Meta-analysis of the Association Between Second-hand Smoke Exposure and Ischaemic Heart Diseases, COPD and Stroke

Florian Fischer; Alexander Kraemer


BMC Public Health. 2015;15(1202) 

In This Article


Systematic Literature Review

As a first step, a systematic literature review was performed in PubMed according to the procedure and requirements described in the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement.[51] The aim of the systematic review was to identify articles dealing with the association between SHS and the three outcomes (IHD, COPD, and stroke). All relevant literature in English or German language was included without any restrictions regarding the year of publication. The search was restricted to studies on the effects of SHS exposure in humans. The search in PubMed was completed in July 2015. Therefore, the systematic literature review contained articles published between 1984 and 2014. The following search algorithm was performed:

(second hand smok* [Title/Abstract] OR second-hand smok* [Title/Abstract] OR passive smok* [Title/Abstract] OR "tobacco smoke pollution" [Title/Abstract] OR environmental tobacco smok* [Title/Abstract]) AND (heart disease* [Title/Abstract] OR COPD [Title/Abstract] OR chronic obstructive pulmonary disease* [Title/Abstract] OR obstructive pulmonary disease* [Title/Abstract] OR chronic obstructive airways disease* [Title/Abstract] OR COAD [Title/Abstract] OR chronic obstructive lung disease* [Title/Abstract] OR COLD [Title/Abstract] OR stroke*[Title/Abstract] OR apople*[Title/Abstract])

Using the search algorithm under the above-mentioned filters led to the identification of 403 records. Among them, 221 were attributable to a combination of the search terms regarding exposure and the outcome IHD, 178 further articles were attributable to the search terms on COPD and 47 on stroke.1 After the screening of title and abstract, 307 of these articles were excluded, because they did not fit the study's objective. Therefore, 96 full-texts were assessed for eligibility. According to this assessment, 71 articles were excluded for the following reasons2:

  • study design

    • survey/cross-sectional study (9)

    • (systematic) review (28)

    • meta-analysis (5)

  • no effect sizes provided (24)

  • other outcomes observed (5)

  • other exposures considered (4)

  • letter to the editor (2)

  • conflict of interest (1)

A manual search was conducted through the reference lists of all full-texts, which led to the inclusion of eight further articles. Finally, 33 articles were included in the qualitative analysis of the systematic review. Before including the studies in the quantitative synthesis in the form of a meta-analysis, a quality assessment was conducted. This quality assessment, which is described in more detail in the following section, led to the exclusion of further 9 studies. The process of the systematic review is presented in a flow chart (Fig. 1).

Figure 1.

Flow chart for study selection

Quality Assessment

A checklist for the quality assessment was compiled on the basis of already existing and well-established instruments, such as the PRISMA guidelines[51] and instruments developed for observational studies.[52–54] The quality score developed for this study consists of three categories, with four items each. The first category was introduced to identify a selection bias. Therefore, the selection of cases and response rate are focused here. Since both case–control and cohort studies were included in the systematic review, two quality scales were developed which differed slightly in the aspects regarding recruitment of the study population. The second category deals with the assessment of misclassification bias. It is asked 1) whether the exposure evaluation was made in relation to the time of diagnosis, 2) whether the exposure was validated by a biomarker, 3) whether specific disease criteria were provided, and 4) whether the disease was validated by histology or another gold standard. The third category focuses on aspects of data analysis. One item was integrated to detect whether or not an adjustment of variables was performed. Additionally, studies with power calculations and sufficient sample size scored higher. A sample size was defined a priori as sufficient if at least 100 subjects were included in the analysis and a minimum of 20 cases occurred, in order to exclude studies with low precision. The last criterion was about the provision of exact p-values and confidence intervals (CI).

Each item of the quality score answered with "yes" received one point, and all items with the labels "uncertain/not reported" or "no" received no points. All points were summed which allows a maximum score of 12 points. A priori, it was decided that all studies with an overall score of 7 points or lower (n = 9) would be excluded from the meta-analysis.

