparison of the burden of infection? Overall, we had no concerns to minor concerns when a modelling study followed all three criteria. If at least one practice was unclear, we had moderate concerns. Lastly, if at least one practice was not followed, we had major concerns. Two investigators independently assessed risk of bias. Discrepancies were solved by mutual agreement.

Certainty of evidence
We rated the certainty of evidence using GRADE [13]. This approach was also used by a Cochrane review on a related topic (quarantines for COVID-19) [10]. Briefly, for observational studies, the evidence starts as low-certainty. For modelling studies, it starts as high-certainty. It can then be adjusted according to risk of bias, indirectness, inconsistency, imprecision, and publication bias. Overall, the evidence can be graded as high-certainty (very confident that the true effect lies close to the estimated effect), moderate-certainty (moderately confident in the effect estimate), low-certainty (limited confidence: true effect may be substantially different), or very low-certainty (very little confidence: true effect likely substantially different). One investigator rated the certainty of evidence. A second investigator checked the ratings.

Result of the search
A total of 544 papers were found. Removing duplicates left 343. We retained 158 based on title, 64 based on abstract, and 27 based on full text. We found one additional study via reference lists and four more via Table 1 Risk of bias criteria for single-arm observational studies of interventions.
the "cited by" and "similar articles" functions of PubMed and Google Scholar (eFigure in the Supplement). Therefore, this systematic review includes 32 studies (Table 2) [3,4,.

Observational studies
Results of observational studies were consistent: 14 out of 14 (100%) reported that contact tracing (alone or in combination with other interventions) was associated with better control of COVID-19 [16][17][18][19][20][21][22][23][24][25][26][27][28][29]. Outcomes measures included changes in daily new cases at the country or regional levels, reproduction numbers, doubling times, and serial intervals. Some studies had no measures of association at all (only descriptive statistics). We found serious risk of bias in 14 out of 14 observational studies (Table 3). Detection bias was present in all studies (14/14 studies). Potential bias due to confounding was the second most common (11/14 studies). In contrast, there was little selection bias (1/ 14 studies) and outcome reporting bias (0/14 studies). Overall, the certainty of this body of evidence was rated as very low (Table 4). Studies were carried out in China [26], Hong Kong [18,24,27], Taiwan [22], Singapore [23], South Korea [17,19] Vietnam [25], France [21], the U.S. [16] in African countries [20], and in international comparisons [28,29]. The six studies with the lowest risk of bias were carried out in China [26], Hong Kong [18,27], Taiwan [22], Singapore [23], and South Korea [17]. In all six studies, contact tracing was implemented early, along with border control measures and enhanced surveillance. In Hong Kong and South Korea, additional social distancing measures were also implemented. In Shenzhen, China, 391 cases and 1286 close contacts were identified over a 1-month period (January 14-February 12, 2020). Cases were isolated on average 4⋅6 days (95% CI 4.1-5.0) after developing symptoms; contact tracing reduced this delay by 1⋅9 days (95% CI 1.1-2.7) [26]. In Hong Kong, the first case was recorded on January 23, 2020. By May 31 (~4.5 months later), the outbreak was controlled (no community transmission). Overall, only 1084 cases and four deaths were recorded in Hong Kong in the first 4.5 months of the pandemic-in a population of 7.4 million with high density and intimate ties with China [18]. In Taiwan, big data and mobile geopositioning were used to help trace 627,386 persons in contact with Diamond Princess cruise ship passengers. Of those, 67 contacts tested negative and no confirmed COVID-19 cases were found, indicating successful containment. In addition, during follow-up, respiratory syndrome and pneumonia were less common among traced contacts than in the general population [22].

Modelling studies
Results of modelling studies (n = 18) depended on their assumptions. Peak et al. [46] illustrated this clearly by examining two sets of assumptions (in a study with a low risk of bias). In the first set of assumptions, 90% of contacts were traced within 0.5 days, quarantine reduced infectiousness by 75%, and isolation of cases reduced infectiousness by 90%. In this "high-feasibility setting", the reproduction number was reduced from 2.2 to a median of 0.49-0.57, and the epidemic was controlled. In a second, "low-feasibility setting", 50% of contacts were traced in 2 days, quarantine reduced infectiousness by 25%, and isolation of cases reduced infectiousness by 50%. In that setting, the epidemic was not controlled in any scenario. We examined a subset of five similar studies that modelled steps of the contact tracing process [3,4,43,44,46]. In those, under optimistic assumptions, efficient contact tracing always led to control of COVID-19. However, when less optimistic assumptions were modelled, inefficient contact tracing could at best slow, but not stop, outbreaks of COVID-19. For efficient tracing, these assumptions included short delays of 2-3 days from the time a new case develops symptoms to isolation of the case and quarantine of its contacts [3,44], at least 80% of contacts traced [3,4,43,44], and no further transmission upon isolation and quarantine [3,4,44,46]. In terms of quality, we had major concerns for 2/18 modelling studies (11%), moderate concerns for 1/18 study (6%) and no to minor concerns for 15/18 studies (83%) ( Table 5). Overall, the certainty of this body of evidence was rated as high (Table 4). This rating should be interpreted with caution, however, as results of modelling studies were highly dependent on their assumptions.

