Data sources

Data analysis

We calculated the proportions of outcomes among subgroups, considering several outcomes and subgroup definitions. Logistic regression was used to estimate the covariate-outcome association. Covariates considered varied for each analysis as not all variables available in each data set: sex, age, symptom (see Table D in S1 Text), symptom number and duration, symptom-to-survey time, self-reported years of education, index of multiple deprivation [IMD] [15], profession/work, self-reported or national rural-urban classification [RUC] [16]. For some analyses, Zoe reports of either loss of taste/smell or altered taste/smell were combined [16]. Zoe and UMD-CTIS analyses were conducted using Python 3.8 and R 3.6.3, respectively.

Ethics statement

Boston Children’s Hospital Institutional Review Board (IRB, P00023700) to use UMD Global CTIS data. The UMD IRB (1587016–10) approved the UMD Global CTIS study. King’s College London ethics committee (REMAS ID 18210; LRS-19/20-18210) to use Zoe data. Informed consent for participation in Zoe was provided by the users through the app upon sign-up, while UMD Global CTIS survey was provided by respondents prior to beginning the survey. Informed consent was documented in the digital platform by the respondent (no witness required). This study did not include minors.

Quantifying the testing gap

The proportion of Zoe participants reporting COVID-19 testing has increased over time (Fig 1), from <20% (1 April 2020) to >70% (1 January 2021). In mid-September, national PCR testing capacity was exceeded [17], coincident with a transient drop in reported tests. In late 2020, despite adequate test capacity, >25% of test-qualifying app users did not report a test. The UMD-CTIS survey time-series mirrors these trends (Fig A in S1 Text), albeit with a lower absolute proportion, likely due, in part, to survey design differences (e.g. ever tests and symptoms in prior 24 hours, vs tests -7 to +14 days from incident symptoms). The UMD-CTIS testing gap is generally lower amongst respondents with a smartphone, and even lower among symptom-tracking app participants (not necessarily Zoe).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Proportion of participants reporting a COVID-19 test among those reporting test-qualifying symptoms.

Proportion of participants reporting at least one test-qualifying symptom for the first time that logged a COVID-19 swab test -7 to +14 days after the onset of the symptom. Dashed vertical lines indicate the three-week study period (14 November 2020 to 8 December 2020) that was used to define participants eligible for a follow-up survey, i.e. those that reported test-qualifying symptoms for the first time but no swab test were invited to the Zoe Follow-Up Survey. Blue line calculated using data logged by 28 January 2020, red line calculated using data logged by 15 December 2020, which was the dataset used to identify target participants for the survey. The difference shows that some participants recorded test data after selection into the follow-up survey cohort. Grey bars indicate the UK PCR testing capacity, while green bars indicate PCR tests performed.

https://doi.org/10.1371/journal.pgph.0000028.g001

Symptom severity and not testing

To better understand the factors contributing to COVID-19 testing, Zoe participants with test-qualifying symptoms during the study period, who did not report testing (N = 20,425), were studied further. During this period, the proportion not tested among those with test-qualifying symptoms was higher for those with 1 vs ≥2 test-qualifying symptoms (27.1% vs 14.6%, p<0.001), and for those with symptoms lasting ≤2 days vs >2 (30.1% vs 14.6% p<0.001), (Table 1). Similarly, the proportion of ever-tested in UMD-CTIS was lowest among those with only one test-qualifying symptom or short symptom duration (Fig A in S1 Text). Females with qualifying symptoms were more likely to report not testing than males (26.2% vs 19.7%, p<0.001). Individuals with test-qualifying symptoms and age < 75 were more likely to report not testing than those age ≥ 75 (24.3% vs 16.9% P<0.001) (Table 1). A total of 1,254 users (26.6%) responded to the follow-up survey. Zoe and follow-up survey participants during this period were younger and more female than the general population, similar to the demographic trends reported previously in Zoe [13] and other digital health studies (Table E in S1 Text) [18, 19].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Characteristics of Zoe CSS app users reporting new-onset test-qualifying symptoms between 14 November and 8 December 2020, and eligible for the Zoe follow-up survey.

