In May 2020, Google and Apple released the Exposure Notification System, which is an application programming interface that logs the following: who a phone user has been in contact with, for how long, and at what distance [1]. This release came 2 months after COVID-19 was declared a pandemic [2], allowing governments to identify and isolate contacts of confirmed cases through a process known as “contact tracing” [3,4].

Less than a year after the first reported cases, over 33 million individuals have tested positive for COVID-19 worldwide and more than 1 million have died [5]. To limit disease spread, over half of the global population has been subjected to lockdowns involving school closures, workplace shutdowns, and movement restrictions [6]. Although these lockdowns are effective in tapering the epidemic curve [7], they are costly to the global economy and are unsustainable [8]. However, allowing the virus to spread unhindered could overwhelm the health care system and result in large-scale mortality [9,10].

To address both infection control and economic concerns, several countries have turned to contact tracing to keep the economy running [11,12]. Epidemiological modeling suggests that if (1) cases are effectively identified (through rigorous testing protocols), (2) contact tracing is comprehensive (identifying all possible exposure), and (3) contacts are quarantined in a timely manner, this strategy can curb the spread of the virus [4,11]. In an optimal scenario, 80% of contacts should be traced on the same day an individual tests positive [3,11].

Early during the pandemic (and in previous infectious disease outbreaks), contact tracing was manually performed [13]. Using a range of interview and surveillance techniques, a human contact tracer would typically identify an average of 36 contacts for each positive case [14]. Although this strategy allows for high levels of case detection when there are few cases [15], its labor-intensive format—requiring ~12 h of tracing for each positive case [16]—is difficult to scale up. Additionally, individuals who test positive may forget whom they have been in contact with, thus undermining the effectiveness of the process [11].

Considering these limitations of manual contact tracing, several mobile apps have been developed to facilitate automated contact tracing [17], for example, COVID Watch in the United States [18], COVIDSafe in Australia [19], and Corona-Warn-App in Germany [20]. These apps primarily track Bluetooth signals from phones in the vicinity [3], capturing contacts without the restraints of staffing or recall biases [4,11]. Further, phone apps can notify individuals swiftly after a contact tests positive, allowing them to be quickly isolated [3].

Despite the potential of digital contact tracing, a recent meta-analysis concluded that owing to implementation barriers, manual contact tracing should remain the order of the day [12]. One major barrier pertains to the uptake of mobile apps. Several modelling studies have assessed parameters needed for the COVID-19 reproduction number (R₀) to fall below 1 [3,11,21]. R₀ refers to the number of infections spread from 1 positive case, and a value less than 1 indicates that the virus has been contained. For this to be achieved, contact tracing apps need to be downloaded by at least 56% of the population [21], which is much higher than the average rate of downloads globally (9%) [22].

To increase uptake, Qatar made it mandatory for residents to use the official contact tracing app [23]. Although this legislation led to high download rates (>90% [24]), the potential backlash from the public (eg, because of privacy concerns [25,26]) implies that few countries are likely to follow suit. Correspondingly, public health agencies would benefit from an understanding of the predictors of voluntary downloads [27], providing an empirical basis to nudge citizens and residents to voluntarily download contact tracing apps [28].

Given the urgent need to boost contact tracing apps, this study is the first to identify sociodemographic factors predicting voluntary uptake. Our study was conducted in Singapore, where the world’s first nationwide contact tracing app TraceTogether was launched in March 2020 [29]. TraceTogether uses a centralized approach adopted by several governments [19]; namely, randomly generated user IDs are generated and shared via Bluetooth with phones in close proximity [30]. When individuals test positive for COVID-19, they consent to add both their own user IDs and those of their contacts to a centralized database. This is used to identify matches, and exposure notifications are then sent from the server to close contacts [31,32]. (As an alternative model, a decentralized approach could be used where both matches and notifications are made through the user’s phone [33].)

As Singapore was the forerunner of this technology, the app has accrued 2.3 million users within 6 months, including approximately 40% of Singapore’s resident population or 50% of all smartphone users (considering a smartphone penetration rate of 82%) [34,35]. Correspondingly, our study represents a “best case scenario” for app uptake after several months have elapsed. In terms of the epidemic curve, our study was conducted between April and July 2020, as the country was in a lockdown (April to May 2020). This period witnessed a peak in daily COVID-19 cases (April: >1000/day or 175 per million population), which gradually tapered over time (July: >100/day or 17.5 per million population).

