The association of smoking status with SARS-CoV-2 infection, hospitalisation and mortality from COVID-19: A living rapid evidence review with Bayesian meta-analyses (version 11)

Aims: To estimate the association of smoking status with rates of i) infection, ii) hospitalisation, iii) disease severity, and iv) mortality from SARS-CoV-2/COVID-19 disease. Design: Living rapid review of observational and experimental studies with random-effects hierarchical Bayesian metaanalyses. Published articles and pre-prints were identified via MEDLINE and medRxiv. Setting: Community or hospital. No restrictions on location. Participants: Adults who received a SARS-CoV-2 test or a COVID-19 diagnosis. Measurements: Outcomes were SARS-CoV-2 infection, hospitalisation, disease severity and mortality stratified by smoking status. Study quality was assessed (i.e. ‘good’, ‘fair’ and ‘poor’). Findings: v11 (searches up to 2021-02-16) included 405 studies with 62 ‘good’ and ‘fair’ quality studies included in unadjusted meta-analyses. 121 studies (29.9%) reported current, former and never smoking status with the remainder using broader categories. Recorded smoking prevalence among people with COVID-19 was generally lower than national prevalence. Current compared with never smokers were at reduced risk of SARS-CoV-2 infection (RR = 0.71, 95% Credible Interval (CrI) = 0.61-0.82, τ = 0.34). Data for former smokers were inconclusive (RR = 1.03, 95% CrI = 0.95-1.11, τ = 0.17) but favoured there being no important association (4% probability of RR ≥1.1). Former compared with never smokers were at increased risk of hospitalisation (RR = 1.19, CrI = 1.1-1.29, τ = 0.13), greater disease severity (RR = 1.8, CrI = 1.27-2.55, τ = 0.46) and mortality (RR = 1.56, CrI = 1.23-2, τ = 0.43). Data for current smokers on hospitalisation, disease severity and mortality were inconclusive (RR = 1.1, 95% CrI = 0.99-1.21, τ = 0.15; RR 1.26, Qeios, CC-BY 4.0 · Article, March 2, 2021 Qeios ID: UJR2AW.13 · https://doi.org/10.32388/UJR2AW.13 1/85 95% CrI = 0.92-1.73, τ = 0.32; RR = 1.12, 95% CrI = 0.84-1.47, τ = 0.42, respectively) but favoured there being no important associations with hospitalisation and mortality (49% and 56% probability of RR ≥1.1, respectively) and a small but important association with disease severity (83% probability of RR ≥1.1). Conclusions: Compared with never smokers, current smokers appear to be at reduced risk of SARS-CoV-2 infection while former smokers appear to be at increased risk of hospitalisation, greater disease severity and mortality from COVID-19. However, it is uncertain whether these associations are causal. v7 of this living review article has been published in Addiction


