Nationwide rollout reveals efficacy of epidemic control through digital contact tracing

When the SARS-CoV-2 virus started spreading globally, many initiatives for the development of digital contact tracing based on mobile phones were initiated, motivated by modeling work (e.g., Ferretti et al.) suggesting that an adequately developed digital contact tracing system could keep the reproduction number below 1. Despite widespread deployment, conclusive real-world evidence of effectiveness has been lacking. Measuring efficacy has been difficult because deployed systems (largely based on the privacy-preserving Exposure Notification System, ENS) limit visibility into contact events, leading most assessments to rely on incomplete telemetry and population surveys. To overcome these limitations, the authors analyze a unique, anonymized, centrally collected contact dataset obtained during the rollout of Norway’s first contact tracing app (Smittestopp) in Spring 2020, encompassing millions of contacts. They use these data to quantify real-world technological tracing efficacy, estimate the proportion of close contacts not identifiable by manual tracing using a learning classifier, relate efficacy to app uptake, and evaluate potential epidemic control impact by feeding efficacy estimates into a standard epidemic spread model.

The paper situates its contribution within a literature that initially proposed digital contact tracing as a potentially decisive intervention (e.g., Ferretti et al.) but has since produced mixed and inconclusive real-world evidence. Prior evaluations of ENS-based apps have been hindered by privacy-driven limitations on data visibility, forcing reliance on partial metrics and surveys. Several studies have questioned the efficacy and need for digital tracing, particularly considering privacy concerns. The authors highlight the absence of a comprehensive contacts database from a large-scale deployed system and position their work as addressing this gap by analyzing centrally collected contact data from a nationwide rollout, thereby enabling direct measurement of technological efficacy and the identification of contacts that manual tracing would likely miss.

Study context and app: Smittestopp, Norway’s first digital contact tracing app, was rolled out in Spring 2020, rapidly reaching 28% installation among the adult population before suspension in June 2020 due to privacy concerns. The app used Bluetooth Low Energy (BLE) to detect proximity events. Received signal strength was used to approximate distance, and contact identification was performed centrally by fusing reports from devices. Contacts of epidemiological interest were defined as within approximately 2 meters and lasting at least 15 minutes. Dataset: The anonymized dataset comprises millions of BLE exposure events and over 26 million close encounters during the rollout, representing interactions among roughly 12.5% of the adult population. Daily categories of exposures were recorded; centralization assumed symmetric contact identification at the system level despite device-level asymmetry. Technological efficacy measurement: The authors estimate directional detection probabilities between device types (iOS and Android). Let p_xy be the probability that a phone of type x detects y when in proximity; detection events are assumed independent between phones. Using centrally collected counts of detected contacts by direction and type pairs (D_xy, D_yx), the authors derive detection probabilities. From the contact dataset they estimate: - Probability iOS detects iOS: P_ii = 0.534 - Probability Android detects iOS: P_ai = 0.353 - Probability iOS detects Android: P_ia = 0.053 - Probability Android detects Android: P_aa = 0.763 These feed into an expression for overall technological tracing efficacy that accounts for market share (ϕ) and heterogeneous app uptake among iOS (α_i) and Android (α_a) users. Tracing efficacy varies with uptake approximately quadratically and depends on the mix of device types. Uptake-efficacy modeling: The authors compute technological tracing efficacy as a function of app uptake in the two user groups using modified pairing probabilities that incorporate α_i and α_a. They report that detection between homogeneous Android pairs can reach ~93%, and that overall technological efficacy is ~80% when the population is equally split across iOS/Android. Identification of random (unknown) contacts: To estimate the proportion of close contacts unlikely to be identified by manual tracing, the authors trained a supervised classifier on the contact graph to distinguish known vs. random contacts. They constructed a training set with plausible positives (device pairs meeting on ≥7 different days) and negatives (pairs never in contact despite both being repeatedly in close contact with a common third device). A random forest classifier (75 trees; grid-searched hyperparameters including minimum depth 8, minimum samples to split 1, and leaf 1) was trained and evaluated for varying close-contact definitions (any duration, ≥5 minutes, ≥15 minutes). Accuracies were 84–89%, with 89% reported for classifying risky close contacts. The model was applied conservatively, counting only contacts likely out of reach of manual tracing as random to avoid overestimating digital tracing’s added value. Epidemic impact modeling: The authors injected empirically estimated technological tracing efficacy as a function of uptake into the epidemic model of Ferretti et al., which describes the effect of case isolation and contact tracing on transmission. They assumed short delays (e.g., ~4 hours) between case identification and isolation/quarantining, explored parameterizations of isolation efficacy, and examined scenarios with R0 = 2.7 and R0 = 1.5. They also assumed 10% environmentally transmitted (non-traceable) infections, consistent with their estimates. Ethics, data, and code: Use of the dataset was approved with privacy safeguards; the contact dataset is not public but available upon request to the institution under strict conditions. Detection-probability data and scripts for analyses (e.g., efficacy modeling, random contact analysis) are publicly available in a GitHub repository as referenced by the article.

