Cardiovascular diseases

We identified 37 studies (59 observations)^{177,207,208,215,225,236,252,262,274–302} quantifying the relationship between SHS exposure and IHD and 20 studies (26 observations)^{176,207,208,225,236,252,262,278,296,297,303–312} assessing the relationship between SHS and stroke (Table ​(Table22 and Supplementary Tables 4 and 5). Our conservative analysis of the effect of SHS on IHD yielded an estimated RR of 1.26 (1.05–1.52) (Table ​(Table2,2, Fig. ​Fig.1a1a and Extended Data Fig. ​Fig.1),1), inclusive of between-study heterogeneity (gamma). We estimated the BPRF—which corresponds to the fifth quantile of RR closest to null and represents the lowest estimate of harmful SHS risk consistent with available evidence—to be 1.08, suggesting that SHS exposure increases an individual’s risk of IHD by a conservative minimum of 8%. In the BPRF framework, this translates to a risk–outcome score (ROS) of 0.04, which distinguishes the SHS–IHD relationship as a two-star risk–outcome pair, which can be interpreted as weak evidence of an association based on the available data (Table ​(Table2).2). Covariates accounting for cases where exposure to SHS was measured at baseline only (rather than multiple times during follow-up) and use of nonprospective cohort design were found to be statistically significant and were adjusted for within our final model (Table ​(Table22).

Open in a separate window

Fig. 1

Modified funnel plots for SHS exposure and two cardiovascular outcomes.

a,b, These modified funnel plots show the residuals of the reported mean RR relative to 0, the null value, on the x axis and the residuals of the standard error, as estimated from both the reported standard error and gamma, relative to 0 on the y axis, for IHD (a) and stroke (b). The light-blue vertical interval corresponds to the 95% uncertainty interval incorporating between-study heterogeneity; the dark-blue vertical interval corresponds to the 95% uncertainty interval (UI) without between-study heterogeneity; the dots are each included observation; the red Xs are outliered observations; the gray dotted line reflects the null log(RR); the blue line is the mean log(RR) for SHS and the outcome of interest; and the red line is the Burden of Proof function at the fifth quantile for these harmful risk–outcome associations.

Similarly, a weak but statistically significant relationship was found between SHS exposure and the risk of stroke. The estimated RR and uncertainty inclusive of between-study heterogeneity was 1.16 (1.03–1.32) (Table ​(Table2,2, Fig. ​Fig.1b1b and Extended Data Fig. ​Fig.2).2). Based on our conservative interpretation of the data, we estimated a BPRF of 1.05, indicating that exposure to tobacco smoke was associated with at least a 5% higher risk of stroke. This corresponds to a ROS of 0.02 and a two-star rating, consistent with weak evidence. In the final model, we adjusted for potential selection bias (based on percentage follow-up for longitudinal study designs and percentages of cases and controls for which exposure data could be ascertained for case–control designs) and for studies based on self-reported outcomes, as these covariates were found to be statistically significant by our bias covariate algorithm (Table ​(Table22).

Open in a separate window

Extended Data Fig. 2

Forest plot of the association between secondhand smoke exposure and stroke.

This forest plot presents the estimated mean relative risk, its 95% uncertainty intervals (UI), and the data points underlying the estimates for ischemic heart disease in association with secondhand smoke exposure (two-star rating of the risk-outcome relationship). The color of the point indicates whether the point was detected and trimmed as an outlier. The light blue interval corresponds to the 95% UI incorporating between-study heterogeneity; the dark blue interval corresponds to the 95% UI without between-study heterogeneity. The black vertical dotted line reflects the null relative risk value (one) and the red vertical line is the burden of proof function at the 5th quantile for this harmful risk-outcome association. The black data points and horizontal lines each correspond to a mean effect size and 95% UI from the included study identified on the y-axis. We included multiple observations from a single study when effects were reported by location or source of exposure and/or separately by sex or other subgroups. See Supplementary Table 5 for more details on included observations from each study (n=20 studies).

