The relative incidence of COVID-19 in healthcare workers versus non-healthcare workers: evidence from a web-based survey of Facebook users in the United States

Study design

We analyzed individual participant data from a large, web-based survey of Facebook users aged 18 and above in the United States (around 300,000 respondents per week). Every day Facebook offered a random sample of US-based users a Qualtrics survey run by the Delphi lab at Carnegie Mellon University who made it rapidly available to other academic researchers⁶. Facebook also provided survey weights to adjust for the demographics of the active Facebook user population^7,8. This sort of survey data has been used previously to perform population based analyses related to COVID-19, though never before at such large scale^9,10. Our analysis relied on the responses to two lines of questions: (1) questions about recent work history, worded as, “In the past 4 weeks, did you do any kind of work for pay?” and if so, “[p]lease select the occupational group that best fits the main kind of work you were doing in the last four weeks”; and (2) questions about COVID-19 testing history, worded as, “Have you ever been tested for coronavirus (COVID-19)?”, “[h]ave you been tested for coronavirus (COVID-19) in the last 14 days?”, “[d]id this test find that you had coronavirus (COVID-19)”, and “[d]o any of the following reasons describe why you were tested for coronavirus (COVID-19) in the last 14 days? Please select all that apply.”

We analyzed the most recently available six weeks of data from September 6, 2020 to October 18, 2020, which provided more than 80% power to detect a 30% difference between COVID-19 prevalence in HCWs and non-HCWs (details below).

Variables

To quantify the relative risk of COVID-19 among healthcare workers (HCWs) versus workers in non-healthcare occupations (non-HCWs), we used the response to the occupational group question as our exposure variable (we coded respondents who selected option “Healthcare practitioners and technicians” or “Healthcare support” as HCWs, and all others, including those with a missing value, as non-HCWs). We identified individuals with COVID-19 as those who reported that they had tested positive for COVID-19 in the last 14 days.

Statistical methods

We calculated the endorsement rate of positive COVID-19 test (ER) for the HCW and non-HCW population as the survey-weighted percent of respondents in either group who reported COVID-19, and calculated the relative COVID-19 incidence ratio (RR) with the equation

   RR = (ER among HCWs) / (ER among non-HCWs).

We quantified the uncertainty in this ratio using non-parametric bootstrap resampling to obtain a 95% uncertainty interval¹¹. To control for confounding due to differential access to COVID-19 testing, we restricted our analysis to only HCWs and non-HCWs who were tested in the last 14 days because their employer or school required it.

As sensitivity analyses, we considered also alternative inclusion criteria and more restrictive subsets of HCWs. The survey provided sample weights that adjust for non-response bias, which we used in our main analysis. As a sensitivity analysis, we repeated our calculation using the unweighted data. To investigate the possibility that workplace testing practices differ between HCW and non-HCW occupational settings, we also repeated our analysis with additional filtering based on the “why you were tested” question. In the main result we used the subset of individuals who responded that they were tested in the last 14 days because of employer/educational requirements, and this question has a “select all that apply” answer type, and also includes “I felt sick” as an option. As a sensitivity analysis, we used only those individuals who were tested because of a workplace requirement and did not feel sick.

Power calculation: To determine the sample size necessary to detect a difference of 30% between the COVID-19 prevalence of HCWs and non-HCWs, we developed a small simulation model where the fraction of HCWs in the general population and the COVID-19 prevalence in the general population both match that observed in the survey data.

Of respondents who were tested in the last 14 days because their employer or school required it, 33.9% were HCWs and 4.9% tested positive for COVID-19, so we simulated populations of size n with these fractions of HCWs and this positive rate among the non-HCW population. We made the positive rate among the HCW population 30% lower:

def sim_data(n_simulants):
    frac_hcw = .339
    frac_cli = .049
    rr_hcw = 0.7
    
    data = pd.DataFrame(index=range(n_simulants))
    data['hcw']  = np.random.uniform(size=n_simulants) < frac_hcw
    cli_pr = np.where(data.hcw, rr_hcw * frac_cli, frac_cli)
    data['cli'] = np.random.uniform(size=n_simulants) < cli_pr
    return data

Then for populations of ranging in size from n = 500 to 9,500, we repeatedly synthesized a simulated population, calculated the RR of COVID-19 between the HCWs and non-HCWs as described in the main text, and checked if the upper bound of the uncertainty interval was less than 1.0. We replicated this experiment 10,000 times for each population size n and found the n where at least 80% of the experimental replications where the uncertainty interval upper bound was less than one.

Ethical statement

These research activities used no identifiable private information and were therefore exempt from institutional board review.

Sensitivity analyses

When we repeated our calculation using the unweighted survey responses to calculate the COVID-19 incidence ratio, we found an even smaller relative incidence ratio of 0.4 (95% UI 0.3 to 0.5).

When we repeated our analysis restricted to only specific types of HCWs, as afforded by the questionnaire, we found a range of risks, usually less than 1.0, with substantially less certainty due to small sample sizes (Table 3).

When we used only those individuals who were tested because of a workplace requirement and did not feel sick, we obtained a relative risk closer to 1.0. Using only those tested because of a workplace requirement who also did feel sick we still obtained a relative risk substantially smaller than 1.0 (Table 4). Although this finding could suggest that differences in testing patterns between healthcare and other work settings are partially responsible for the different positivity rates among HCWs and non-HCWs, it could also be driven by greater access to COVID-19 testing for confirmation of illness among HCWs experiencing symptoms. The recall period of 14 days provides ample time for an individual to receive a workplace test without symptoms, then develop symptoms, and then receive another test to determine if the symptoms are due to COVID-19, and HCWs might have more opportunity to access such a follow-up test, since they are visiting a healthcare setting for work already.

Limitations

This work has at least three limitations. First, our results are based on self-reported data and therefore subject to both recall bias and social desirability bias, although the questions we relied on did not seem particularly at risk for either of these biases; the question “have you been tested for COVID-19 in the last 14 days?” likely included positive responses from individuals who received seroprevalence testing as well as PCR testing as well, which could also introduce a small amount of bias. Second, our approach required a large sample size to obtain a sufficiently precise estimate of RR, but this seems safer than including respondents who did not report receiving a required test, as that could introduce confounding. Third, it is possible that there was still uncontrolled confounding due to differential access to tests between HCWs and non-HCWs. Our sensitivity analysis found substantively similar results when restricted only to individuals who had workplace testing when they did not feel sick, but since we have only considered respondents with tests required by their employer or school, this might focus on non-HCW setting with better-than-average infection control policies (for example, they are doing asymptomatic testing) and therefore the relative risk for HCWs might be even lower than our method estimated.

Underlying data

The underlying data used in this study are available to academic researchers for research purposes from Facebook at: https://www.facebook.com/research-operations/rfp/?title=covid19-symptom-survey-data-access. Conditions of access and instructions for applications can be found at https://dataforgood.fb.com/docs/covid-19-symptom-survey-request-for-data-access/.