The epidemiological impact of the NHS COVID-19 app – Nature

No statistical strategies have been used to predetermine pattern measurement. The experiments weren’t randomized. The investigators weren’t blinded to allocation all through experiments and consequence evaluate.

Estimating app uptake

To watch the secure serve as of the app and allow its analysis, a restricted quantity of knowledge are shared with a protected NHS server. Every lively app sends a unmarried knowledge packet day-to-day. The fields in those packets include no delicate or figuring out knowledge, and are authorized and publicly indexed by means of the Data Commissioner ( The uncooked knowledge fields we used are described in Supplementary Desk 2; additional variables derived from those are described in Supplementary Desk 3. A schematic representation of knowledge amassing is proven in Prolonged Information Fig. 5. For the reported numbers of downloads, repeat downloads to the similar telephone are counted simplest as soon as. The selection of lively customers on a daily basis is outlined because the selection of knowledge packets gained by means of the NHS server; for a unmarried consultant worth of this amount, we took the imply over all days from 1 November to 11 December 2020 (previous knowledge used to be deemed much less dependable). We word that there proceed to be unexplained fluctuations in reported consumer numbers on Android telephones. To estimate uptake inside an LTLA, every postcode district used to be mapped to the LTLA wherein the vast majority of its inhabitants live, and we took the ratio (selection of lively customers in postcode districts mapped to this LTLA)/(overall inhabitants in postcode districts mapped to this LTLA). The inhabitants of England and Wales is 59.4 million, of whom 48.1 million are 16 or over and thus eligible to make use of the app ( We assumed that England and Wales are consultant of the United Kingdom, wherein 82% of folks elderly 16 years and over have smartphones (OFCOM, private communique), and that of smartphones in movement, 87% enhance the Google Apple Publicity Notification machine (Division of Well being and Social Care, private communique). The denominators for measuring uptake on the nationwide stage are subsequently 59.4 million (overall inhabitants) and 34.3 million (eligible inhabitants with appropriate telephones).

Defining numbers of instances

The COVID-19 case numbers consistent with day we used listed below are the ones reported at, by means of specimen date and LTLA. We got per-capita case numbers on the LTLA stage by means of dividing by means of LTLA populations reported by means of ONS. Those consistent with capita case numbers by means of section are proven in Prolonged Information Fig. 6. Trying out has been to be had in the course of the NHS Take a look at and Hint machine in all spaces all over the length, with a mean lengthen of not up to 2 days from reserving a check to receiving the end result. Trying out capability has most commonly exceeded call for, except for for 2 weeks in early September. We assumed that case ascertainment has been reasonably consistent over the length of study, an assumption qualitatively supported by means of the independent ONS and REACT research26,27.

Estimating the SAR

We desirous about a length in December 2020 and January 2021 when the selection of sure check ends up in app customers might be disaggregated by means of whether or not the consumer have been lately notified or now not. Even with this knowledge, successive knowledge packets despatched by means of the similar tool aren’t connected to one another. Because of this when a given selection of notifications are despatched on a selected day, the precise selection of the ones people notified who later obtain a good check result’s unknown, as a result of the loss of linkage over the years. We subsequently used a probabilistic style for what number of sure check effects we might be expecting amongst the ones lately notified, as a serve as of the selection of notifications on earlier days, of the estimated lengthen from notification to checking out sure, and of the SAR. We estimated the SAR by means of maximising the chance of this style. Intimately: let  fNP(t) be the likelihood that a person notified on a given day then exams sure t days later (conditional on their checking out sure at some later time, this is, the serve as is normalized to at least one). Let N(t) be the selection of people notified on day t, and IN(t) the selection of people reporting a good check on day t having been notified lately (both they’re these days within the quarantine length really useful by means of the app, or the next 14 days). The quantity anticipated for the latter is  ({rm{S}}{rm{A}}{rm{R}}instances sum _{{t}^{{top} }le t}{f}_{NP}(t-{t}^{{top} })N({t}^{{top} })), and we maximised a Poisson probability for the quantity noticed, IN(t) (proven in Prolonged Information Fig. 1d), given the quantity anticipated, treating observations from other days as unbiased. The boldness period used to be got by means of probability profiling; then again, sensitivity analyses steered higher uncertainty (see Supplementary Information).  fNP(t) used to be calculated as a convolution of the distributions for instances from publicity to signs, from signs to checking out sure, and from publicity to notification (Supplementary Information). Our SAR calculation used simplest knowledge from iPhones, with the exception of Android telephones, for extra strong day-to-day numbers of study packets.