The two-star rating for IHD was consistent with sensitivity analyses in which we restricted the models to studies with a prospective cohort design (Supplementary Table 13), subset to observations of never smokers only (Supplementary Table 14), and applied both these restrictions at the same time (Supplementary Table 15). When restricted to prospective cohort data for never smokers only, the association between SHS and stroke was downgraded to one star (ROS −0.001) (Extended Data Fig. ​Fig.10).10). We did not detect publication bias, as identified by Egger’s regression test, in the primary analysis or in any of the sensitivity analyses for the cardiovascular outcomes (Table ​(Table22 and Supplementary Tables 13–15).

Open in a separate window

Extended Data Fig. 10

Summarized results of the primary model and sensitivity analyses conducted across all nine health outcomes.

This heatmap reports the summarized results of the main model and the sensitivity analyses (columns) conducted for each of the nine health outcomes (rows) reported in this study. Detailed results for each of the sensitivity models are presented in the Supplementary Information (Supplementary Tables 13–16). Sensitivity analyses reflect the impact of restricting the input data to 1) prospective cohort studies, 2) observations associated with never-smokers, and 3) both prospective cohort studies and never-smoking samples. For asthma, we additionally restrict the data to children population aged 16 or less. General model parameters remained constant across models; we trimmed 10% of the data if more than 10 observations were available for the specific model. The color of the blue boxes and the number depicted in each box corresponds to the resulting risk-outcome score (ROS) calculated for models in which the estimates of association without incorporating between-study heterogeneity were statistically significant. Grey boxes depict models that did not pass this threshold and, thus, ROS did not apply (NA). For models that did pass this threshold, the ROS reflects a conservative interpretation of the data that aligns with the Burden of Proof approach incorporating between-study heterogeneity and other sources of uncertainty. The ROS is translated into a star rating from 1 to 5 stars based on thresholds outlined in Zheng et al. The star rating for each model result is reported as the yellow stars in each box. A one-star association suggests that there is weak evidence supporting estimates of an association between the risk and the outcome. A two-star association reflects that there is weak-to-moderate evidence suggesting an association between the risk and outcome, and additional stars illustrate increasing strength of evidence.

Cancer

The conservative BPRF analysis indicated that passive smoking was weakly associated with an increased risk of lung cancer, based on a BPRF of 1.00 and a corresponding ROS of 0.001 (Table ​(Table2),2), which translates to a two-star rating at the lower threshold of the two-star range and suggests that SHS exposure was associated with at least around 1% higher risk of lung cancer. When between-study heterogeneity and other sources of uncertainty were accounted for, the estimated RR was 1.37 (0.94–1.99) (Table ​(Table2,2, Fig. ​Fig.2a2a and Extended Data Fig. ​Fig.3).3). The bias covariate algorithm selected observations that did not originally control for smoking to be adjusted in the final model (Table ​(Table2).2). In a sensitivity analysis in which we restricted the data to prospective cohort studies, the strength of the association was even lower (BPRF 0.95, ROS −0.03), downgrading the relationship to a one-star rating (Extended Data Fig. ​Fig.1010 and Supplementary Table 13).

Open in a separate window

Fig. 2

Modified funnel plots for SHS exposure and two cancer outcomes.

a,b, These modified funnel plots show the residuals of the reported mean RR relative to 0, the null value, on the x axis and the residuals of the standard error, as estimated from both the reported standard error and gamma, relative to 0 on the y axis, for lung cancer (a) and breast cancer (b). The light-blue vertical interval corresponds to the 95% uncertainty interval incorporating between-study heterogeneity; the dark-blue vertical interval corresponds to the 95% uncertainty interval (UI) without between-study heterogeneity; the dots are each included observation; the red Xs are outliered observations; the gray dotted line reflects the null log(RR); the blue line is the mean log(RR) for SHS and the outcome of interest; the red line is the Burden of Proof function at the fifth quantile for these harmful risk–outcome associations.