Between April 3 and July 17, 2020, we recruited 505 adults who met the following eligibility criteria: (1) at least 21 years of age and (2) had lived in Singapore for a minimum of 2 years. All participants responded to online advertisements. Within the constraints of online sampling owing to the pandemic, we strove to obtain a representative sample by placing advertisements in a wide range of online community groups (eg, Facebook or WhatsApp groups among individuals in residential estates, universities, and workplaces) and by using paid online advertisements targeting the broad spectrum of Singapore residents.

Prior to study enrolment, participants provided informed consent in accordance with a protocol approved by the Yale-NUS College Ethics Review Committee (#2020-CERC-001; ClinicalTrials.gov ID NCT04468581). They then completed a 10-min online survey hosted on the platform Qualtrics [36]. Data were collected in accordance with the second phase of a larger study tracking COVID-19 responses, and findings from the first phase have been described previously [37,38].

As the primary outcome variable, participants were asked to indicate whether they had downloaded the government’s contact tracing app TraceTogether (binary variables: 1=they had, 0=they had not).

For primary analysis, binary logistic regression was used to identify predictors TraceTogether uptake. In the first model (model 1), participants' demographics were included as predictors (age, citizenship, gender, marital status, education level, ethnicity, household type, and household size). Citizenship (base=others), gender (base=female), marital status (base=single), and ethnicity (base=Chinese) were coded as dummy variables. In the second model (model 2), we repeated the first model with the inclusion of situational variables (log-transformed total number of COVID-19 cases to date and lockdown status). Finally, in the third model (model 3), we repeated the second model with the inclusion of the total number of behavioral modifications as a predictor. All data were analyzed using SPSS (version 23, IBM Corp) and R (version 3.6.0, The R Foundation), with the type 1 familywise error rate controlled at α=.05 via Bonferroni correction (Bonferroni-adjusted α=.003 [.05/17 predictors]).

Table 1 shows the wide range of demographic characteristics of our study cohort (N=505). Compared to the resident population, the sample was matched in the following characteristics: ethnic composition, household size, and housing type (a proxy of socioeconomic status in Singapore) (≤10% difference). However, compared to the resident population, the present participants were more likely to be female (n=313, 62.0% vs 51.1%), single or dating (n=234, 46.4% vs 31.6%), to have a higher level of education or no tertiary education (n=65, 12.9% vs 51.7%), and to be citizens of Singapore or of other countries (n=456, 90.3% vs 61.4%).

Of the 505 participants, 274 (54.3%; 95% CI 49.8%-58.7%) reported having downloaded TraceTogether. The download rate in this sample matches that of smartphone users in the resident population [34], and Table 2 describes the characteristics of users and nonusers.

Table 3 shows parameter estimates from logistic regression analyses of the predictors of TraceTogether uptake. No demographic or situational variable significantly predicted downloads (models 1 and 2). After controlling for these variables, the number of behavioral modifications emerged as a significant predictor (model 3); that is, with each unit increase in the number of behavioral modifications adopted, participants were 1.10 times more likely to download TraceTogether (z=13.56; P<.001).

In the logistic regression analyses, TraceTogether downloads were predicted from the number of behavioral modifications because of the pandemic. To understand this association better, we conducted further exploratory analyses.

As shown in Figure 1, the majority of participants had modified their behaviors to curb the spread of COVID-19. The use of TraceTogether ranked 10th in the frequency of adoption (274/505, 54.3%), similar to the frequency of voluntary mask wearing (n=276, 54.7%).

A corollary question is how TraceTogether usage is associated with other health protective behaviors; that is, how likely were people to download TraceTogether if they had modified their behavior in other ways? To address this question, we conducted network analyses by estimating a mixed graphical model (MGM) with the R package mgm [41]. MGM constructs weighted and undirected networks where the pathways among behaviors represent conditionally dependent associations, having controlled for the other associations in the network. Similar to partial correlations, each association (or “edge”) is the average regression coefficient of two nodes. To avoid false-positive findings, we set small associations to 0 for the main models.

As shown in Figure 2, TraceTogether usage was associated with hand sanitizer use, avoidance of public transport, and a preference for outdoor vs indoor venues. The adjacency matrix (ie, numerical values for the average regression coefficient between two nodes) for Figure 2 is presented in Multimedia Appendix 1.

For sensitivity analysis, we performed logistic regression analysis using TraceTogether downloads as the dependent variable, and 18 other behavioral modifications (see Methods) as the predictors. Our conclusions did not change, as indicated in Multimedia Appendix 2.

