We registered our protocol on Open Science Framework and present our results in accordance with the PRISMA guidelines: https://osf.io/jrdzp .
We included published and unpublished (abstracts, conferences, pre-prints) cohort studies that compared ICS with placebo/standard of care or different dosing regimens of ICS in patients with COPD. We also included mixed cohorts of asthma and COPD patients but excluded studies enrolling only asthma patients. We did not restrict study eligibility based on language or year of publication.
An experienced research librarian searched EMBASE, MEDLINE, Cochrane Controlled Register of Trials (CENTRAL), Web of Science, and MedRxiv databases from inception to January 2022. Additional file 1: Appendix A1 describes our search strategy.
Data management and selection process
We uploaded citations to COVIDENCE, an online citation manager . Pairs of reviewers, following calibration exercises to ensure sufficient agreement, worked independently and in duplicate to screen titles and abstracts of search records and subsequently the full texts of records determined potentially eligible at the title and abstract screening stage. Reviewers resolved discrepancies by discussion or, when necessary, by third party adjudication.
Data collection process
Pairs of reviewers, following calibration exercises to ensure sufficient agreement, worked independently and in duplicate to collect data from eligible studies. Reviewers resolved discrepancies by discussion or, when necessary, by third party adjudication.
We collected data on study characteristics (time and country of recruitment), patient demographics (age, sex), clinical characteristics (emphysema, bronchitis, mixed, COPD/asthma overlap), and factors potentially predictive of lung cancer (smoking status, duration of smoking, duration of COPD, history of cancer, long acting muscarinic antagonist/long acting beta agonist (LAMA/LABA) use, chronic antibiotics therapies, home oxygen therapy, non-invasive ventilation, and treatment with roflumilast, theophylline, oral steroids and type and dose of ICS). Our choice of co-variates was based on factors highly associated with the development of lung cancer .
Outcomes and prioritization
We collected data on all-cause mortality, cancer-associated mortality, and serious adverse events. However, we only found data on the incidence of lung malignancy for analysis.
Risk of bias
We assessed the risk of bias independently and in duplicate for each outcome using the risk of bias in non-randomised studies of interventions (ROBINS-I) tool . We rated each outcome as either (1) low risk of bias, (2) moderate risk of bias, (3) serious risk of bias, and (4) critical risk of bias, across the following domains: bias due to confounding, bias in selection of participants into the study, bias in classification of interventions, bias due to deviations from intended interventions, bias due to missing data, bias in measurement of outcomes, and bias in selection of the reported result.
For studies to be rated as low risk of bias for confounding required at a minimum, adjustment for: age, sex, smoking (duration, pack years, quantity), COPD duration, socioeconomic status (employment, income, education), history of previous lung cancer, obesity, other lung disease (bronchiectasis, asthma, interstitial lung disease, obstructive sleep apnea), use of LAMA, LABA or both, treatment with oral corticosteroids and exposure to radon, radiation, or asbestosis. Additional file 1: Appendix A2 presents additional details on our assessment of risk of bias.
We report relative risk (RR) with 95% confidence intervals (CI) and risk differences per 1000 patients. To calculate risk differences, we used the baseline risk in a study we found most credible based on our assessment of risk of bias .
To compare the effects of lower versus higher doses of ICS and risk of lung cancer, we conducted a random-effects dose–response meta-analysis with the restricted maximum likelihood estimator (REML) using methods proposed by Greenland and Longnecker and Crippa and Orsini [14, 15]. Dose–response meta-analysis summarizes the quantitative relationship between doses of an exposure and the outcome across studies. We tested for nonlinearity using restricted cubic splines with knots at 10%, 50%, and 90% and a Wald-type test.
Because dose–response meta-analysis requires knowledge of the total number of participants or person-years, number of events, and mean or median dose across each dose category, not all studies were eligible for dose–response meta-analysis. Hence, we also present a random-effects meta-analysis with the REML estimator comparing the highest reported dose of ICS with the lowest reported dose across studies.
Where studies reported other types of ICS, we converted them to fluticasone equivalents. We used dose equivalents from data published by the Canadian Thoracic Society . We made assumptions about dosing based on conversions and expert opinion from respirologist and consensus of the authors. For studies reporting doses per prescription, we assumed one prescription to be equivalent to 500 ug/day of fluticasone and two prescriptions to be equivalent to approximately 1000 ug/day. For studies reporting the dose of ICS as a range of values, we assigned the midpoint of upper and lower boundaries in each category as the average dose. If the highest or lowest category were open ended, we assumed that the open-ended interval is the same size as the most adjacent interval.
We evaluated heterogeneity in part by inspecting the I2 values: 0–39% as unimportant, 40–59% as moderate, 60–74% as substantial, and 75–100% as considerable heterogeneity. We performed a subgroup analysis for COPD only and asthma/COPD mixed cohorts. We also performed a meta-regression using reported sex as a moderator. No data was available on severity of COPD to perform subgroup analysis. We used the ICEMAN tool to assess the credibility of subgroups if the result was statistically significant .
We conducted all analysis using the meta, dosresmeta, and rcs packages in R, version 4.0.3 .
Certainty of the evidence
We assessed the certainty of the evidence using the GRADE framework for observational studies and ROBINS-I [18, 19]. According to this approach, evidence starts at high certainty and may be further downgraded for risk of bias, inconsistency, indirectness, imprecision, or publication bias and may be upgraded for large effect, if suspected biases work against the observed direction of effect, or for dose–response gradient.