Open in a separate window

Extended Data Fig. 3

Forest plot of the association between secondhand smoke exposure and lung cancer.

This forest plot presents the estimated mean relative risk, its 95% uncertainty intervals (UI), and the data points underlying the estimates for ischemic heart disease in association with secondhand smoke exposure (two-star rating of the risk-outcome relationship). The color of the point indicates whether the point was detected and trimmed as an outlier. The light blue interval corresponds to the 95% UI incorporating between-study heterogeneity; the dark blue interval corresponds to the 95% UI without between-study heterogeneity. The black vertical dotted line reflects the null relative risk value (one) and the red vertical line is the burden of proof function at the 5th quantile for this harmful risk-outcome association. The black data points and horizontal lines each correspond to a mean effect size and 95% UI from the included study identified on the y-axis. We included multiple observations from a single study when effects were reported by location or source of exposure and/or separately by sex or other subgroups. See Supplementary Table 6 for more details on included observations from each study (n=104 studies).

Our conservative BPRF analysis also found weak evidence of a harmful association between exposure to tobacco smoke and risk of breast cancer (BPRF 0.81, ROS −0.11, one-star rating; Table ​Table2).2). The meta-analysis, which is supported by 51 unique studies^{170,220,313–361} and 79 observations (Supplementary Table 7), yielded an RR of 1.22 (0.75–1.98), inclusive of between-study heterogeneity (Table ​(Table2,2, Fig. ​Fig.2b2b and Extended Data Fig. ​Fig.4).4). In our model, observations that did not control for smoking and those from study designs other than prospective cohorts were adjusted since these covariates were found to be significant by our algorithm (Table ​(Table2).2). In further sensitivity analyses, the one-star relationship was still observed when we restricted to observations from never smokers only (Extended Data Fig. ​Fig.11 and Supplementary Table 14). However, when restricting to prospective cohort studies, we found no statistically significant evidence of an association between exposure to SHS and the risk of breast cancer in our fixed-effect model without between-study heterogeneity; that is, the estimated RR and associated uncertainty without gamma includes the null. These risk–outcome pairs are automatically assigned a zero-star rating, and the BPRF and ROS are not computed (Extended Data Fig. ​Fig.1010 and Supplementary Table 13).

Open in a separate window

Extended Data Fig. 4

Forest plot of the association between secondhand smoke exposure and breast cancer.

This forest plot presents the estimated mean relative risk, its 95% uncertainty intervals (UI), and the data points underlying the estimates for ischemic heart disease in association with secondhand smoke exposure (one-star rating of the risk-outcome relationship). The color of the point indicates whether the point was detected and trimmed as an outlier. The light blue interval corresponds to the 95% UI incorporating between-study heterogeneity; the dark blue interval corresponds to the 95% UI without between-study heterogeneity. The black vertical dotted line reflects the null relative risk value (one) and the red vertical line is the burden of proof function at the 5th quantile for this harmful risk-outcome association. The black data points and horizontal lines each correspond to a mean effect size and 95% UI from the included study identified on the y-axis. We included multiple observations from a single study when effects were reported by location or source of exposure and/or separately by sex or other subgroups. See Supplementary Table 7 for more details on included observations from each study (n=51 studies).

Based on Egger’s regression test, no significant evidence of publication bias was found for the main lung cancer and breast cancer models or the exploratory models (Table ​(Table22 and Supplementary Tables 13–15). Visual inspection of the funnel plots supported this finding (Fig. ​(Fig.22).