Introduction
COVID-19 is a respiratory disease caused by the SARS-CoV-2 virus. Large age and gender differences in case severity and mortality have been observed in the ongoing COVID-19 pandemic (Guan, Ni, et al., 2020); however, these differences are currently unexplained. SARS-CoV-2 enters epithelial cells through the angiotensin-converting enzyme 2 (ACE-2) receptor (Hoffmann et al., 2020). Some evidence suggests that gene expression and subsequent receptor levels are elevated in the airway and oral epithelium of current smokers (Brake et al., 2020;Cai, 2020), which could put smokers at higher risk of contracting SARS-CoV-2. Other studies, however, suggest that nicotine downregulates the ACE-2 receptor (Oakes et al., 2018). These uncertainties notwithstanding, both former and current smoking is known to increase the risk of respiratory viral (Abadom et al., 2016;Denholm et al., 2010) and bacterial (Almirall et al., 1999;Feldman and Anderson, 2013) infections and is associated with worse outcomes once infected. Cigarette smoke reduces the respiratory immune defence through peri-bronchiolar inflammation and fibrosis, impaired mucociliary clearance and disruption of the respiratory epithelium (Dye and Adler, 1994). There is also reason to believe that behavioural factors (e.g. regular hand-tomouth movements) involved in smoking may increase SARS-CoV-2 infection and transmission in current smokers.
However, early data from the COVID-19 pandemic have not provided clear evidence for a negative impact of current or former smoking on SARS-CoV-2 infection or COVID-19 disease outcomes, such as hospitalisation or mortality (Vardavas and Nikitara, 2020). It has also been hypothesised that nicotine might protect against a hyper-inflammatory response to SARS-CoV-2 infection, which may lead to adverse outcomes in patients with COVID-19 disease .
In a living review where new data are regularly added to the analyses, it may be more appropriate to use a Bayesian (as opposed to frequentist) approach where prior knowledge is used in combination with new data to estimate a posterior risk distribution. A Bayesian approach mitigates the issue of performing multiple statistical tests, which can inflate family-wise error. A series of random-effects hierarchical Bayesian meta-analyses were performed with the brms (Bürkner, 2018) package to estimate the relative risk for each comparison with accompanying 95% credible intervals (CrIs). We first defined prior distributions for the true pooled effect size (µ) and the between-study heterogeneity (τ), with µ specified as a normal distribution with a mean equal to the derived point estimate from each comparison of interest in the immediately preceding version of this living review, and τ specified as a half-Cauchy distribution with a mean of 0 and standard deviation of 1.
The half-Cauchy distribution was selected to reflect prior knowledge that high levels of between-study heterogeneity are more likely than lower levels. Markov Chain Monte Carlo methods (20,000 burn-ins followed by 80,000 iterations) were then used to generate a risk distribution for each study, in addition to a pooled effect for the posterior risk distribution. We report forest plots with the pooled effect for the posterior risk distribution displayed as the median relative risk with an accompanying 95% CrIs. We used the empirical cumulative distribution function (ECDF) to estimate the probability of there being a 10% reduction or 10% increase in relative risk (RR) (i.e. RR ≥1.1 or RR ≤0.9). Due to a lack of indication as to what constitutes a clinically or epidemiologically meaningful effect (e.g. with regards to onward disease transmission or requirements for intensive care beds), we deemed a 10% change in risk as small but important. Where data were inconclusive (as indicated by CrIs crossing RR = 1.0), to disambiguate whether data favoured no effect or there being a small but important association, we estimated whether there was ≥75% probability of RR ≥1.1 or RR ≤0.9.
Two sensitivity analyses were performed. First, a minimally informative prior for µ was specified as a normal distribution with a mean of 0 and standard deviation of 1 and τ as described above. Second, an informative prior as described above for µ was used with τ specified as a half-Cauchy distribution with a mean of 0.3 and standard deviation of 1 to reflect greater between-study heterogeneity.
To aid in the visualisation of smoking prevalence in the included studies, the weighted mean prevalence of current and former smoking was calculated for countries with ≥3 studies and plotted for comparison with national prevalence estimates. It should be noted that prevalence estimates in the included studies were not adjusted for age, sex, socioeconomic position, or geographic region within countries.

Results
In the current review version (v11) with searches up to 2021-02-16, a total of 1133 records were identified, with 405 studies included in a narrative synthesis and 62 studies included in meta-analyses (see Figure 1). Qeios,.0 · Article, March 2, 2021 Qeios ID: UJR2AW.13 · https://doi.org/10.32388/UJR2AW.13 5/85 Characteristics of included studies are presented in Table 1. Studies were conducted across 41 countries. 109 studies were conducted in the USA, 70 in China, 44 in the UK, 28 in Spain,20 in France,18 in Mexico,16 in Italy,13 in Multiple,8 in Turkey, 7 in Brazil and Iran, 5 in Israel and Switzerland,4 in Finland and India, with 3 each from Australia, Austria, Japan, Saudi Arabia, and South Korea and a single study from 10 further countries. The majority of studies used observational designs (see Supplementary table S1). 256 (63.2%) were conducted in hospital settings, 106 studies (26.2%) included individuals from community and hospital settings, 40 studies (9.9%) were conducted exclusively in the community, with one study each conducted in a homeless shelter and a quarantine centre, and one study that did not state the study setting. Studies had a median of 502 (interquartile range = 146-2038) participants. The majority of studies (87.8%) used reverse transcriptase polymerase chain reaction (RT-PCR) for confirmation of SARS-CoV-2 infection, 12.2% used an antibody test to confirm prior infection and 6.8% of studies relied on a combination of RT-PCR or antibody assays.

Smoking status
Categorisation of smoking status was heterogeneous (see Table 1). 236 studies collected data on smoking status through routine electronic health records (EHRs), 129 studies used a bespoke case report form for COVID-19 and 40 studies did not state the source for information on smoking status. None of the studies verified smoking status biochemically. Notably, Qeios,.0 · Article, March 2, 2021 Qeios ID: UJR2AW.13 · https://doi.org/10.32388/UJR2AW.13 6/85 only 121 (29.9%) studies reported current, former and never smoking status (see Supplementary table S2a), with a further 26 studies reporting only ever and never smoking status (see Supplementary table S2b). The remaining 252 studies reported current, current/former or current and former smoking status but did not explicitly state whether remaining participants were never smokers or if data were missing on smoking status (see Supplementary table S2c). 133 studies explicitly reported the proportion with missing data on smoking status, which ranged from 0% to 97.6%.

