Association of smoking status with hospitalisation for COVID-19 compared with other respiratory viruses a year previous: a case-control study at a single UK National Health Service trust

Ethics statement

This study was approved by the UCL/UCLH Joint Research Office Research Strategy Group and UCLH Data Access Committee. Approval to conduct research limited to pseudonymised patient data was provided by the NHS Health Research Authority (IRAS_282704). The requirement for informed consent was waived by the NHS Health Research Authority due to the observational nature of the study.

Study design

This was an observational case-control study with historical controls, performed at a single National Health Service (NHS) hospital trust (comprising two hospital sites) in London, UK. The study protocol and analysis plan were pre-registered on the Open Science Framework in April 2020 (Simons et al., 2020a). The pre-registered protocol stipulated a non-inferiority design (i.e. a one-tailed statistical test) to maximise statistical power to detect a significantly lower proportion of current smokers (i.e. <10%) among patients hospitalised with COVID-19 compared with patients hospitalised with another respiratory viral infection a year previous (i.e. 20%). The protocol was amended after data collection but prior to statistical analysis in September 2020 to implement a traditional case-control design (i.e. a two-tailed statistical test), as a delay in study approval meant that the number of eligible cases and controls exceeded our expectations; providing sufficient power for a two-tailed test. We had also planned to compare current smoking in cases with age- and sex-matched London prevalence, with data obtained from the representative Annual Population Survey. However, following an external review on an earlier manuscript draft, we decided against presenting data from this comparison due to smoking rates in hospitalised populations typically being greater than in the general population (Benowitz et al., 2009).

A sample size calculation, updated after data collection but prior to data analysis, indicated that 363 cases and 109 controls would provide 80% power to detect a 10% difference in current smoking prevalence in cases compared with controls (e.g. 10% in cases and 20% in controls) with alpha set to 5%. We included all cases from 1^st March 2020 to the 26^th August 2020 (the date on which data were obtained) and all controls from the 1^st January 2019 and the 31^st December 2019.

Eligibility criteria

Inclusion criteria

Cases

Controls

Exclusion criteria

Outcome variable

The outcome of interest was the type of hospital admission (i.e. with COVID-19 vs. other respiratory viral infections a year previous).

Exposure variable

Smoking status (i.e. current, former, never) was obtained from the summary EHR or the medical notes. A number of cases were recorded as ‘non-smokers’ without distinguishing between ‘former smokers’ and ‘never smokers’. For the primary analysis, patients categorised as a ‘non-smoker’ were treated as ‘never smokers’. Where possible, information on use of smokeless tobacco, waterpipe and/or alternative nicotine products (e.g. e-cigarettes) was extracted. We searched within the contemporaneous medical records for free-text entries of smoking status. The most recent record of smoking status, obtained from either the summary EHR or the contemporaneous medical notes, was extracted. Where there was discordance between the summary EHR and the contemporaneous medical notes, smoking status was classified based on the contemporaneous medical notes (as this clinician obtained information was assumed to be more recent and therefore accurate). Where available, data on pack-year history of smoking (i.e. the number of packs of cigarettes smoked per day multiplied by the number of years of smoking, with a pack equal to 20 cigarettes) were extracted.

Covariates

Covariates included age, sex, ethnicity, socioeconomic position (SEP; with post codes linked by the research team to the Index of Multiple Deprivation (IMD) (Department for Communities and Local Government, 2019)) and comorbidities (classified by organ system, including cardiac, metabolic and respiratory diseases). Medical conditions not expected to be strongly associated with COVID-19 hospitalisation were not considered in the analyses (e.g. sciatica and fibromyalgia; see Extended data). Age was treated as a continuous variable in the primary analysis, with banded age groups (i.e. 18-29 years, 30-44 years, 45-59 years, 60-74 years, 75-89 years and > 90 years) used in exploratory analyses. The IMD was categorised as quintiles to reduce the impact of sparse data.

Strengths and limitations

To our knowledge, this is one of few studies specifically designed to examine the association between smoking status and hospitalisation with COVID-19. It was further strengthened by an assessment of the quality of data on smoking status gleaned from summary EHRs.