Respiratory conditions

We evaluated the association between exposure to SHS and three respiratory conditions: asthma, lower respiratory infections and COPD. Based on the conservative BPRF framework, the evidence supporting each of these relationships was weak (one-star rating), when between-study heterogeneity and other sources of bias were taken into account. Across these outcomes, no significant publication bias was detected in the primary models (Table ​(Table2)2) or in the sensitivity analyses (Supplementary Tables 13–16). For SHS and asthma, a risk–outcome pair not yet included in the GBD, the estimated RR incorporating between-study heterogeneity into the uncertainty was 1.21 (0.88–1.66) (Table ​(Table2,2, Fig. ​Fig.3a3a and Extended Data Fig. ​Fig.5).5). Data points associated with a self-reported diagnosis and those restricted to children (age ≤16years) were adjusted for in our main model, as the corresponding bias covariates were found to be statistically significant (Table ​(Table2).2). The BPRF and ROS were 0.93 and −0.04, respectively, which equates to a one-star risk classification. When restricting to prospective cohort studies, a two-star rating for the relationship between SHS and asthma was observed (Extended Data Fig. ​Fig.1010 and Supplementary Tables 13).

Open in a separate window

Fig. 3

Modified funnel plots for SHS exposure and three respiratory outcomes.

These modified funnel plots show the residuals of the reported mean RR relative to 0, the null value, on the x axis and the residuals of the standard error, as estimated from both the reported standard error and gamma, relative to 0 on the y axis, for asthma (a), lower respiratory infections (b) and COPD (c). The light-blue vertical interval corresponds to the 95% uncertainty interval incorporating between-study heterogeneity; the dark-blue vertical interval corresponds to the 95% uncertainty interval (UI) without between-study heterogeneity; the dots are each included observation; the red Xs are outliered observations; the gray dotted line reflects the null log(RR); the blue line is the mean log(RR) for SHS and the outcome of interest; the red line is the Burden of Proof function at the fifth quantile for these harmful risk–outcome associations.

Open in a separate window

Extended Data Fig. 5

Forest plot of the association between secondhand smoke exposure and asthma.

This forest plot presents the estimated mean relative risk, its 95% uncertainty intervals (UI), and the data points underlying the estimates for ischemic heart disease in association with secondhand smoke exposure (one-star rating of the risk-outcome relationship). The color of the point indicates whether the point was detected and trimmed as an outlier. The light blue interval corresponds to the 95% UI incorporating between-study heterogeneity; the dark blue interval corresponds to the 95% UI without between-study heterogeneity. The black vertical dotted line reflects the null relative risk value (one) and the red vertical line is the burden of proof function at the 5th quantile for this harmful risk-outcome association. The black data points and horizontal lines each correspond to a mean effect size and 95% UI from the included study identified on the y-axis. We included multiple observations from a single study when effects were reported by location or source of exposure and/or separately by sex or other subgroups. See Supplementary Table 8 for more details on included observations from each study (n=125 studies).

The meta-analysis of the risk of lower respiratory infections associated with SHS exposure included 50 studies^{53,64,91,134,362–407} and 66 observations (Supplementary Table 9) and yielded an RR and uncertainty interval inclusive of between-study heterogeneity of 1.34 (0.81–2.19) (Table ​(Table2,2, Fig. ​Fig.3b3b and Extended Data Fig. ​Fig.6).6). The BPRF (0.88) and corresponding ROS (−0.06) translated into a one-star rating, consistent with weak evidence of an association between passive smoking and increased risk of lower respiratory infections. The covariate selection algorithm flagged studies performed among populations that were not generalizable and those that used exposure definitions other than current SHS (for example, ever exposure to SHS) to be adjusted in our final model (Table ​(Table2).2). The strength of association as measured in the BPRF framework was not sensitive to any additional restrictions we applied to the input data, meaning that the one-star rating was still observed when we subset the data to prospective cohorts, never-smoking samples and a combination of the two (Extended Data Fig. ​Fig.1010 and Supplementary Tables 13–15).

Open in a separate window

Extended Data Fig. 6

Forest plot of the association between secondhand smoke exposure and lower respiratory infections.

