The strength of our study is that centralised, routinely collected national data sources were used which recorded all positive COVID-19 test results from Pillars 1 and 2 of the UK government’s public testing programmes, and all NHS hospital attendances and admissions in England. The latter represented approximately 98% of all hospital activity in the country.³⁰ Pillar 1 tests were provided by NHS and PHE laboratories for community cases from January 16^th to March 12^th, 2020, for hospitalised cases and the investigation of care home outbreaks from March 12^th to March 31^st, 2020, and for healthcare workers and their families from April 1^st onwards (where additional capacity was available). Pillar 2 testing was delivered by central government through academic, public and private partnerships: from April 1^st 2020 it was progressively rolled out to key workers in the NHS, social care and other critical sectors; from May 23^rd the general population could also access testing from this route; from August 1^st it was used to test asymptomatic staff and residents in care homes and for asymptomatic contacts in outbreak investigations. We saw a large expected decrease in the proportion of cases with a temporal link to a hospital record (from 79% in March to 7% in August), as testing policy across the UK expanded from an initial focus on testing in hospitals to community testing.

The main limitation of our study is that our case numbers reflect national testing activity, not the true number of cases in the population, and this activity was severely constrained by testing capacity during the phase of the pandemic covered by our study. As of September 8^th, 2020, it was estimated from household survey data that approximately 2.8 million people (95% CI 2.4 to 3.2 million people) aged 16 years and over would have antibodies to COVID-19.³¹ Therefore, the 15,564 HO.pHA and HO.HA cases during this six-month period represent approximately 1% of all cases in England at the time of this estimate. While numbers of hospital-onset cases should be closer to the true number of cases, assuming that patients in hospital were more likely to be tested, an unknown number of asymptomatic cases will have been missed where inpatients were not routinely swabbed. Systematic testing of inpatients in hospitals in England did not start until 24^th June. Similarly, cases classified as probable or definite hospital-onset may have been infected before admission or during the first 7 days of admission but were not tested until they became symptomatic, or they may have had negative test results, which were not available for our analysis. The overall effect of these limitations will likely have been to over-estimate probable and definite hospital-onset case numbers. Linkage of national surveillance and hospital activity data will be imperfect, and our algorithm made assumptions in assigning a priority order when test result data linked to more than one attendance or admission. For CO.pHA cases, we treated all hospital admissions equally regardless of duration of healthcare exposure. Conversely, we would not have captured infections potentially acquired from primary care, outpatient and emergency department attendances. Our data did not allow us to identify healthcare workers whose infection may have been acquired in the workplace, and we would have misclassified these cases as community-acquired. Our analysis used test and admission dates rather than dates and times because time of day was not recorded in our test or inpatient data sources, therefore our results will not be exactly comparable with classifications based on exact time, i.e., ≤48 hours rather than ≤2 days.