Discussion
This systematic review aimed to examine contact tracing effectiveness and to identify characteristics of effective tracing programmes. Observational studies (n = 14) provided consistent, very-low certainty evidence that contact tracing (alone or in combination with other interventions) was associated with better control of COVID-19. Modelling studies (n = 18) provided consistent, high-certainty evidence that under assumptions of prompt and thorough tracing with effective quarantines, contact tracing could stop the spread of COVID-19. This conclusion was supported by a number of preprints and other unpublished work [38][39][40][41][42][43][44][45][46][47][48][49][50]. However, under assumptions of slower, less efficient tracing, modelling studies found that tracing could slow, but not stop COVID-19.
Our results are in line with those of other reviews. A recent Cochrane review of quarantine (alone or in combination with other public health measures) found that modelling studies consistently reported a benefit of quarantine to control COVID-19, and that early implementation of quarantine and combining quarantine with other public health measures were important to ensure effectiveness [10]. Likewise, a narrative review of contact tracing in patients infected with SARS-CoV-2 concluded that this "classic" strategy could be applied, but that it should be accelerated [51].
According to modelling studies, if an outbreak is left unchecked for >2-3 days, it becomes practically impossible to stop with contact tracing alone (at this stage, tracing alone may only slow its spread) [3,54]. Therefore, timing and speed are crucial early on in an outbreak, and it stands to reason that as part of pandemic preparedness, contact tracers should be trained ahead of time, and proper settings and monitoring should be put in place to react in a timely manner. Ideally, tracing of contacts should begin as soon as the first case is detected in a jurisdiction. New cases can transmit the virus to others while presymptomatic or asymptomatic, further highlighting the importance of speed [3,4,53,54,56]. Cases should be tested and their results communicated in the shortest possible time. If new cases are not tested, wait to be tested, or receive test results after much delay, contact tracing is less effective. Thus, testing should be widely available, and the public could be reminded to seek testing at the earliest signs of symptoms. Likewise, systemic delays in communicating test results should be minimised. In the US, one report of drive-through testing found that when outsourced, tests (n = 476) were turned around in a median of 9.21 days [42]. Contact tracing may not be effective with such delays [3,54].
Once a case is confirmed, its contacts must be traced. When tracing contacts manually over the telephone, cases may not disclose, remember, or have contact information for all contacts. For example, a pilot programme in Sheffield, UK, found that two thirds of people contacted through tracing did not fully cooperate [53]. Moreover, during a peak of COVID-19, thousands of new contacts may need to be traced daily. Large contact tracing efforts with new staff are costly, and may not be able to maintain the level of effectiveness of smaller programmes with experienced staff. Mobile phone apps and other technologies can circumvent these shortcomings, but raise a number of ethical issues [54,55] and pose technological challenges [56]. Moreover, early evidence suggests that contact tracing apps, despite wide encouragement, have limited adoption [57,58] Thus, they may slow (but not stop) COVID-19.
Cases should isolate, and contacts should quarantine. Four of the five modelling studies we have reviewed assumed perfect prevention of transmission at these steps [3,4,53,54] This may be difficult to achieve in practice, even if cases and contacts never leave the home. Indeed, Bi et al. [26] found that household secondary attack rate was 11.2% (95% Table 2 Characteristics of the 32 studies included in this systematic review. As of February 2020 in France, initial questionning of cases for contact tracing was done by clinicians in collaboration with regional health entities and Santé Publique France. Contacts were stratified in different risk categories and measures were taken accordingly. Active surveillance and home quarantine was applied to moderate-high risk contacts. -Most contacts at risk were identified for the first 3 cases, but some low risk contacts could not be traced (e.g co-travellers on public transportation). No secondary transmission was detected.
-Contact tracing and follow-up was rapid and collaborative.
Chen 05-2020 Observationnal study Taiwan Contact tracing using data from GPS in the shuttle bus, credit card transaction log, closed-circuit television (CCTV), and mobile position data -Tracing of taiwanese citizens possibly exposed to passengers on the Diamon Princess was done by using travelling itinerary arranged by the agency, GPS in shuttle buses, credit card transaction logs, closed circuit televisions, vehicle license plate recognition system, and mobile positioning data.
-627,386 citizens were sent syndrome monitoring and self-quarantine information via SMS messaging.
-National Health Insurance Claims data was used to track the health status of all contacts. No COVID positive case was identfied.
(continued on next page) -Close contacts were those who lived in the same apartment, shared a meal, travelled, or socially interacted with an index case 2 days before symptom onset. Casual contacts (eg, other clinic patients) and some close contacts (eg, nurses) who wore a mask during exposure were not included.
-All close contacts were tested at the beginning and end of isolation.
-Contact-based surveillance was associated with a 2⋅3-day decrease in time to confirmation and a 1⋅9-day decrease in time to isolation.
-The reduction of time during which cases are infectious in the community should reduce the R. However, the overall impact is highly dependent on the number of asymptomatic cases.   Tang

