Discussion
It is crucial to remain impartial in our interpretation when faced with a lack of statistical significance, as relying solely on hypothesis testing can lead to the rejection of studies that may nevertheless provide valuable insights.23 In the current investigation, notwithstanding the absence of statistically significant associations between individuals with a history of smoking and adverse clinical outcomes in COVID-19 patients, notable indications of a potential inclination towards unfavourable clinical outcomes were observed. Specifically, these were evident in the context of the WHO ordinal COVID-19 scale, use of oxygen therapy (non-invasive or invasive), ICU admission and hospital LOS among patients with a smoking history. These associations require careful interpretation within the clinical context. They consistently emerged across various analytical approaches, including crude analysis, matching weights and cautious adjustments for confounding factors.
By adjusting for the confounding factors, we were able to make sense of the counterintuitive initial results produced by the crude analysis. For instance, as depicted in figure 3, the crude analysis results indicate that the death rate among never-smokers was higher than that among ever-smokers by approximately 11%. However, when the two groups were matched by weight, the death rates, although still slightly higher among never-smokers, varied considerably less. This discrepancy is likely due to the imbalance in baseline characteristics among the groups, with diabetes and hypertension being more prevalent in the never-smoker group. Had we relied solely on the crude analysis, we might have concluded that smoking protects against COVID-19 mortality.
At a mechanistic level, in the case of SARS-CoV-2 infection, smoking is thought to enhance the expression of ACE2 receptors. It is also associated with immune suppression, oxidative stress inflammation and vascular injury. Consequently, smoking increases susceptibility to more severe forms of COVID-19 disease.4 Although one single study may not fully address the question under discussion, it is important to note that earlier data from the first wave of the pandemic presented paradoxical findings, particularly regarding the population of current smokers.
For instance, in a UK study conducted by Gao et al24 during the first wave, current smokers exhibited a reduced risk of severe COVID-19. However, mortality outcomes increased after careful adjustment of potential confounders. Several researchers have highlighted methodological issues with this research, including reliance on record linkage, usage of data not specifically designed to answer the question, the risk of information and classification errors during the pandemic and non-random outcome misclassification, as one-fifth of those who died were never hospitalised and were, therefore, less likely to be tested and diagnosed if they were current smokers. To their credit, Gao et al supplemented their paper with post-hoc analyses that addressed some of the issues raised by referees and other researchers.25 Furthermore, these contradictory findings appeared in the OpenSAFELY study, where authors analysed data from the UK national primary care EHR linked to COVID-19 death data.26 It is also important to note that this study was not specifically designed to examine the association between smoking and COVID-19 mortality; however, as indicated by other researchers, interpreting the HRs in their tables as risk factors may represent a classic table 2 Fallacy.26
Notably, a recent meta-analysis conducted by Gallus et al found that the pooled OR estimate for the risk of severe COVID-19 diseases was 1.44 (95% CI 1.31 to 1.58) for an ever-smoker group versus a never-smoker group, which was close to the findings in the current study.27 These findings were confirmed by a robust Mendelian randomisation study conducted by Clift et al using UK Biobank data. This was undertaken in the general population, making it less susceptible to selection bias. The findings suggest that smoking increases the risk of acquiring SARS-CoV-2 infection, hospitalisation and death.28
To the best of our knowledge, this study offered the first opportunity to directly study the association between smoking and COVID-19 severity in Saudi Arabia. Previous studies in our population have discussed the smoking behavioural impact of COVID-19 or quality of life measures.29 30 Furthermore, the current study attempts to address some of the methodological limitations in the literature. First, not unique to our setting, the reliance on EHRs to determine smoking status has challenged researchers.31 In a study by Polubriaginof et al conducted in New York-Presbyterian Hospital/Columbia University Medical Centre, retrospective data were collected to illustrate the challenges of obtaining accurate information on smoking status. Despite the use of EHR systems, reporting smoking status was often dispersed in clinical notes, and there was no standardisation of reporting by the clinicians. In addition, their analysis revealed that when reported smoking status was treated as a dynamic concept (ie, it changes with time), implausible changes in smoking status were observed in 54.5% of cases. For example, some patients were reported as current smokers, but in other visits, they were reported as being never smokers.31 Determining exposure status in our study was essential to avoid or minimise misclassification bias.32 In particular, we were aware that such information may not be accurate if taken directly from the EHR system without confirmation. In addition, given the urgency of the pandemic, clinicians may have committed omission errors when acquiring such important information. Therefore, we contacted all eligible patients to confirm their exposure status prior to COVID-19 admission, which could have introduced a degree of recall bias, although this was less likely given the weight of the event (hospital admission due to COVID-19). Moreover, this recall bias was always present, as the nature of the exposure variable requires subjective data from the patient. Furthermore, because the study was conducted in a COVID-19 centre, the availability of SARS-CoV testing was not a limitation, as access to healthcare was not an issue.