https://doi.org/10.1371/journal.pgph.0000028.t001

Journey to successful COVID-19 testing

In the Zoe follow-up survey, only 42.1% survey respondents recalled having experienced at least one test-qualifying symptom in the past month (Fig 2). Of those who recalled their symptoms, 54.7% recognised that these symptoms qualified them for a COVID-19 test. Among participants who recognized test-qualifying symptoms, they were likely to go on to attempt (85.6%) and then successfully obtain a COVID-19 test (93.0%, or 18.4% of all survey respondents).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Sankey flow diagram of Zoe follow-up survey responses by stage of the testing journey.

https://doi.org/10.1371/journal.pgph.0000028.g002

Recall of previously-reported test-qualifying symptoms

Among those who reported a single test-qualifying symptom (Table 2), recall of absent or altered taste/smell (45.8%) and cough (40.8%), was higher than fever (25.5%), the last of which, as queried in-app (“Do you have a fever or feel too hot?”) was an admixture of two symptoms that may not be recalled equally well. Recall of having experienced test-qualifying symptoms was associated with number of symptoms experienced (per test-qualifying symptom OR = 1.302 [95% CI 1.220–1.391]), symptom duration (per day OR = 1.065 [95% CI 1.054–1.076]) and recency (per symptom-to-survey days OR = 0.995 [95% CI 0.991–1.000]). Number of symptoms and symptom duration remained significantly associated with recall after adjusting for age, sex, and recency of symptom onset (model results in Table F in S1 Text).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Recall of test-qualifying symptoms in the Zoe follow-up survey.

https://doi.org/10.1371/journal.pgph.0000028.t002

Recognizing COVID-19 test-qualifying symptoms

Just 54.7% of those who recalled experiencing test-qualifying symptoms indicated their symptoms qualified them for a COVID-19 test. We queried the remaining respondents (N = 809, i.e. those who did not recall or who did not indicate the symptoms they recalled qualified them for testing) which symptoms would qualify them for a test. These respondents were similarly able to recognize fever (63.4%), cough (67.6%), and loss of smell (65.4%), and much less so for altered smell (29.7%). Only 59.6% identified the triad of fever, cough, and loss of smell as qualifying for a test, with lower recognition amongst the oldest age groups (Table 3). In univariate analyses, each decade older reduced the odds of recognizing the triad (OR = 0.908 [95% CI 0.883–0.933]). This finding remained largely unchanged after adjustment for sex, IMD, and rural-urban living. We found similar associations with the outcome of identifying each individual symptoms (model results in Table G in S1 Text). No associations were found for sex, age, IMD and rural-/urban living.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Understanding of symptoms qualifying for COVID-19 testing among respondents who did not recall having experienced a test-qualifying symptom.

https://doi.org/10.1371/journal.pgph.0000028.t003

Reasons for not testing among those who qualified for and wanted a test

In the follow-up survey, the proportion of respondents who recognised their symptoms qualified them for testing, and attempted, but did not succeed at testing was low (7%, N = 17 of 244, Fig 2). Due to limited sample size of this path to testing, other barriers could not be evaluated. We therefore evaluated complementary data from a subcohort of UMD-CTIS respondents from 21 December 2020 to 21 February 2021, who endorsed test-qualifying symptoms, who had never tested, and who indicated “yes” to the question “Have you wanted to get tested for coronavirus (COVID-19) at any time in the last 14 days?” (N = 1,956, Table 4). Among those who wanted testing, “I don’t know where to go” was the most frequently selected option (32.4%). The other multi-choice reasons were: “I am unable to travel to a testing location” (29.1%), “I tried to get a test but was not able to get one” (25.6%), “I am worried about bad things happening to me or my family (including discrimination, government policies, and social stigma)” (18.4%), “I can’t afford the cost of the test” (17.9%), and “I don’t have time to get tested”(13.3%). Given the scope of this study, we have focused on the knowledge-based response, though we acknowledge the logistical barriers are important.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Characteristics of UK Global COVID-19 Trends and Impacts Survey (UMD-CTIS) survey respondents reporting test-qualifying symptoms, never testing, but wanting to test in the prior 14 days, for surveys between 21 December 2020 through February 21, 2021.

https://doi.org/10.1371/journal.pgph.0000028.t004

Not knowing where to test among the symptomatic wanting testing.

