Explaining Support for COVID-19 Cell Phone Contact Tracing

Contact tracing applications have been deployed at a fast pace around the world to stop the spread of COVID-19 and may be key to containing future pandemics. This study aims to explain public opinion toward cell phone contact tracing using a survey experiment. We build upon a theory in evolutionary psychology—disease avoidance—to predict how media coverage of the pandemic affects public support for containment measures. We report three key findings. First, exposure to a news item that shows people ignoring social distancing rules causes an increase in support for cell phone contact tracing. Second, pre-treatment covariates such as anxiety and a belief that other people are not following the rules rank among the strongest predictors of support for COVID-19 apps. And third, while a majority of respondents approve of the reliance on cell phone contact tracing, concerns for rights and freedoms remain a salient preoccupation.

Notes: All census proportions come from the 2016 Canadian Census Pro le tables. Proportions by ethnic group are from the item "Visible minority for the population in private households, " with the exception of the "Registered and Treaty Indian" proportion used for the indigenous group, and marked in parentheses. The proportion with a university degree is for the Canadian population aged 25 and older. The household after-tax income is used for census income groups. The p-values are for Pearson chi-square tests of di erence in proportions with Yates continuity correction. Signi cant di erences indicate that the proportion is unlikely to be observed with random sampling. * : p < 0.005.

Questionnaire and Treatment Vignettes
The survey questionnaire presented to respondents contains ve blocks. The rst block included pretreatment covariates (the variables labelled "Not Serious Enough" and "Worried" in the main text). The second block randomly assigned respondents to one of three groups and presented the media framing vignettes to the two treated subgroups. The third block queried respondents about their opinion toward cell phone contact tracing; the outcome variable of interest. The fourth block asked additional questions regarding containment measures. The fth and nal block contained demographic questions. Figures A1 and A2 display the vignettes used for the two framing treatments administered during the survey. Each version invited respondents to read the excerpt from a news headline that included title, author, picture and lead. The two news articles were selected speci cally because they captured the theoretical concept of interest, and since they had an equivalent version published in French media. As explained in the manuscript, we devoted considerable time to ensure that the question used to ask respondents about COVID-19 apps remained as neutral as possible. One reason for this was the existence of con icting results in previous polls. In the introduction to the main text, we mentioned the Senate study and the Mainstreet Research poll, two sources of data on public opinion on cell phone contact tracing that featured in Canadian news media. The phrasing of survey questions di ered in these two studies, which raises the issue of how sensitive public opinion is to question wording. The Senate study's query contained a preamble explaining the purpose and bene ts of contact tracing apps. While it provided respondents with background information about the technology, this choice may have painted the issue in a more positive light. The preamble and question read: Today's smart phones have location and proximity tracking capabilities. Used together with rapid testing capabilities, this technology could help public health professionals to more rapidly, accurately and completely trace the possible spread of COVID-19. This would allow them to protect public health and help to better manage the easing of social and economic restrictions.
If the tracking capabilities of smart phones provided public health o cials with the ability to anonymously and automatically notify all those who have been close to someone who tested positive for COVID-19, how supportive would you be of using this capability in Canada? (Moodie et al., 2020, 34).
The Mainstreet Research question, on the other hand, relied on a di erent language. It contained the verb "track", which may have primed the privacy implications of contact tracing apps. The question was: [...] Please tell us if you think it is acceptable or not: The government asking you to download an app on your smart-phone to track who you might come into contact with, otherwise known as contact tracing (?, 13).
It is not clear whether "the government asking you" means making the use of a COVID-19 app mandatorya practice currently used in other countries-but some respondents could interpret the query as such.
As discussed in the main text, our study relied on a question that described the app using plain language, and we avoided to prime the merits versus the risks. To assess the sensitivity of our results to question wording, we randomly assigned respondents to one of three variants of the same question. The text appears in the next section. These variations in question wording, however, had no signi cant impact on the response.
In the next section we also report the question wording for the variables used in the main text. For the "Not Serious Enough" and "Worried" variables, we relied on the phrasing used in COVID-19 surveys from the Washington Post and the Center for Democracy and Civic Engagement at the University of Maryland, to allow comparisons with existing datasets. For the purpose of our analysis, these variables were recoded to equal 1 for respondents who answered "Not seriously enough" and "Very worried", respectively, and 0 otherwise. Using the multinomial version of these variables did not a ect the substantive conclusions reported in the text. The "Trudeau Approval" variable is coded 1 for respondent who believe the federal government did an excellent job handling the pandemic, 0 otherwise. The "Lost Job" variable equals 1 for respondents who reported having lost their job during the pandemic, and 0 otherwise.