Modelling instances avoided in response to notifications and SAR

The impact of notifications gained at time t on instances avoided can also be modelled because the fabricated from (i) the selection of notifications, (ii) the secondary assault charge, this is, a conservative underestimate of the likelihood that notified persons are in fact inflamed, (iii) the predicted fraction of transmissions preventable by means of strict quarantine of an infectious particular person after a notification, (iv) the real adherence to quarantine, and (v) the predicted measurement of the whole transmission chain that will be originated by means of the touch if now not notified. Ahead of every notification, the touch’s app sends a request for permission to the central NHS server. We estimated the whole selection of notifications consistent with day on every working machine (OS; being both Android or iOS) from those requests. We estimated the selection of notifications consistent with LTLA from the selection of partial days of quarantine (normally akin to the primary day of quarantine, this is, the day of notification) consistent with day, OS and LTLA, rescaling it by means of a time- and OS-dependent issue to check the selection of notifications consistent with day and OS. The geographical variability in notifications after summing over the years is proven in Supplementary Fig. 1. The lengthen between final publicity and notification is believed to observe a standard distribution, with time-dependent parameters estimated by means of least squares from the day-to-day selection of notifications and people in quarantine. The fraction of preventable transmissions is estimated from the lengthen distribution the usage of the technology time distribution in28 with imply 5.5 days. For the effectiveness of quarantine in decreasing transmission from traced contacts, we assumed as our central worth that 45.5% of traced contacts quarantine completely (100% aid in transmission), 31% of traced contacts quarantine imperfectly with 50% aid in transmission, and 23.5% of traced contacts don’t quarantine in any respect (0% aid in transmission). That is an identical to a median effectiveness of quarantine of 61%. In spite of everything, the dimensions of the epidemic chain caused by means of a unmarried case is computed assuming that native epidemics don’t combine and that the additional instances don’t have an effect on the epidemic dynamic. See Supplementary Information for additional main points.

Statistical research

The primary statistical research in comparison statistics for every LTLA, labelled x, to these of the set comprising all of its ‘matched’ neighbours N(x) = {n1, n2, n3,…, and so forth}. The matched neighbours N(x) have been outlined as different LTLAs that percentage a border with x and have been in the similar quintile for selection of instances consistent with capita in section 0. Distributions appearing the range between LTLAs within the selection of neighbours and selection of matched neighbours are proven in Supplementary Fig. 2. Stratification into quintiles (versus deciles and so forth) used to be selected to steadiness energy and enough adjustment; no different risk used to be attempted, to protect towards investigator bias.

Every statistic of passion used to be averaged over the matched neighbours, weighting by means of inhabitants measurement, to acquire the imply worth within the matched neighbours of x. This used to be in comparison to the statistic for x. Linear regression used to be performed the usage of, for every statistic of passion, the variation between its worth in x and in its matched neighbours N(x). The statistics we thought to be have been: consistent with capita selection of instances in every section; the fraction of the inhabitants the usage of the app; a measure of rural/city combine on a scale from 1 to five, from the Place of job of Nationwide Statistics (ONS); a measure of native GDP consistent with capita from the ONS, adjusted for rural/city ranking; and a measure of the fraction of the inhabitants residing in poverty ahead of housing prices, from the ONS.

Our major regression used to be

log(cumulative instances consistent with capita in x) – log(cumulative instances consistent with capita in N(x)) =

beta_rural_urban × (rural/city ranking of x − rural/city ranking of N(x)) +

beta_gdp_band × (native GDP band of x − native GDP band of N(x)) +

beta_poverty × (consistent with cent of the inhabitants residing in poverty in x – consistent with cent of the inhabitants residing in poverty in N(x)) +

beta_users × (consistent with cent of the inhabitants the usage of the app in x − consistent with cent of the inhabitants the usage of the app in N(x)) +


the place the other knowledge issues for the regression (the other values of x) have been the set of LTLAs with no less than one matched neighbour, with the exception of LTLAs with out a matched neighbours. Cumulative instances have been thought to be in every of the 3 levels one after the other or with levels 1 and a pair of, as reported in our effects. The values of the beta coefficients we estimated are proven in Prolonged Information Desk 2. We used a logarithmic turn out to be for the reaction variable in our regression, as a result of instances are generated by means of an exponential procedure (transmission) and so the velocity at which the selection of instances varies with the dose of a remedy (this is, the level of an intervention) is very confounded with absolutely the selection of instances. A regression with quadratic impact of uptake and intercept at 0 produced very equivalent findings to the above regression with linear impact of uptake (now not proven). We thought to be further uncertainty within the regression because of redundancy within the variations manner, for instance, in evaluating each LTLA x with LTLA n and LTLA n with LTLA x, described within the bootstrapping phase of Supplementary Information.

Predictions for instances avoided have been discovered the usage of the regression coefficient beta_uptake to linearly extrapolate log(cumulative instances consistent with capita) for every LTLA to that anticipated for an uptake of 15% (or holding instances counts as they have been, if uptake used to be already not up to 15%). Right here we assumed that there’s negligible advantage of app uptake beneath 15% (although this isn’t anticipated to be the case in settings the place utilization is clustered into high-uptake communities29). The definition of beta_users within the regression equation above manner it’s the anticipated build up in log(cumulative instances consistent with capita) related to a one-percentage-point build up in app uptake, when holding consistent GDP, rural/city combine, and stage of poverty. Our central estimate of beta_users on this research used to be −0.023 for section 1 and a pair of blended; this implies an build up in uptake of p proportion issues is anticipated to be related to an build up by means of the issue e−0.023p within the cumulative selection of instances consistent with capita in levels 1 and a pair of. An build up of p = 1 proportion issues in uptake manner a lower of two.3% in instances as we reported above. We estimated the selection of deaths avoided by means of multiplying the selection of instances avoided by means of the crude case fatality charge.

Choice regressions are described in Supplementary Information; their effects are in Prolonged Information Tables 3 and 4, and Prolonged Information Fig. 2.

Case fatality charge

The case fatality charge used to be estimated because the ratio of overall deaths (27,922) to instances (1,891,777) for levels 1 and a pair of blended. To check for heterogeneity, it used to be additionally estimated because the regression of native deaths to instances, however no considerable heterogeneity used to be noticed (now not proven). This is a lower-bound because of proper censoring of the time sequence of deaths.

Reporting abstract

Additional knowledge on analysis design is to be had within the Nature Research Reporting Summary connected to this paper.

Similar Posts