We further investigated demographic factors associated with not knowing where to go to obtain a test (Fig 3). Not knowing where to go to obtain a test (“yes” vs referent “no”) was associated with older age (per decade OR = 1.207 [1.129–1.292]) and less education (per 4-years OR = 0.685 [0.599–0.783]). Male sex (OR = 1.334 [1.064–1.675]) and living outside a city (OR = 1.201 [0.926–1.562]) were not significant with Bonferroni correction for multiple hypothesis testing (p-threshold 0.0125 = 0.05/4). Education was similarly protective for not knowing where to test adjusting for age and sex, use of survey weights, or assuming missing responses were “no” (models results in Table H in S1 Text).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Not knowing where to get a COVID-19 test stratified by age and education.

Proportion (95% confidence interval) of respondents who indicated “yes” (per “yes” plus “no”) for “I don’t know where to go” as an option to the question “Do any of the following reasons describe why you haven’t been tested for coronavirus (COVID-19) in the last X days?”, where X is the self-reported duration of symptom capped at 14 days. This was restricted to untested respondents with test-qualifying symptoms. Proportions stratified by age (top) and education (bottom).

https://doi.org/10.1371/journal.pgph.0000028.g003

Acknowledging our limited sample size, we conducted qualitative, hypothesis-generating analyses of other demographic factors correlated with knowledge barriers (Fig B in S1 Text). While cities were not protective in the regression model, the proportion not knowing where to test was slightly lower in London than elsewhere (Fig C in S1 Text). Having a smartphone and using a symptom-tracking app qualitatively had a bigger impact on the knowledge gap than the relatively small urban-rural and regional differences. There were modest qualitative differences in not knowing where to test by profession, with the highest proportions among those in transportation, tourism and construction and the lowest in finance, public administration and health.

Persistent testing gap

Our analysis finds that in December 2020 approximately one quarter of symptomatic UK Zoe participants who qualified for a COVID-19 test did not undergo testing. The proportion of ever-tested recently-symptomatic UK UMD-CTIS respondents echoes this trend. While we show a substantial improvement from April, 2020, the persistent testing gap is problematic for pandemic management in the UK and elsewhere. Non-pharmaceutical mitigation strategies are likely to be required [20–22], despite effective vaccines, because of COVID-19 transmissibility and the anticipated time to reach herd immunity, even with a one-dose immunization strategy [23]. The lower the proportion of identified infections, such as through insufficient testing of symptomatic cases, the more likely transmission events will go unchecked.

Knowledge barriers to testing

With this testing gap in mind, we sought to characterize barriers to testing that might inform improvements to the UK testing programme using data from two large, complementary surveillance platforms. Through analysis of prospective self-reported testing outcomes, along with follow-up and population-sampled surveys, we identified three key knowledge barriers along the path to successful COVID-19 testing.

Firstly, the association of less testing with brief and/or single test-qualifying symptoms suggests an implied severity threshold for testing that is inconsistent with guidance [5], and with the spectrum of infectious COVID-19; given those with asymptomatic disease can transmit, it is reasonable to expect that those experiencing just a single symptom, or brief duration of symptoms, are able to transmit and need to be tested [24]. Individuals may minimise their symptoms, particularly if they perceive their risk of contracting COVID-19 as low, or possibly hold the misconceived notion that COVID-19 manifests in a stereotypical manner [25]. In a preprint DHSC report utilizing online surveys May 25 to August 5, 2020, those with test-qualifying symptoms did not request testing because symptoms were mild (16.0%), improved (16.1%), or they did not think symptoms were due to COVID-19 (20%) [8]. Secondly, we show that four of ten did not recognize all three of the triad of UK test-qualifying—i.e. fever, cough and loss of smell. The DHSC report [8] estimated 51.1% of all respondents failed to recognize the triad. Despite increased press coverage, across newspapers, TV, and radio, and a second wave with mitigation intensification [26], recognition of these test-qualifying symptoms in our survey around six months later increased by ~10% and therefore remains alarmingly low. Lastly, one third of those who wanted a COVID-19 test cited not knowing where to test as a contributing factor to their not getting one. This was the most frequently cited reason among the six options. Thus, despite the fact that survey respondents are possibly more aware of health information, they acknowledge challenges to finding testing that are comparable to, if not greater than, logistical barriers to testing (e.g. travel, time), even within the framework of the more centralized UK testing programme.