As our primary outcome, we observed that the number of behavioral modifications significantly predicted the use of TraceTogether. In other words, a person who had already changed his/her lifestyle on account of the pandemic was also likely to download a contact tracing app. Network analyses revealed that downloads clustered with (1) using hand sanitizers, (2) avoiding public transport, and (3) preferring outdoor to indoor venues. This finding may suggest that public health campaigns could capitalize on other behavioral modifications when seeking to promote app downloads, for example, by printing information regarding a contact tracing app on the packaging of hand sanitizers or by framing the use of digital contact tracing as a preventive behavior. Policy makers might also expect app download rates to track behavioral modifications, anticipating, for example, higher download rates when the public fears an increase in COVID-19 cases (leading to more behavioral modifications) [42].

Theoretically, our findings further corroborate those of previous studies on how individuals change their behaviors during a pandemic. Based on prior outbreaks, a taxonomy of modifications had been identified whereby (1) “avoidant behaviors” are measures taken to avoid contact with potential carriers (eg, avoiding crowded places), while (2) “prevention behaviors” are those associated with maintaining hygiene (eg, regular hand washing) [42]. Extrapolating to the technological realm, our findings suggest that the use of a contact tracing app cuts across this taxonomy, since downloads were associated with both avoidant (avoiding public transport and preferring outdoor venues) and prevention behaviors (using hand sanitizers). Moving forward, we urge further studies to revise these classification systems in light of new technological developments.

Apart from behavioral modifications, it is notable that no demographic (eg, age, gender, etc) or situational variable (eg, number of COVID-19 cases and lockdown status) significantly predicted TraceTogether uptake. Prior to our study, it would have been conceivable that only a subset of the population would download a contact tracing app (eg, demographic groups based on gender, educational level, or age) [27,38,43]. By contrast, our findings highlight how uptake of digital contact tracing apps cuts across demographic groups.

While the lack of significant associations may be counterintuitive, a recent study reported similar results when predicting COVID-19–related behavioral modifications [42]. In a multinational survey, Harper et al [42] similarly observed that demographic and situational variables were unrelated to behavioral modifications owing to the pandemic. Since behavioral modifications predicted the use of a digital contact tracing app in our study, it seems reasonable to observe an analogous pattern here; that is, our results are unlikely to be based on false-negative outcomes.

As public health agencies develop strategies to promote downloads for contact tracing apps, the pattern of our findings may in turn suggest that demographic-specific messages are not needed. This is encouraging because the behavioral sciences offer widespread measures to “nudge” the general population [44]. In this case, the general public simply needs a one-off nudge to download the contact tracing app, after which the app functions independently in the background. Thus, if governments can nudge users in this first step (eg, by introducing incentives to download or by introducing contact tracing as an opt-out feature of existing government apps), it may be possible to attain the download rates necessary for contact tracing to be effective. Simultaneously, we urge further studies on the acceptance of such strategies; considering public concerns regarding privacy [25,26], any widespread intervention would need to be introduced cautiously.

Our study has several limitations of note. As the first study of its kind, we made several choices at the exclusion of others. First, we opted for a cross-sectional design that precludes strong conclusions regarding causality. Second, we included an online sample to minimize person-to-person contact during the pandemic. Although we sampled individuals from a wide array of demographic groups, respondents were not representative of the general nationwide population; this may have deterred the establishment of potential associations among variables (eg, by including a high proportion of educated participants). Third, our survey relied on participants’ self-reported use of a contact tracing app. Although our download rate is similar to that of the general population, further studies may seek to verify actual usage (eg, by incorporating survey questions in a contact tracing app). Fourth, our survey was not intended to measure every aspect of TraceTogether usage, and there were several notable omissions (eg, reasons why individuals chose to use or not use the app, phone ownership, and usage-related questions). Indeed, our model metrics (eg, small Nagelkerke R²) indicate small effect sizes, highlighting the need for further studies to include a more comprehensive set of variables that may account for app downloads. Finally, we examined TraceTogether—an app with a centralized contact tracing protocol. Future studies are required to assess whether our findings extend to apps with decentralized protocols or to other forms of digital contact tracing that do not rely on mobile apps (eg, public acceptance of cloud-based contact in South Korea).

In conclusion, the potential contribution of digital technology to pandemic management is receiving increasing attention. What remains unclear, however, is how this technology is received and how best to promote its uptake. Focusing on contact tracing, this study shows that downloads of a mobile app was best predicted from the adoption of other infection control measures such as increased hand hygiene. In other words, the introduction of digital contact tracing is not merely a call to “trace together” but rather to “modify together,” to use contact tracing apps as part of the broader spectrum of behavioral modifications during a pandemic.