This forest plot presents the estimated mean relative risk, its 95% uncertainty intervals (UI), and the data points underlying the estimates for ischemic heart disease in association with secondhand smoke exposure (one-star rating of the risk-outcome relationship). The color of the point indicates whether the point was detected and trimmed as an outlier. The light blue interval corresponds to the 95% UI incorporating between-study heterogeneity; the dark blue interval corresponds to the 95% UI without between-study heterogeneity. The black vertical dotted line reflects the null relative risk value (one) and the red vertical line is the burden of proof function at the 5th quantile for this harmful risk-outcome association. The black data points and horizontal lines each correspond to a mean effect size and 95% UI from the included study identified on the y-axis. We included multiple observations from a single study when effects were reported by location or source of exposure and/or separately by sex or other subgroups. See Supplementary Table 9 for more details on included observations from each study (n=50 studies).

Similar to the results for asthma and lower respiratory infections, the ROS for COPD was also negative (−0.14), equating to a one-star rating, indicating weak evidence of an association between SHS exposure and the risk of COPD. When accounting for between-study heterogeneity, the RR was 1.44 (0.67–3.12) (Table ​(Table2,2, Fig. ​Fig.3c3c and Extended Data Fig. ​Fig.7).7). Covariates representing studies that did not control for smoking and those with potential selection bias were found to be significant in our primary model and were adjusted for accordingly (Table ​(Table2).2). When including observations from seven prospective cohorts only, we found no statistically significant evidence of an association between SHS exposure and COPD when not including between-study heterogeneity (RR 1.21 (0.93–1.57, without gamma)). This was similar to the result we found when subsetting the data to never-smoking populations (RR 1.15 (0.95–1.40, without gamma)). The one-star association was observed, however, in a sensitivity analysis in which we applied both data restrictions simultaneously (Extended Data Fig. ​Fig.1010 and Supplementary Tables 13–15).

Open in a separate window

Extended Data Fig. 7

Forest plot of the association between secondhand smoke exposure and chronic obstructive pulmonary disease.

This forest plot presents the estimated mean relative risk, its 95% uncertainty intervals (UI), and the data points underlying the estimates for ischemic heart disease in association with secondhand smoke exposure (one-star rating of the risk-outcome relationship). The color of the point indicates whether the point was detected and trimmed as an outlier. The light blue interval corresponds to the 95% UI incorporating between-study heterogeneity; the dark blue interval corresponds to the 95% UI without between-study heterogeneity. The black vertical dotted line reflects the null relative risk value (one) and the red vertical line is the burden of proof function at the 5th quantile for this harmful risk-outcome association. The black data points and horizontal lines each correspond to a mean effect size and 95% UI from the included study identified on the y-axis. We included multiple observations from a single study when effects were reported by location or source of exposure and/or separately by sex or other subgroups. See Supplementary Table 10 for more details on included observations from each study (n=21 studies).

Health outcomes of interest

We selected outcomes on the basis of the availability of epidemiological evidence on their potential relationship with SHS. Eight out of the nine outcomes of interest—lung and breast cancer, IHD, stroke, COPD, lower respiratory infections, type 2 diabetes and otitis media—constitute SHS risk–outcome pairs considered in previous iterations of the GBD and were initially selected using the World Cancer Research Fund criteria for convincing or probable evidence as detailed in Murray et al.¹. Through review of published meta-analyses and systematic reviews and consultations with key external experts, we identified asthma as an additional health outcome of interest to SHS researchers and one for which sufficient literature was available to enable BPRF analytic methods; we therefore included it in our analysis. Reference and alternative definitions of each of the outcomes are listed in Supplementary Table 1.

Systematic review

We conducted separate systematic reviews to identify peer-reviewed literature reporting relative measures of association quantifying the relationship between SHS exposure and each health outcome of interest. We searched PubMed and Web of Science for studies published between 1 January 1970 and 31 July 2022. Furthermore, we reviewed the citation lists of the systematic reviews and meta-analyses captured in our searches to identify additional pertinent studies.