Using our methods, we also avoided overadjustments or adjusting potential colliders or mediators, which are known to attenuate the effect of the estimates or lead to reverse causality.33 34 It was imperative to navigate the causal pathway and carefully select variables to ascertain the total causal effect of smoking status on COVID-19 severity. In addition, our analytical approach used matching weights, in which we balanced all groups in terms of comorbid conditions. This was illustrated in the love’s plot (figure 2). One advantage of using the SMD to examine covariate balance between the two groups was that it was independent of the sample size, thereby allowing a comparison to be made of mean differences in units of the pooled SD. As argued by Imai et al, the use of significance testing to assess baseline balance in a matched sample is inappropriate due to reduced statistical power and the property of balance being specific to a particular sample.35 Moreover, given that there is value in examining outcome variables as ordinal rather than binary, using ordinal regression models was appropriate. There are also several notable advantages to using an ordinal regression model. For instance, it increases the power of the analysis because it respects the ordinality of the outcome. Furthermore, it can provide more nuanced information about the relationship between predictor and outcome variables when the covariate balance between two groups is examined.36
The current study, however, involves several limitations that need to be addressed. First, the calculation of the sample size was based on the assumption of identifying an association that carries twice the risk of COVID-19 severity.37 Increasing the sample size could have produced statistically significant results for the severity outcome and given a more precise estimate of the smoking effect. However, we were limited by the number of patients to be screened who were admitted prior to the introduction of vaccines in our centre. Nonetheless, our results should be interpreted in the broader clinical context and the existing literature, as they provide clinical insights into the risk associated with smoking and COVID-19 severity. Second, because we relied on receiving confirmation of smoking status, we necessarily excluded many patients from the original sample pool due to unconfirmed smoking status, the inclusion of which could have increased the power of the study. Third, due to the low number of e-cigarette smokers, we could not perform an analysis to address this type of smoking. Fourth, only 33 patients were current smokers. No deaths occurred in this population, which could have resulted in a bias in the true estimates for COVID-19 severity in this cohort. We estimated COVID-19 severity to be similar in most of the WHO categories, with an emphasis on the non-significantly higher odds of the use of non-invasive mechanical ventilation or high-flow nasal cannula (adjusted OR=1.07 with a 95% CI 0.99 to 1.15, p=0.062, (online supplemental table S1). Using the DAGs method, we carefully selected the adjustments set for this analysis. The results disagree with some of the literature, which suggests that current smoking is protective against severe COVID-19, confirming the concerns raised by Westreich and colleagues on the possibility of a table 2 Fallacy in interpreting some of these data.26
Another limitation of our study is that because we used multiple significance tests, our results are vulnerable to a type I error, which occurs when researchers conduct multiple significance tests and emphasise the test with the highest statistical significance. This increases the probability of a statistically significant ‘false-positive’ result.38 Although adjustments for the multiple testing problem exist, conducting these would have exposed our results to a type II error or led to a failure to detect significant effects, even when they are present.
Finally, the generalisability of our study may be limited as new strains of the coronavirus continue to emerge and ignite new waves of infection around the globe. It is imperative that we reach a better understanding of the way in which risk factors such as cigarette smoking interact with disease progression and outcomes. Therefore, we recommend that larger, in-depth studies be conducted to investigate COVID-19 outcomes among ever-smokers and examine smoking behaviour, including current versus former status, type of smoking and frequency. Moreover, we recommend further investigation into the long-term impact of e-cigarette use, given the increasing prevalence of their usage on a local and global scale.39 40