02-2020
Modelling study China Quarantine rate of contacts -Increasing quarantine rate by 10 or 20 times could bring forward the peak by 6.5 or 9 days, and lead to a reduction of the peak number of infected individuals by 87% or 93%.
-The number of contacts traced in Wuhan as of 22 January 2020 (estimated to be 5897) was insufficient compared to the population size and appears to have had a limited impact on the epidemic control.
-The quarantine rate needs to be very high for a city to avoid an outbreak.
-The duration of travel restriction depends on a combination of effective quarantine and reduction of contact within the city. Lai

05-2020 Modelling study China
Travel restrictions, early identification and isolation of cases (including contact tracing), and social distancing -Early detection and isolation of cases was estimated to quickly and substantially prevent more infections than contact reduction and social distancing across the country (5-fold versus 2.6-fold).
-Combined interventions achieved the strongest and most rapid efect. Tang

03-2020
Modelling study Wuhan Lock-down, contact tracing followed by quarantine and isolation -Under the strict prevention and control measures, the effective daily reproduction ratio has been <1 since January 26th, 2020 -Our updated findings suggest that the best measure is persistent and strict self-isolation. Currie
-The effect of the app increases as its uptake increases, to a disproportionately greater extent than the increment in uptake.
-Some scenarios show that an app uptake of 61% has the potential to reduce the cumulative total number of new cases by >50%.
-COVIDSafe would have the capacity to contribute susbtantially to contact tracing in a second wave and serve as an adjunct to testing and social distancing. Ferretti
-The delay to isolation and contact quarantine is key to containing the pandemic -No delays to isolation and contact quarantine is associated with the greatest epidemic decline, whereas a 3 days delay would be associated with no decline.
-A mobile phone app implementing instantaneous contact tracing could reduce transmission enough to achieve R < 1 and sustained epidemic suppression. Hellewell
-A high number of cases and contacts can overwhelm the contact-tracing system and affect the quality of the contact-tracing effort.
-In most plausible outbreaks scenarios, case isolation and contact tracing alone is insufficient to control outbreaks within 3 months, but can contribute to reduce the (continued on next page) CI 9.1-13.8) in Shenzhen, China. Similarly, Park et al. [59] found that 11.8% of household contacts had COVID-19 in South Korea. Moreover, Wu and McGoogan [35] report that in 20 Chinese provinces outside of Hubei, a total of 1183 case clusters were found, of which 64% were within familial households. Thus, it may be more effective to isolate and quarantine outside the home, in hotels or central locations, especially considering many homes (1 in 5 in the US) will lack sufficient space to comply with recommendations [60]. If home isolation and quarantine are used nonetheless, information could be provided to new cases and contacts to reduce household transmission. Li et al. [61] followed 105 index patients and 392 household contacts in Wuhan, China. They found 14 cases had isolated by themselves at home immediately after the onset of symptoms-with masks, dining separately, and residing alone. No households contacts were infected. A final consideration for this step of the process is financial and social support. Contacts may need support for lost wages, daily activities carried out outside the home (e.g. groceries), or questions related to their health (e.g. telephone helpline).