Calculation of Relative Risks

To allow for comparability between the results of the single studies, those results in which regular SHS exposure was investigated were focused upon. The definition of regular exposure varied between studies. Most commonly, spousal smoking or being exposed to about 20 cigarettes or more per day was interpreted as regular SHS exposure. In case studies divided between SHS exposure at home or at work, only the results for exposure at home were chosen. Nevertheless, several studies only provided information for SHS exposure at home and work combined.

The RR from the cohort studies were directly transferred to the summary of studies presented in Table 1. For case–control studies RR had to be derived from the provided odds ratios (OR). This was done for reasons of comparability of the results and because a single measurement unit was needed for the meta-analysis. For the calculation of RR based on OR an approach introduced by Barendregt[55] was selected. This approach describes the OR as a function of the RR, the average risk of disease in the population (s), and the prevalence of the risk factor (p). The equation uses the assumptions of the common definitions of RR and OR, and the observation that the average risk of a disease in any population is a linear combination of the risk in the exposed and non-exposed sub-populations:

The reciprocal conversion from OR to RR requires a numerical optimization procedure. The detailed derivation of the equation and the Excel add-in for the calculation of RR is provided by Barendregt.[55]


The provided or calculated RRs from the primary studies with high methodological quality were used for the meta-analysis. The meta-analysis was conducted in MIX 2.0 Pro, which is a statistical add-in to perform meta-analysis with Microsoft Excel.[56] As a first step, the RRs and CIs from all the studies were converted into the logarithm function of the RR (log (rr)) and standard errors (se). This information, including the sample size, was used to calculate effect sizes for each of the three outcomes, stratified by sex. The precision was set to an alpha-level of 0.05 and a z-distribution as the standard distribution was chosen. For the analysis, a generic inverse-variance method random effects model was chosen, to provide estimates for the association between SHS exposure and the outcomes IHD, COPD and stroke. In this model, weight is given to each study according to the inverse variance of the effect, to minimize uncertainty about the summarized effect estimates, according to the widely used approach developed by DerSimonian and Laird.[57]

Statistical Analysis

The random effects model was chosen, because the data were expected to be heterogeneous across studies. The advantage of a random effects model is that it incorporates variation in the underlying effect sizes between studies. It is assumed that each single study has its own (true) effect and that there is a random distribution of these effects around a central effect.[58] In contrast, using a fixed effect model under conditions of heterogeneity, the CI for the overall effects reflects the random variation within each study, but not the potential heterogeneity across studies, which would lead to artificially narrow CIs.[59] Furthermore, random effects models are more sensitive to publication bias, due to the larger relative weight given to smaller studies. This implies that a random effects model may still be worth considering as it cannot be assumed that true homogeneity exists across the studies.[60]

In order to consider the sensitivity of results, potential publication and study bias were assessed visually using a heterogeneity funnel plot (see Additional file 1 Additionally, heterogeneity was quantified using two statistical measures: The Q- and I2-statistics reflect a certain dimension of the extent of heterogeneity between the studies. The Q-statistic is the sum of the weighted squared differences between each individual study's estimate and the overall (inverse variance) summary estimates. This statistic follows a χ2-distribution with k–1 degrees of freedom, under the null-hypothesis of homogeneity. The Q-test is defined by Hedges and Olkin[61] as:

In this equation, w i is the weighting factor for the ith study, T i is the ith effect estimate in a collection of k studies and T̄ is the estimate of the mean effect size, which consists of weighting every effect estimate Ti by its inverse variance. A p-value < 0.1 for the Q-statistic indicates heterogeneity.[61]

Afterwards, the I2 is derived from the Q-statistic. The I2-index measures the extent of true heterogeneity by dividing the difference between the results of the Q test and its degrees of freedom by the Q-value itself, and multiplying by 100:

The I2-index quantifies the proportion of inconsistency among the study results. It is commonly expressed as a percentage and is therefore interpreted as the percentage of the total variability in a set of effect sizes due to between-study variation that is not attributable to random sampling from a fixed parameter.[62] Higgins and Thompson[62] proposed a tentative classification of I2-values to help in the interpretation of the heterogeneity's magnitude: according to this classification, percentages of around 25 %, 50 % and 75 % would mean low, medium, and high heterogeneity, respectively.