Survey Questionnaire
[Not Serious] "During the COVID-19 lockdown, do you think most people have taken social distancing measures too seriously, not seriously enough, or are most people striking the right balance?" • "Not taking seriously enough" • "Striking the right balance" • "Taking too seriously" [Worried] "How worried, if at all, are you about close family members or friends becoming infected and seriously ill from the coronavirus?" • "Very worried" • "Somewhat worried" • "Not worried" [COVID Apps (Base Wording)] "Many COVID-19 apps are being used around the world to notify people who were in contact with someone infected (contact tracing apps). These apps record the interactions between users by detecting when two cell phones are close to each other. These apps require the participation of health agencies to con rm who tested positive for COVID-19. Do you support the government's participation in a COVID-19 contact tracing app?" • "Yes" • "Yes, but only if using the app is voluntary" [COVID Apps (Health Agency Wording)] "Many COVID-19 apps are being used around the world to notify people who were in contact with someone infected (contact tracing apps). These apps record the interactions between users by detecting when two cell phones are close to each other. These apps require the participation of health agencies to con rm who tested positive for COVID-19. In most cases, COVID-19 apps are designed to notify health agencies when someone was in contact with an infected individual. Do you support the government's participation in a COVID-19 contact tracing app?" • "Yes" • "Yes, but only if using the app is voluntary" [COVID Apps (Dilemma Wording)] "Many COVID-19 apps are being used around the world to notify people who were in contact with someone infected (contact tracing apps). These apps record the interactions between users by detecting when two cell phones are close to each other. These apps require the participation of health agencies to con rm who tested positive for COVID-19. Some people claim that COVID-19 apps may pose a risk to fundamental rights, such as the right to privacy. Others claim these apps are needed to help reopen the economy while protecting public health. Do you support the government's participation in a COVID-19 contact tracing app?" • "Yes" • "Yes, but only if using the app is voluntary" [Open-Ended] "We would like to understand public opinion about COVID-19 apps. Could you please give us the main reason for your previous answer, in one or two sentences, using the box below. " [Trudeau Approval] "How would you rate the Canadian government's overall response to the coronavirus outbreak?" • "Excellent" • "Good" • "Not so good" • "Poor" [Lost Job.] "Have you lost your job due to the COVID-19 lockdown?" • "Yes" • "No, but my hours were reduced" • "No" • "Not applicable/I did not work/I am retired" Table A2 reports the full results for the logistic regression models used to calculate the di erences in predicted probabilities reported in Figure 3 of the main text. The di erences in predicted probabilities are averaged across all respondents after leaving all other covariates at their observed values. While the trend in experimental research is to model average treatment e ects with linear regression, we should point out that the conclusions are the same using linear probability models. Table A3 shows the full models used to create Figure 4 in the main text, where the outcome variables are two types of argument mentioned in the answers to the open-ended question on COVID-19 cell phone contact tracing apps. Once again, we used logistic regression models and the gure in the main text reports di erences in predicted probabilities for a change from 0 to 1 in each independent variable, leaving other variables in the sample at their observed values. The e ect sizes are virtually the same as coe cients from linear regression models. Note that the treatment e ects on the other argument types are not statistically signi cant.

Extended Results
Finally, we report alternative speci cations of the main models used to measure treatment e ects in Figure 3. Table A4 shows the output from regression models without covariates, as well as a speci cation including an indicator of general support for "tougher" policies, which is an additive composite of four other survey questions asking respondents whether they support 1) mandatory facial masks in public transit, 2) the use of infrared cameras in public spaces, 3) mandatory COVID-19 testing in the workplace, and 4) stronger nes for people violating quarantine rules. In all cases, the dependent variable is the binary indicator of unconditional support for COVID-19 apps. These alternative speci cations produce results that are consistent with those reported in the main text, although the e ect of the "Not Serious Enough" covariate is not statistically signi cant. Notes: The table shows the full output of logistic regressions used to compute the di erences in predicted probabilities reported in Figure 3 of the main text. The dependent variable equals 1 if the respondent supports COVID apps unconditionally, and 0 otherwise. The rst model is unweighted. The last two models are computed using raking weights for interlocking quotas by age, gender and region. * p<0.05; * * p<0.01; * * * p<0.001 Notes: The table shows the full output of logistic regressions used to compute the di erences in predicted probabilities reported in Figure 4 of the main text. The dependent variable equals 1 if the respondent invoked the argument indicated in the column header, and 0 otherwise. The models are computed using raking weights for interlocking quotas by age, gender and region. * p<0.05; * * p<0.01; * * * p<0.001 Notes: Alternative speci cations using the unconditional support for COVID-19 apps as the binary outcome variable. All models are computed using raking weights. * p<0.05; * * p<0.01; * * * p<0.001