Briefly, after deduplicating the search results, each study’s title and abstract were manually screened by a single reviewer for inclusion eligibility. Subsequently, the full text was retrieved and screened, and data were extracted from those studies that passed our inclusion criteria of being published in English; being a case–control, cohort, case-cohort or case-crossover study conducted in participant groups likely to be generalizable; using suitable exposure and outcome definitions; and reporting both a relative measure of association (that is, RR, odds ratio or hazard ratio) and some measure of uncertainty (for example, sample size, standard error or confidence intervals). In terms of outcome definitions, studies using either a reference or an alternative health outcome definition met our inclusion criteria (Supplementary Table 1). As for SHS exposure, we included studies with varied SHS definitions, including proxies, but restricted to those reporting dichotomous current or ever exposure (that is, yes/no exposure). We excluded studies reporting only former exposure to SHS and those only assessing exposure in specific public settings. To better match our SHS definition, we also excluded studies and observations reporting health risk for current smokers. Finally, for all outcomes but otitis media, lower respiratory infections and asthma, we excluded studies that exclusively assessed childhood exposure to SHS to best account for the exposure temporality reflected in the SHS definition in GBD. In the case that multiple studies provided estimates from the same cohort, we included only the study with the largest sample or follow-up period so as not to duplicate data. The search strings used in each database, detailed inclusion and exclusion criteria, and outcome-specific PRISMA flow diagrams are available in Supplementary Figs. 1–9.

Data from eligible publications were manually extracted into a template designed to capture information about study and sample characteristics, exposure and outcome definitions, ascertainment methods, effect size and corresponding uncertainty reported for each model/population, and covariates included in the statistical analyses. We also assessed each study for risk of potential bias following the Grading of Recommendations, Assessment, Development and Evaluations (GRADE) approach and recorded the information in the extraction template⁴⁶⁹. As part of the exposure definition review, we cataloged multiple aspects of SHS exposure linked to each reported effect size, including the location of exposure (home or/and work combined; home; work; or any/unspecified location), the source of exposure (family; parental; maternal; paternal; spouse; or any/unspecified source), the timing of exposure (current or ever), and the smoking status of the exposed population (nonsmoker; never smoker; former smoker; or any/unspecified). Those studies performed only among children aged 15 years or less with an original ‘unspecified’ smoking status were reassigned to ‘adjusted never smokers’ and treated as ‘never smokers’ and ‘controlled for smoking’ in our analyses. In the GBD, we assume no smoking prevalence for ages under 10years; given the small prevalence for ages 10–15 and since most of the identified childhood studies included those past age 10, we believe this classification best reflects the smoking status of the studied population in these cases. All extracted data underwent manual quality assurance by the research team to verify accuracy. For a full list of extracted variables, with corresponding definitions, see Supplementary Table 2.

Estimating pooled RRs for each risk–outcome pair

We selected the effect sizes to be used in our meta-analytic approach within each included study and health outcome based on a prioritization cascade. All included effect sizes are reported in Supplementary Tables 4–12. Starting with the exposure definition, we chose the data points that closest matched the GBD risk definition in terms of the smoking status of the exposed population, followed by the location of exposure, the source of exposure, and the temporality. Thus, data points for nonsmokers currently exposed to SHS at home or work combined were prioritized over the other ones. In the absence of this exact definition, we prioritized the inclusion of effect sizes for each/any of the components of the GBD risk definition (that is, never smoker; former smoker; home; work) over those associated with a broader definition (that is, any/unspecified location or smoking status). Due to data sparsity, ‘ever exposure’ definitions were accepted for inclusion if results for ‘current exposure’ were not available. We did not include observations referring to exposure in specific settings other than home or work (for example, public settings or public transportation) or exposure among current smokers. Bias covariates were created to capture the impact of using alternate exposure definitions.