Quality appraisal
Three studies were performed in random or representative population samples and were rated as 'good' quality, and 94 studies were rated as 'fair' quality, of which 62 studies reported results stratified by smoking status for the outcomes of interest and could be included in meta-analyses. The remaining 308 studies were rated as 'poor' quality (see Table 1). years using any tobacco product between 2010-2015 (Odani, 2018). In the UK Biobank cohort , a multivariable analysis showed former (RR = 1.29, 95% CI = 1.14-1.45, p < .001) and current (RR = 1.44, 95% CI = 1.20-1.71, p < .001) compared with never smokers to be more likely to receive a test. In an Australian rapid assessment screening clinic for COVID-19 (Trubiano et al., 2020), 9.4% (397/4,226) of the self-referred sample (subsequently assessed by a healthcare professional to decide on testing) were current smokers. Of these self-referrals, healthcare professionals decided that current compared with former or never smokers were less likely to require a test (RR = 0.93, 95% CI = 0.86-1.0, p = 0.045). In a further study using the UK Biobank cohort (Didikoglu et al., 2021), current (RR = 1.23, 95% CI = 1.19-1.26, p < 0.001) and former smokers (RR = 1.20, 95% CI = 1.18-1.23, p < 0.001) were more likely to receive a test compared with never smokers.

SARS-CoV-2 infection by smoking status
76 studies provided data on SARS-CoV-2 infection for people meeting local testing criteria by smoking status (see Table   2). Meta-analyses were performed for 3 'good' and 27 'fair' quality studies (see Figure 3 and 4). Current smokers were at reduced risk of testing positive for SARS-CoV-2 compared with never smokers (RR = 0.71, 95% Credible Interval (CrI) = 0.61-0.82, τ = 0.34). The three good quality studies each reported point estimates less than 1, although the CrI for one of the three studies crossed 1. The probability of current smokers being at reduced risk of infection compared with never smokers (RR ≤0.9) was >99%. Former compared with never smokers were at increased risk of testing positive, but data were inconclusive (RR = 1.03, 95% CrI = 0.95-1.11, τ = 0.17) and favoured there being no important association. The probability of former smokers being at increased risk of infection (RR ≥1.1) compared with never smokers was 4%. Results were materially unchanged in sensitivity analyses.     Hospitalisation for COVID-19 by smoking status 48 studies examined hospitalisation for COVID-19 disease, stratified by smoking status (see Table 3). Meta-analyses were performed for 16 'fair' quality studies (see Figure 5 and 6). Current (RR = 1.1, 95% CrI = 0.99-1.21, τ = 0.15) and former (RR = 1.19, CrI = 1.1-1.29, τ = 0.13) compared with never smokers were at increased risk of hospitalisation with COVID-19. However, data for current smokers were inconclusive and favoured there being no important association. The probability of current and former smokers being at increased risk of hospitalisation (RR ≥1.1) compared with never smokers was 49% and 98%, respectively. Results were materially unchanged in two sensitivity analyses.   Disease severity by smoking status 85 studies reported disease severity in hospitalised patients stratified by smoking status (see Table 4). Severe (as opposed to non-severe) disease was broadly defined as requiring intensive treatment unit (ITU) admission, requiring oxygen as a hospital inpatient or in-hospital death. Meta-analyses were performed for 11 'fair' quality studies (see Figure 7 and 8). Current (RR = 1.26, 95% CrI = 0.92-1.73, τ = 0.32) and former (RR = 1.8, 95% CrI = 1.27-2.55, τ = 0.46) compared with never smokers were at increased risk of greater disease severity. However, while data for current smokers only were inconclusive, they favoured there being a small but important association. The probability of current and former smokers having increased risk of greater disease severity (RR ≥1.1) compared with never smokers was 80% and >99%, respectively. Results were materially unchanged in two sensitivity analyses.    Mortality by smoking status 89 studies reported mortality from COVID-19 by smoking status (see Table 5), with 19 'fair' quality studies included in meta-analyses (see Figure 9 and 10). Current (RR = 1.12, 95% CrI = 0.84-1.47, τ = 0.42) and former (RR = 1.56, 95% CrI = 1.23-2, τ = 0.43) compared with never smokers were at increased risk of in-hospital mortality from COVID-19. However, data for current smokers were inconclusive and favoured there being no important association. The probability of current and former smokers being at greater risk of in-hospital mortality (RR ≥1.1) compared with never smokers was 60% and >99%, respectively. Results were materially unchanged in two sensitivity analyses.