Content Analysis
In this section, we report the coding scheme used to annotate the written answers to the open-ended question on COVID-19 contact tracing applications. The comments were annotated independently by three coders. The coding scheme was created a priori (i.e. before commencing the manual annotation of comments) and is based on theory. The arguments against COVID-19 apps (concerns for privacy, the need to impose conditions, doubts about the e ectiveness, and dismissal of the threat) were all part of the public debate and prominent in the literature cited in the main text. The arguments in favour of COVID-19 apps were selected according to our theory. The three categories are meant to measure whether the respondent's considerations are oriented toward society ("Societal Concerns"), the threat posed by others ("Others as a Threat") or the application itself ("App is E ective"). In particular, the second category is the mechanism expected in the disease avoidance hypothesis, whereby individuals come to support the app because of the threat posed by other groups, either the infected themselves or people whose behaviour pose a risk of contagion.
Our coding scheme originally included a separate category for supportive arguments evoking economic concerns-the need for a contact tracing app to help reopen the economy. However, that argument was seldom evoked by respondents, and we ultimately merged it with the "Societal Concerns" category for simplicity of presentation. The third coder used the coding scheme with seven substantive categories from the start.
The coding scheme below corresponds to the document used as instructions for the classi cation of comments. Each category is binary. The classi cation is not mutually exclusive: for each comment, we indicate if an argument is present or not. As a result, some comments may include more than one of the arguments from the coding scheme. Some respondents also expressed their ambivalence and explicitly mentioned arguments for and against the use of COVID apps.

Coding Scheme
What are the considerations/arguments emphasized by the respondent to explain their position regarding cell phone contact tracing apps? For each comment, check all categories that apply.
1. Risk for privacy; impact on civil liberties e.g. the app infringes on privacy; app violates civil liberties; government may use the data to track people ("Civil Liberties" in Table 1) 2. The app must be restricted in scope; conditions must be in place e.g. the app must be voluntary; only a part of the population should be required to use it; government must guarantee the app won't be used after the pandemic ("With Conditions" in Table 1) 3. The app will not work e.g. not everyone has a phone; people may just stop carrying their phone; other methods are more useful ("Not Going to Work" in Table 1) 4. The threat isn't real e.g. the media/government exaggerated the threat of the virus; COVID is a conspiracy ("Threat Is Not Real" in Table 1) 5. Social bene ts (societal considerations) e.g. we must protect the vulnerable; we need to stop the virus; we need to reopen the economy; public health is more important than anything else ("Societal Concerns" in Table 1) 6. Other people are a source of risk (focusing on others as a threat) e.g. people not following the rules pose a threat; app is needed because people won't quarantine; we need to know where infected people are; we need to avoid hot spots ("Others as a Threat" in Table 1) 7. The app is useful (considerations focusing on the app itself) e.g. it's an e ective technology, the app will give useful information, the app worked well in other countries ("App is E ective" in Table 1) 8. Not applicable (other arguments) When the text makes another type of argument that does not t any of the categories.

No response
Non-response (blank text box) or gibberish comment.
The three human coders annotated all English language comments (978) and the average Cohen's Kappa coe cients (Cohen, 1960) are calculated on that common sample. The nal categories used for analysis are based on the majority choice for English language comments. Two out of the three coders annotated the French language comments; the few cases of disagreements were resolved manually by discussion among the two coders. To further assess the robustness of our results, we replicated the analysis reported in the main text using a unanimity rule for English language comments (i.e. an argument is considered to be present in the written answer only if all three coders agree). The ndings remain consistent to those reported in the text when using the unanimity rule (in fact, the con dence in our inferences improves slightly when using unanimity).