After this first selection stage, we proceeded with identifying the least granular analyses to be used in our models. For example, within each study and outcome, sex- and age-specific results were dropped in favor of aggregated data points, and results associated with the entire study population were retained over those for subgroup analyses when possible. We also favored observations reporting the risk of incidence and mortality combined over those that estimated each outcome type separately in cases where both were available. Moreover, for stroke, we dropped observations for subtypes (ischemic and hemorrhagic stroke) in favor of those for overall stroke due to data availability restrictions and to allow for best comparability across studies. In our last data selection step, the most-adjusted remainder data points within each study outcome were selected for inclusion in our analyses. This selection process is described in more detail in Supplementary Information.

To reduce the influence on our model of multiple observations coming from the same study, we adjusted the standard errors of effect sizes reported for multiple non-mutually exclusive exposure groups in each study by a factor matching the number of repeated measurements within each age–sex–smoking status group (Supplementary Information Section 2.2).

Finally, we used the MR-BRT tool to conduct each risk–outcome meta-regression analysis with the log-space RR of the outcome modeled as the dependent variable and exposure to SHS as the dichotomous independent variable (exposed to SHS versus not exposed to SHS). These analyses generated a single estimate of pooled RR of the given health outcome occurring for those exposed to SHS relative to unexposed counterparts. Following the BPRF methodology, we applied a 10% likelihood-based data-trimming algorithm to detect and remove outliers that may otherwise over-influence the model. This approach is suggested for all analyses with more than ten data points; therefore, it was implemented across all of our primary risk–outcome assessments and most of our sensitivity analyses⁴⁷⁰.

Testing and adjusting for biases across study designs and characteristics

Following the GRADE approach, we used the extracted data related to specific study characteristics to create binary covariates that captured potential sources of systematic bias within our input datasets. These covariates reflected the risk of bias associated with study design (prospective cohorts versus others), representativeness of the study population, exposure measurement (measured at baseline only versus multiple times during follow-up), outcome assessment method (self-report versus medical records), degree of control for confounding, and potential for selection bias (based on percentage follow-up for longitudinal study designs and percentages of cases and controls for which exposure data could be ascertained for case–control designs). Additionally, given SHS-specific characteristics, we created covariates to indicate whether a study controlled for smoking, regardless of other confounders, and whether the definition of SHS matched the one in GBD in terms of the location of exposure (home or work exposure versus broader definitions). A covariate reflecting studies performed among females only was also created. For the stroke models, we created two bias covariates to account for possible differences between studies reporting subtype-specific effect size only and those reporting stroke as an aggregated outcome; for asthma we created a specific covariate to indicate if a study was performed among children only (≤16years old). Detailed information about each of the bias covariates is provided in Supplementary Information Section 5 (Supplementary Table 20). We systematically tested for the effect of bias covariates using a selection algorithm, which uses a step-wise Lasso strategy to identify statistically significant covariates at a threshold of 0.05, and adjusted for the selected bias covariates in the final model used to generate the RR estimates. Covariates were eligible for testing if there was a minimum of two data points in the model associated with each covariate value. If multiple covariates had the same distribution of values within a model, we randomly selected one of the covariates to be tested.

Quantifying remaining between-study heterogeneity

After adjusting for study-level bias covariates, we used a linear mixed-effects model to capture the remaining unexplained between-study heterogeneity, in which we included a study-level random slope (gamma) and a study-level random intercept for within-study correlation. We derived the uncertainty of gamma using the inverse Fisher information matrix, which is sensitive to the number of studies, study design and reported uncertainty. The draws of gamma are used to derive the conservative uncertainty interval estimate for our RR (with gamma), estimated from both the uncertainty surrounding the mean effect and the 95th quantile of between-study heterogeneity. The RR without gamma, as reported in Table ​Table2,2, is reported with an uncertainty derived without fully accounting for between-study heterogeneity and reflects the RR estimates that are typically reported in traditional meta-analyses, while that with gamma better reflects the degree of consistency across the underlying studies. In this study, the RR metric of primary interest was the pooled RR with 95% uncertainty intervals that are inclusive (using gamma) of the effect of between-study heterogeneity. The estimated gamma for each risk–outcome primary assessment is presented in Supplementary Table 21.