Discussion
This living rapid review found uncertainty in the majority of 405 studies arising from the recording of smoking status.
Notwithstanding this uncertainty, compared with overall adult national prevalence estimates, recorded current smoking rates in most studies were lower than expected. In a subset of good and fair quality studies (n = 30), current but not former smokers had a reduced risk of testing positive for SARS-CoV-2 but current smokers appeared somewhat more likely to present for testing and/or receive a test. Data for current smokers on the risk of hospitalisation, disease severity and mortality were inconclusive, and favoured there being no important associations with hospitalisation and mortality and a small but important increase in the risk of severe disease. Former smokers were at increased risk of hospitalisation,

Issues complicating interpretation
Interpretation of results from studies conducted during the first phase of the SARS-CoV-2 pandemic is complicated by several factors (see Figure 11): 1. Exposure to SARS-CoV-2 i) Exposure to the SARS-CoV-2 virus is heterogeneous with different subgroups at heightened risk of infection at different stages of the pandemic, at least partly due to differential contact matrices by age, sex and socioeconomic position (CMMID COVID-19 working group et al., 2020), which are associated with smoking status.
ii) The probability of viral exposure depends largely on local prevalence, which varies over time. This likely introduces bias in studies assessing the rate of infection by smoking status conducted in the early phase of the pandemic.

Infection with SARS-CoV-2
i) Infection following viral exposure depends on individual differences in, for example, genetic susceptibility or immunocompetence, which are poorly understood at present and may be confounded with smoking. For example, the household secondary attack rate for COVID-19 is estimated at 17% (Fung et al., 2020).
ii) Heated and humidified air may act to disrupt the ability of the virus to persist in the airway mucosa of smokers. There is some evidence that transient localised hyperthermia can inhibit replication of rhinoviruses, a non-enveloped virus that causes the common cold (Conti et al., 1999). However, as SARS-CoV-2 is an enveloped virus (Schoeman and Fielding, 2019), it is unclear whether a similar protective effect against viral replication or invasion by heated and humidified air may occur.

Symptomatic COVID-19
i) An estimated 20% (95% CI = 17-25%) of COVID-19 cases are asymptomatic (Buitrago-Garcia et al., 2020), with some evidence suggesting younger people are more likely to be asymptomatic (Kronbichler et al., 2020). Testing is hence likely limited in some subgroups, with the potential for these groups to include an overrepresentation of current smokers.
ii) On the other hand, current and former smokers may be more likely to meet local criteria for community testing due to increased prevalence of symptoms consistent with SARS-CoV-2 infection, such as cough, increased sputum production or altered sense of smell or taste (Hopkinson et al., 2020). Evidence from a small number of studies indicates that current smokers may be more likely to present for testing, hence increasing the denominator in comparisons with never smokers and potentially inflating the rate of negative tests in current smokers. Infection positivity rates estimated among random samples are more informative. We identified one population study conducted in Hungary reporting on seroprevalence and smoking status (Merkely et al., 2020); however, the response rate was only 58.8% and the current smoking rate was 10 percentage points below national prevalence estimates, which raises some doubt about representativeness of the final sample. Similarly, two further representative population surveys (Carrat et al., 2020;Richard et al., 2020)  to healthcare and may be more likely to die in the community from sudden complications (i.e. self-selection bias) and thus not be recorded (Brown, 2020).
ii) Diagnostic criteria for SARS-CoV-2 infection and COVID-19 have changed during the course of the pandemic (Organisation, n.d.). It was not possible to extract details on the specific RT-PCR or antibody-based techniques or platforms used across the included studies due to reporting gaps. Different platforms have varying sensitivity and specificity to detect SARS-CoV-2 infection. In addition, testing for acute infection requires swabbing of the mucosal epithelium, which may be disrupted in current smokers, potentially altering the sensitivity of assays (Lusignan et al., 2020).