Strengths and limitations
This study has a number of strengths. To our knowledge, it is the first to systematically review the effectiveness of contact tracing efforts. It included both observational and modelling studies, hence examining contact tracing in two complementary ways. Observational studies examined how contact tracing (alone or in combination with other interventions) was associated with the spread of COVID-19. Modelling studies examined how varying assumptions about each step of the tracing process might influence its effectiveness.
This study also has a number of limitations. It was not feasible to perform a meta-analysis, so no synthesised estimates of the magnitude of associations are available. Moreover, observational studies reported results of contact tracing efforts in combination with other interventions. So, its independent potential effect could only be examined in modelling studies. Another limitation is the rapidly evolving nature of the pandemic and related contact tracing policies and regulations, as well as public perceptions, which are all likely to influence future tracing efforts. A final limitation is we did not discuss implementation. Rajan et al. [62] describe specific challenges and solutions when implementing contact tracing programmes.
This body of literature also has a number of limitations. These may be inherent, as studying contact tracing in the midst of a pandemic poses substantial methodological challenges. Indeed, we found no randomised controlled trials. This is perhaps not surprising, as lack of feasibility and ethical considerations arguably precludes them. We could not therefore draw conclusions about causality. The strength of our conclusions was further limited by the quality of the observational studies we found. All had a single arm, and none mentioned blinding of assessors. This increased risk of bias, according to the Cochrane tool we used. While we aimed to enhance replicability and standardization by using this tool, one may wonder if multiple arms and blinding of assessors are truly feasible during a pandemic. If not, specific tools may be required to assess the quality of pandemic studies, taking into account the exceptional circumstances in which they are carried out. Another limitation was that most observational studies (11 out of 14) did not account for potential confounders. What were the roles of age and gender? Of varying levels of compliance in naming contacts? Of other concurrent public health interventions, like masking and lockdowns? The complex interplay of these potential confounders and covariates limit our understanding of the true underlying dynamics of transmission from person to person and from group to group, which may change across settings, times, and concurrent interventions (e.g., from children in school to their home and play environment; from residents in nursing facilities to staff and home environments that are geographically dispersed, etc.). For all these reasons, while results of modelling studies may be encouraging, the connection between contact tracing and epidemic trajectory is difficult to demonstrate in the real world. Indeed, while some modelling studies provided strong support for tracing, their most supportive results were based on optimistic assumptions. Observational studies did also provide some support for contact tracing, but the certainty of this body of evidence was rated as very low. Policymakers should keep these limitations in mind before enacting policies based on optimistic modelling studies, while researchers may wish to examine these and other potential confounders, mediators, and effect modifiers in future studies.

Conclusions
Observational studies (n = 14) provided consistent, very-low certainty evidence that contact tracing (alone or in combination with other interventions) was associated with better control of COVID-19. Modelling studies (n = 18) provided consistent, high-certainty evidence that under assumptions of prompt and thorough tracing with effective quarantines, contact tracing could stop the spread of COVID-19. A cautious interpretation suggests that to stop the spread of COVID-19 with contact tracing, public health practitioners have 2-3 days from the time a new case develops symptoms to isolate the case and quarantine at least 80% of its contacts, and that once isolated, cases and contacts should infect zero new cases. Unfortunately, under assumptions of slower, less efficient tracing, contact tracing may slow, but not stop COVID-19. In those cases, given the limitations of this body of literature, it is unclear whether the benefits of tracing outweigh its costs, and practitioners may consider scaling down efforts as the US CDC advise [5], and turning instead to other more promising evidence-based, costeffective interventions [2]. Future research may improve our understanding of contact tracing effectiveness by assessing emerging empirical evidence from ongoing efforts, best practices and policy responses, and differences in outcomes across jurisdictions with more or less efficient tracing.

Funding
None.

Authors' contributions
CEJ and ASB designed the study. CEJ and ASB searched the literature. CEJ, ASB, PC, and US analysed the literature. All authors interpreted the findings. CEJ and ASB wrote the first draft. All authors revised drafts and approved the final manuscript.

Transparency declaration
The corresponding author (PC) affirms that this manuscript is an honest, accurate, and transparent account of the study being reported; Table 3 Risk of bias assessment of observational studies (n = 14).

Study first author
Year  that no important aspects of the study have been omitted; and that any discrepancies from the study as planned have been explained.

Declaration of Competing Interest
The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: CEJ has contractual agreements with the Centre intégré universitaire de santé et de services sociaux du Centre-Sud-de-l'Île-de-Montréal and is founder of the COVID-19 Science Updates.

Acknowledgments
We thank Dr. Louise Potvin (Université de Montréal) for insightful comments on the manuscript.  Table 5 Risk of bias assessment of modelling studies (n = 18).