Evaluating publication and reporting bias

To assess the presence of publication or reporting bias, we visually inspected the funnel plots (Figs. ​(Figs.11–4) produced for each risk–outcome evaluation, which show the residuals of the reported mean RR against the residuals of the standard error from each individual study. Visual inspection of the plots was accompanied by Egger’s regression tests to test for significant correlation between the standard error and the reported effect size. We did not find evidence of publication or reporting bias across any of the risk–outcome pairs in our primary models. We found publication bias for otitis media in one of our sensitivity analyses. We flagged the potential publication bias but did not correct for it in the model.

Estimating the BPRF

In our final step, we estimated the BPRF, which reflects the most conservative estimate of the association between exposure to SHS and the selected health outcomes that is consistent with the available evidence. For dichotomous harmful risk factors, the BPRF corresponds to the fifth quantile of RR closest to null, derived from the RR model inclusive of between-study heterogeneity. For each risk–outcome pair, the BPRF can be used to compute measures of increased or decreased risk of developing the health outcome due to exposure to the risk factor. BPRF values can be converted into ROSs, defined as the signed value of the average log RR of the BPRF. Large positive ROSs correspond to strong and consistent evidence of an association, while small positive ROSs and negative ROSs reflect weak evidence for an association, based on the available data. To facilitate the interpretation and comparison of the ROS results, the BPRF framework translates the ROS into star rating categories ranging from one to five (one star, ≤0.0 ROS; two stars, >0.0–0.14 ROS; three stars, >0.14–0.41 ROS; four stars, >0.41–0.62 ROS; five stars, >0.62 ROS). A one-star rating indicates weak evidence of association, while a five-star rating indicates very strong evidence. Zero-star risk–outcome pairs are not based on ROSs values but are defined as pairs for which there is no evidence of a statistically significant association between the risk and the health outcome when not accounting for between-study heterogeneity (that is, the 95% uncertainty interval without gamma crosses the null). Risk–outcome pairs receiving a one- through five-star rating are eligible for inclusion in the GBD.

Model validation

The validity of the BPRF approach to meta-analyze data extracted across studies has been extensively and rigorously evaluated by Zheng and colleagues⁴³⁰. For the present study, we conducted three main sensitivity analyses to examine the robustness of our primary findings to our data input in which we kept most of the model parameters consistent but (1) restricted our analysis to studies with a prospective cohort design; (2) subset our input data to never-smoking samples only; and (3) applied both these restrictions in conjunction. For asthma, specifically, we ran an additional model in which we restrict the data to those studies performed among children only (≤16years old). The only modification in our model parameters was related to the implementation of the 10% data trimming, which is dependent on the number of observations available for each outcome model (that is, data are trimmed only if ten observations or more are included). We present the detailed results of these sensitivity analyses in Supplementary Tables 13–16.

Statistical analysis and reproducibility

Analyses were carried out using R version 4.0.5 and Python version 3.10.9.

This investigation relied on existing published data. No statistical method was used to predetermine sample size. For each health outcome, we included all studies that met our inclusion criteria. This study did not engage in primary data collection, randomization or blinding. Therefore, data exclusions were not relevant to the present study, and, as such, no data were excluded from the analyses. We have made our data and code available to foster reproducibility.

Research reported in this publication was supported by the Bill & Melinda Gates Foundation (award OPP1152504, E.G.) and Bloomberg Philanthropies (award 47386, E.G.). The funders of the study had no role in study design, data collection, data analysis, data interpretation, writing of the final report or the decision to publish.