However, this study has several important limitations, the majority of which pertain to the selection of the controls. First, current smoking is expected a priori to be associated with hospitalisation for non-COVID-19 respiratory viruses (Stämpfli & Anderson, 2009). As concluded in the US Surgeon General’s 2014 report, smoking and infection with acute respiratory illnesses (including pneumonia) appear causally related. However, evidence as to the association between smoking and increased disease severity (once infected) is currently lacking. Ideally, hospital-based case-control studies should avoid selecting a control disease which is associated with the exposure of interest (i.e. smoking status) (Vandenbroucke & Pearce, 2012). However, to our knowledge, there is no other control disease with a similar route of acquisition and mechanism for hospitalisation/severe disease that is not a priori also associated with smoking status. The greater smoking prevalence in controls compared with the general population from which the cases emerged (Vandenbroucke & Pearce, 2012) therefore likely contributes to the significantly reduced odds of current smoking in our cases.

Second, the risk profile for controls likely differs from cases in that there is prior immunity to other respiratory viruses (e.g. influenza, respiratory syncytial virus), with no prior immunity in the population to SARS-CoV-2.

Third, we selected the controls on the basis of sharing a similar route of transmission and risk factors for hospitalisation as cases. However, at the time of writing (March 2021), we now suspect that COVID-19 differs from other respiratory viruses in several ways. For example, COVID-19 gains cell entry via the ACE-2 receptor (Hoffmann et al., 2020), with unknown receptor binding in flu (Killingley & Nguyen-Van-Tam, 2013) and appears to display less fomite and physical contact transmission than flu (Ben-Shmuel et al., 2020). In addition, emerging evidence suggests that COVID-19 has a significantly different pathological process compared with other respiratory viruses. For example, mortality rates from COVID-19 differ widely from those due to epidemic influenza (Office for National Statistics, 2020a). Although we currently do not know the importance of these factors, taken together, emerging observations may mean that direct comparison of risk profiles in cases and controls is limited.

Fourth, while no known behavioural restrictions were implemented during the control period, London was under lockdown restrictions from March to July 2020, which likely impacted the risk of viral exposure in cases (Davies et al., 2020). This may further have impacted the different risk profiles of controls and cases beyond the adjustments made in this analysis for sex, age and SEP.

Fifth, hospital admission routines were severely disrupted in many countries during the early stages of the COVID-19 pandemic (e.g., patient transfer between hospitals, altered admission criteria). This may have influenced patient selection and/or the recording of smoking status. With regards to patient selection, albeit possible, we are unaware of evidence from UK hospitals of patient characteristics correlated with smoking (e.g., age, SEP) determining whether a patient would be admitted, rather than their clinical situation on presentation. For example, smoking status has not been included in a widely used hospital-based algorithm for predicting clinical deterioration from COVID-19 (i.e., the ISARIC-4C deterioration score; Gupta et al., 2021). With regards to the recording of smoking status, to the best of our knowledge, the admission routines (including the medical history taking) within the selected hospital trust remained stable during the pandemic. However, patients admitted with COVID-19 may generally have been more unwell at the point of admission compared with controls, which may have led to less accurate documentation of smoking status. This potential bias was somewhat reduced by those surviving their COVID-19 illness being followed up through a specialised respiratory medicine clinic implemented during the pandemic, with smoking status ascertained during follow-up calls/visits. However, the finding that COVID-19 patients with unknown (compared with those with recorded) smoking status had greater in-hospital mortality means that residual bias cannot be ruled out. A sensitivity analysis with all missing assumed to be current smokers made the negative association of smoking status and COVID-19 go away. However, it should be noted that this is a very strong assumption, unlikely to hold true.

Sixth, the selection of historical controls may mean that there are non-trivial differences in smoking status between controls and cases due to a declining trend in London smoking prevalence (Office for National Statistics, 2020b). However, a single year was used for the selection of controls, and there was no important change in smoking prevalence among all (Jackson et al., 2021), so we expect any potential impact of pandemic-related trends in smoking prevalence on the results to be minimal. We considered using a contemporaneous control (i.e. patients hospitalised with other respiratory viral infections in 2020), which would have mitigated against this potential bias. However, due to factors such as reduced national and international travel, physical distancing, increased hand hygiene and potential viral dominance by COVID-19, exposure to and hospitalisation with other respiratory viruses has been substantially reduced in 2020 (GOV.UK, 2020), which would have limited the sample size for controls.