- Rapid uptake: Smittestopp reached 28% installation among the adult population during the initial rollout. - Scale of data: Over 26 million close encounters were recorded. - Technological efficacy: Overall technological detection efficacy reached about 80% with an equal iOS/Android split; detection among Android-only pairs was as high as ~93%. - Added value over manual tracing: At least 11% of risky close contacts (and up to 16% in some estimates) were random/unknown and likely not identifiable by manual contact tracing, demonstrating a measurable supplementary benefit of digital tracing. - Random contact characteristics: The fraction of risky close contacts classified as random was at least 11%; without a duration threshold, the fraction increased to about 33.3%. Daily fractions of random contacts across all contacts were substantial and varied with holidays/weekends. Random contacts were typically shorter (often 20–40 minutes) but still epidemiologically relevant. - Highly connected users: A subset of users exhibited very high numbers of daily close contacts and tended to connect to other highly connected users, suggesting potential roles in superspreading contexts (e.g., occupations with frequent close contact). Digital tracing can help identify such individuals/events. - Uptake dependence and epidemic control: Tracing efficacy scales strongly with app uptake. Modeling indicates that app-based tracing alone would require ~95% uptake to control an epidemic with R0 = 2.7, which is likely unrealistic. However, with R0 = 1.5 (representative of concurrent control measures), an uptake of ~40% could reduce the growth rate below zero, contributing to epidemic control, assuming rapid isolation/quarantine and other reasonable parameters. - Policy relevance: Even moderate uptake (e.g., 30–40%) yields tangible benefits and can supplement manual tracing to improve overall contact identification by roughly 7.5–10.5% in contexts where 25–35% of contacts are random.

The findings provide real-world evidence that smartphone-based digital contact tracing can achieve substantial technological efficacy and contribute meaningfully to epidemic control when combined with manual tracing and other measures. The system identified a significant set of risky close contacts that manual tracing would likely miss, particularly random interactions and contacts among highly connected individuals who may drive superspreading. The dependence of effectiveness on uptake underscores the need for public engagement and policies that encourage adoption. While ENS-based systems were not directly evaluated here, the authors note that ENS is expected to perform comparably or better than Smittestopp technically, suggesting the results generalize or may even underestimate performance. Incorporating empirically grounded efficacy into epidemic models indicates that, although digital tracing alone is unlikely to control high-R0 spread, it can be instrumental in scenarios where other interventions reduce R0 (e.g., to ~1.5), with moderate uptake (~40%) helping to achieve declining transmission. The work highlights digital tracing’s role as a complementary tool that can speed isolation, capture otherwise untraceable contacts, and help contain superspreading-related outbreaks.

This study leverages a unique, large-scale, real-world contact dataset from the nationwide rollout of Norway’s Smittestopp app to quantify technological tracing efficacy, characterize contacts unreachable by manual tracing, and assess potential epidemiological impact. The app achieved approximately 80% technological detection efficacy (with an equal iOS/Android mix), and at least 11% of risky close contacts were random and likely unidentifiable manually. Modeling suggests substantial public health benefits at moderate uptake levels, particularly when combined with other control measures that reduce R0. The primary barrier to realizing digital tracing’s full potential is uptake rather than technical performance. Future work should focus on increasing adoption, integrating digital tracing with broader public health systems, further validating performance in ENS-based deployments, and, where feasible and privacy-protecting, linking contact data with infection outcomes to quantify direct impact.

- Data from a pre-ENS app: The dataset derives from Smittestopp, developed before the Apple/Google Exposure Notification System; while ENS may perform as well or better, this could not be verified in deployed systems due to privacy constraints. - No direct linkage to infection outcomes: The study does not link detected contacts to confirmed transmission, limiting inference about direct reductions in infection. - False positives/negatives: The technological efficacy estimate focuses on detection (false negatives); false positives are not directly captured. The authors argue modest secondary attack rates limit the impact of technological false positives on unnecessary isolation. - Centralized architecture: Smittestopp’s centralized design differs from ENS’s decentralized approach, potentially affecting generalizability. Centralization improved detectability at a potential privacy cost. - Parameter uncertainties: Epidemic impact modeling depends on assumptions (e.g., isolation efficacy, delays, environmentally transmitted fraction). Results for alternative parameters are provided in supplementary materials. - App uptake and device heterogeneity: Efficacy depends on population uptake and on the iOS/Android mix; results may vary with market shares and behavior in other settings.