Public health implications

Our findings have significant public health implications. The UK NHS testing programme offers free COVID-19 tests to those with test-qualifying symptoms, with the list of qualifying symptoms unchanged since loss/alteration to taste and smell were included on 18 May 2020 [27], and tests accessed through a central booking system [5]. Risk mitigation and public health principles generally would agree with these key features of the UK program i.e. the use of concise and consistent guidance, and limiting logistical barriers to following guidance. In this sense, the UK is a sort of case study of the “best case scenario”, and yet there is still a significant gap in understanding. Other testing barriers do exist and should be addressed. Nevertheless, we highlight through this study that, for many with test-qualifying symptoms, knowing when to test is an early essential factor and knowing where to test is a later contributing factor to inadequate testing of potential COVID-19 cases. Not only are greater efforts needed to educate the UK public, it is likely that comparable efforts to mind the knowledge gap will be needed in countries with regionally varying testing criteria or methods of accessing testing.

Our work suggests there is a need for messaging improvements to the UK testing campaign. In our study, among the untested who qualified for a test, older age was associated with not knowing when and where to test. In the earlier DHSC report [8], older age was generally protective with respect to testing knowledge and behaviors, perhaps suggesting knowledge gains in the young over the past six months. Fewer years of education was also associated with not knowing where to test in our study. We found suggestive evidence that the knowledge gap may be more pronounced among those who do not have smartphones. Older populations in pre-pandemic studies have slower adoption of certain technologies, yet the abrupt social isolation resulting from mitigation strategies may be leaving important segments of the population behind [28]. Education attained and age are likely not the root cause of not knowing when and where to test. Rather they likely highlight pre-existing health-information disparities that have been exacerbated by a year of unprecedented changes in how individuals interface with each other and the world.

Our findings support the need for targeted messaging to certain at-risk demographic groups, possibly in a non-digital format (e.g. radio, community signage). These efforts would be aided by further research seeking to understand where people obtain their information about COVID-19—information we did not collect in our study. Our findings are particularly timely in light of work showing that expansion of the symptoms that qualify for a test would help detect more cases, assuming those who qualify do indeed successfully test [29, 30]. This theoretical gain in case detection could be lost if the change in tack leaves vulnerable populations behind. There is overlap between knowledge risk factors and COVID-19 risk, such as older age [31], though we did see a higher absolute rate of testing in the oldest age group. Overlap with vaccine hesitancy risk factors may further amplify disparities in healthcare access, leaving some groups both less tested and less protected.

Furthermore, messaging could also emphasise that even individuals with mild or transient symptoms may have COVID-19 and should get tested. COVID-19 has a broad spectrum of disease severity with a substantial number of cases being fully asymptomatic, and with asymptomatic carriers still being able to transmit, albeit at reduced rates [24].

Strengths and limitations

The Zoe platform affords a unique opportunity to prospectively link testing behaviours with incident symptoms in a large user base comprising ~6% of the UK population. The UMD-CTIS platform, though smaller in size and slightly different in survey design, corroborates temporal trends over in the broader population. To our knowledge, this has enabled the first time-varying estimate of testing rates amongst individuals that qualify for COVID-19 tests over the course of the pandemic. Both platforms could be leveraged to track the testing and knowledge gaps, in real time, allowing the effectiveness of interventions, such as improved messaging on when and where to test, to be assessed.

We acknowledge a number of limitations to this study. Digital surveys include selected populations not necessarily representative of the wider population. Such platforms have well-documented biases in demographic age, sex, and socioeconomic factors which we adjusted for in our analyses [18, 32]. In addition, digital surveys may not be generalizable, as they may be enriched for health-conscious internet-connected participants, and thus underestimate disparities in at-risk demographic groups. We show that symptom-tracking app participants and those with smartphones have higher testing rates than all UMD-CTIS survey respondents.

Confounding and measurement bias in this observational study using self-reported covariates and outcomes may also cause us to miss other important issues related to testing. We adjusted for common confounders, and attempted to identify proxies for the knowledge gap rather than attribute causality. There is no timely, efficient trial to conduct analyses of this scale. Self-report could introduce non-differential and differential measurement error, including the possibility of some events being omitted, or recorded inaccurately or inappropriately. Furthermore, the financial implications of having to self-isolate disproportionately affect the poorest, and may increase unwillingness to test [33] and respondents may be wary of self-reporting socially stigmatized reasons for not complying with guidance. We examined the association between knowledge barriers and level of education and profession, but were not able to examine other socioeconomic factors, such as income. Furthermore, this study does not explore the joint effect of multiple barriers (knowledge, logistical, financial or other), nor the potential benefit of facilitators. Lastly, selection can theoretically induce collider bias [34] if the exposure and outcome are both causes of participation or subpopulation selection.