Hospitalisation with COVID-19
i) Reasons for hospitalisation vary by country and time in the pandemic. For example, early cases may have been hospitalised for isolation and quarantine reasons and not due to medical necessity. It is plausible this may have skewed early data towards less severe cases. In addition, the observed association between former smoking and greater disease severity may be explained by collider bias , where conditioning on a collider (e.g. testing or hospitalisation) by design or analysis may introduce a spurious association between current or former smoking (a potential cause of testing or hospitalisation) and SARS-CoV-2 infection/adverse outcomes from COVID-19 (potentially exacerbated by smoking) (Murray, 2020).
ii) The majority of included studies relied on EHRs as the source of information on smoking status. Research shows large discrepancies between EHRs and actual behaviour (Polubriaginof et al., 2018). Known failings of EHRs include implausible longitudinal changes, such as former smokers being recorded as never smokers at subsequent hospital visits (Polubriaginof et al., 2018). Misreporting on the part of the patient (perhaps due to perceived stigma) has also been observed, with biochemical measures showing higher rates of smoking compared with self-report in hospitalised patients in the US (Benowitz et al., 2009). It is hence possible that under-reporting of current and former smoking status in hospitals occurred across the included studies.
iii) The majority of included studies were conducted in hospital settings. It is plausible that a non-trivial proportion of patients were infected with SARS-CoV-2 while being an inpatient for a different medical reason. If so, this may have biased the hospitalised populations towards older and more frail groups, who are less likely to be smokers (Mangera et al., 2017). iv) Individuals with severe COVID-19 symptoms may have stopped smoking immediately before admission to hospital and may therefore not have been recorded as current smokers (i.e. reverse causality).
6. COVID-19 disease severity and death i) Given lack of knowledge of the disease progression and long-term outcomes of COVID-19, it is unclear whether studies conducted thus far in the pandemic have monitored patients for a sufficient time period to report complete survival outcomes or whether they are subject to early censoring. Adding to this, COVID-19 related mortality has been differentially defined across countries and epidemic phases. For example, in some UK reporting, death within 28 days Qeios, CC-BY 4.0 · Article, March 2, 2021 This living rapid evidence review was limited by having a single reviewer extracting data with a second independently verifying the data extracted to minimise errors, restricting the search to one electronic database and one pre-print server and by not including at least three large population surveys due to their reliance on self-reported suspected or confirmed SARS-CoV-2 infection (which means they do not meet our eligibility criteria) (Bowyer et al., 2020;Hopkinson et al., 2020;Jackson et al., 2020). We also did not include a large, UK-based, representative seroprevalence study (Ward et al., 2020) in our meta-analyses as the odds of testing positive in former smokers was not reported. However, the odds of infection for current smokers (OR = 0.64, 95% CI = 0.58-0.71) was in concordance with the pooled estimate in our meta-analysis.
Population surveys -particularly with linked data on confirmed infection or antibodies -will be included in future review versions to help mitigate some of the limitations of healthcare based observational studies. The comparisons of current and former smoking prevalence in the included studies with national prevalence estimates did not adjust observed prevalence for the demographic profile of those tested/admitted to hospital. Other reviews focused on this comparison have applied adjustments for sex and age, and continue to find lower than expected prevalence -notwithstanding the issues complicating interpretation described above (Farsalinos, Barbouni, et al., 2020).

Implications for research, policy and practice
Further scientific research is needed to resolve the mixed findings summarised in our review. First, clinical trials of the posited therapeutic effect of nicotine could have important implications both for smokers and for improved understanding of how the SARS-CoV-2 virus causes disease in humans. Such trials should focus on medicinal nicotine (as smoked tobacco is a dirty delivery mechanism that could mask beneficial effects) and potentially differentiate between different modes of delivery (i.e. inhaled vs. ingested) since this can affect pharmacokinetics (Shahab et al., 2013) and potential therapeutic effects. A second research priority would be a large, representative (randomly sampled) population survey with a validated assessment of smoking status which distinguishes between recent and long-term ex-smokers -ideally biochemically verified -and assesses seroprevalence and links to health records.
In the meantime, public-facing messages about the possible protective effect of smoking or nicotine are premature. In our view, until there is further research, the quality of the evidence does not justify the huge risk associated with a message likely to reach millions of people that a lethal activity, such as smoking, may protect against COVID-19. It continues to be appropriate to recommend smoking cessation and emphasise the role of alternative nicotine products to support smokers to stop as part of public health efforts during COVID-19. At the very least, smoking cessation reduces acute risks from cardiovascular disease and could reduce demands on the healthcare system (Stead et al., 2013). GPs and other healthcare providers can play a crucial role -brief, high-quality and free online training is available at National Centre for Smoking Cessation and Training.

Conclusion
Across 405 studies, recorded current but not past smoking prevalence was generally lower than national prevalence Data availability All data contributing to the current and future review versions are available here All code required to reproduce the current and future analyses are available here