How Good is Good Enough? Quantifying the Impact of Benefits, Accuracy, and Privacy on Willingness to Adopt COVID-19 Decision Aids

Abstract

An increasing number of data-driven decision aids are being developed to provide humans with advice to improve decision-making around important issues such as personal health and criminal justice. For algorithmic systems to support human decision-making effectively, people must be willing to use them. We expand upon prior research by empirically modeling how accuracy and privacy influence intent to adopt algorithmic systems, focusing on an globally-relevant decision context with tangible consequences: the COVID-19 pandemic. We analyze surveys of 4,615 Americans to (1) evaluate the effect of both accuracy and privacy concerns on reported willingness to install COVID-19 apps; (2) examine how different groups of users weigh accuracy relative to privacy; and (3) we empirically develop the first statistical models, to our knowledge, of how the amount of benefit (e.g., error rate) and degree of privacy risk in a data-driven decision aid may influence willingness to adopt.

Decentralized is not risk-free: Understanding public perceptions of privacy-utility trade-offs in COVID-19 contact-tracing apps

“…These people seemed to prefer decentralized designs over centralized designs. This finding aligns with a study by Kaptchuk et al [7] that found 27% of people do not want to install COVID contact-tracing apps even if they are perfectly private (Section 4.3). • The most popular app design (centralized, release infection hotspots in public places) had around 55% of our participants willing to install (Section 4.1).…”

“…They can also gather information to help epidemiologists monitor the spread of the disease, discover disease hotspots, and contact exposed individuals. However, their effectiveness is highly dependent on the adoption rate, which has been demonstrated to be challenging due to people's concerns about issues such as privacy [5,6,7]. Ferretti et al [8,9] suggested that if 60% of the population installed the app, the estimated number of coronavirus cases would go down.…”

“…Although there are many attempts to create COVID-19 contact-tracing apps in a privacy-preserving way [10,3], it is less clear whether sufficient people are willing to install the apps, and what design choices can lead to higher adoption rates. Some recent studies have started looking into this problem by projecting the installation rate of contact-tracing apps in the U.S. in general [11,12] or examining how the willingness to adopt COVID-19 apps varies with accuracy levels, public health benefits, and who the data may be leaked to [7]. However, none of them has connected these abstract factors to concrete app-design choices to provide guidelines on designing and deploying COVID-19 apps that can achieve optimal adoption rates.…”

Deploying wearable sensors for pandemic mitigation

“…In a baseline scenario, without improvements to detection specificity, every user would spend over two days a month on average incorrectly quarantining, and ∼40,000 to ∼65,000 additional confirmatory lab-based tests would be required each day. The social and economic harm caused by solely promoting uptake or adherence without improvements to detection specificity would likely undermine public confidence in and compliance with a wearable-based pandemic mitigation strategy 22 . Alavi et al found that many false positives were due to the detection algorithm identifying lifestyle- driven changes in resting heart rate (e.g., after intense exercise or alcohol consumption); accounting for these factors using more advanced algorithms may be one way to target improved detection specificity 10 .…”

“…Third, we made the simplifying assumption that all users without symptoms (and that no users with symptoms) could benefit from wearable-informed prompts to take a confirmatory test and self-isolate. Fourth, we did not consider how uptake or adherence might vary over time or with detection accuracy 18,22 . Fifth, we used median values for SARS-CoV-2 infection parameters (e.g., latent period) and did not account for reinfections or vaccinations.…”