Seventh, given the broad and currently poorly understood COVID-19 pathology following SARS-CoV-2 infection, it was not possible to differentiate patients presenting to hospital due to COVID-19 or those admitted for other reasons. This may have introduced additional bias in that individuals may have been admitted to hospital with co-incident SARS-CoV-2 infection that was not associated with their current health need. We tried to mitigate this by only including cases if they had a positive COVID-19 test result less than 5 days from their date of hospital admission.

Eighth, a history of current or past cancer was high in both groups at greater than 20% and was significantly greater in controls compared with cases. This reflects a bias in the population that regularly interacts with the selected NHS hospital trust, which is a specialist cancer referral centre. We visualised the geographic regions where patients were admitted from to examine any systemic differences between cases and controls, and caution that the differing catchment areas of the two samples may have led to important differences in the underlying populations. In addition, during the peak of the first wave of the pandemic in the UK (i.e. March-April 2020), many cases were transferred across hospital sites due to bed pressures (Dunhill, 2020).

Finally, we have argued elsewhere that studies that aim to elucidate the potential causal relationship between smoking status and COVID-19 should first clarify the mechanism(s) through which smoking and/or nicotine use is expected to influence COVID-19 infection and/or disease outcomes contingent on infection and plan their studies accordingly (Perski et al., 2021). Considering these reflections, and following constructive peer review, we acknowledge that the present study through its selection of the control sample does not distinguish between two potential mechanisms: comparable exposure to SARS-CoV-2 virus with differential infection in current compared with never smokers, and comparable SARS-CoV-2 infection with differential hospitalisation in current compared with never smokers. Future research should aim to isolate these potential mechanisms by design, thus getting closer to unbiased estimates of the association of smoking/nicotine use with COVID-19 outcomes. However, we note that – given the time-varying dynamics of respiratory viruses – it would be incredibly challenging to design a representative study to compare rates of infection in those with similar exposure using available datasets. Similarly, as data linkage of representative infection surveys with hospital records is also limited at present, designing a study to examine the second potential pathway (i.e., examining hospitalisation rates in a representative sample of those infected) would also be very challenging.

Despite these limitations, alternative designs were impracticable or would have had different limitations. In the future, the current study can be considered alongside findings across multiple such alternative methodological approaches, each with different sources of bias, to triangulate on the extent to which associations between smoking and COVID-19 are causal.

Implications for policy and practice

COVID-19 will continue to place a large burden on healthcare services in the UK and internationally over the coming months and years. To mitigate against this, multiple non-pharmacological interventions are being implemented to reduce the intensity of demand on acute and intensive services. Irrespective of any direct link between smoking and COVID-19 disease outcomes, smoking is a significant cause for healthcare demand globally. We have argued elsewhere for the need to ramp up smoking cessation support to reduce the current and future burden on healthcare and social services (Simons et al., 2020a).

Avenues for future research

The selection of appropriate controls in hospital-based case-control studies is very challenging for a novel respiratory virus such as COVID-19 (which means we converged on a hybrid approach, combining elements from hospital-based case series and case-control designs with historical controls). We recommend the use of representative population-studies with data from multiple sites and with purposeful acquisition of smoking status, to better understand the role of smoking as a potential risk or protective factor for COVID-19 hospitalisation and disease severity.

Underlying data

Due to the sensitive nature of the data, we do not have ethical approval to release the individual-level data underpinning the analyses. Anonymised and de-identified individual-level data are available upon request from the corresponding author to bona fide researchers and following approval from the Biomedical Research Centre Clinical and Research Informatics Unit at University College London Hospital NHS foundation trust.

Extended data

OSF: Association of smoking status with hospitalisation for COVID-19 compared with other respiratory viruses a year previous: a case-control study at a single UK National Health Service trust: Protocol. https://doi.org/10.17605/OSF.IO/84VYD (Simons et al., 2021b).

This project contains the following extended data:

Data are available under the terms of the Creative Commons Attribution 4.0 International (CC-By Attribution 4.0 International).